blob: 40c00dacbe4b9a478ea2655f86979bd00307f9e4 [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
Simon Arlott183ff222007-10-20 01:27:18 +020029 * slabs and you must pass objects with the same initializations to
Linus Torvalds1da177e2005-04-16 15:20:36 -070030 * 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 *)
David Woodhouse87a927c2007-07-04 21:26:44 -0400140#define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700141
142#ifndef cache_line_size
143#define cache_line_size() L1_CACHE_BYTES
144#endif
145
146#ifndef ARCH_KMALLOC_MINALIGN
147/*
148 * Enforce a minimum alignment for the kmalloc caches.
149 * Usually, the kmalloc caches are cache_line_size() aligned, except when
150 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
151 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
David Woodhouseb46b8f12007-05-08 00:22:59 -0700152 * alignment larger than the alignment of a 64-bit integer.
153 * ARCH_KMALLOC_MINALIGN allows that.
154 * Note that increasing this value may disable some debug features.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700155 */
David Woodhouseb46b8f12007-05-08 00:22:59 -0700156#define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700157#endif
158
159#ifndef ARCH_SLAB_MINALIGN
160/*
161 * Enforce a minimum alignment for all caches.
162 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
163 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
164 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
165 * some debug features.
166 */
167#define ARCH_SLAB_MINALIGN 0
168#endif
169
170#ifndef ARCH_KMALLOC_FLAGS
171#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
172#endif
173
174/* Legal flag mask for kmem_cache_create(). */
175#if DEBUG
Christoph Lameter50953fe2007-05-06 14:50:16 -0700176# define CREATE_MASK (SLAB_RED_ZONE | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700177 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
Christoph Lameterac2b8982006-03-22 00:08:15 -0800178 SLAB_CACHE_DMA | \
Christoph Lameter5af60832007-05-06 14:49:56 -0700179 SLAB_STORE_USER | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700180 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
Paul Jackson101a5002006-03-24 03:16:07 -0800181 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700182#else
Christoph Lameterac2b8982006-03-22 00:08:15 -0800183# define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
Christoph Lameter5af60832007-05-06 14:49:56 -0700184 SLAB_CACHE_DMA | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700185 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
Paul Jackson101a5002006-03-24 03:16:07 -0800186 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700187#endif
188
189/*
190 * kmem_bufctl_t:
191 *
192 * Bufctl's are used for linking objs within a slab
193 * linked offsets.
194 *
195 * This implementation relies on "struct page" for locating the cache &
196 * slab an object belongs to.
197 * This allows the bufctl structure to be small (one int), but limits
198 * the number of objects a slab (not a cache) can contain when off-slab
199 * bufctls are used. The limit is the size of the largest general cache
200 * that does not use off-slab slabs.
201 * For 32bit archs with 4 kB pages, is this 56.
202 * This is not serious, as it is only for large objects, when it is unwise
203 * to have too many per slab.
204 * Note: This limit can be raised by introducing a general cache whose size
205 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
206 */
207
Kyle Moffettfa5b08d2005-09-03 15:55:03 -0700208typedef unsigned int kmem_bufctl_t;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700209#define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
210#define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
Al Viro871751e2006-03-25 03:06:39 -0800211#define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
212#define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700213
Linus Torvalds1da177e2005-04-16 15:20:36 -0700214/*
215 * struct slab
216 *
217 * Manages the objs in a slab. Placed either at the beginning of mem allocated
218 * for a slab, or allocated from an general cache.
219 * Slabs are chained into three list: fully used, partial, fully free slabs.
220 */
221struct slab {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800222 struct list_head list;
223 unsigned long colouroff;
224 void *s_mem; /* including colour offset */
225 unsigned int inuse; /* num of objs active in slab */
226 kmem_bufctl_t free;
227 unsigned short nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700228};
229
230/*
231 * struct slab_rcu
232 *
233 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
234 * arrange for kmem_freepages to be called via RCU. This is useful if
235 * we need to approach a kernel structure obliquely, from its address
236 * obtained without the usual locking. We can lock the structure to
237 * stabilize it and check it's still at the given address, only if we
238 * can be sure that the memory has not been meanwhile reused for some
239 * other kind of object (which our subsystem's lock might corrupt).
240 *
241 * rcu_read_lock before reading the address, then rcu_read_unlock after
242 * taking the spinlock within the structure expected at that address.
243 *
244 * We assume struct slab_rcu can overlay struct slab when destroying.
245 */
246struct slab_rcu {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800247 struct rcu_head head;
Pekka Enberg343e0d72006-02-01 03:05:50 -0800248 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800249 void *addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700250};
251
252/*
253 * struct array_cache
254 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700255 * Purpose:
256 * - LIFO ordering, to hand out cache-warm objects from _alloc
257 * - reduce the number of linked list operations
258 * - reduce spinlock operations
259 *
260 * The limit is stored in the per-cpu structure to reduce the data cache
261 * footprint.
262 *
263 */
264struct array_cache {
265 unsigned int avail;
266 unsigned int limit;
267 unsigned int batchcount;
268 unsigned int touched;
Christoph Lametere498be72005-09-09 13:03:32 -0700269 spinlock_t lock;
Robert P. J. Daybda5b652007-10-16 23:30:05 -0700270 void *entry[]; /*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800271 * Must have this definition in here for the proper
272 * alignment of array_cache. Also simplifies accessing
273 * the entries.
Andrew Mortona737b3e2006-03-22 00:08:11 -0800274 */
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 */
Pekka Enberg556a1692008-01-25 08:20:51 +0200307#define NUM_INIT_LISTS (3 * MAX_NUMNODES)
Christoph Lametere498be72005-09-09 13:03:32 -0700308struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
309#define CACHE_CACHE 0
Pekka Enberg556a1692008-01-25 08:20:51 +0200310#define SIZE_AC MAX_NUMNODES
311#define SIZE_L3 (2 * 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 */
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -0700410 void (*ctor)(struct kmem_cache *, void *);
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) -
David Woodhouse87a927c2007-07-04 21:26:44 -0400550 REDZONE_ALIGN);
David Woodhouseb46b8f12007-05-08 00:22:59 -0700551 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/*
Gautham R Shenoy95402b32008-01-25 21:08:02 +0100733 * Guard access to the cache-chain.
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -0800734 */
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800735static DEFINE_MUTEX(cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700736static struct list_head cache_chain;
737
738/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700739 * chicken and egg problem: delay the per-cpu array allocation
740 * until the general caches are up.
741 */
742static enum {
743 NONE,
Christoph Lametere498be72005-09-09 13:03:32 -0700744 PARTIAL_AC,
745 PARTIAL_L3,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700746 FULL
747} g_cpucache_up;
748
Mike Kravetz39d24e62006-05-15 09:44:13 -0700749/*
750 * used by boot code to determine if it can use slab based allocator
751 */
752int slab_is_available(void)
753{
754 return g_cpucache_up == FULL;
755}
756
David Howells52bad642006-11-22 14:54:01 +0000757static DEFINE_PER_CPU(struct delayed_work, reap_work);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700758
Pekka Enberg343e0d72006-02-01 03:05:50 -0800759static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700760{
761 return cachep->array[smp_processor_id()];
762}
763
Andrew Mortona737b3e2006-03-22 00:08:11 -0800764static inline struct kmem_cache *__find_general_cachep(size_t size,
765 gfp_t gfpflags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700766{
767 struct cache_sizes *csizep = malloc_sizes;
768
769#if DEBUG
770 /* This happens if someone tries to call
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800771 * kmem_cache_create(), or __kmalloc(), before
772 * the generic caches are initialized.
773 */
Alok Katariac7e43c72005-09-14 12:17:53 -0700774 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700775#endif
Christoph Lameter6cb8f912007-07-17 04:03:22 -0700776 if (!size)
777 return ZERO_SIZE_PTR;
778
Linus Torvalds1da177e2005-04-16 15:20:36 -0700779 while (size > csizep->cs_size)
780 csizep++;
781
782 /*
Martin Hicks0abf40c2005-09-03 15:54:54 -0700783 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
Linus Torvalds1da177e2005-04-16 15:20:36 -0700784 * has cs_{dma,}cachep==NULL. Thus no special case
785 * for large kmalloc calls required.
786 */
Christoph Lameter4b51d662007-02-10 01:43:10 -0800787#ifdef CONFIG_ZONE_DMA
Linus Torvalds1da177e2005-04-16 15:20:36 -0700788 if (unlikely(gfpflags & GFP_DMA))
789 return csizep->cs_dmacachep;
Christoph Lameter4b51d662007-02-10 01:43:10 -0800790#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700791 return csizep->cs_cachep;
792}
793
Adrian Bunkb2213852006-09-25 23:31:02 -0700794static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700795{
796 return __find_general_cachep(size, gfpflags);
797}
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700798
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800799static size_t slab_mgmt_size(size_t nr_objs, size_t align)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700800{
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800801 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
802}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700803
Andrew Mortona737b3e2006-03-22 00:08:11 -0800804/*
805 * Calculate the number of objects and left-over bytes for a given buffer size.
806 */
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800807static void cache_estimate(unsigned long gfporder, size_t buffer_size,
808 size_t align, int flags, size_t *left_over,
809 unsigned int *num)
810{
811 int nr_objs;
812 size_t mgmt_size;
813 size_t slab_size = PAGE_SIZE << gfporder;
814
815 /*
816 * The slab management structure can be either off the slab or
817 * on it. For the latter case, the memory allocated for a
818 * slab is used for:
819 *
820 * - The struct slab
821 * - One kmem_bufctl_t for each object
822 * - Padding to respect alignment of @align
823 * - @buffer_size bytes for each object
824 *
825 * If the slab management structure is off the slab, then the
826 * alignment will already be calculated into the size. Because
827 * the slabs are all pages aligned, the objects will be at the
828 * correct alignment when allocated.
829 */
830 if (flags & CFLGS_OFF_SLAB) {
831 mgmt_size = 0;
832 nr_objs = slab_size / buffer_size;
833
834 if (nr_objs > SLAB_LIMIT)
835 nr_objs = SLAB_LIMIT;
836 } else {
837 /*
838 * Ignore padding for the initial guess. The padding
839 * is at most @align-1 bytes, and @buffer_size is at
840 * least @align. In the worst case, this result will
841 * be one greater than the number of objects that fit
842 * into the memory allocation when taking the padding
843 * into account.
844 */
845 nr_objs = (slab_size - sizeof(struct slab)) /
846 (buffer_size + sizeof(kmem_bufctl_t));
847
848 /*
849 * This calculated number will be either the right
850 * amount, or one greater than what we want.
851 */
852 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
853 > slab_size)
854 nr_objs--;
855
856 if (nr_objs > SLAB_LIMIT)
857 nr_objs = SLAB_LIMIT;
858
859 mgmt_size = slab_mgmt_size(nr_objs, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700860 }
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800861 *num = nr_objs;
862 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700863}
864
865#define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
866
Andrew Mortona737b3e2006-03-22 00:08:11 -0800867static void __slab_error(const char *function, struct kmem_cache *cachep,
868 char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700869{
870 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800871 function, cachep->name, msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700872 dump_stack();
873}
874
Paul Menage3395ee02006-12-06 20:32:16 -0800875/*
876 * By default on NUMA we use alien caches to stage the freeing of
877 * objects allocated from other nodes. This causes massive memory
878 * inefficiencies when using fake NUMA setup to split memory into a
879 * large number of small nodes, so it can be disabled on the command
880 * line
881 */
882
883static int use_alien_caches __read_mostly = 1;
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -0700884static int numa_platform __read_mostly = 1;
Paul Menage3395ee02006-12-06 20:32:16 -0800885static int __init noaliencache_setup(char *s)
886{
887 use_alien_caches = 0;
888 return 1;
889}
890__setup("noaliencache", noaliencache_setup);
891
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800892#ifdef CONFIG_NUMA
893/*
894 * Special reaping functions for NUMA systems called from cache_reap().
895 * These take care of doing round robin flushing of alien caches (containing
896 * objects freed on different nodes from which they were allocated) and the
897 * flushing of remote pcps by calling drain_node_pages.
898 */
899static DEFINE_PER_CPU(unsigned long, reap_node);
900
901static void init_reap_node(int cpu)
902{
903 int node;
904
905 node = next_node(cpu_to_node(cpu), node_online_map);
906 if (node == MAX_NUMNODES)
Paul Jackson442295c2006-03-22 00:09:11 -0800907 node = first_node(node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800908
Daniel Yeisley7f6b8872006-11-02 22:07:14 -0800909 per_cpu(reap_node, cpu) = node;
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800910}
911
912static void next_reap_node(void)
913{
914 int node = __get_cpu_var(reap_node);
915
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800916 node = next_node(node, node_online_map);
917 if (unlikely(node >= MAX_NUMNODES))
918 node = first_node(node_online_map);
919 __get_cpu_var(reap_node) = node;
920}
921
922#else
923#define init_reap_node(cpu) do { } while (0)
924#define next_reap_node(void) do { } while (0)
925#endif
926
Linus Torvalds1da177e2005-04-16 15:20:36 -0700927/*
928 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
929 * via the workqueue/eventd.
930 * Add the CPU number into the expiration time to minimize the possibility of
931 * the CPUs getting into lockstep and contending for the global cache chain
932 * lock.
933 */
Adrian Bunk897e6792007-07-15 23:38:20 -0700934static void __cpuinit start_cpu_timer(int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700935{
David Howells52bad642006-11-22 14:54:01 +0000936 struct delayed_work *reap_work = &per_cpu(reap_work, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700937
938 /*
939 * When this gets called from do_initcalls via cpucache_init(),
940 * init_workqueues() has already run, so keventd will be setup
941 * at that time.
942 */
David Howells52bad642006-11-22 14:54:01 +0000943 if (keventd_up() && reap_work->work.func == NULL) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800944 init_reap_node(cpu);
David Howells65f27f32006-11-22 14:55:48 +0000945 INIT_DELAYED_WORK(reap_work, cache_reap);
Arjan van de Ven2b284212006-12-10 02:21:28 -0800946 schedule_delayed_work_on(cpu, reap_work,
947 __round_jiffies_relative(HZ, cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700948 }
949}
950
Christoph Lametere498be72005-09-09 13:03:32 -0700951static struct array_cache *alloc_arraycache(int node, int entries,
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800952 int batchcount)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700953{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800954 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700955 struct array_cache *nc = NULL;
956
Christoph Lametere498be72005-09-09 13:03:32 -0700957 nc = kmalloc_node(memsize, GFP_KERNEL, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700958 if (nc) {
959 nc->avail = 0;
960 nc->limit = entries;
961 nc->batchcount = batchcount;
962 nc->touched = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700963 spin_lock_init(&nc->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700964 }
965 return nc;
966}
967
Christoph Lameter3ded1752006-03-25 03:06:44 -0800968/*
969 * Transfer objects in one arraycache to another.
970 * Locking must be handled by the caller.
971 *
972 * Return the number of entries transferred.
973 */
974static int transfer_objects(struct array_cache *to,
975 struct array_cache *from, unsigned int max)
976{
977 /* Figure out how many entries to transfer */
978 int nr = min(min(from->avail, max), to->limit - to->avail);
979
980 if (!nr)
981 return 0;
982
983 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
984 sizeof(void *) *nr);
985
986 from->avail -= nr;
987 to->avail += nr;
988 to->touched = 1;
989 return nr;
990}
991
Christoph Lameter765c4502006-09-27 01:50:08 -0700992#ifndef CONFIG_NUMA
993
994#define drain_alien_cache(cachep, alien) do { } while (0)
995#define reap_alien(cachep, l3) do { } while (0)
996
997static inline struct array_cache **alloc_alien_cache(int node, int limit)
998{
999 return (struct array_cache **)BAD_ALIEN_MAGIC;
1000}
1001
1002static inline void free_alien_cache(struct array_cache **ac_ptr)
1003{
1004}
1005
1006static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1007{
1008 return 0;
1009}
1010
1011static inline void *alternate_node_alloc(struct kmem_cache *cachep,
1012 gfp_t flags)
1013{
1014 return NULL;
1015}
1016
Christoph Hellwig8b98c162006-12-06 20:32:30 -08001017static inline void *____cache_alloc_node(struct kmem_cache *cachep,
Christoph Lameter765c4502006-09-27 01:50:08 -07001018 gfp_t flags, int nodeid)
1019{
1020 return NULL;
1021}
1022
1023#else /* CONFIG_NUMA */
1024
Christoph Hellwig8b98c162006-12-06 20:32:30 -08001025static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
Paul Jacksonc61afb12006-03-24 03:16:08 -08001026static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
Christoph Lameterdc85da12006-01-18 17:42:36 -08001027
Pekka Enberg5295a742006-02-01 03:05:48 -08001028static struct array_cache **alloc_alien_cache(int node, int limit)
Christoph Lametere498be72005-09-09 13:03:32 -07001029{
1030 struct array_cache **ac_ptr;
Christoph Lameter8ef82862007-02-20 13:57:52 -08001031 int memsize = sizeof(void *) * nr_node_ids;
Christoph Lametere498be72005-09-09 13:03:32 -07001032 int i;
1033
1034 if (limit > 1)
1035 limit = 12;
1036 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
1037 if (ac_ptr) {
1038 for_each_node(i) {
1039 if (i == node || !node_online(i)) {
1040 ac_ptr[i] = NULL;
1041 continue;
1042 }
1043 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
1044 if (!ac_ptr[i]) {
Akinobu Mitacc550de2007-11-14 16:58:35 -08001045 for (i--; i >= 0; i--)
Christoph Lametere498be72005-09-09 13:03:32 -07001046 kfree(ac_ptr[i]);
1047 kfree(ac_ptr);
1048 return NULL;
1049 }
1050 }
1051 }
1052 return ac_ptr;
1053}
1054
Pekka Enberg5295a742006-02-01 03:05:48 -08001055static void free_alien_cache(struct array_cache **ac_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -07001056{
1057 int i;
1058
1059 if (!ac_ptr)
1060 return;
Christoph Lametere498be72005-09-09 13:03:32 -07001061 for_each_node(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001062 kfree(ac_ptr[i]);
Christoph Lametere498be72005-09-09 13:03:32 -07001063 kfree(ac_ptr);
1064}
1065
Pekka Enberg343e0d72006-02-01 03:05:50 -08001066static void __drain_alien_cache(struct kmem_cache *cachep,
Pekka Enberg5295a742006-02-01 03:05:48 -08001067 struct array_cache *ac, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07001068{
1069 struct kmem_list3 *rl3 = cachep->nodelists[node];
1070
1071 if (ac->avail) {
1072 spin_lock(&rl3->list_lock);
Christoph Lametere00946f2006-03-25 03:06:45 -08001073 /*
1074 * Stuff objects into the remote nodes shared array first.
