blob: cceb347a354ec01e39fce84a58e2c164c99845ff [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * linux/mm/slab.c
3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
5 *
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
7 *
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
10 *
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
13 *
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
21 *
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
27 *
28 * This means, that your constructor is used only for newly allocated
29 * slabs and you must pass objects with the same intializations to
30 * kmem_cache_free.
31 *
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
35 *
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
38 * partial slabs
39 * empty slabs with no allocated objects
40 *
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
43 *
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
46 *
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
52 *
Andrew Mortona737b3e2006-03-22 00:08:11 -080053 * The c_cpuarray may not be read with enabled local interrupts -
Linus Torvalds1da177e2005-04-16 15:20:36 -070054 * it's changed with a smp_call_function().
55 *
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
Pekka Enberg343e0d72006-02-01 03:05:50 -080058 * Several members in struct kmem_cache and struct slab never change, they
Linus Torvalds1da177e2005-04-16 15:20:36 -070059 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
63 *
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
66 * his patch.
67 *
68 * Further notes from the original documentation:
69 *
70 * 11 April '97. Started multi-threading - markhe
Ingo Molnarfc0abb12006-01-18 17:42:33 -080071 * The global cache-chain is protected by the mutex 'cache_chain_mutex'.
Linus Torvalds1da177e2005-04-16 15:20:36 -070072 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
75 *
76 * At present, each engine can be growing a cache. This should be blocked.
77 *
Christoph Lametere498be72005-09-09 13:03:32 -070078 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
83 *
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
Linus Torvalds1da177e2005-04-16 15:20:36 -070087 */
88
Linus Torvalds1da177e2005-04-16 15:20:36 -070089#include <linux/slab.h>
90#include <linux/mm.h>
Randy Dunlapc9cf5522006-06-27 02:53:52 -070091#include <linux/poison.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070092#include <linux/swap.h>
93#include <linux/cache.h>
94#include <linux/interrupt.h>
95#include <linux/init.h>
96#include <linux/compiler.h>
Paul Jackson101a5002006-03-24 03:16:07 -080097#include <linux/cpuset.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070098#include <linux/seq_file.h>
99#include <linux/notifier.h>
100#include <linux/kallsyms.h>
101#include <linux/cpu.h>
102#include <linux/sysctl.h>
103#include <linux/module.h>
104#include <linux/rcupdate.h>
Paulo Marques543537b2005-06-23 00:09:02 -0700105#include <linux/string.h>
Andrew Morton138ae662006-12-06 20:36:41 -0800106#include <linux/uaccess.h>
Christoph Lametere498be72005-09-09 13:03:32 -0700107#include <linux/nodemask.h>
Christoph Lameterdc85da12006-01-18 17:42:36 -0800108#include <linux/mempolicy.h>
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800109#include <linux/mutex.h>
Akinobu Mita8a8b6502006-12-08 02:39:44 -0800110#include <linux/fault-inject.h>
Ingo Molnare7eebaf2006-06-27 02:54:55 -0700111#include <linux/rtmutex.h>
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800112#include <linux/reciprocal_div.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700113
Linus Torvalds1da177e2005-04-16 15:20:36 -0700114#include <asm/cacheflush.h>
115#include <asm/tlbflush.h>
116#include <asm/page.h>
117
118/*
Christoph Lameter50953fe2007-05-06 14:50:16 -0700119 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700120 * 0 for faster, smaller code (especially in the critical paths).
121 *
122 * STATS - 1 to collect stats for /proc/slabinfo.
123 * 0 for faster, smaller code (especially in the critical paths).
124 *
125 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
126 */
127
128#ifdef CONFIG_DEBUG_SLAB
129#define DEBUG 1
130#define STATS 1
131#define FORCED_DEBUG 1
132#else
133#define DEBUG 0
134#define STATS 0
135#define FORCED_DEBUG 0
136#endif
137
Linus Torvalds1da177e2005-04-16 15:20:36 -0700138/* Shouldn't this be in a header file somewhere? */
139#define BYTES_PER_WORD sizeof(void *)
140
141#ifndef cache_line_size
142#define cache_line_size() L1_CACHE_BYTES
143#endif
144
145#ifndef ARCH_KMALLOC_MINALIGN
146/*
147 * Enforce a minimum alignment for the kmalloc caches.
148 * Usually, the kmalloc caches are cache_line_size() aligned, except when
149 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
150 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
David Woodhouseb46b8f12007-05-08 00:22:59 -0700151 * alignment larger than the alignment of a 64-bit integer.
152 * ARCH_KMALLOC_MINALIGN allows that.
153 * Note that increasing this value may disable some debug features.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700154 */
David Woodhouseb46b8f12007-05-08 00:22:59 -0700155#define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700156#endif
157
158#ifndef ARCH_SLAB_MINALIGN
159/*
160 * Enforce a minimum alignment for all caches.
161 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
162 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
163 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
164 * some debug features.
165 */
166#define ARCH_SLAB_MINALIGN 0
167#endif
168
169#ifndef ARCH_KMALLOC_FLAGS
170#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
171#endif
172
173/* Legal flag mask for kmem_cache_create(). */
174#if DEBUG
Christoph Lameter50953fe2007-05-06 14:50:16 -0700175# define CREATE_MASK (SLAB_RED_ZONE | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700176 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
Christoph Lameterac2b8982006-03-22 00:08:15 -0800177 SLAB_CACHE_DMA | \
Christoph Lameter5af60832007-05-06 14:49:56 -0700178 SLAB_STORE_USER | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700179 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
Paul Jackson101a5002006-03-24 03:16:07 -0800180 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700181#else
Christoph Lameterac2b8982006-03-22 00:08:15 -0800182# define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
Christoph Lameter5af60832007-05-06 14:49:56 -0700183 SLAB_CACHE_DMA | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700184 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
Paul Jackson101a5002006-03-24 03:16:07 -0800185 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700186#endif
187
188/*
189 * kmem_bufctl_t:
190 *
191 * Bufctl's are used for linking objs within a slab
192 * linked offsets.
193 *
194 * This implementation relies on "struct page" for locating the cache &
195 * slab an object belongs to.
196 * This allows the bufctl structure to be small (one int), but limits
197 * the number of objects a slab (not a cache) can contain when off-slab
198 * bufctls are used. The limit is the size of the largest general cache
199 * that does not use off-slab slabs.
200 * For 32bit archs with 4 kB pages, is this 56.
201 * This is not serious, as it is only for large objects, when it is unwise
202 * to have too many per slab.
203 * Note: This limit can be raised by introducing a general cache whose size
204 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
205 */
206
Kyle Moffettfa5b08d2005-09-03 15:55:03 -0700207typedef unsigned int kmem_bufctl_t;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700208#define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
209#define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
Al Viro871751e2006-03-25 03:06:39 -0800210#define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
211#define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700212
Linus Torvalds1da177e2005-04-16 15:20:36 -0700213/*
214 * struct slab
215 *
216 * Manages the objs in a slab. Placed either at the beginning of mem allocated
217 * for a slab, or allocated from an general cache.
218 * Slabs are chained into three list: fully used, partial, fully free slabs.
219 */
220struct slab {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800221 struct list_head list;
222 unsigned long colouroff;
223 void *s_mem; /* including colour offset */
224 unsigned int inuse; /* num of objs active in slab */
225 kmem_bufctl_t free;
226 unsigned short nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700227};
228
229/*
230 * struct slab_rcu
231 *
232 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
233 * arrange for kmem_freepages to be called via RCU. This is useful if
234 * we need to approach a kernel structure obliquely, from its address
235 * obtained without the usual locking. We can lock the structure to
236 * stabilize it and check it's still at the given address, only if we
237 * can be sure that the memory has not been meanwhile reused for some
238 * other kind of object (which our subsystem's lock might corrupt).
239 *
240 * rcu_read_lock before reading the address, then rcu_read_unlock after
241 * taking the spinlock within the structure expected at that address.
242 *
243 * We assume struct slab_rcu can overlay struct slab when destroying.
244 */
245struct slab_rcu {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800246 struct rcu_head head;
Pekka Enberg343e0d72006-02-01 03:05:50 -0800247 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800248 void *addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700249};
250
251/*
252 * struct array_cache
253 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700254 * Purpose:
255 * - LIFO ordering, to hand out cache-warm objects from _alloc
256 * - reduce the number of linked list operations
257 * - reduce spinlock operations
258 *
259 * The limit is stored in the per-cpu structure to reduce the data cache
260 * footprint.
261 *
262 */
263struct array_cache {
264 unsigned int avail;
265 unsigned int limit;
266 unsigned int batchcount;
267 unsigned int touched;
Christoph Lametere498be72005-09-09 13:03:32 -0700268 spinlock_t lock;
Andrew Mortona737b3e2006-03-22 00:08:11 -0800269 void *entry[0]; /*
270 * Must have this definition in here for the proper
271 * alignment of array_cache. Also simplifies accessing
272 * the entries.