1075 * That way we could avoid the overhead of putting the objects
1076 * into the free lists and getting them back later.
1077 */
shin, jacob693f7d32006-04-28 10:54:37 -05001078 if (rl3->shared)
1079 transfer_objects(rl3->shared, ac, ac->limit);
Christoph Lametere00946f2006-03-25 03:06:45 -08001080
Christoph Lameterff694162005-09-22 21:44:02 -07001081 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001082 ac->avail = 0;
1083 spin_unlock(&rl3->list_lock);
1084 }
1085}
1086
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001087/*
1088 * Called from cache_reap() to regularly drain alien caches round robin.
1089 */
1090static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1091{
1092 int node = __get_cpu_var(reap_node);
1093
1094 if (l3->alien) {
1095 struct array_cache *ac = l3->alien[node];
Christoph Lametere00946f2006-03-25 03:06:45 -08001096
1097 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001098 __drain_alien_cache(cachep, ac, node);
1099 spin_unlock_irq(&ac->lock);
1100 }
1101 }
1102}
1103
Andrew Mortona737b3e2006-03-22 00:08:11 -08001104static void drain_alien_cache(struct kmem_cache *cachep,
1105 struct array_cache **alien)
Christoph Lametere498be72005-09-09 13:03:32 -07001106{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001107 int i = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07001108 struct array_cache *ac;
1109 unsigned long flags;
1110
1111 for_each_online_node(i) {
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001112 ac = alien[i];
Christoph Lametere498be72005-09-09 13:03:32 -07001113 if (ac) {
1114 spin_lock_irqsave(&ac->lock, flags);
1115 __drain_alien_cache(cachep, ac, i);
1116 spin_unlock_irqrestore(&ac->lock, flags);
1117 }
1118 }
1119}
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001120
Ingo Molnar873623d2006-07-13 14:44:38 +02001121static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001122{
1123 struct slab *slabp = virt_to_slab(objp);
1124 int nodeid = slabp->nodeid;
1125 struct kmem_list3 *l3;
1126 struct array_cache *alien = NULL;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001127 int node;
1128
1129 node = numa_node_id();
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001130
1131 /*
1132 * Make sure we are not freeing a object from another node to the array
1133 * cache on this cpu.
1134 */
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001135 if (likely(slabp->nodeid == node))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001136 return 0;
1137
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001138 l3 = cachep->nodelists[node];
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001139 STATS_INC_NODEFREES(cachep);
1140 if (l3->alien && l3->alien[nodeid]) {
1141 alien = l3->alien[nodeid];
Ingo Molnar873623d2006-07-13 14:44:38 +02001142 spin_lock(&alien->lock);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001143 if (unlikely(alien->avail == alien->limit)) {
1144 STATS_INC_ACOVERFLOW(cachep);
1145 __drain_alien_cache(cachep, alien, nodeid);
1146 }
1147 alien->entry[alien->avail++] = objp;
1148 spin_unlock(&alien->lock);
1149 } else {
1150 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1151 free_block(cachep, &objp, 1, nodeid);
1152 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1153 }
1154 return 1;
1155}
Christoph Lametere498be72005-09-09 13:03:32 -07001156#endif
1157
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001158static void __cpuinit cpuup_canceled(long cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001159{
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001160 struct kmem_cache *cachep;
1161 struct kmem_list3 *l3 = NULL;
1162 int node = cpu_to_node(cpu);
1163
1164 list_for_each_entry(cachep, &cache_chain, next) {
1165 struct array_cache *nc;
1166 struct array_cache *shared;
1167 struct array_cache **alien;
1168 cpumask_t mask;
1169
1170 mask = node_to_cpumask(node);
1171 /* cpu is dead; no one can alloc from it. */
1172 nc = cachep->array[cpu];
1173 cachep->array[cpu] = NULL;
1174 l3 = cachep->nodelists[node];
1175
1176 if (!l3)
1177 goto free_array_cache;
1178
1179 spin_lock_irq(&l3->list_lock);
1180
1181 /* Free limit for this kmem_list3 */
1182 l3->free_limit -= cachep->batchcount;
1183 if (nc)
1184 free_block(cachep, nc->entry, nc->avail, node);
1185
1186 if (!cpus_empty(mask)) {
1187 spin_unlock_irq(&l3->list_lock);
1188 goto free_array_cache;
1189 }
1190
1191 shared = l3->shared;
1192 if (shared) {
1193 free_block(cachep, shared->entry,
1194 shared->avail, node);
1195 l3->shared = NULL;
1196 }
1197
1198 alien = l3->alien;
1199 l3->alien = NULL;
1200
1201 spin_unlock_irq(&l3->list_lock);
1202
1203 kfree(shared);
1204 if (alien) {
1205 drain_alien_cache(cachep, alien);
1206 free_alien_cache(alien);
1207 }
1208free_array_cache:
1209 kfree(nc);
1210 }
1211 /*
1212 * In the previous loop, all the objects were freed to
1213 * the respective cache's slabs, now we can go ahead and
1214 * shrink each nodelist to its limit.
1215 */
1216 list_for_each_entry(cachep, &cache_chain, next) {
1217 l3 = cachep->nodelists[node];
1218 if (!l3)
1219 continue;
1220 drain_freelist(cachep, l3, l3->free_objects);
1221 }
1222}
1223
1224static int __cpuinit cpuup_prepare(long cpu)
1225{
Pekka Enberg343e0d72006-02-01 03:05:50 -08001226 struct kmem_cache *cachep;
Christoph Lametere498be72005-09-09 13:03:32 -07001227 struct kmem_list3 *l3 = NULL;
1228 int node = cpu_to_node(cpu);
David Howellsea02e3d2007-07-19 01:49:09 -07001229 const int memsize = sizeof(struct kmem_list3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001230
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001231 /*
1232 * We need to do this right in the beginning since
1233 * alloc_arraycache's are going to use this list.
1234 * kmalloc_node allows us to add the slab to the right
1235 * kmem_list3 and not this cpu's kmem_list3
1236 */
1237
1238 list_for_each_entry(cachep, &cache_chain, next) {
1239 /*
1240 * Set up the size64 kmemlist for cpu before we can
1241 * begin anything. Make sure some other cpu on this
1242 * node has not already allocated this
1243 */
1244 if (!cachep->nodelists[node]) {
1245 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1246 if (!l3)
1247 goto bad;
1248 kmem_list3_init(l3);
1249 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1250 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1251
1252 /*
1253 * The l3s don't come and go as CPUs come and
1254 * go. cache_chain_mutex is sufficient
1255 * protection here.
1256 */
1257 cachep->nodelists[node] = l3;
1258 }
1259
1260 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1261 cachep->nodelists[node]->free_limit =
1262 (1 + nr_cpus_node(node)) *
1263 cachep->batchcount + cachep->num;
1264 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1265 }
1266
1267 /*
1268 * Now we can go ahead with allocating the shared arrays and
1269 * array caches
1270 */
1271 list_for_each_entry(cachep, &cache_chain, next) {
1272 struct array_cache *nc;
1273 struct array_cache *shared = NULL;
1274 struct array_cache **alien = NULL;
1275
1276 nc = alloc_arraycache(node, cachep->limit,
1277 cachep->batchcount);
1278 if (!nc)
1279 goto bad;
1280 if (cachep->shared) {
1281 shared = alloc_arraycache(node,
1282 cachep->shared * cachep->batchcount,
1283 0xbaadf00d);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001284 if (!shared) {
1285 kfree(nc);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001286 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001287 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001288 }
1289 if (use_alien_caches) {
1290 alien = alloc_alien_cache(node, cachep->limit);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001291 if (!alien) {
1292 kfree(shared);
1293 kfree(nc);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001294 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001295 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001296 }
1297 cachep->array[cpu] = nc;
1298 l3 = cachep->nodelists[node];
1299 BUG_ON(!l3);
1300
1301 spin_lock_irq(&l3->list_lock);
1302 if (!l3->shared) {
1303 /*
1304 * We are serialised from CPU_DEAD or
1305 * CPU_UP_CANCELLED by the cpucontrol lock
1306 */
1307 l3->shared = shared;
1308 shared = NULL;
1309 }
1310#ifdef CONFIG_NUMA
1311 if (!l3->alien) {
1312 l3->alien = alien;
1313 alien = NULL;
1314 }
1315#endif
1316 spin_unlock_irq(&l3->list_lock);
1317 kfree(shared);
1318 free_alien_cache(alien);
1319 }
1320 return 0;
1321bad:
Akinobu Mita12d00f62007-10-18 03:05:11 -07001322 cpuup_canceled(cpu);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001323 return -ENOMEM;
1324}
1325
1326static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1327 unsigned long action, void *hcpu)
1328{
1329 long cpu = (long)hcpu;
1330 int err = 0;
1331
Linus Torvalds1da177e2005-04-16 15:20:36 -07001332 switch (action) {
Heiko Carstens38c3bd92007-05-09 02:34:05 -07001333 case CPU_UP_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001334 case CPU_UP_PREPARE_FROZEN:
Gautham R Shenoy95402b32008-01-25 21:08:02 +01001335 mutex_lock(&cache_chain_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001336 err = cpuup_prepare(cpu);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01001337 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001338 break;
1339 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001340 case CPU_ONLINE_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001341 start_cpu_timer(cpu);
1342 break;
1343#ifdef CONFIG_HOTPLUG_CPU
Christoph Lameter5830c592007-05-09 02:34:22 -07001344 case CPU_DOWN_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001345 case CPU_DOWN_PREPARE_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001346 /*
1347 * Shutdown cache reaper. Note that the cache_chain_mutex is
1348 * held so that if cache_reap() is invoked it cannot do
1349 * anything expensive but will only modify reap_work
1350 * and reschedule the timer.
1351 */
1352 cancel_rearming_delayed_work(&per_cpu(reap_work, cpu));
1353 /* Now the cache_reaper is guaranteed to be not running. */
1354 per_cpu(reap_work, cpu).work.func = NULL;
1355 break;
1356 case CPU_DOWN_FAILED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001357 case CPU_DOWN_FAILED_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001358 start_cpu_timer(cpu);
1359 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001360 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001361 case CPU_DEAD_FROZEN:
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001362 /*
1363 * Even if all the cpus of a node are down, we don't free the
1364 * kmem_list3 of any cache. This to avoid a race between
1365 * cpu_down, and a kmalloc allocation from another cpu for
1366 * memory from the node of the cpu going down. The list3
1367 * structure is usually allocated from kmem_cache_create() and
1368 * gets destroyed at kmem_cache_destroy().
1369 */
Simon Arlott183ff222007-10-20 01:27:18 +02001370 /* fall through */
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08001371#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001372 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001373 case CPU_UP_CANCELED_FROZEN:
Gautham R Shenoy95402b32008-01-25 21:08:02 +01001374 mutex_lock(&cache_chain_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001375 cpuup_canceled(cpu);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001376 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001377 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001378 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001379 return err ? NOTIFY_BAD : NOTIFY_OK;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001380}
1381
Chandra Seetharaman74b85f32006-06-27 02:54:09 -07001382static struct notifier_block __cpuinitdata cpucache_notifier = {
1383 &cpuup_callback, NULL, 0
1384};
Linus Torvalds1da177e2005-04-16 15:20:36 -07001385
Christoph Lametere498be72005-09-09 13:03:32 -07001386/*
1387 * swap the static kmem_list3 with kmalloced memory
1388 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001389static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1390 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001391{
1392 struct kmem_list3 *ptr;
1393
Christoph Lametere498be72005-09-09 13:03:32 -07001394 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1395 BUG_ON(!ptr);
1396
1397 local_irq_disable();
1398 memcpy(ptr, list, sizeof(struct kmem_list3));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001399 /*
1400 * Do not assume that spinlocks can be initialized via memcpy:
1401 */
1402 spin_lock_init(&ptr->list_lock);
1403
Christoph Lametere498be72005-09-09 13:03:32 -07001404 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1405 cachep->nodelists[nodeid] = ptr;
1406 local_irq_enable();
1407}
1408
Andrew Mortona737b3e2006-03-22 00:08:11 -08001409/*
Pekka Enberg556a1692008-01-25 08:20:51 +02001410 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1411 * size of kmem_list3.
1412 */
1413static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1414{
1415 int node;
1416
1417 for_each_online_node(node) {
1418 cachep->nodelists[node] = &initkmem_list3[index + node];
1419 cachep->nodelists[node]->next_reap = jiffies +
1420 REAPTIMEOUT_LIST3 +
1421 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1422 }
1423}
1424
1425/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08001426 * Initialisation. Called after the page allocator have been initialised and
1427 * before smp_init().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001428 */
1429void __init kmem_cache_init(void)
1430{
1431 size_t left_over;
1432 struct cache_sizes *sizes;
1433 struct cache_names *names;
Christoph Lametere498be72005-09-09 13:03:32 -07001434 int i;
Jack Steiner07ed76b2006-03-07 21:55:46 -08001435 int order;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001436 int node;
Christoph Lametere498be72005-09-09 13:03:32 -07001437
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -07001438 if (num_possible_nodes() == 1) {
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001439 use_alien_caches = 0;
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -07001440 numa_platform = 0;
1441 }
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001442
Christoph Lametere498be72005-09-09 13:03:32 -07001443 for (i = 0; i < NUM_INIT_LISTS; i++) {
1444 kmem_list3_init(&initkmem_list3[i]);
1445 if (i < MAX_NUMNODES)
1446 cache_cache.nodelists[i] = NULL;
1447 }
Pekka Enberg556a1692008-01-25 08:20:51 +02001448 set_up_list3s(&cache_cache, CACHE_CACHE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001449
1450 /*
1451 * Fragmentation resistance on low memory - only use bigger
1452 * page orders on machines with more than 32MB of memory.
1453 */
1454 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1455 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1456
Linus Torvalds1da177e2005-04-16 15:20:36 -07001457 /* Bootstrap is tricky, because several objects are allocated
1458 * from caches that do not exist yet:
Andrew Mortona737b3e2006-03-22 00:08:11 -08001459 * 1) initialize the cache_cache cache: it contains the struct
1460 * kmem_cache structures of all caches, except cache_cache itself:
1461 * cache_cache is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001462 * Initially an __init data area is used for the head array and the
1463 * kmem_list3 structures, it's replaced with a kmalloc allocated
1464 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001465 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001466 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001467 * An __init data area is used for the head array.
1468 * 3) Create the remaining kmalloc caches, with minimally sized
1469 * head arrays.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001470 * 4) Replace the __init data head arrays for cache_cache and the first
1471 * kmalloc cache with kmalloc allocated arrays.
Christoph Lametere498be72005-09-09 13:03:32 -07001472 * 5) Replace the __init data for kmem_list3 for cache_cache and
1473 * the other cache's with kmalloc allocated memory.
1474 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001475 */
1476
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001477 node = numa_node_id();
1478
Linus Torvalds1da177e2005-04-16 15:20:36 -07001479 /* 1) create the cache_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001480 INIT_LIST_HEAD(&cache_chain);
1481 list_add(&cache_cache.next, &cache_chain);
1482 cache_cache.colour_off = cache_line_size();
1483 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001484 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001485
Eric Dumazet8da34302007-05-06 14:49:29 -07001486 /*
1487 * struct kmem_cache size depends on nr_node_ids, which
1488 * can be less than MAX_NUMNODES.
1489 */
1490 cache_cache.buffer_size = offsetof(struct kmem_cache, nodelists) +
1491 nr_node_ids * sizeof(struct kmem_list3 *);
1492#if DEBUG
1493 cache_cache.obj_size = cache_cache.buffer_size;
1494#endif
Andrew Mortona737b3e2006-03-22 00:08:11 -08001495 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1496 cache_line_size());
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08001497 cache_cache.reciprocal_buffer_size =
1498 reciprocal_value(cache_cache.buffer_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001499
Jack Steiner07ed76b2006-03-07 21:55:46 -08001500 for (order = 0; order < MAX_ORDER; order++) {
1501 cache_estimate(order, cache_cache.buffer_size,
1502 cache_line_size(), 0, &left_over, &cache_cache.num);
1503 if (cache_cache.num)
1504 break;
1505 }
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02001506 BUG_ON(!cache_cache.num);
Jack Steiner07ed76b2006-03-07 21:55:46 -08001507 cache_cache.gfporder = order;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001508 cache_cache.colour = left_over / cache_cache.colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001509 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1510 sizeof(struct slab), cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001511
1512 /* 2+3) create the kmalloc caches */
1513 sizes = malloc_sizes;
1514 names = cache_names;
1515
Andrew Mortona737b3e2006-03-22 00:08:11 -08001516 /*
1517 * Initialize the caches that provide memory for the array cache and the
1518 * kmem_list3 structures first. Without this, further allocations will
1519 * bug.