273 * [0] is for gcc 2.95. It should really be [].
274 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700275};
276
Andrew Mortona737b3e2006-03-22 00:08:11 -0800277/*
278 * bootstrap: The caches do not work without cpuarrays anymore, but the
279 * cpuarrays are allocated from the generic caches...
Linus Torvalds1da177e2005-04-16 15:20:36 -0700280 */
281#define BOOT_CPUCACHE_ENTRIES 1
282struct arraycache_init {
283 struct array_cache cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800284 void *entries[BOOT_CPUCACHE_ENTRIES];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700285};
286
287/*
Christoph Lametere498be72005-09-09 13:03:32 -0700288 * The slab lists for all objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700289 */
290struct kmem_list3 {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800291 struct list_head slabs_partial; /* partial list first, better asm code */
292 struct list_head slabs_full;
293 struct list_head slabs_free;
294 unsigned long free_objects;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800295 unsigned int free_limit;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800296 unsigned int colour_next; /* Per-node cache coloring */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800297 spinlock_t list_lock;
298 struct array_cache *shared; /* shared per node */
299 struct array_cache **alien; /* on other nodes */
Christoph Lameter35386e32006-03-22 00:09:05 -0800300 unsigned long next_reap; /* updated without locking */
301 int free_touched; /* updated without locking */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700302};
303
Christoph Lametere498be72005-09-09 13:03:32 -0700304/*
305 * Need this for bootstrapping a per node allocator.
306 */
307#define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
308struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
309#define CACHE_CACHE 0
310#define SIZE_AC 1
311#define SIZE_L3 (1 + MAX_NUMNODES)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700312
Christoph Lametered11d9e2006-06-30 01:55:45 -0700313static int drain_freelist(struct kmem_cache *cache,
314 struct kmem_list3 *l3, int tofree);
315static void free_block(struct kmem_cache *cachep, void **objpp, int len,
316 int node);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -0700317static int enable_cpucache(struct kmem_cache *cachep);
David Howells65f27f32006-11-22 14:55:48 +0000318static void cache_reap(struct work_struct *unused);
Christoph Lametered11d9e2006-06-30 01:55:45 -0700319
Christoph Lametere498be72005-09-09 13:03:32 -0700320/*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800321 * This function must be completely optimized away if a constant is passed to
322 * it. Mostly the same as what is in linux/slab.h except it returns an index.
Christoph Lametere498be72005-09-09 13:03:32 -0700323 */
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700324static __always_inline int index_of(const size_t size)
Christoph Lametere498be72005-09-09 13:03:32 -0700325{
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800326 extern void __bad_size(void);
327
Christoph Lametere498be72005-09-09 13:03:32 -0700328 if (__builtin_constant_p(size)) {
329 int i = 0;
330
331#define CACHE(x) \
332 if (size <=x) \
333 return i; \
334 else \
335 i++;
336#include "linux/kmalloc_sizes.h"
337#undef CACHE
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800338 __bad_size();
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700339 } else
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800340 __bad_size();
Christoph Lametere498be72005-09-09 13:03:32 -0700341 return 0;
342}
343
Ingo Molnare0a42722006-06-23 02:03:46 -0700344static int slab_early_init = 1;
345
Christoph Lametere498be72005-09-09 13:03:32 -0700346#define INDEX_AC index_of(sizeof(struct arraycache_init))
347#define INDEX_L3 index_of(sizeof(struct kmem_list3))
348
Pekka Enberg5295a742006-02-01 03:05:48 -0800349static void kmem_list3_init(struct kmem_list3 *parent)
Christoph Lametere498be72005-09-09 13:03:32 -0700350{
351 INIT_LIST_HEAD(&parent->slabs_full);
352 INIT_LIST_HEAD(&parent->slabs_partial);
353 INIT_LIST_HEAD(&parent->slabs_free);
354 parent->shared = NULL;
355 parent->alien = NULL;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800356 parent->colour_next = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700357 spin_lock_init(&parent->list_lock);
358 parent->free_objects = 0;
359 parent->free_touched = 0;
360}
361
Andrew Mortona737b3e2006-03-22 00:08:11 -0800362#define MAKE_LIST(cachep, listp, slab, nodeid) \
363 do { \
364 INIT_LIST_HEAD(listp); \
365 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
Christoph Lametere498be72005-09-09 13:03:32 -0700366 } while (0)
367
Andrew Mortona737b3e2006-03-22 00:08:11 -0800368#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
369 do { \
Christoph Lametere498be72005-09-09 13:03:32 -0700370 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
371 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
372 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
373 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700374
375/*
Pekka Enberg343e0d72006-02-01 03:05:50 -0800376 * struct kmem_cache
Linus Torvalds1da177e2005-04-16 15:20:36 -0700377 *
378 * manages a cache.