Christoph Lametere498be72005-09-09 13:03:32 -07001520 */
1521
1522 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001523 sizes[INDEX_AC].cs_size,
1524 ARCH_KMALLOC_MINALIGN,
1525 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001526 NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07001527
Andrew Mortona737b3e2006-03-22 00:08:11 -08001528 if (INDEX_AC != INDEX_L3) {
Christoph Lametere498be72005-09-09 13:03:32 -07001529 sizes[INDEX_L3].cs_cachep =
Andrew Mortona737b3e2006-03-22 00:08:11 -08001530 kmem_cache_create(names[INDEX_L3].name,
1531 sizes[INDEX_L3].cs_size,
1532 ARCH_KMALLOC_MINALIGN,
1533 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001534 NULL);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001535 }
Christoph Lametere498be72005-09-09 13:03:32 -07001536
Ingo Molnare0a42722006-06-23 02:03:46 -07001537 slab_early_init = 0;
1538
Linus Torvalds1da177e2005-04-16 15:20:36 -07001539 while (sizes->cs_size != ULONG_MAX) {
Christoph Lametere498be72005-09-09 13:03:32 -07001540 /*
1541 * For performance, all the general caches are L1 aligned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001542 * This should be particularly beneficial on SMP boxes, as it
1543 * eliminates "false sharing".
1544 * Note for systems short on memory removing the alignment will
Christoph Lametere498be72005-09-09 13:03:32 -07001545 * allow tighter packing of the smaller caches.
1546 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001547 if (!sizes->cs_cachep) {
Christoph Lametere498be72005-09-09 13:03:32 -07001548 sizes->cs_cachep = kmem_cache_create(names->name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001549 sizes->cs_size,
1550 ARCH_KMALLOC_MINALIGN,
1551 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001552 NULL);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001553 }
Christoph Lameter4b51d662007-02-10 01:43:10 -08001554#ifdef CONFIG_ZONE_DMA
1555 sizes->cs_dmacachep = kmem_cache_create(
1556 names->name_dma,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001557 sizes->cs_size,
1558 ARCH_KMALLOC_MINALIGN,
1559 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1560 SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001561 NULL);
Christoph Lameter4b51d662007-02-10 01:43:10 -08001562#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001563 sizes++;
1564 names++;
1565 }
1566 /* 4) Replace the bootstrap head arrays */
1567 {
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001568 struct array_cache *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001569
Linus Torvalds1da177e2005-04-16 15:20:36 -07001570 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001571
Linus Torvalds1da177e2005-04-16 15:20:36 -07001572 local_irq_disable();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001573 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1574 memcpy(ptr, cpu_cache_get(&cache_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001575 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001576 /*
1577 * Do not assume that spinlocks can be initialized via memcpy:
1578 */
1579 spin_lock_init(&ptr->lock);
1580
Linus Torvalds1da177e2005-04-16 15:20:36 -07001581 cache_cache.array[smp_processor_id()] = ptr;
1582 local_irq_enable();
Christoph Lametere498be72005-09-09 13:03:32 -07001583
Linus Torvalds1da177e2005-04-16 15:20:36 -07001584 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001585
Linus Torvalds1da177e2005-04-16 15:20:36 -07001586 local_irq_disable();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001587 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001588 != &initarray_generic.cache);
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001589 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001590 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001591 /*
1592 * Do not assume that spinlocks can be initialized via memcpy:
1593 */
1594 spin_lock_init(&ptr->lock);
1595
Christoph Lametere498be72005-09-09 13:03:32 -07001596 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001597 ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001598 local_irq_enable();
1599 }
Christoph Lametere498be72005-09-09 13:03:32 -07001600 /* 5) Replace the bootstrap kmem_list3's */
1601 {
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001602 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001603
Mel Gorman9c09a952008-01-24 05:49:54 -08001604 for_each_online_node(nid) {
Pekka Enberg556a1692008-01-25 08:20:51 +02001605 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE], nid);
1606
Christoph Lametere498be72005-09-09 13:03:32 -07001607 init_list(malloc_sizes[INDEX_AC].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001608 &initkmem_list3[SIZE_AC + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001609
1610 if (INDEX_AC != INDEX_L3) {
1611 init_list(malloc_sizes[INDEX_L3].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001612 &initkmem_list3[SIZE_L3 + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001613 }
1614 }
1615 }
1616
1617 /* 6) resize the head arrays to their final sizes */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001618 {
Pekka Enberg343e0d72006-02-01 03:05:50 -08001619 struct kmem_cache *cachep;
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001620 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001621 list_for_each_entry(cachep, &cache_chain, next)
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07001622 if (enable_cpucache(cachep))
1623 BUG();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001624 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001625 }
1626
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001627 /* Annotate slab for lockdep -- annotate the malloc caches */
1628 init_lock_keys();
1629
1630
Linus Torvalds1da177e2005-04-16 15:20:36 -07001631 /* Done! */
1632 g_cpucache_up = FULL;
1633
Andrew Mortona737b3e2006-03-22 00:08:11 -08001634 /*
1635 * Register a cpu startup notifier callback that initializes
1636 * cpu_cache_get for all new cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07001637 */
1638 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001639
Andrew Mortona737b3e2006-03-22 00:08:11 -08001640 /*
1641 * The reap timers are started later, with a module init call: That part
1642 * of the kernel is not yet operational.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001643 */
1644}
1645
1646static int __init cpucache_init(void)
1647{
1648 int cpu;
1649
Andrew Mortona737b3e2006-03-22 00:08:11 -08001650 /*
1651 * Register the timers that return unneeded pages to the page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -07001652 */
Christoph Lametere498be72005-09-09 13:03:32 -07001653 for_each_online_cpu(cpu)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001654 start_cpu_timer(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001655 return 0;
1656}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001657__initcall(cpucache_init);
1658
1659/*
1660 * Interface to system's page allocator. No need to hold the cache-lock.
1661 *
1662 * If we requested dmaable memory, we will get it. Even if we
1663 * did not request dmaable memory, we might get it, but that
1664 * would be relatively rare and ignorable.
1665 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001666static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001667{
1668 struct page *page;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001669 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001670 int i;
1671
Luke Yangd6fef9d2006-04-10 22:52:56 -07001672#ifndef CONFIG_MMU
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001673 /*
1674 * Nommu uses slab's for process anonymous memory allocations, and thus
1675 * requires __GFP_COMP to properly refcount higher order allocations
Luke Yangd6fef9d2006-04-10 22:52:56 -07001676 */
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001677 flags |= __GFP_COMP;
Luke Yangd6fef9d2006-04-10 22:52:56 -07001678#endif
Christoph Lameter765c4502006-09-27 01:50:08 -07001679
Christoph Lameter3c517a62006-12-06 20:33:29 -08001680 flags |= cachep->gfpflags;
Mel Gormane12ba742007-10-16 01:25:52 -07001681 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1682 flags |= __GFP_RECLAIMABLE;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001683
1684 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001685 if (!page)
1686 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001687
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001688 nr_pages = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001689 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
Christoph Lameter972d1a72006-09-25 23:31:51 -07001690 add_zone_page_state(page_zone(page),
1691 NR_SLAB_RECLAIMABLE, nr_pages);
1692 else
1693 add_zone_page_state(page_zone(page),
1694 NR_SLAB_UNRECLAIMABLE, nr_pages);
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001695 for (i = 0; i < nr_pages; i++)
1696 __SetPageSlab(page + i);
1697 return page_address(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001698}
1699
1700/*
1701 * Interface to system's page release.
1702 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001703static void kmem_freepages(struct kmem_cache *cachep, void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001704{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001705 unsigned long i = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001706 struct page *page = virt_to_page(addr);
1707 const unsigned long nr_freed = i;
1708
Christoph Lameter972d1a72006-09-25 23:31:51 -07001709 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1710 sub_zone_page_state(page_zone(page),
1711 NR_SLAB_RECLAIMABLE, nr_freed);
1712 else
1713 sub_zone_page_state(page_zone(page),
1714 NR_SLAB_UNRECLAIMABLE, nr_freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001715 while (i--) {
Nick Pigginf205b2f2006-03-22 00:08:02 -08001716 BUG_ON(!PageSlab(page));
1717 __ClearPageSlab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001718 page++;
1719 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001720 if (current->reclaim_state)
1721 current->reclaim_state->reclaimed_slab += nr_freed;
1722 free_pages((unsigned long)addr, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001723}
1724
1725static void kmem_rcu_free(struct rcu_head *head)
1726{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001727 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001728 struct kmem_cache *cachep = slab_rcu->cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001729
1730 kmem_freepages(cachep, slab_rcu->addr);
1731 if (OFF_SLAB(cachep))
1732 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1733}
1734
1735#if DEBUG
1736
1737#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001738static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001739 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001740{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001741 int size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001742
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001743 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001744
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001745 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001746 return;
1747
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001748 *addr++ = 0x12345678;
1749 *addr++ = caller;
1750 *addr++ = smp_processor_id();
1751 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001752 {
1753 unsigned long *sptr = &caller;
1754 unsigned long svalue;
1755
1756 while (!kstack_end(sptr)) {
1757 svalue = *sptr++;
1758 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001759 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001760 size -= sizeof(unsigned long);
1761 if (size <= sizeof(unsigned long))
1762 break;
1763 }
1764 }
1765
1766 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001767 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001768}
1769#endif
1770
Pekka Enberg343e0d72006-02-01 03:05:50 -08001771static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001772{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001773 int size = obj_size(cachep);
1774 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001775
1776 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001777 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001778}
1779
1780static void dump_line(char *data, int offset, int limit)
1781{
1782 int i;
Dave Jonesaa83aa42006-09-29 01:59:51 -07001783 unsigned char error = 0;
1784 int bad_count = 0;
1785
Linus Torvalds1da177e2005-04-16 15:20:36 -07001786 printk(KERN_ERR "%03x:", offset);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001787 for (i = 0; i < limit; i++) {
1788 if (data[offset + i] != POISON_FREE) {
1789 error = data[offset + i];
1790 bad_count++;
1791 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001792 printk(" %02x", (unsigned char)data[offset + i]);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001793 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001794 printk("\n");
Dave Jonesaa83aa42006-09-29 01:59:51 -07001795
1796 if (bad_count == 1) {
1797 error ^= POISON_FREE;
1798 if (!(error & (error - 1))) {
1799 printk(KERN_ERR "Single bit error detected. Probably "
1800 "bad RAM.\n");
1801#ifdef CONFIG_X86
1802 printk(KERN_ERR "Run memtest86+ or a similar memory "
1803 "test tool.\n");
1804#else
1805 printk(KERN_ERR "Run a memory test tool.\n");
1806#endif
1807 }
1808 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001809}
1810#endif
1811
1812#if DEBUG
1813
Pekka Enberg343e0d72006-02-01 03:05:50 -08001814static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001815{
1816 int i, size;
1817 char *realobj;
1818
1819 if (cachep->flags & SLAB_RED_ZONE) {
David Woodhouseb46b8f12007-05-08 00:22:59 -07001820 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001821 *dbg_redzone1(cachep, objp),
1822 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001823 }
1824
1825 if (cachep->flags & SLAB_STORE_USER) {
1826 printk(KERN_ERR "Last user: [<%p>]",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001827 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001828 print_symbol("(%s)",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001829 (unsigned long)*dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001830 printk("\n");
1831 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001832 realobj = (char *)objp + obj_offset(cachep);
1833 size = obj_size(cachep);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001834 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001835 int limit;
1836 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001837 if (i + limit > size)
1838 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001839 dump_line(realobj, i, limit);
1840 }
1841}
1842
Pekka Enberg343e0d72006-02-01 03:05:50 -08001843static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001844{
1845 char *realobj;
1846 int size, i;
1847 int lines = 0;
1848
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001849 realobj = (char *)objp + obj_offset(cachep);
1850 size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001851
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001852 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001853 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001854 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001855 exp = POISON_END;
1856 if (realobj[i] != exp) {
1857 int limit;
1858 /* Mismatch ! */
1859 /* Print header */
1860 if (lines == 0) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001861 printk(KERN_ERR
David Howellse94a40c2007-04-02 23:46:28 +01001862 "Slab corruption: %s start=%p, len=%d\n",
1863 cachep->name, realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001864 print_objinfo(cachep, objp, 0);
1865 }
1866 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001867 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001868 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001869 if (i + limit > size)
1870 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001871 dump_line(realobj, i, limit);
1872 i += 16;
1873 lines++;
1874 /* Limit to 5 lines */
1875 if (lines > 5)
1876 break;
1877 }
1878 }
1879 if (lines != 0) {
1880 /* Print some data about the neighboring objects, if they
1881 * exist:
1882 */
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08001883 struct slab *slabp = virt_to_slab(objp);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001884 unsigned int objnr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001885
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001886 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001887 if (objnr) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001888 objp = index_to_obj(cachep, slabp, objnr - 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001889 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001890 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001891 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001892 print_objinfo(cachep, objp, 2);
1893 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001894 if (objnr + 1 < cachep->num) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001895 objp = index_to_obj(cachep, slabp, objnr + 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001896 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001897 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001898 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001899 print_objinfo(cachep, objp, 2);
1900 }
1901 }
1902}
1903#endif
1904
Linus Torvalds1da177e2005-04-16 15:20:36 -07001905#if DEBUG
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001906/**
Randy Dunlap911851e2006-03-22 00:08:14 -08001907 * slab_destroy_objs - destroy a slab and its objects
1908 * @cachep: cache pointer being destroyed
1909 * @slabp: slab pointer being destroyed
1910 *
1911 * Call the registered destructor for each object in a slab that is being
1912 * destroyed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001913 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001914static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001915{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001916 int i;
1917 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001918 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001919
1920 if (cachep->flags & SLAB_POISON) {
1921#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08001922 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1923 OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001924 kernel_map_pages(virt_to_page(objp),
Andrew Mortona737b3e2006-03-22 00:08:11 -08001925 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001926 else
1927 check_poison_obj(cachep, objp);
1928#else
1929 check_poison_obj(cachep, objp);
1930#endif
1931 }
1932 if (cachep->flags & SLAB_RED_ZONE) {
1933 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1934 slab_error(cachep, "start of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001935 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001936 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1937 slab_error(cachep, "end of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001938 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001939 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001940 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001941}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001942#else
Pekka Enberg343e0d72006-02-01 03:05:50 -08001943static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001944{
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001945}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001946#endif
1947
Randy Dunlap911851e2006-03-22 00:08:14 -08001948/**
1949 * slab_destroy - destroy and release all objects in a slab
1950 * @cachep: cache pointer being destroyed
1951 * @slabp: slab pointer being destroyed
1952 *
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001953 * Destroy all the objs in a slab, and release the mem back to the system.
Andrew Mortona737b3e2006-03-22 00:08:11 -08001954 * Before calling the slab must have been unlinked from the cache. The
1955 * cache-lock is not held/needed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001956 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001957static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001958{
1959 void *addr = slabp->s_mem - slabp->colouroff;
1960
1961 slab_destroy_objs(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001962 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1963 struct slab_rcu *slab_rcu;
1964
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001965 slab_rcu = (struct slab_rcu *)slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001966 slab_rcu->cachep = cachep;
1967 slab_rcu->addr = addr;
1968 call_rcu(&slab_rcu->head, kmem_rcu_free);
1969 } else {
1970 kmem_freepages(cachep, addr);
Ingo Molnar873623d2006-07-13 14:44:38 +02001971 if (OFF_SLAB(cachep))
1972 kmem_cache_free(cachep->slabp_cache, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001973 }
1974}
1975
Christoph Lameter117f6eb2006-09-25 23:31:37 -07001976static void __kmem_cache_destroy(struct kmem_cache *cachep)
1977{
1978 int i;
1979 struct kmem_list3 *l3;
1980
1981 for_each_online_cpu(i)
1982 kfree(cachep->array[i]);
1983
1984 /* NUMA: free the list3 structures */
1985 for_each_online_node(i) {
1986 l3 = cachep->nodelists[i];
1987 if (l3) {
1988 kfree(l3->shared);
1989 free_alien_cache(l3->alien);
1990 kfree(l3);
1991 }
1992 }
1993 kmem_cache_free(&cache_cache, cachep);
1994}
1995
1996
Linus Torvalds1da177e2005-04-16 15:20:36 -07001997/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08001998 * calculate_slab_order - calculate size (page order) of slabs
1999 * @cachep: pointer to the cache that is being created
2000 * @size: size of objects to be created in this cache.
2001 * @align: required alignment for the objects.
2002 * @flags: slab allocation flags
2003 *
2004 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002005 *
2006 * This could be made much more intelligent. For now, try to avoid using
2007 * high order pages for slabs. When the gfp() functions are more friendly
2008 * towards high-order requests, this should be changed.
2009 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002010static size_t calculate_slab_order(struct kmem_cache *cachep,
Randy Dunlapee13d782006-02-01 03:05:53 -08002011 size_t size, size_t align, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002012{
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002013 unsigned long offslab_limit;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002014 size_t left_over = 0;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002015 int gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002016
Christoph Lameter0aa817f2007-05-16 22:11:01 -07002017 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002018 unsigned int num;
2019 size_t remainder;
2020
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002021 cache_estimate(gfporder, size, align, flags, &remainder, &num);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002022 if (!num)
2023 continue;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002024
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002025 if (flags & CFLGS_OFF_SLAB) {
2026 /*
2027 * Max number of objs-per-slab for caches which
2028 * use off-slab slabs. Needed to avoid a possible
2029 * looping condition in cache_grow().