379 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800380
Pekka J Enberg2109a2d2005-11-07 00:58:01 -0800381struct kmem_cache {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700382/* 1) per-cpu data, touched during every alloc/free */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800383 struct array_cache *array[NR_CPUS];
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800384/* 2) Cache tunables. Protected by cache_chain_mutex */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800385 unsigned int batchcount;
386 unsigned int limit;
387 unsigned int shared;
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800388
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800389 unsigned int buffer_size;
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800390 u32 reciprocal_buffer_size;
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800391/* 3) touched by every alloc & free from the backend */
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800392
Andrew Mortona737b3e2006-03-22 00:08:11 -0800393 unsigned int flags; /* constant flags */
394 unsigned int num; /* # of objs per slab */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700395
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800396/* 4) cache_grow/shrink */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700397 /* order of pgs per slab (2^n) */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800398 unsigned int gfporder;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700399
400 /* force GFP flags, e.g. GFP_DMA */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800401 gfp_t gfpflags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700402
Andrew Mortona737b3e2006-03-22 00:08:11 -0800403 size_t colour; /* cache colouring range */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800404 unsigned int colour_off; /* colour offset */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800405 struct kmem_cache *slabp_cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800406 unsigned int slab_size;
Andrew Mortona737b3e2006-03-22 00:08:11 -0800407 unsigned int dflags; /* dynamic flags */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700408
409 /* constructor func */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800410 void (*ctor) (void *, struct kmem_cache *, unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700411
412 /* de-constructor func */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800413 void (*dtor) (void *, struct kmem_cache *, unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700414
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800415/* 5) cache creation/removal */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800416 const char *name;
417 struct list_head next;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700418
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800419/* 6) statistics */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700420#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800421 unsigned long num_active;
422 unsigned long num_allocations;
423 unsigned long high_mark;
424 unsigned long grown;
425 unsigned long reaped;
426 unsigned long errors;
427 unsigned long max_freeable;
428 unsigned long node_allocs;
429 unsigned long node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700430 unsigned long node_overflow;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800431 atomic_t allochit;
432 atomic_t allocmiss;
433 atomic_t freehit;
434 atomic_t freemiss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700435#endif
436#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800437 /*
438 * If debugging is enabled, then the allocator can add additional
439 * fields and/or padding to every object. buffer_size contains the total
440 * object size including these internal fields, the following two
441 * variables contain the offset to the user object and its size.
442 */
443 int obj_offset;
444 int obj_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700445#endif
Eric Dumazet8da34302007-05-06 14:49:29 -0700446 /*
447 * We put nodelists[] at the end of kmem_cache, because we want to size
448 * this array to nr_node_ids slots instead of MAX_NUMNODES
449 * (see kmem_cache_init())
450 * We still use [MAX_NUMNODES] and not [1] or [0] because cache_cache
451 * is statically defined, so we reserve the max number of nodes.
452 */
453 struct kmem_list3 *nodelists[MAX_NUMNODES];
454 /*
455 * Do not add fields after nodelists[]
456 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700457};
458
459#define CFLGS_OFF_SLAB (0x80000000UL)
460#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
461
462#define BATCHREFILL_LIMIT 16
Andrew Mortona737b3e2006-03-22 00:08:11 -0800463/*
464 * Optimization question: fewer reaps means less probability for unnessary
465 * cpucache drain/refill cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700466 *
Adrian Bunkdc6f3f22005-11-08 16:44:08 +0100467 * OTOH the cpuarrays can contain lots of objects,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700468 * which could lock up otherwise freeable slabs.