2030 */
2031 offslab_limit = size - sizeof(struct slab);
2032 offslab_limit /= sizeof(kmem_bufctl_t);
2033
2034 if (num > offslab_limit)
2035 break;
2036 }
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002037
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002038 /* Found something acceptable - save it away */
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002039 cachep->num = num;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002040 cachep->gfporder = gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002041 left_over = remainder;
2042
2043 /*
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002044 * A VFS-reclaimable slab tends to have most allocations
2045 * as GFP_NOFS and we really don't want to have to be allocating
2046 * higher-order pages when we are unable to shrink dcache.
2047 */
2048 if (flags & SLAB_RECLAIM_ACCOUNT)
2049 break;
2050
2051 /*
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002052 * Large number of objects is good, but very large slabs are
2053 * currently bad for the gfp()s.
2054 */
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002055 if (gfporder >= slab_break_gfp_order)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002056 break;
2057
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002058 /*
2059 * Acceptable internal fragmentation?
2060 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002061 if (left_over * 8 <= (PAGE_SIZE << gfporder))
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002062 break;
2063 }
2064 return left_over;
2065}
2066
Sam Ravnborg38bdc322007-05-17 23:48:19 +02002067static int __init_refok setup_cpu_cache(struct kmem_cache *cachep)
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002068{
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002069 if (g_cpucache_up == FULL)
2070 return enable_cpucache(cachep);
2071
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002072 if (g_cpucache_up == NONE) {
2073 /*
2074 * Note: the first kmem_cache_create must create the cache
2075 * that's used by kmalloc(24), otherwise the creation of
2076 * further caches will BUG().
2077 */
2078 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2079
2080 /*
2081 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2082 * the first cache, then we need to set up all its list3s,
2083 * otherwise the creation of further caches will BUG().
2084 */
2085 set_up_list3s(cachep, SIZE_AC);
2086 if (INDEX_AC == INDEX_L3)
2087 g_cpucache_up = PARTIAL_L3;
2088 else
2089 g_cpucache_up = PARTIAL_AC;
2090 } else {
2091 cachep->array[smp_processor_id()] =
2092 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
2093
2094 if (g_cpucache_up == PARTIAL_AC) {
2095 set_up_list3s(cachep, SIZE_L3);
2096 g_cpucache_up = PARTIAL_L3;
2097 } else {
2098 int node;
Pekka Enberg556a1692008-01-25 08:20:51 +02002099 for_each_online_node(node) {
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002100 cachep->nodelists[node] =
2101 kmalloc_node(sizeof(struct kmem_list3),
2102 GFP_KERNEL, node);
2103 BUG_ON(!cachep->nodelists[node]);
2104 kmem_list3_init(cachep->nodelists[node]);
2105 }
2106 }
2107 }
2108 cachep->nodelists[numa_node_id()]->next_reap =
2109 jiffies + REAPTIMEOUT_LIST3 +
2110 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2111
2112 cpu_cache_get(cachep)->avail = 0;
2113 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2114 cpu_cache_get(cachep)->batchcount = 1;
2115 cpu_cache_get(cachep)->touched = 0;
2116 cachep->batchcount = 1;
2117 cachep->limit = BOOT_CPUCACHE_ENTRIES;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002118 return 0;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002119}
2120
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002121/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07002122 * kmem_cache_create - Create a cache.
2123 * @name: A string which is used in /proc/slabinfo to identify this cache.
2124 * @size: The size of objects to be created in this cache.
2125 * @align: The required alignment for the objects.
2126 * @flags: SLAB flags
2127 * @ctor: A constructor for the objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002128 *
2129 * Returns a ptr to the cache on success, NULL on failure.
2130 * Cannot be called within a int, but can be interrupted.
Paul Mundt20c2df82007-07-20 10:11:58 +09002131 * The @ctor is run when new pages are allocated by the cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002132 *
2133 * @name must be valid until the cache is destroyed. This implies that
Andrew Mortona737b3e2006-03-22 00:08:11 -08002134 * the module calling this has to destroy the cache before getting unloaded.
2135 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002136 * The flags are
2137 *
2138 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2139 * to catch references to uninitialised memory.
2140 *
2141 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2142 * for buffer overruns.
2143 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002144 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2145 * cacheline. This can be beneficial if you're counting cycles as closely
2146 * as davem.
2147 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002148struct kmem_cache *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002149kmem_cache_create (const char *name, size_t size, size_t align,
Andrew Mortona737b3e2006-03-22 00:08:11 -08002150 unsigned long flags,
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07002151 void (*ctor)(struct kmem_cache *, void *))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002152{
2153 size_t left_over, slab_size, ralign;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07002154 struct kmem_cache *cachep = NULL, *pc;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002155
2156 /*
2157 * Sanity checks... these are all serious usage bugs.
2158 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002159 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
Paul Mundt20c2df82007-07-20 10:11:58 +09002160 size > KMALLOC_MAX_SIZE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002161 printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__,
2162 name);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002163 BUG();
2164 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002165
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002166 /*
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002167 * We use cache_chain_mutex to ensure a consistent view of
2168 * cpu_online_map as well. Please see cpuup_callback
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002169 */
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002170 get_online_cpus();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002171 mutex_lock(&cache_chain_mutex);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002172
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07002173 list_for_each_entry(pc, &cache_chain, next) {
Andrew Morton4f12bb42005-11-07 00:58:00 -08002174 char tmp;
2175 int res;
2176
2177 /*
2178 * This happens when the module gets unloaded and doesn't
2179 * destroy its slab cache and no-one else reuses the vmalloc
2180 * area of the module. Print a warning.
2181 */
Andrew Morton138ae662006-12-06 20:36:41 -08002182 res = probe_kernel_address(pc->name, tmp);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002183 if (res) {
matzeb4169522007-05-06 14:49:52 -07002184 printk(KERN_ERR
2185 "SLAB: cache with size %d has lost its name\n",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002186 pc->buffer_size);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002187 continue;
2188 }
2189
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002190 if (!strcmp(pc->name, name)) {
matzeb4169522007-05-06 14:49:52 -07002191 printk(KERN_ERR
2192 "kmem_cache_create: duplicate cache %s\n", name);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002193 dump_stack();
2194 goto oops;
2195 }
2196 }
2197
Linus Torvalds1da177e2005-04-16 15:20:36 -07002198#if DEBUG
2199 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002200#if FORCED_DEBUG
2201 /*
2202 * Enable redzoning and last user accounting, except for caches with
2203 * large objects, if the increased size would increase the object size
2204 * above the next power of two: caches with object sizes just above a
2205 * power of two have a significant amount of internal fragmentation.
2206 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002207 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2208 2 * sizeof(unsigned long long)))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002209 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002210 if (!(flags & SLAB_DESTROY_BY_RCU))
2211 flags |= SLAB_POISON;
2212#endif
2213 if (flags & SLAB_DESTROY_BY_RCU)
2214 BUG_ON(flags & SLAB_POISON);
2215#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002216 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002217 * Always checks flags, a caller might be expecting debug support which
2218 * isn't available.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002219 */
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002220 BUG_ON(flags & ~CREATE_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002221
Andrew Mortona737b3e2006-03-22 00:08:11 -08002222 /*
2223 * Check that size is in terms of words. This is needed to avoid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002224 * unaligned accesses for some archs when redzoning is used, and makes
2225 * sure any on-slab bufctl's are also correctly aligned.
2226 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002227 if (size & (BYTES_PER_WORD - 1)) {
2228 size += (BYTES_PER_WORD - 1);
2229 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002230 }
2231
Andrew Mortona737b3e2006-03-22 00:08:11 -08002232 /* calculate the final buffer alignment: */
2233
Linus Torvalds1da177e2005-04-16 15:20:36 -07002234 /* 1) arch recommendation: can be overridden for debug */
2235 if (flags & SLAB_HWCACHE_ALIGN) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002236 /*
2237 * Default alignment: as specified by the arch code. Except if
2238 * an object is really small, then squeeze multiple objects into
2239 * one cacheline.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002240 */
2241 ralign = cache_line_size();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002242 while (size <= ralign / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002243 ralign /= 2;
2244 } else {
2245 ralign = BYTES_PER_WORD;
2246 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002247
2248 /*
David Woodhouse87a927c2007-07-04 21:26:44 -04002249 * Redzoning and user store require word alignment or possibly larger.
2250 * Note this will be overridden by architecture or caller mandated
2251 * alignment if either is greater than BYTES_PER_WORD.
Pekka Enbergca5f9702006-09-25 23:31:25 -07002252 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002253 if (flags & SLAB_STORE_USER)
2254 ralign = BYTES_PER_WORD;
2255
2256 if (flags & SLAB_RED_ZONE) {
2257 ralign = REDZONE_ALIGN;
2258 /* If redzoning, ensure that the second redzone is suitably
2259 * aligned, by adjusting the object size accordingly. */
2260 size += REDZONE_ALIGN - 1;
2261 size &= ~(REDZONE_ALIGN - 1);
2262 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002263
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002264 /* 2) arch mandated alignment */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002265 if (ralign < ARCH_SLAB_MINALIGN) {
2266 ralign = ARCH_SLAB_MINALIGN;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002267 }
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002268 /* 3) caller mandated alignment */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002269 if (ralign < align) {
2270 ralign = align;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002271 }
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002272 /* disable debug if necessary */
David Woodhouseb46b8f12007-05-08 00:22:59 -07002273 if (ralign > __alignof__(unsigned long long))
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002274 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002275 /*
Pekka Enbergca5f9702006-09-25 23:31:25 -07002276 * 4) Store it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002277 */
2278 align = ralign;
2279
2280 /* Get cache's description obj. */
Christoph Lametere94b1762006-12-06 20:33:17 -08002281 cachep = kmem_cache_zalloc(&cache_cache, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002282 if (!cachep)
Andrew Morton4f12bb42005-11-07 00:58:00 -08002283 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002284
2285#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002286 cachep->obj_size = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002287
Pekka Enbergca5f9702006-09-25 23:31:25 -07002288 /*
2289 * Both debugging options require word-alignment which is calculated
2290 * into align above.
2291 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002292 if (flags & SLAB_RED_ZONE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002293 /* add space for red zone words */
David Woodhouseb46b8f12007-05-08 00:22:59 -07002294 cachep->obj_offset += sizeof(unsigned long long);
2295 size += 2 * sizeof(unsigned long long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002296 }
2297 if (flags & SLAB_STORE_USER) {
Pekka Enbergca5f9702006-09-25 23:31:25 -07002298 /* user store requires one word storage behind the end of
David Woodhouse87a927c2007-07-04 21:26:44 -04002299 * the real object. But if the second red zone needs to be
2300 * aligned to 64 bits, we must allow that much space.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002301 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002302 if (flags & SLAB_RED_ZONE)
2303 size += REDZONE_ALIGN;
2304 else
2305 size += BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002306 }
2307#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002308 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002309 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2310 cachep->obj_offset += PAGE_SIZE - size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002311 size = PAGE_SIZE;
2312 }
2313#endif
2314#endif
2315
Ingo Molnare0a42722006-06-23 02:03:46 -07002316 /*
2317 * Determine if the slab management is 'on' or 'off' slab.
2318 * (bootstrapping cannot cope with offslab caches so don't do
2319 * it too early on.)
2320 */
2321 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002322 /*
2323 * Size is large, assume best to place the slab management obj
2324 * off-slab (should allow better packing of objs).
2325 */
2326 flags |= CFLGS_OFF_SLAB;
2327
2328 size = ALIGN(size, align);
2329
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002330 left_over = calculate_slab_order(cachep, size, align, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002331
2332 if (!cachep->num) {
matzeb4169522007-05-06 14:49:52 -07002333 printk(KERN_ERR
2334 "kmem_cache_create: couldn't create cache %s.\n", name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002335 kmem_cache_free(&cache_cache, cachep);
2336 cachep = NULL;
Andrew Morton4f12bb42005-11-07 00:58:00 -08002337 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002338 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002339 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2340 + sizeof(struct slab), align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002341
2342 /*
2343 * If the slab has been placed off-slab, and we have enough space then
2344 * move it on-slab. This is at the expense of any extra colouring.
2345 */
2346 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2347 flags &= ~CFLGS_OFF_SLAB;
2348 left_over -= slab_size;
2349 }
2350
2351 if (flags & CFLGS_OFF_SLAB) {
2352 /* really off slab. No need for manual alignment */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002353 slab_size =
2354 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002355 }
2356
2357 cachep->colour_off = cache_line_size();
2358 /* Offset must be a multiple of the alignment. */
2359 if (cachep->colour_off < align)
2360 cachep->colour_off = align;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002361 cachep->colour = left_over / cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002362 cachep->slab_size = slab_size;
2363 cachep->flags = flags;
2364 cachep->gfpflags = 0;
Christoph Lameter4b51d662007-02-10 01:43:10 -08002365 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002366 cachep->gfpflags |= GFP_DMA;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002367 cachep->buffer_size = size;
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08002368 cachep->reciprocal_buffer_size = reciprocal_value(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002369
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002370 if (flags & CFLGS_OFF_SLAB) {
Victor Fuscob2d55072005-09-10 00:26:36 -07002371 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002372 /*
2373 * This is a possibility for one of the malloc_sizes caches.
2374 * But since we go off slab only for object size greater than
2375 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2376 * this should not happen at all.
2377 * But leave a BUG_ON for some lucky dude.
2378 */
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002379 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002380 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002381 cachep->ctor = ctor;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002382 cachep->name = name;
2383
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002384 if (setup_cpu_cache(cachep)) {
2385 __kmem_cache_destroy(cachep);
2386 cachep = NULL;
2387 goto oops;
2388 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002389
Linus Torvalds1da177e2005-04-16 15:20:36 -07002390 /* cache setup completed, link it into the list */
2391 list_add(&cachep->next, &cache_chain);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002392oops:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002393 if (!cachep && (flags & SLAB_PANIC))
2394 panic("kmem_cache_create(): failed to create slab `%s'\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002395 name);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002396 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002397 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002398 return cachep;
2399}
2400EXPORT_SYMBOL(kmem_cache_create);
2401
2402#if DEBUG
2403static void check_irq_off(void)
2404{
2405 BUG_ON(!irqs_disabled());
2406}
2407
2408static void check_irq_on(void)
2409{
2410 BUG_ON(irqs_disabled());
2411}
2412
Pekka Enberg343e0d72006-02-01 03:05:50 -08002413static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002414{
2415#ifdef CONFIG_SMP
2416 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002417 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002418#endif
2419}
Christoph Lametere498be72005-09-09 13:03:32 -07002420
Pekka Enberg343e0d72006-02-01 03:05:50 -08002421static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002422{
2423#ifdef CONFIG_SMP
2424 check_irq_off();
2425 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2426#endif
2427}
2428
Linus Torvalds1da177e2005-04-16 15:20:36 -07002429#else
2430#define check_irq_off() do { } while(0)
2431#define check_irq_on() do { } while(0)
2432#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002433#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002434#endif
2435
Christoph Lameteraab22072006-03-22 00:09:06 -08002436static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2437 struct array_cache *ac,
2438 int force, int node);
2439
Linus Torvalds1da177e2005-04-16 15:20:36 -07002440static void do_drain(void *arg)
2441{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002442 struct kmem_cache *cachep = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002443 struct array_cache *ac;
Christoph Lameterff694162005-09-22 21:44:02 -07002444 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002445
2446 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002447 ac = cpu_cache_get(cachep);
Christoph Lameterff694162005-09-22 21:44:02 -07002448 spin_lock(&cachep->nodelists[node]->list_lock);
2449 free_block(cachep, ac->entry, ac->avail, node);
2450 spin_unlock(&cachep->nodelists[node]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002451 ac->avail = 0;
2452}
2453
Pekka Enberg343e0d72006-02-01 03:05:50 -08002454static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002455{
Christoph Lametere498be72005-09-09 13:03:32 -07002456 struct kmem_list3 *l3;
2457 int node;
2458
Andrew Mortona07fa392006-03-22 00:08:17 -08002459 on_each_cpu(do_drain, cachep, 1, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002460 check_irq_on();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002461 for_each_online_node(node) {
Christoph Lametere498be72005-09-09 13:03:32 -07002462 l3 = cachep->nodelists[node];
Roland Dreiera4523a82006-05-15 11:41:00 -07002463 if (l3 && l3->alien)
2464 drain_alien_cache(cachep, l3->alien);
2465 }
2466
2467 for_each_online_node(node) {
2468 l3 = cachep->nodelists[node];
2469 if (l3)
Christoph Lameteraab22072006-03-22 00:09:06 -08002470 drain_array(cachep, l3, l3->shared, 1, node);
Christoph Lametere498be72005-09-09 13:03:32 -07002471 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002472}
2473
Christoph Lametered11d9e2006-06-30 01:55:45 -07002474/*
2475 * Remove slabs from the list of free slabs.
2476 * Specify the number of slabs to drain in tofree.
2477 *
2478 * Returns the actual number of slabs released.
2479 */
2480static int drain_freelist(struct kmem_cache *cache,
2481 struct kmem_list3 *l3, int tofree)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002482{
Christoph Lametered11d9e2006-06-30 01:55:45 -07002483 struct list_head *p;
2484 int nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002485 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002486
Christoph Lametered11d9e2006-06-30 01:55:45 -07002487 nr_freed = 0;
2488 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002489
Christoph Lametered11d9e2006-06-30 01:55:45 -07002490 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002491 p = l3->slabs_free.prev;
Christoph Lametered11d9e2006-06-30 01:55:45 -07002492 if (p == &l3->slabs_free) {
2493 spin_unlock_irq(&l3->list_lock);
2494 goto out;
2495 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002496
Christoph Lametered11d9e2006-06-30 01:55:45 -07002497 slabp = list_entry(p, struct slab, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002498#if DEBUG
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002499 BUG_ON(slabp->inuse);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002500#endif
2501 list_del(&slabp->list);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002502 /*
2503 * Safe to drop the lock. The slab is no longer linked
2504 * to the cache.
2505 */
2506 l3->free_objects -= cache->num;
Christoph Lametere498be72005-09-09 13:03:32 -07002507 spin_unlock_irq(&l3->list_lock);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002508 slab_destroy(cache, slabp);
2509 nr_freed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002510 }
Christoph Lametered11d9e2006-06-30 01:55:45 -07002511out:
2512 return nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002513}
2514
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002515/* Called with cache_chain_mutex held to protect against cpu hotplug */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002516static int __cache_shrink(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07002517{
2518 int ret = 0, i = 0;
2519 struct kmem_list3 *l3;
2520
2521 drain_cpu_caches(cachep);
2522
2523 check_irq_on();
2524 for_each_online_node(i) {
2525 l3 = cachep->nodelists[i];
Christoph Lametered11d9e2006-06-30 01:55:45 -07002526 if (!l3)
2527 continue;
2528
2529 drain_freelist(cachep, l3, l3->free_objects);
2530
2531 ret += !list_empty(&l3->slabs_full) ||
2532 !list_empty(&l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07002533 }
2534 return (ret ? 1 : 0);
2535}
2536
Linus Torvalds1da177e2005-04-16 15:20:36 -07002537/**
2538 * kmem_cache_shrink - Shrink a cache.
2539 * @cachep: The cache to shrink.
2540 *
2541 * Releases as many slabs as possible for a cache.
2542 * To help debugging, a zero exit status indicates all slabs were released.
2543 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002544int kmem_cache_shrink(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002545{
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002546 int ret;
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002547 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002548
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002549 get_online_cpus();
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002550 mutex_lock(&cache_chain_mutex);
2551 ret = __cache_shrink(cachep);
2552 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002553 put_online_cpus();
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002554 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002555}
2556EXPORT_SYMBOL(kmem_cache_shrink);
2557
2558/**
2559 * kmem_cache_destroy - delete a cache
2560 * @cachep: the cache to destroy
2561 *
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08002562 * Remove a &struct kmem_cache object from the slab cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002563 *
2564 * It is expected this function will be called by a module when it is
2565 * unloaded. This will remove the cache completely, and avoid a duplicate
2566 * cache being allocated each time a module is loaded and unloaded, if the
2567 * module doesn't have persistent in-kernel storage across loads and unloads.
2568 *
2569 * The cache must be empty before calling this function.
2570 *
2571 * The caller must guarantee that noone will allocate memory from the cache
2572 * during the kmem_cache_destroy().
2573 */
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002574void kmem_cache_destroy(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002575{
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002576 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002577
Linus Torvalds1da177e2005-04-16 15:20:36 -07002578 /* Find the cache in the chain of caches. */
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002579 get_online_cpus();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002580 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002581 /*
2582 * the chain is never empty, cache_cache is never destroyed
2583 */
2584 list_del(&cachep->next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002585 if (__cache_shrink(cachep)) {
2586 slab_error(cachep, "Can't free all objects");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002587 list_add(&cachep->next, &cache_chain);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002588 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002589 put_online_cpus();
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002590 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002591 }
2592
2593 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
Paul E. McKenneyfbd568a3e2005-05-01 08:59:04 -07002594 synchronize_rcu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002595
Christoph Lameter117f6eb2006-09-25 23:31:37 -07002596 __kmem_cache_destroy(cachep);
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002597 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002598 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002599}
2600EXPORT_SYMBOL(kmem_cache_destroy);
2601
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002602/*
2603 * Get the memory for a slab management obj.
2604 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2605 * always come from malloc_sizes caches. The slab descriptor cannot
2606 * come from the same cache which is getting created because,
2607 * when we are searching for an appropriate cache for these
2608 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2609 * If we are creating a malloc_sizes cache here it would not be visible to
2610 * kmem_find_general_cachep till the initialization is complete.
2611 * Hence we cannot have slabp_cache same as the original cache.
2612 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002613static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002614 int colour_off, gfp_t local_flags,
2615 int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002616{
2617 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002618
Linus Torvalds1da177e2005-04-16 15:20:36 -07002619 if (OFF_SLAB(cachep)) {
2620 /* Slab management obj is off-slab. */
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002621 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
Christoph Lameter3c517a62006-12-06 20:33:29 -08002622 local_flags & ~GFP_THISNODE, nodeid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002623 if (!slabp)
2624 return NULL;
2625 } else {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002626 slabp = objp + colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002627 colour_off += cachep->slab_size;
2628 }
2629 slabp->inuse = 0;
2630 slabp->colouroff = colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002631 slabp->s_mem = objp + colour_off;
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002632 slabp->nodeid = nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002633 return slabp;
2634}
2635
2636static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2637{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002638 return (kmem_bufctl_t *) (slabp + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002639}
2640
Pekka Enberg343e0d72006-02-01 03:05:50 -08002641static void cache_init_objs(struct kmem_cache *cachep,
Christoph Lametera35afb82007-05-16 22:10:57 -07002642 struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002643{
2644 int i;
2645
2646 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002647 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002648#if DEBUG
2649 /* need to poison the objs? */
2650 if (cachep->flags & SLAB_POISON)
2651 poison_obj(cachep, objp, POISON_FREE);
2652 if (cachep->flags & SLAB_STORE_USER)
2653 *dbg_userword(cachep, objp) = NULL;
2654
2655 if (cachep->flags & SLAB_RED_ZONE) {
2656 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2657 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2658 }
2659 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002660 * Constructors are not allowed to allocate memory from the same
2661 * cache which they are a constructor for. Otherwise, deadlock.
2662 * They must also be threaded.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002663 */
2664 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07002665 cachep->ctor(cachep, objp + obj_offset(cachep));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002666
2667 if (cachep->flags & SLAB_RED_ZONE) {
2668 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2669 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002670 " end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002671 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2672 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002673 " start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002674 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08002675 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2676 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002677 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002678 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002679#else
2680 if (cachep->ctor)
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07002681 cachep->ctor(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002682#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002683 slab_bufctl(slabp)[i] = i + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002684 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002685 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002686 slabp->free = 0;
2687}
2688
Pekka Enberg343e0d72006-02-01 03:05:50 -08002689static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002690{
Christoph Lameter4b51d662007-02-10 01:43:10 -08002691 if (CONFIG_ZONE_DMA_FLAG) {
2692 if (flags & GFP_DMA)
2693 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2694 else
2695 BUG_ON(cachep->gfpflags & GFP_DMA);
2696 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002697}
2698
Andrew Mortona737b3e2006-03-22 00:08:11 -08002699static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2700 int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002701{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002702 void *objp = index_to_obj(cachep, slabp, slabp->free);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002703 kmem_bufctl_t next;
2704
2705 slabp->inuse++;
2706 next = slab_bufctl(slabp)[slabp->free];
2707#if DEBUG
2708 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2709 WARN_ON(slabp->nodeid != nodeid);
2710#endif
2711 slabp->free = next;
2712
2713 return objp;
2714}
2715
Andrew Mortona737b3e2006-03-22 00:08:11 -08002716static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2717 void *objp, int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002718{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002719 unsigned int objnr = obj_to_index(cachep, slabp, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002720
2721#if DEBUG
2722 /* Verify that the slab belongs to the intended node */
2723 WARN_ON(slabp->nodeid != nodeid);
2724
Al Viro871751e2006-03-25 03:06:39 -08002725 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
Matthew Dobson78d382d2006-02-01 03:05:47 -08002726 printk(KERN_ERR "slab: double free detected in cache "
Andrew Mortona737b3e2006-03-22 00:08:11 -08002727 "'%s', objp %p\n", cachep->name, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002728 BUG();
2729 }
2730#endif
2731 slab_bufctl(slabp)[objnr] = slabp->free;
2732 slabp->free = objnr;
2733 slabp->inuse--;
2734}
2735
Pekka Enberg47768742006-06-23 02:03:07 -07002736/*
2737 * Map pages beginning at addr to the given cache and slab. This is required
2738 * for the slab allocator to be able to lookup the cache and slab of a
2739 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2740 */
2741static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2742 void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002743{
Pekka Enberg47768742006-06-23 02:03:07 -07002744 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002745 struct page *page;
2746
Pekka Enberg47768742006-06-23 02:03:07 -07002747 page = virt_to_page(addr);
Nick Piggin84097512006-03-22 00:08:34 -08002748
Pekka Enberg47768742006-06-23 02:03:07 -07002749 nr_pages = 1;
Nick Piggin84097512006-03-22 00:08:34 -08002750 if (likely(!PageCompound(page)))
Pekka Enberg47768742006-06-23 02:03:07 -07002751 nr_pages <<= cache->gfporder;
2752
Linus Torvalds1da177e2005-04-16 15:20:36 -07002753 do {
Pekka Enberg47768742006-06-23 02:03:07 -07002754 page_set_cache(page, cache);
2755 page_set_slab(page, slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002756 page++;
Pekka Enberg47768742006-06-23 02:03:07 -07002757 } while (--nr_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002758}
2759
2760/*
2761 * Grow (by 1) the number of slabs within a cache. This is called by
2762 * kmem_cache_alloc() when there are no active objs left in a cache.
2763 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002764static int cache_grow(struct kmem_cache *cachep,
2765 gfp_t flags, int nodeid, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002766{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002767 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002768 size_t offset;
2769 gfp_t local_flags;
Christoph Lametere498be72005-09-09 13:03:32 -07002770 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002771
Andrew Mortona737b3e2006-03-22 00:08:11 -08002772 /*
2773 * Be lazy and only check for valid flags here, keeping it out of the
2774 * critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002775 */
Christoph Lameter6cb06222007-10-16 01:25:41 -07002776 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2777 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002778
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002779 /* Take the l3 list lock to change the colour_next on this node */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002780 check_irq_off();
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002781 l3 = cachep->nodelists[nodeid];
2782 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002783
2784 /* Get colour for the slab, and cal the next value. */
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002785 offset = l3->colour_next;
2786 l3->colour_next++;
2787 if (l3->colour_next >= cachep->colour)
2788 l3->colour_next = 0;
2789 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002790
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002791 offset *= cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002792
2793 if (local_flags & __GFP_WAIT)
2794 local_irq_enable();
2795
2796 /*
2797 * The test for missing atomic flag is performed here, rather than
2798 * the more obvious place, simply to reduce the critical path length
2799 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2800 * will eventually be caught here (where it matters).
2801 */
2802 kmem_flagcheck(cachep, flags);
2803
Andrew Mortona737b3e2006-03-22 00:08:11 -08002804 /*
2805 * Get mem for the objs. Attempt to allocate a physical page from
2806 * 'nodeid'.
Christoph Lametere498be72005-09-09 13:03:32 -07002807 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002808 if (!objp)
Andrew Mortonb8c1c5d2007-07-24 12:02:40 -07002809 objp = kmem_getpages(cachep, local_flags, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002810 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002811 goto failed;
2812
2813 /* Get slab management. */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002814 slabp = alloc_slabmgmt(cachep, objp, offset,
Christoph Lameter6cb06222007-10-16 01:25:41 -07002815 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002816 if (!slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002817 goto opps1;
2818
Christoph Lametere498be72005-09-09 13:03:32 -07002819 slabp->nodeid = nodeid;
Pekka Enberg47768742006-06-23 02:03:07 -07002820 slab_map_pages(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002821
Christoph Lametera35afb82007-05-16 22:10:57 -07002822 cache_init_objs(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002823
2824 if (local_flags & __GFP_WAIT)
2825 local_irq_disable();
2826 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002827 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002828
2829 /* Make slab active. */
Christoph Lametere498be72005-09-09 13:03:32 -07002830 list_add_tail(&slabp->list, &(l3->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002831 STATS_INC_GROWN(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002832 l3->free_objects += cachep->num;
2833 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002834 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002835opps1:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002836 kmem_freepages(cachep, objp);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002837failed:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002838 if (local_flags & __GFP_WAIT)
2839 local_irq_disable();
2840 return 0;
2841}
2842
2843#if DEBUG
2844
2845/*
2846 * Perform extra freeing checks:
2847 * - detect bad pointers.
2848 * - POISON/RED_ZONE checking
Linus Torvalds1da177e2005-04-16 15:20:36 -07002849 */
2850static void kfree_debugcheck(const void *objp)
2851{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002852 if (!virt_addr_valid(objp)) {
2853 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002854 (unsigned long)objp);
2855 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002856 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002857}
2858
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002859static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2860{
David Woodhouseb46b8f12007-05-08 00:22:59 -07002861 unsigned long long redzone1, redzone2;
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002862
2863 redzone1 = *dbg_redzone1(cache, obj);
2864 redzone2 = *dbg_redzone2(cache, obj);
2865
2866 /*
2867 * Redzone is ok.
2868 */
2869 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2870 return;
2871
2872 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2873 slab_error(cache, "double free detected");
2874 else
2875 slab_error(cache, "memory outside object was overwritten");
2876
David Woodhouseb46b8f12007-05-08 00:22:59 -07002877 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002878 obj, redzone1, redzone2);
2879}
2880
Pekka Enberg343e0d72006-02-01 03:05:50 -08002881static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002882 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002883{
2884 struct page *page;
2885 unsigned int objnr;
2886 struct slab *slabp;
2887
Matthew Wilcox80cbd912007-11-29 12:05:13 -07002888 BUG_ON(virt_to_cache(objp) != cachep);
2889
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002890 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002891 kfree_debugcheck(objp);
Christoph Lameterb49af682007-05-06 14:49:41 -07002892 page = virt_to_head_page(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002893
Pekka Enberg065d41c2005-11-13 16:06:46 -08002894 slabp = page_get_slab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002895
2896 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002897 verify_redzone_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002898 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2899 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2900 }
2901 if (cachep->flags & SLAB_STORE_USER)
2902 *dbg_userword(cachep, objp) = caller;
2903
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002904 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002905
2906 BUG_ON(objnr >= cachep->num);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002907 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002908
Al Viro871751e2006-03-25 03:06:39 -08002909#ifdef CONFIG_DEBUG_SLAB_LEAK
2910 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2911#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002912 if (cachep->flags & SLAB_POISON) {
2913#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08002914 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002915 store_stackinfo(cachep, objp, (unsigned long)caller);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002916 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002917 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002918 } else {
2919 poison_obj(cachep, objp, POISON_FREE);
2920 }
2921#else
2922 poison_obj(cachep, objp, POISON_FREE);
2923#endif
2924 }
2925 return objp;
2926}
2927
Pekka Enberg343e0d72006-02-01 03:05:50 -08002928static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002929{
2930 kmem_bufctl_t i;
2931 int entries = 0;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002932
Linus Torvalds1da177e2005-04-16 15:20:36 -07002933 /* Check slab's freelist to see if this obj is there. */
2934 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2935 entries++;
2936 if (entries > cachep->num || i >= cachep->num)
2937 goto bad;
2938 }
2939 if (entries != cachep->num - slabp->inuse) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002940bad:
2941 printk(KERN_ERR "slab: Internal list corruption detected in "
2942 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2943 cachep->name, cachep->num, slabp, slabp->inuse);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002944 for (i = 0;
Linus Torvalds264132b2006-03-06 12:10:07 -08002945 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002946 i++) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002947 if (i % 16 == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002948 printk("\n%03x:", i);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002949 printk(" %02x", ((unsigned char *)slabp)[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002950 }
2951 printk("\n");
2952 BUG();
2953 }
2954}
2955#else
2956#define kfree_debugcheck(x) do { } while(0)
2957#define cache_free_debugcheck(x,objp,z) (objp)
2958#define check_slabp(x,y) do { } while(0)
2959#endif
2960
Pekka Enberg343e0d72006-02-01 03:05:50 -08002961static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002962{
2963 int batchcount;
2964 struct kmem_list3 *l3;
2965 struct array_cache *ac;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002966 int node;
2967
2968 node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002969
2970 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002971 ac = cpu_cache_get(cachep);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002972retry:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002973 batchcount = ac->batchcount;
2974 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002975 /*
2976 * If there was little recent activity on this cache, then
2977 * perform only a partial refill. Otherwise we could generate
2978 * refill bouncing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002979 */
2980 batchcount = BATCHREFILL_LIMIT;
2981 }
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002982 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002983
Christoph Lametere498be72005-09-09 13:03:32 -07002984 BUG_ON(ac->avail > 0 || !l3);
2985 spin_lock(&l3->list_lock);
2986
Christoph Lameter3ded1752006-03-25 03:06:44 -08002987 /* See if we can refill from the shared array */
2988 if (l3->shared && transfer_objects(ac, l3->shared, batchcount))
2989 goto alloc_done;
2990
Linus Torvalds1da177e2005-04-16 15:20:36 -07002991 while (batchcount > 0) {
2992 struct list_head *entry;
2993 struct slab *slabp;
2994 /* Get slab alloc is to come from. */
2995 entry = l3->slabs_partial.next;
2996 if (entry == &l3->slabs_partial) {
2997 l3->free_touched = 1;
2998 entry = l3->slabs_free.next;
2999 if (entry == &l3->slabs_free)
3000 goto must_grow;
3001 }
3002
3003 slabp = list_entry(entry, struct slab, list);
3004 check_slabp(cachep, slabp);
3005 check_spinlock_acquired(cachep);
Pekka Enberg714b8172007-05-06 14:49:03 -07003006
3007 /*
3008 * The slab was either on partial or free list so
3009 * there must be at least one object available for
3010 * allocation.
3011 */
3012 BUG_ON(slabp->inuse < 0 || slabp->inuse >= cachep->num);
3013
Linus Torvalds1da177e2005-04-16 15:20:36 -07003014 while (slabp->inuse < cachep->num && batchcount--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003015 STATS_INC_ALLOCED(cachep);
3016 STATS_INC_ACTIVE(cachep);
3017 STATS_SET_HIGH(cachep);
3018
Matthew Dobson78d382d2006-02-01 03:05:47 -08003019 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07003020 node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003021 }
3022 check_slabp(cachep, slabp);
3023
3024 /* move slabp to correct slabp list: */
3025 list_del(&slabp->list);
3026 if (slabp->free == BUFCTL_END)
3027 list_add(&slabp->list, &l3->slabs_full);
3028 else
3029 list_add(&slabp->list, &l3->slabs_partial);
3030 }
3031
Andrew Mortona737b3e2006-03-22 00:08:11 -08003032must_grow:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003033 l3->free_objects -= ac->avail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003034alloc_done:
Christoph Lametere498be72005-09-09 13:03:32 -07003035 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003036
3037 if (unlikely(!ac->avail)) {
3038 int x;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003039 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07003040
Andrew Mortona737b3e2006-03-22 00:08:11 -08003041 /* cache_grow can reenable interrupts, then ac could change. */
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003042 ac = cpu_cache_get(cachep);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003043 if (!x && ac->avail == 0) /* no objects in sight? abort */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003044 return NULL;
3045
Andrew Mortona737b3e2006-03-22 00:08:11 -08003046 if (!ac->avail) /* objects refilled by interrupt? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003047 goto retry;
3048 }
3049 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07003050 return ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003051}
3052
Andrew Mortona737b3e2006-03-22 00:08:11 -08003053static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3054 gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003055{
3056 might_sleep_if(flags & __GFP_WAIT);
3057#if DEBUG
3058 kmem_flagcheck(cachep, flags);
3059#endif
3060}
3061
3062#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003063static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3064 gfp_t flags, void *objp, void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003065{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003066 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003067 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003068 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003069#ifdef CONFIG_DEBUG_PAGEALLOC
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003070 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003071 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003072 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003073 else
3074 check_poison_obj(cachep, objp);
3075#else
3076 check_poison_obj(cachep, objp);
3077#endif
3078 poison_obj(cachep, objp, POISON_INUSE);
3079 }
3080 if (cachep->flags & SLAB_STORE_USER)
3081 *dbg_userword(cachep, objp) = caller;
3082
3083 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003084 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3085 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3086 slab_error(cachep, "double free, or memory outside"
3087 " object was overwritten");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003088 printk(KERN_ERR
David Woodhouseb46b8f12007-05-08 00:22:59 -07003089 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08003090 objp, *dbg_redzone1(cachep, objp),
3091 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003092 }
3093 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3094 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3095 }
Al Viro871751e2006-03-25 03:06:39 -08003096#ifdef CONFIG_DEBUG_SLAB_LEAK
3097 {
3098 struct slab *slabp;
3099 unsigned objnr;
3100
Christoph Lameterb49af682007-05-06 14:49:41 -07003101 slabp = page_get_slab(virt_to_head_page(objp));
Al Viro871751e2006-03-25 03:06:39 -08003102 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3103 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3104 }
3105#endif
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003106 objp += obj_offset(cachep);
Christoph Lameter4f104932007-05-06 14:50:17 -07003107 if (cachep->ctor && cachep->flags & SLAB_POISON)
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07003108 cachep->ctor(cachep, objp);
Kevin Hilmana44b56d2006-12-06 20:32:11 -08003109#if ARCH_SLAB_MINALIGN
3110 if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) {
3111 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3112 objp, ARCH_SLAB_MINALIGN);
3113 }
3114#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003115 return objp;
3116}
3117#else
3118#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3119#endif
3120
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003121#ifdef CONFIG_FAILSLAB
3122
3123static struct failslab_attr {
3124
3125 struct fault_attr attr;
3126
3127 u32 ignore_gfp_wait;
3128#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3129 struct dentry *ignore_gfp_wait_file;
3130#endif
3131
3132} failslab = {
3133 .attr = FAULT_ATTR_INITIALIZER,
Don Mullis6b1b60f2006-12-08 02:39:53 -08003134 .ignore_gfp_wait = 1,
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003135};
3136
3137static int __init setup_failslab(char *str)
3138{
3139 return setup_fault_attr(&failslab.attr, str);
3140}
3141__setup("failslab=", setup_failslab);
3142
3143static int should_failslab(struct kmem_cache *cachep, gfp_t flags)
3144{
3145 if (cachep == &cache_cache)
3146 return 0;
3147 if (flags & __GFP_NOFAIL)
3148 return 0;
3149 if (failslab.ignore_gfp_wait && (flags & __GFP_WAIT))
3150 return 0;
3151
3152 return should_fail(&failslab.attr, obj_size(cachep));
3153}
3154
3155#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3156
3157static int __init failslab_debugfs(void)
3158{
3159 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
3160 struct dentry *dir;
3161 int err;
3162
Akinobu Mita824ebef2007-05-06 14:49:58 -07003163 err = init_fault_attr_dentries(&failslab.attr, "failslab");
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003164 if (err)
3165 return err;
3166 dir = failslab.attr.dentries.dir;
3167
3168 failslab.ignore_gfp_wait_file =
3169 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3170 &failslab.ignore_gfp_wait);
3171
3172 if (!failslab.ignore_gfp_wait_file) {
3173 err = -ENOMEM;
3174 debugfs_remove(failslab.ignore_gfp_wait_file);
3175 cleanup_fault_attr_dentries(&failslab.attr);
3176 }
3177
3178 return err;
3179}
3180
3181late_initcall(failslab_debugfs);
3182
3183#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3184
3185#else /* CONFIG_FAILSLAB */
3186
3187static inline int should_failslab(struct kmem_cache *cachep, gfp_t flags)
3188{
3189 return 0;
3190}
3191
3192#endif /* CONFIG_FAILSLAB */
3193
Pekka Enberg343e0d72006-02-01 03:05:50 -08003194static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003195{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003196 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003197 struct array_cache *ac;
3198
Alok N Kataria5c382302005-09-27 21:45:46 -07003199 check_irq_off();
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003200
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003201 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003202 if (likely(ac->avail)) {
3203 STATS_INC_ALLOCHIT(cachep);
3204 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07003205 objp = ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003206 } else {
3207 STATS_INC_ALLOCMISS(cachep);
3208 objp = cache_alloc_refill(cachep, flags);
3209 }
Alok N Kataria5c382302005-09-27 21:45:46 -07003210 return objp;
3211}
3212
Christoph Lametere498be72005-09-09 13:03:32 -07003213#ifdef CONFIG_NUMA
3214/*
Paul Jacksonb2455392006-03-24 03:16:12 -08003215 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
Paul Jacksonc61afb12006-03-24 03:16:08 -08003216 *
3217 * If we are in_interrupt, then process context, including cpusets and
3218 * mempolicy, may not apply and should not be used for allocation policy.
3219 */
3220static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3221{
3222 int nid_alloc, nid_here;
3223
Christoph Lameter765c4502006-09-27 01:50:08 -07003224 if (in_interrupt() || (flags & __GFP_THISNODE))
Paul Jacksonc61afb12006-03-24 03:16:08 -08003225 return NULL;
3226 nid_alloc = nid_here = numa_node_id();
3227 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3228 nid_alloc = cpuset_mem_spread_node();
3229 else if (current->mempolicy)
3230 nid_alloc = slab_node(current->mempolicy);
3231 if (nid_alloc != nid_here)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003232 return ____cache_alloc_node(cachep, flags, nid_alloc);
Paul Jacksonc61afb12006-03-24 03:16:08 -08003233 return NULL;
3234}
3235
3236/*
Christoph Lameter765c4502006-09-27 01:50:08 -07003237 * Fallback function if there was no memory available and no objects on a
Christoph Lameter3c517a62006-12-06 20:33:29 -08003238 * certain node and fall back is permitted. First we scan all the
3239 * available nodelists for available objects. If that fails then we
3240 * perform an allocation without specifying a node. This allows the page
3241 * allocator to do its reclaim / fallback magic. We then insert the
3242 * slab into the proper nodelist and then allocate from it.
Christoph Lameter765c4502006-09-27 01:50:08 -07003243 */
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003244static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
Christoph Lameter765c4502006-09-27 01:50:08 -07003245{
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003246 struct zonelist *zonelist;
3247 gfp_t local_flags;
Christoph Lameter765c4502006-09-27 01:50:08 -07003248 struct zone **z;
3249 void *obj = NULL;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003250 int nid;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003251
3252 if (flags & __GFP_THISNODE)
3253 return NULL;
3254
3255 zonelist = &NODE_DATA(slab_node(current->mempolicy))
3256 ->node_zonelists[gfp_zone(flags)];
Christoph Lameter6cb06222007-10-16 01:25:41 -07003257 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Christoph Lameter765c4502006-09-27 01:50:08 -07003258
Christoph Lameter3c517a62006-12-06 20:33:29 -08003259retry:
3260 /*
3261 * Look through allowed nodes for objects available
3262 * from existing per node queues.
3263 */
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003264 for (z = zonelist->zones; *z && !obj; z++) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003265 nid = zone_to_nid(*z);
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003266
Paul Jackson02a0e532006-12-13 00:34:25 -08003267 if (cpuset_zone_allowed_hardwall(*z, flags) &&
Christoph Lameter3c517a62006-12-06 20:33:29 -08003268 cache->nodelists[nid] &&
3269 cache->nodelists[nid]->free_objects)
3270 obj = ____cache_alloc_node(cache,
3271 flags | GFP_THISNODE, nid);
3272 }
3273
Christoph Lametercfce6602007-05-06 14:50:17 -07003274 if (!obj) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003275 /*
3276 * This allocation will be performed within the constraints
3277 * of the current cpuset / memory policy requirements.
3278 * We may trigger various forms of reclaim on the allowed
3279 * set and go into memory reserves if necessary.
3280 */
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003281 if (local_flags & __GFP_WAIT)
3282 local_irq_enable();
3283 kmem_flagcheck(cache, flags);
Christoph Lameter3c517a62006-12-06 20:33:29 -08003284 obj = kmem_getpages(cache, flags, -1);
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003285 if (local_flags & __GFP_WAIT)
3286 local_irq_disable();
Christoph Lameter3c517a62006-12-06 20:33:29 -08003287 if (obj) {
3288 /*
3289 * Insert into the appropriate per node queues
3290 */
3291 nid = page_to_nid(virt_to_page(obj));
3292 if (cache_grow(cache, flags, nid, obj)) {
3293 obj = ____cache_alloc_node(cache,
3294 flags | GFP_THISNODE, nid);
3295 if (!obj)
3296 /*
3297 * Another processor may allocate the
3298 * objects in the slab since we are
3299 * not holding any locks.
3300 */
3301 goto retry;
3302 } else {
Hugh Dickinsb6a60452007-01-05 16:36:36 -08003303 /* cache_grow already freed obj */
Christoph Lameter3c517a62006-12-06 20:33:29 -08003304 obj = NULL;
3305 }
3306 }
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003307 }
Christoph Lameter765c4502006-09-27 01:50:08 -07003308 return obj;
3309}
3310
3311/*
Christoph Lametere498be72005-09-09 13:03:32 -07003312 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07003313 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003314static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003315 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07003316{
3317 struct list_head *entry;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003318 struct slab *slabp;
3319 struct kmem_list3 *l3;
3320 void *obj;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003321 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003322
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003323 l3 = cachep->nodelists[nodeid];
3324 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07003325
Andrew Mortona737b3e2006-03-22 00:08:11 -08003326retry:
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003327 check_irq_off();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003328 spin_lock(&l3->list_lock);
3329 entry = l3->slabs_partial.next;
3330 if (entry == &l3->slabs_partial) {
3331 l3->free_touched = 1;
3332 entry = l3->slabs_free.next;
3333 if (entry == &l3->slabs_free)
3334 goto must_grow;
3335 }
Christoph Lametere498be72005-09-09 13:03:32 -07003336
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003337 slabp = list_entry(entry, struct slab, list);
3338 check_spinlock_acquired_node(cachep, nodeid);
3339 check_slabp(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07003340
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003341 STATS_INC_NODEALLOCS(cachep);
3342 STATS_INC_ACTIVE(cachep);
3343 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003344
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003345 BUG_ON(slabp->inuse == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07003346
Matthew Dobson78d382d2006-02-01 03:05:47 -08003347 obj = slab_get_obj(cachep, slabp, nodeid);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003348 check_slabp(cachep, slabp);
3349 l3->free_objects--;
3350 /* move slabp to correct slabp list: */
3351 list_del(&slabp->list);
Christoph Lametere498be72005-09-09 13:03:32 -07003352
Andrew Mortona737b3e2006-03-22 00:08:11 -08003353 if (slabp->free == BUFCTL_END)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003354 list_add(&slabp->list, &l3->slabs_full);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003355 else
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003356 list_add(&slabp->list, &l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07003357
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003358 spin_unlock(&l3->list_lock);
3359 goto done;
Christoph Lametere498be72005-09-09 13:03:32 -07003360
Andrew Mortona737b3e2006-03-22 00:08:11 -08003361must_grow:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003362 spin_unlock(&l3->list_lock);
Christoph Lameter3c517a62006-12-06 20:33:29 -08003363 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
Christoph Lameter765c4502006-09-27 01:50:08 -07003364 if (x)
3365 goto retry;
Christoph Lametere498be72005-09-09 13:03:32 -07003366
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003367 return fallback_alloc(cachep, flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003368
Andrew Mortona737b3e2006-03-22 00:08:11 -08003369done:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003370 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07003371}
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003372
3373/**
3374 * kmem_cache_alloc_node - Allocate an object on the specified node
3375 * @cachep: The cache to allocate from.
3376 * @flags: See kmalloc().
3377 * @nodeid: node number of the target node.
3378 * @caller: return address of caller, used for debug information
3379 *
3380 * Identical to kmem_cache_alloc but it will allocate memory on the given
3381 * node, which can improve the performance for cpu bound structures.
3382 *
3383 * Fallback to other node is possible if __GFP_THISNODE is not set.
3384 */
3385static __always_inline void *
3386__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3387 void *caller)
3388{
3389 unsigned long save_flags;
3390 void *ptr;
3391
Akinobu Mita824ebef2007-05-06 14:49:58 -07003392 if (should_failslab(cachep, flags))
3393 return NULL;
3394
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003395 cache_alloc_debugcheck_before(cachep, flags);
3396 local_irq_save(save_flags);
3397
3398 if (unlikely(nodeid == -1))
3399 nodeid = numa_node_id();
3400
3401 if (unlikely(!cachep->nodelists[nodeid])) {
3402 /* Node not bootstrapped yet */
3403 ptr = fallback_alloc(cachep, flags);
3404 goto out;
3405 }
3406
3407 if (nodeid == numa_node_id()) {
3408 /*
3409 * Use the locally cached objects if possible.
3410 * However ____cache_alloc does not allow fallback
3411 * to other nodes. It may fail while we still have
3412 * objects on other nodes available.
3413 */
3414 ptr = ____cache_alloc(cachep, flags);
3415 if (ptr)
3416 goto out;
3417 }
3418 /* ___cache_alloc_node can fall back to other nodes */
3419 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3420 out:
3421 local_irq_restore(save_flags);
3422 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3423
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003424 if (unlikely((flags & __GFP_ZERO) && ptr))
3425 memset(ptr, 0, obj_size(cachep));
3426
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003427 return ptr;
3428}
3429
3430static __always_inline void *
3431__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3432{
3433 void *objp;
3434
3435 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3436 objp = alternate_node_alloc(cache, flags);
3437 if (objp)
3438 goto out;
3439 }
3440 objp = ____cache_alloc(cache, flags);
3441
3442 /*
3443 * We may just have run out of memory on the local node.
3444 * ____cache_alloc_node() knows how to locate memory on other nodes
3445 */
3446 if (!objp)
3447 objp = ____cache_alloc_node(cache, flags, numa_node_id());
3448
3449 out:
3450 return objp;
3451}
3452#else
3453
3454static __always_inline void *
3455__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3456{
3457 return ____cache_alloc(cachep, flags);
3458}
3459
3460#endif /* CONFIG_NUMA */
3461
3462static __always_inline void *
3463__cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3464{
3465 unsigned long save_flags;
3466 void *objp;
3467
Akinobu Mita824ebef2007-05-06 14:49:58 -07003468 if (should_failslab(cachep, flags))
3469 return NULL;
3470
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003471 cache_alloc_debugcheck_before(cachep, flags);
3472 local_irq_save(save_flags);
3473 objp = __do_cache_alloc(cachep, flags);
3474 local_irq_restore(save_flags);
3475 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3476 prefetchw(objp);
3477
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003478 if (unlikely((flags & __GFP_ZERO) && objp))
3479 memset(objp, 0, obj_size(cachep));
3480
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003481 return objp;
3482}
Christoph Lametere498be72005-09-09 13:03:32 -07003483
3484/*
3485 * Caller needs to acquire correct kmem_list's list_lock
3486 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003487static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003488 int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003489{
3490 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07003491 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003492
3493 for (i = 0; i < nr_objects; i++) {
3494 void *objp = objpp[i];
3495 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003496
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08003497 slabp = virt_to_slab(objp);
Christoph Lameterff694162005-09-22 21:44:02 -07003498 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003499 list_del(&slabp->list);
Christoph Lameterff694162005-09-22 21:44:02 -07003500 check_spinlock_acquired_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003501 check_slabp(cachep, slabp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08003502 slab_put_obj(cachep, slabp, objp, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003503 STATS_DEC_ACTIVE(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003504 l3->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003505 check_slabp(cachep, slabp);
3506
3507 /* fixup slab chains */
3508 if (slabp->inuse == 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07003509 if (l3->free_objects > l3->free_limit) {
3510 l3->free_objects -= cachep->num;
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07003511 /* No need to drop any previously held
3512 * lock here, even if we have a off-slab slab
3513 * descriptor it is guaranteed to come from
3514 * a different cache, refer to comments before
3515 * alloc_slabmgmt.
3516 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003517 slab_destroy(cachep, slabp);
3518 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07003519 list_add(&slabp->list, &l3->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003520 }
3521 } else {
3522 /* Unconditionally move a slab to the end of the
3523 * partial list on free - maximum time for the
3524 * other objects to be freed, too.
3525 */
Christoph Lametere498be72005-09-09 13:03:32 -07003526 list_add_tail(&slabp->list, &l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003527 }
3528 }
3529}
3530
Pekka Enberg343e0d72006-02-01 03:05:50 -08003531static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003532{
3533 int batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07003534 struct kmem_list3 *l3;
Christoph Lameterff694162005-09-22 21:44:02 -07003535 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003536
3537 batchcount = ac->batchcount;
3538#if DEBUG
3539 BUG_ON(!batchcount || batchcount > ac->avail);
3540#endif
3541 check_irq_off();
Christoph Lameterff694162005-09-22 21:44:02 -07003542 l3 = cachep->nodelists[node];
Ingo Molnar873623d2006-07-13 14:44:38 +02003543 spin_lock(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003544 if (l3->shared) {
3545 struct array_cache *shared_array = l3->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003546 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003547 if (max) {
3548 if (batchcount > max)
3549 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07003550 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003551 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003552 shared_array->avail += batchcount;
3553 goto free_done;
3554 }
3555 }
3556
Christoph Lameterff694162005-09-22 21:44:02 -07003557 free_block(cachep, ac->entry, batchcount, node);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003558free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003559#if STATS
3560 {
3561 int i = 0;
3562 struct list_head *p;
3563
Christoph Lametere498be72005-09-09 13:03:32 -07003564 p = l3->slabs_free.next;
3565 while (p != &(l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003566 struct slab *slabp;
3567
3568 slabp = list_entry(p, struct slab, list);
3569 BUG_ON(slabp->inuse);
3570
3571 i++;
3572 p = p->next;
3573 }
3574 STATS_SET_FREEABLE(cachep, i);
3575 }
3576#endif
Christoph Lametere498be72005-09-09 13:03:32 -07003577 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003578 ac->avail -= batchcount;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003579 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003580}
3581
3582/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08003583 * Release an obj back to its cache. If the obj has a constructed state, it must
3584 * be in this state _before_ it is released. Called with disabled ints.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003585 */
Ingo Molnar873623d2006-07-13 14:44:38 +02003586static inline void __cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003587{
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003588 struct array_cache *ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003589
3590 check_irq_off();
3591 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3592
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -07003593 /*
3594 * Skip calling cache_free_alien() when the platform is not numa.
3595 * This will avoid cache misses that happen while accessing slabp (which
3596 * is per page memory reference) to get nodeid. Instead use a global
3597 * variable to skip the call, which is mostly likely to be present in
3598 * the cache.
3599 */
3600 if (numa_platform && cache_free_alien(cachep, objp))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07003601 return;
Christoph Lametere498be72005-09-09 13:03:32 -07003602
Linus Torvalds1da177e2005-04-16 15:20:36 -07003603 if (likely(ac->avail < ac->limit)) {
3604 STATS_INC_FREEHIT(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003605 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003606 return;
3607 } else {
3608 STATS_INC_FREEMISS(cachep);
3609 cache_flusharray(cachep, ac);
Christoph Lametere498be72005-09-09 13:03:32 -07003610 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003611 }
3612}
3613
3614/**
3615 * kmem_cache_alloc - Allocate an object
3616 * @cachep: The cache to allocate from.
3617 * @flags: See kmalloc().
3618 *
3619 * Allocate an object from this cache. The flags are only relevant
3620 * if the cache has no available objects.
3621 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003622void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003623{
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003624 return __cache_alloc(cachep, flags, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003625}
3626EXPORT_SYMBOL(kmem_cache_alloc);
3627
3628/**
3629 * kmem_ptr_validate - check if an untrusted pointer might
3630 * be a slab entry.
3631 * @cachep: the cache we're checking against
3632 * @ptr: pointer to validate
3633 *
3634 * This verifies that the untrusted pointer looks sane:
3635 * it is _not_ a guarantee that the pointer is actually
3636 * part of the slab cache in question, but it at least
3637 * validates that the pointer can be dereferenced and
3638 * looks half-way sane.
3639 *
3640 * Currently only used for dentry validation.
3641 */
Christoph Lameterb7f869a2006-12-22 01:06:44 -08003642int kmem_ptr_validate(struct kmem_cache *cachep, const void *ptr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003643{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003644 unsigned long addr = (unsigned long)ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003645 unsigned long min_addr = PAGE_OFFSET;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003646 unsigned long align_mask = BYTES_PER_WORD - 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003647 unsigned long size = cachep->buffer_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003648 struct page *page;
3649
3650 if (unlikely(addr < min_addr))
3651 goto out;
3652 if (unlikely(addr > (unsigned long)high_memory - size))
3653 goto out;
3654 if (unlikely(addr & align_mask))
3655 goto out;
3656 if (unlikely(!kern_addr_valid(addr)))
3657 goto out;
3658 if (unlikely(!kern_addr_valid(addr + size - 1)))
3659 goto out;
3660 page = virt_to_page(ptr);
3661 if (unlikely(!PageSlab(page)))
3662 goto out;
Pekka Enberg065d41c2005-11-13 16:06:46 -08003663 if (unlikely(page_get_cache(page) != cachep))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003664 goto out;
3665 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003666out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003667 return 0;
3668}
3669
3670#ifdef CONFIG_NUMA
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003671void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3672{
3673 return __cache_alloc_node(cachep, flags, nodeid,
3674 __builtin_return_address(0));
3675}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003676EXPORT_SYMBOL(kmem_cache_alloc_node);
3677
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003678static __always_inline void *
3679__do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003680{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003681 struct kmem_cache *cachep;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003682
3683 cachep = kmem_find_general_cachep(size, flags);
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003684 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3685 return cachep;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003686 return kmem_cache_alloc_node(cachep, flags, node);
3687}
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003688
3689#ifdef CONFIG_DEBUG_SLAB
3690void *__kmalloc_node(size_t size, gfp_t flags, int node)
3691{
3692 return __do_kmalloc_node(size, flags, node,
3693 __builtin_return_address(0));
3694}
Christoph Hellwigdbe5e692006-09-25 23:31:36 -07003695EXPORT_SYMBOL(__kmalloc_node);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003696
3697void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3698 int node, void *caller)
3699{
3700 return __do_kmalloc_node(size, flags, node, caller);
3701}
3702EXPORT_SYMBOL(__kmalloc_node_track_caller);
3703#else
3704void *__kmalloc_node(size_t size, gfp_t flags, int node)
3705{
3706 return __do_kmalloc_node(size, flags, node, NULL);
3707}
3708EXPORT_SYMBOL(__kmalloc_node);
3709#endif /* CONFIG_DEBUG_SLAB */
3710#endif /* CONFIG_NUMA */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003711
3712/**
Paul Drynoff800590f2006-06-23 02:03:48 -07003713 * __do_kmalloc - allocate memory
Linus Torvalds1da177e2005-04-16 15:20:36 -07003714 * @size: how many bytes of memory are required.
Paul Drynoff800590f2006-06-23 02:03:48 -07003715 * @flags: the type of memory to allocate (see kmalloc).
Randy Dunlap911851e2006-03-22 00:08:14 -08003716 * @caller: function caller for debug tracking of the caller
Linus Torvalds1da177e2005-04-16 15:20:36 -07003717 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003718static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3719 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003720{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003721 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003722
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003723 /* If you want to save a few bytes .text space: replace
3724 * __ with kmem_.
3725 * Then kmalloc uses the uninlined functions instead of the inline
3726 * functions.
3727 */
3728 cachep = __find_general_cachep(size, flags);
Linus Torvaldsa5c96d82007-07-19 13:17:15 -07003729 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3730 return cachep;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003731 return __cache_alloc(cachep, flags, caller);
3732}
3733
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003734
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003735#ifdef CONFIG_DEBUG_SLAB
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003736void *__kmalloc(size_t size, gfp_t flags)
3737{
Al Viro871751e2006-03-25 03:06:39 -08003738 return __do_kmalloc(size, flags, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003739}
3740EXPORT_SYMBOL(__kmalloc);
3741
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003742void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller)
3743{
3744 return __do_kmalloc(size, flags, caller);
3745}
3746EXPORT_SYMBOL(__kmalloc_track_caller);
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003747
3748#else
3749void *__kmalloc(size_t size, gfp_t flags)
3750{
3751 return __do_kmalloc(size, flags, NULL);
3752}
3753EXPORT_SYMBOL(__kmalloc);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003754#endif
3755
Linus Torvalds1da177e2005-04-16 15:20:36 -07003756/**
3757 * kmem_cache_free - Deallocate an object
3758 * @cachep: The cache the allocation was from.
3759 * @objp: The previously allocated object.
3760 *
3761 * Free an object which was previously allocated from this
3762 * cache.
3763 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003764void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003765{
3766 unsigned long flags;
3767
3768 local_irq_save(flags);
Ingo Molnar898552c2007-02-10 01:44:57 -08003769 debug_check_no_locks_freed(objp, obj_size(cachep));
Ingo Molnar873623d2006-07-13 14:44:38 +02003770 __cache_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003771 local_irq_restore(flags);
3772}
3773EXPORT_SYMBOL(kmem_cache_free);
3774
3775/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003776 * kfree - free previously allocated memory
3777 * @objp: pointer returned by kmalloc.
3778 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003779 * If @objp is NULL, no operation is performed.
3780 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003781 * Don't free memory not originally allocated by kmalloc()
3782 * or you will run into trouble.
3783 */
3784void kfree(const void *objp)
3785{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003786 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003787 unsigned long flags;
3788
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003789 if (unlikely(ZERO_OR_NULL_PTR(objp)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003790 return;
3791 local_irq_save(flags);
3792 kfree_debugcheck(objp);
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08003793 c = virt_to_cache(objp);
Ingo Molnarf9b84042006-06-27 02:54:49 -07003794 debug_check_no_locks_freed(objp, obj_size(c));
Ingo Molnar873623d2006-07-13 14:44:38 +02003795 __cache_free(c, (void *)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003796 local_irq_restore(flags);
3797}
3798EXPORT_SYMBOL(kfree);
3799
Pekka Enberg343e0d72006-02-01 03:05:50 -08003800unsigned int kmem_cache_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003801{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003802 return obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003803}
3804EXPORT_SYMBOL(kmem_cache_size);
3805
Pekka Enberg343e0d72006-02-01 03:05:50 -08003806const char *kmem_cache_name(struct kmem_cache *cachep)
Arnaldo Carvalho de Melo19449722005-06-18 22:46:19 -07003807{
3808 return cachep->name;
3809}
3810EXPORT_SYMBOL_GPL(kmem_cache_name);
3811
Christoph Lametere498be72005-09-09 13:03:32 -07003812/*
Simon Arlott183ff222007-10-20 01:27:18 +02003813 * This initializes kmem_list3 or resizes various caches for all nodes.
Christoph Lametere498be72005-09-09 13:03:32 -07003814 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003815static int alloc_kmemlist(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07003816{
3817 int node;
3818 struct kmem_list3 *l3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003819 struct array_cache *new_shared;
Paul Menage3395ee02006-12-06 20:32:16 -08003820 struct array_cache **new_alien = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003821
Mel Gorman9c09a952008-01-24 05:49:54 -08003822 for_each_online_node(node) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003823
Paul Menage3395ee02006-12-06 20:32:16 -08003824 if (use_alien_caches) {
3825 new_alien = alloc_alien_cache(node, cachep->limit);
3826 if (!new_alien)
3827 goto fail;
3828 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003829
Eric Dumazet63109842007-05-06 14:49:28 -07003830 new_shared = NULL;
3831 if (cachep->shared) {
3832 new_shared = alloc_arraycache(node,
Christoph Lameter0718dc22006-03-25 03:06:47 -08003833 cachep->shared*cachep->batchcount,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003834 0xbaadf00d);
Eric Dumazet63109842007-05-06 14:49:28 -07003835 if (!new_shared) {
3836 free_alien_cache(new_alien);
3837 goto fail;
3838 }
Christoph Lameter0718dc22006-03-25 03:06:47 -08003839 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003840
Andrew Mortona737b3e2006-03-22 00:08:11 -08003841 l3 = cachep->nodelists[node];
3842 if (l3) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003843 struct array_cache *shared = l3->shared;
3844
Christoph Lametere498be72005-09-09 13:03:32 -07003845 spin_lock_irq(&l3->list_lock);
3846
Christoph Lametercafeb022006-03-25 03:06:46 -08003847 if (shared)
Christoph Lameter0718dc22006-03-25 03:06:47 -08003848 free_block(cachep, shared->entry,
3849 shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07003850
Christoph Lametercafeb022006-03-25 03:06:46 -08003851 l3->shared = new_shared;
3852 if (!l3->alien) {
Christoph Lametere498be72005-09-09 13:03:32 -07003853 l3->alien = new_alien;
3854 new_alien = NULL;
3855 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003856 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003857 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003858 spin_unlock_irq(&l3->list_lock);
Christoph Lametercafeb022006-03-25 03:06:46 -08003859 kfree(shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003860 free_alien_cache(new_alien);
3861 continue;
3862 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08003863 l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node);
Christoph Lameter0718dc22006-03-25 03:06:47 -08003864 if (!l3) {
3865 free_alien_cache(new_alien);
3866 kfree(new_shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003867 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003868 }
Christoph Lametere498be72005-09-09 13:03:32 -07003869
3870 kmem_list3_init(l3);
3871 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Andrew Mortona737b3e2006-03-22 00:08:11 -08003872 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003873 l3->shared = new_shared;
Christoph Lametere498be72005-09-09 13:03:32 -07003874 l3->alien = new_alien;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003875 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003876 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003877 cachep->nodelists[node] = l3;
3878 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003879 return 0;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003880
Andrew Mortona737b3e2006-03-22 00:08:11 -08003881fail:
Christoph Lameter0718dc22006-03-25 03:06:47 -08003882 if (!cachep->next.next) {
3883 /* Cache is not active yet. Roll back what we did */
3884 node--;
3885 while (node >= 0) {
3886 if (cachep->nodelists[node]) {
3887 l3 = cachep->nodelists[node];
3888
3889 kfree(l3->shared);
3890 free_alien_cache(l3->alien);
3891 kfree(l3);
3892 cachep->nodelists[node] = NULL;
3893 }
3894 node--;
3895 }
3896 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003897 return -ENOMEM;
Christoph Lametere498be72005-09-09 13:03:32 -07003898}
3899
Linus Torvalds1da177e2005-04-16 15:20:36 -07003900struct ccupdate_struct {
Pekka Enberg343e0d72006-02-01 03:05:50 -08003901 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003902 struct array_cache *new[NR_CPUS];
3903};
3904
3905static void do_ccupdate_local(void *info)
3906{
Andrew Mortona737b3e2006-03-22 00:08:11 -08003907 struct ccupdate_struct *new = info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003908 struct array_cache *old;
3909
3910 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003911 old = cpu_cache_get(new->cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003912
Linus Torvalds1da177e2005-04-16 15:20:36 -07003913 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3914 new->new[smp_processor_id()] = old;
3915}
3916
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003917/* Always called with the cache_chain_mutex held */
Andrew Mortona737b3e2006-03-22 00:08:11 -08003918static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3919 int batchcount, int shared)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003920{
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003921 struct ccupdate_struct *new;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003922 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003923
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003924 new = kzalloc(sizeof(*new), GFP_KERNEL);
3925 if (!new)
3926 return -ENOMEM;
3927
Christoph Lametere498be72005-09-09 13:03:32 -07003928 for_each_online_cpu(i) {
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003929 new->new[i] = alloc_arraycache(cpu_to_node(i), limit,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003930 batchcount);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003931 if (!new->new[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003932 for (i--; i >= 0; i--)
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003933 kfree(new->new[i]);
3934 kfree(new);
Christoph Lametere498be72005-09-09 13:03:32 -07003935 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003936 }
3937 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003938 new->cachep = cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003939
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003940 on_each_cpu(do_ccupdate_local, (void *)new, 1, 1);
Christoph Lametere498be72005-09-09 13:03:32 -07003941
Linus Torvalds1da177e2005-04-16 15:20:36 -07003942 check_irq_on();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003943 cachep->batchcount = batchcount;
3944 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07003945 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003946
Christoph Lametere498be72005-09-09 13:03:32 -07003947 for_each_online_cpu(i) {
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003948 struct array_cache *ccold = new->new[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003949 if (!ccold)
3950 continue;
Christoph Lametere498be72005-09-09 13:03:32 -07003951 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -07003952 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
Christoph Lametere498be72005-09-09 13:03:32 -07003953 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003954 kfree(ccold);
3955 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003956 kfree(new);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003957 return alloc_kmemlist(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003958}
3959
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003960/* Called with cache_chain_mutex held always */
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003961static int enable_cpucache(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003962{
3963 int err;
3964 int limit, shared;
3965
Andrew Mortona737b3e2006-03-22 00:08:11 -08003966 /*
3967 * The head array serves three purposes:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003968 * - create a LIFO ordering, i.e. return objects that are cache-warm
3969 * - reduce the number of spinlock operations.
Andrew Mortona737b3e2006-03-22 00:08:11 -08003970 * - reduce the number of linked list operations on the slab and
Linus Torvalds1da177e2005-04-16 15:20:36 -07003971 * bufctl chains: array operations are cheaper.
3972 * The numbers are guessed, we should auto-tune as described by
3973 * Bonwick.
3974 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003975 if (cachep->buffer_size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003976 limit = 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003977 else if (cachep->buffer_size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003978 limit = 8;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003979 else if (cachep->buffer_size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003980 limit = 24;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003981 else if (cachep->buffer_size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003982 limit = 54;
3983 else
3984 limit = 120;
3985
Andrew Mortona737b3e2006-03-22 00:08:11 -08003986 /*
3987 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
Linus Torvalds1da177e2005-04-16 15:20:36 -07003988 * allocation behaviour: Most allocs on one cpu, most free operations
3989 * on another cpu. For these cases, an efficient object passing between
3990 * cpus is necessary. This is provided by a shared array. The array
3991 * replaces Bonwick's magazine layer.
3992 * On uniprocessor, it's functionally equivalent (but less efficient)
3993 * to a larger limit. Thus disabled by default.
3994 */
3995 shared = 0;
Eric Dumazet364fbb22007-05-06 14:49:27 -07003996 if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003997 shared = 8;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003998
3999#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08004000 /*
4001 * With debugging enabled, large batchcount lead to excessively long
4002 * periods with disabled local interrupts. Limit the batchcount
Linus Torvalds1da177e2005-04-16 15:20:36 -07004003 */
4004 if (limit > 32)
4005 limit = 32;
4006#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004007 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004008 if (err)
4009 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004010 cachep->name, -err);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07004011 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004012}
4013
Christoph Lameter1b552532006-03-22 00:09:07 -08004014/*
4015 * Drain an array if it contains any elements taking the l3 lock only if
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004016 * necessary. Note that the l3 listlock also protects the array_cache
4017 * if drain_array() is used on the shared array.
Christoph Lameter1b552532006-03-22 00:09:07 -08004018 */
4019void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
4020 struct array_cache *ac, int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004021{
4022 int tofree;
4023
Christoph Lameter1b552532006-03-22 00:09:07 -08004024 if (!ac || !ac->avail)
4025 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004026 if (ac->touched && !force) {
4027 ac->touched = 0;
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004028 } else {
Christoph Lameter1b552532006-03-22 00:09:07 -08004029 spin_lock_irq(&l3->list_lock);
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004030 if (ac->avail) {
4031 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4032 if (tofree > ac->avail)
4033 tofree = (ac->avail + 1) / 2;
4034 free_block(cachep, ac->entry, tofree, node);
4035 ac->avail -= tofree;
4036 memmove(ac->entry, &(ac->entry[tofree]),
4037 sizeof(void *) * ac->avail);
4038 }
Christoph Lameter1b552532006-03-22 00:09:07 -08004039 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004040 }
4041}
4042
4043/**
4044 * cache_reap - Reclaim memory from caches.
Randy Dunlap05fb6bf2007-02-28 20:12:13 -08004045 * @w: work descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07004046 *
4047 * Called from workqueue/eventd every few seconds.
4048 * Purpose:
4049 * - clear the per-cpu caches for this CPU.
4050 * - return freeable pages to the main free memory pool.
4051 *
Andrew Mortona737b3e2006-03-22 00:08:11 -08004052 * If we cannot acquire the cache chain mutex then just give up - we'll try
4053 * again on the next iteration.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004054 */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004055static void cache_reap(struct work_struct *w)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004056{
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004057 struct kmem_cache *searchp;
Christoph Lametere498be72005-09-09 13:03:32 -07004058 struct kmem_list3 *l3;
Christoph Lameteraab22072006-03-22 00:09:06 -08004059 int node = numa_node_id();
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004060 struct delayed_work *work =
4061 container_of(w, struct delayed_work, work);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004062
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004063 if (!mutex_trylock(&cache_chain_mutex))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004064 /* Give up. Setup the next iteration. */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004065 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004066
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004067 list_for_each_entry(searchp, &cache_chain, next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004068 check_irq_on();
4069
Christoph Lameter35386e32006-03-22 00:09:05 -08004070 /*
4071 * We only take the l3 lock if absolutely necessary and we
4072 * have established with reasonable certainty that
4073 * we can do some work if the lock was obtained.
4074 */
Christoph Lameteraab22072006-03-22 00:09:06 -08004075 l3 = searchp->nodelists[node];
Christoph Lameter35386e32006-03-22 00:09:05 -08004076
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004077 reap_alien(searchp, l3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004078
Christoph Lameteraab22072006-03-22 00:09:06 -08004079 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004080
Christoph Lameter35386e32006-03-22 00:09:05 -08004081 /*
4082 * These are racy checks but it does not matter
4083 * if we skip one check or scan twice.
4084 */
Christoph Lametere498be72005-09-09 13:03:32 -07004085 if (time_after(l3->next_reap, jiffies))
Christoph Lameter35386e32006-03-22 00:09:05 -08004086 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004087
Christoph Lametere498be72005-09-09 13:03:32 -07004088 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004089
Christoph Lameteraab22072006-03-22 00:09:06 -08004090 drain_array(searchp, l3, l3->shared, 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004091
Christoph Lametered11d9e2006-06-30 01:55:45 -07004092 if (l3->free_touched)
Christoph Lametere498be72005-09-09 13:03:32 -07004093 l3->free_touched = 0;
Christoph Lametered11d9e2006-06-30 01:55:45 -07004094 else {
4095 int freed;
4096
4097 freed = drain_freelist(searchp, l3, (l3->free_limit +
4098 5 * searchp->num - 1) / (5 * searchp->num));
4099 STATS_ADD_REAPED(searchp, freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004100 }
Christoph Lameter35386e32006-03-22 00:09:05 -08004101next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004102 cond_resched();
4103 }
4104 check_irq_on();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004105 mutex_unlock(&cache_chain_mutex);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004106 next_reap_node();
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004107out:
Andrew Mortona737b3e2006-03-22 00:08:11 -08004108 /* Set up the next iteration */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004109 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004110}
4111
Linus Torvalds158a9622008-01-02 13:04:48 -08004112#ifdef CONFIG_SLABINFO
Linus Torvalds1da177e2005-04-16 15:20:36 -07004113
Pekka Enberg85289f92006-01-08 01:00:36 -08004114static void print_slabinfo_header(struct seq_file *m)
4115{
4116 /*
4117 * Output format version, so at least we can change it
4118 * without _too_ many complaints.
4119 */
4120#if STATS
4121 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4122#else
4123 seq_puts(m, "slabinfo - version: 2.1\n");
4124#endif
4125 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4126 "<objperslab> <pagesperslab>");
4127 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4128 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4129#if STATS
4130 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004131 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
Pekka Enberg85289f92006-01-08 01:00:36 -08004132 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4133#endif
4134 seq_putc(m, '\n');
4135}
4136
Linus Torvalds1da177e2005-04-16 15:20:36 -07004137static void *s_start(struct seq_file *m, loff_t *pos)
4138{
4139 loff_t n = *pos;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004140
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004141 mutex_lock(&cache_chain_mutex);
Pekka Enberg85289f92006-01-08 01:00:36 -08004142 if (!n)
4143 print_slabinfo_header(m);
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004144
4145 return seq_list_start(&cache_chain, *pos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004146}
4147
4148static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4149{
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004150 return seq_list_next(p, &cache_chain, pos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004151}
4152
4153static void s_stop(struct seq_file *m, void *p)
4154{
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004155 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004156}
4157
4158static int s_show(struct seq_file *m, void *p)
4159{
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004160 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004161 struct slab *slabp;
4162 unsigned long active_objs;
4163 unsigned long num_objs;
4164 unsigned long active_slabs = 0;
4165 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004166 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004167 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07004168 int node;
4169 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004170
Linus Torvalds1da177e2005-04-16 15:20:36 -07004171 active_objs = 0;
4172 num_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004173 for_each_online_node(node) {
4174 l3 = cachep->nodelists[node];
4175 if (!l3)
4176 continue;
4177
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004178 check_irq_on();
4179 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07004180
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004181 list_for_each_entry(slabp, &l3->slabs_full, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004182 if (slabp->inuse != cachep->num && !error)
4183 error = "slabs_full accounting error";
4184 active_objs += cachep->num;
4185 active_slabs++;
4186 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004187 list_for_each_entry(slabp, &l3->slabs_partial, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004188 if (slabp->inuse == cachep->num && !error)
4189 error = "slabs_partial inuse accounting error";
4190 if (!slabp->inuse && !error)
4191 error = "slabs_partial/inuse accounting error";
4192 active_objs += slabp->inuse;
4193 active_slabs++;
4194 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004195 list_for_each_entry(slabp, &l3->slabs_free, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004196 if (slabp->inuse && !error)
4197 error = "slabs_free/inuse accounting error";
4198 num_slabs++;
4199 }
4200 free_objects += l3->free_objects;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08004201 if (l3->shared)
4202 shared_avail += l3->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07004203
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004204 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004205 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004206 num_slabs += active_slabs;
4207 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07004208 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004209 error = "free_objects accounting error";
4210
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004211 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004212 if (error)
4213 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4214
4215 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004216 name, active_objs, num_objs, cachep->buffer_size,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004217 cachep->num, (1 << cachep->gfporder));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004218 seq_printf(m, " : tunables %4u %4u %4u",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004219 cachep->limit, cachep->batchcount, cachep->shared);
Christoph Lametere498be72005-09-09 13:03:32 -07004220 seq_printf(m, " : slabdata %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004221 active_slabs, num_slabs, shared_avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004222#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004223 { /* list3 stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004224 unsigned long high = cachep->high_mark;
4225 unsigned long allocs = cachep->num_allocations;
4226 unsigned long grown = cachep->grown;
4227 unsigned long reaped = cachep->reaped;
4228 unsigned long errors = cachep->errors;
4229 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004230 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07004231 unsigned long node_frees = cachep->node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004232 unsigned long overflows = cachep->node_overflow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004233
Christoph Lametere498be72005-09-09 13:03:32 -07004234 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004235 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown,
Andrew Mortona737b3e2006-03-22 00:08:11 -08004236 reaped, errors, max_freeable, node_allocs,
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004237 node_frees, overflows);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004238 }
4239 /* cpu stats */
4240 {
4241 unsigned long allochit = atomic_read(&cachep->allochit);
4242 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4243 unsigned long freehit = atomic_read(&cachep->freehit);
4244 unsigned long freemiss = atomic_read(&cachep->freemiss);
4245
4246 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004247 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004248 }
4249#endif
4250 seq_putc(m, '\n');
Linus Torvalds1da177e2005-04-16 15:20:36 -07004251 return 0;
4252}
4253
4254/*
4255 * slabinfo_op - iterator that generates /proc/slabinfo
4256 *
4257 * Output layout:
4258 * cache-name
4259 * num-active-objs
4260 * total-objs
4261 * object size
4262 * num-active-slabs
4263 * total-slabs
4264 * num-pages-per-slab
4265 * + further values on SMP and with statistics enabled
4266 */
4267
Helge Deller15ad7cd2006-12-06 20:40:36 -08004268const struct seq_operations slabinfo_op = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004269 .start = s_start,
4270 .next = s_next,
4271 .stop = s_stop,
4272 .show = s_show,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004273};
4274
4275#define MAX_SLABINFO_WRITE 128
4276/**
4277 * slabinfo_write - Tuning for the slab allocator
4278 * @file: unused
4279 * @buffer: user buffer
4280 * @count: data length
4281 * @ppos: unused
4282 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004283ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4284 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004285{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004286 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004287 int limit, batchcount, shared, res;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004288 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004289
Linus Torvalds1da177e2005-04-16 15:20:36 -07004290 if (count > MAX_SLABINFO_WRITE)
4291 return -EINVAL;
4292 if (copy_from_user(&kbuf, buffer, count))
4293 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004294 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07004295
4296 tmp = strchr(kbuf, ' ');
4297 if (!tmp)
4298 return -EINVAL;
4299 *tmp = '\0';
4300 tmp++;
4301 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4302 return -EINVAL;
4303
4304 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004305 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004306 res = -EINVAL;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004307 list_for_each_entry(cachep, &cache_chain, next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004308 if (!strcmp(cachep->name, kbuf)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08004309 if (limit < 1 || batchcount < 1 ||
4310 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07004311 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004312 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07004313 res = do_tune_cpucache(cachep, limit,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004314 batchcount, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004315 }
4316 break;
4317 }
4318 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004319 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004320 if (res >= 0)
4321 res = count;
4322 return res;
4323}
Al Viro871751e2006-03-25 03:06:39 -08004324
4325#ifdef CONFIG_DEBUG_SLAB_LEAK
4326
4327static void *leaks_start(struct seq_file *m, loff_t *pos)
4328{
Al Viro871751e2006-03-25 03:06:39 -08004329 mutex_lock(&cache_chain_mutex);
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004330 return seq_list_start(&cache_chain, *pos);
Al Viro871751e2006-03-25 03:06:39 -08004331}
4332
4333static inline int add_caller(unsigned long *n, unsigned long v)
4334{
4335 unsigned long *p;
4336 int l;
4337 if (!v)
4338 return 1;
4339 l = n[1];
4340 p = n + 2;
4341 while (l) {
4342 int i = l/2;
4343 unsigned long *q = p + 2 * i;
4344 if (*q == v) {
4345 q[1]++;
4346 return 1;
4347 }
4348 if (*q > v) {
4349 l = i;
4350 } else {
4351 p = q + 2;
4352 l -= i + 1;
4353 }
4354 }
4355 if (++n[1] == n[0])
4356 return 0;
4357 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4358 p[0] = v;
4359 p[1] = 1;
4360 return 1;
4361}
4362
4363static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4364{
4365 void *p;
4366 int i;
4367 if (n[0] == n[1])
4368 return;
4369 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4370 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4371 continue;
4372 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4373 return;
4374 }
4375}
4376
4377static void show_symbol(struct seq_file *m, unsigned long address)
4378{
4379#ifdef CONFIG_KALLSYMS
Al Viro871751e2006-03-25 03:06:39 -08004380 unsigned long offset, size;
Tejun Heo9281ace2007-07-17 04:03:51 -07004381 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
Al Viro871751e2006-03-25 03:06:39 -08004382
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004383 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
Al Viro871751e2006-03-25 03:06:39 -08004384 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004385 if (modname[0])
Al Viro871751e2006-03-25 03:06:39 -08004386 seq_printf(m, " [%s]", modname);
4387 return;
4388 }
4389#endif
4390 seq_printf(m, "%p", (void *)address);
4391}
4392
4393static int leaks_show(struct seq_file *m, void *p)
4394{
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004395 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
Al Viro871751e2006-03-25 03:06:39 -08004396 struct slab *slabp;
4397 struct kmem_list3 *l3;
4398 const char *name;
4399 unsigned long *n = m->private;
4400 int node;
4401 int i;
4402
4403 if (!(cachep->flags & SLAB_STORE_USER))
4404 return 0;
4405 if (!(cachep->flags & SLAB_RED_ZONE))
4406 return 0;
4407
4408 /* OK, we can do it */
4409
4410 n[1] = 0;
4411
4412 for_each_online_node(node) {
4413 l3 = cachep->nodelists[node];
4414 if (!l3)
4415 continue;
4416
4417 check_irq_on();
4418 spin_lock_irq(&l3->list_lock);
4419
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004420 list_for_each_entry(slabp, &l3->slabs_full, list)
Al Viro871751e2006-03-25 03:06:39 -08004421 handle_slab(n, cachep, slabp);
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004422 list_for_each_entry(slabp, &l3->slabs_partial, list)
Al Viro871751e2006-03-25 03:06:39 -08004423 handle_slab(n, cachep, slabp);
Al Viro871751e2006-03-25 03:06:39 -08004424 spin_unlock_irq(&l3->list_lock);
4425 }
4426 name = cachep->name;
4427 if (n[0] == n[1]) {
4428 /* Increase the buffer size */
4429 mutex_unlock(&cache_chain_mutex);
4430 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4431 if (!m->private) {
4432 /* Too bad, we are really out */
4433 m->private = n;
4434 mutex_lock(&cache_chain_mutex);
4435 return -ENOMEM;
4436 }
4437 *(unsigned long *)m->private = n[0] * 2;
4438 kfree(n);
4439 mutex_lock(&cache_chain_mutex);
4440 /* Now make sure this entry will be retried */
4441 m->count = m->size;
4442 return 0;
4443 }
4444 for (i = 0; i < n[1]; i++) {
4445 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4446 show_symbol(m, n[2*i+2]);
4447 seq_putc(m, '\n');
4448 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004449
Al Viro871751e2006-03-25 03:06:39 -08004450 return 0;
4451}
4452
Helge Deller15ad7cd2006-12-06 20:40:36 -08004453const struct seq_operations slabstats_op = {
Al Viro871751e2006-03-25 03:06:39 -08004454 .start = leaks_start,
4455 .next = s_next,
4456 .stop = s_stop,
4457 .show = leaks_show,
4458};
4459#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004460#endif
4461
Manfred Spraul00e145b2005-09-03 15:55:07 -07004462/**
4463 * ksize - get the actual amount of memory allocated for a given object
4464 * @objp: Pointer to the object
4465 *
4466 * kmalloc may internally round up allocations and return more memory
4467 * than requested. ksize() can be used to determine the actual amount of
4468 * memory allocated. The caller may use this additional memory, even though
4469 * a smaller amount of memory was initially specified with the kmalloc call.
4470 * The caller must guarantee that objp points to a valid object previously
4471 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4472 * must not be freed during the duration of the call.
4473 */
Pekka Enbergfd76bab2007-05-06 14:48:40 -07004474size_t ksize(const void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004475{
Christoph Lameteref8b4522007-10-16 01:24:46 -07004476 BUG_ON(!objp);
4477 if (unlikely(objp == ZERO_SIZE_PTR))
Manfred Spraul00e145b2005-09-03 15:55:07 -07004478 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004479
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08004480 return obj_size(virt_to_cache(objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004481}
Tetsuo Handaf8fcc932007-12-04 23:45:08 -08004482EXPORT_SYMBOL(ksize);