469 */
470#define REAPTIMEOUT_CPUC (2*HZ)
471#define REAPTIMEOUT_LIST3 (4*HZ)
472
473#if STATS
474#define STATS_INC_ACTIVE(x) ((x)->num_active++)
475#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
476#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
477#define STATS_INC_GROWN(x) ((x)->grown++)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700478#define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
Andrew Mortona737b3e2006-03-22 00:08:11 -0800479#define STATS_SET_HIGH(x) \
480 do { \
481 if ((x)->num_active > (x)->high_mark) \
482 (x)->high_mark = (x)->num_active; \
483 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700484#define STATS_INC_ERR(x) ((x)->errors++)
485#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
Christoph Lametere498be72005-09-09 13:03:32 -0700486#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700487#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800488#define STATS_SET_FREEABLE(x, i) \
489 do { \
490 if ((x)->max_freeable < i) \
491 (x)->max_freeable = i; \
492 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700493#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
494#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
495#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
496#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
497#else
498#define STATS_INC_ACTIVE(x) do { } while (0)
499#define STATS_DEC_ACTIVE(x) do { } while (0)
500#define STATS_INC_ALLOCED(x) do { } while (0)
501#define STATS_INC_GROWN(x) do { } while (0)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700502#define STATS_ADD_REAPED(x,y) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700503#define STATS_SET_HIGH(x) do { } while (0)
504#define STATS_INC_ERR(x) do { } while (0)
505#define STATS_INC_NODEALLOCS(x) do { } while (0)
Christoph Lametere498be72005-09-09 13:03:32 -0700506#define STATS_INC_NODEFREES(x) do { } while (0)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700507#define STATS_INC_ACOVERFLOW(x) do { } while (0)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800508#define STATS_SET_FREEABLE(x, i) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700509#define STATS_INC_ALLOCHIT(x) do { } while (0)
510#define STATS_INC_ALLOCMISS(x) do { } while (0)
511#define STATS_INC_FREEHIT(x) do { } while (0)
512#define STATS_INC_FREEMISS(x) do { } while (0)
513#endif
514
515#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -0700516
Andrew Mortona737b3e2006-03-22 00:08:11 -0800517/*
518 * memory layout of objects:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700519 * 0 : objp
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800520 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
Linus Torvalds1da177e2005-04-16 15:20:36 -0700521 * the end of an object is aligned with the end of the real
522 * allocation. Catches writes behind the end of the allocation.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800523 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700524 * redzone word.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800525 * cachep->obj_offset: The real object.
526 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
Andrew Mortona737b3e2006-03-22 00:08:11 -0800527 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
528 * [BYTES_PER_WORD long]
Linus Torvalds1da177e2005-04-16 15:20:36 -0700529 */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800530static int obj_offset(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700531{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800532 return cachep->obj_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700533}
534
Pekka Enberg343e0d72006-02-01 03:05:50 -0800535static int obj_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700536{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800537 return cachep->obj_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700538}
539
David Woodhouseb46b8f12007-05-08 00:22:59 -0700540static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700541{
542 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
David Woodhouseb46b8f12007-05-08 00:22:59 -0700543 return (unsigned long long*) (objp + obj_offset(cachep) -
544 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700545}
546
David Woodhouseb46b8f12007-05-08 00:22:59 -0700547static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700548{
549 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
550 if (cachep->flags & SLAB_STORE_USER)
David Woodhouseb46b8f12007-05-08 00:22:59 -0700551 return (unsigned long long *)(objp + cachep->buffer_size -
552 sizeof(unsigned long long) -
553 BYTES_PER_WORD);
554 return (unsigned long long *) (objp + cachep->buffer_size -
555 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700556}
557
Pekka Enberg343e0d72006-02-01 03:05:50 -0800558static void **dbg_userword(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700559{
560 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800561 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700562}
563
564#else
565
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800566#define obj_offset(x) 0
567#define obj_size(cachep) (cachep->buffer_size)
David Woodhouseb46b8f12007-05-08 00:22:59 -0700568#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
569#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
Linus Torvalds1da177e2005-04-16 15:20:36 -0700570#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
571
572#endif
573
574/*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800575 * Maximum size of an obj (in 2^order pages) and absolute limit for the gfp
576 * order.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700577 */
578#if defined(CONFIG_LARGE_ALLOCS)
579#define MAX_OBJ_ORDER 13 /* up to 32Mb */
580#define MAX_GFP_ORDER 13 /* up to 32Mb */
581#elif defined(CONFIG_MMU)
582#define MAX_OBJ_ORDER 5 /* 32 pages */
583#define MAX_GFP_ORDER 5 /* 32 pages */
584#else
585#define MAX_OBJ_ORDER 8 /* up to 1Mb */
586#define MAX_GFP_ORDER 8 /* up to 1Mb */
587#endif
588
589/*
590 * Do not go above this order unless 0 objects fit into the slab.
591 */
592#define BREAK_GFP_ORDER_HI 1
593#define BREAK_GFP_ORDER_LO 0
594static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
595
Andrew Mortona737b3e2006-03-22 00:08:11 -0800596/*
597 * Functions for storing/retrieving the cachep and or slab from the page
598 * allocator. These are used to find the slab an obj belongs to. With kfree(),
599 * these are used to find the cache which an obj belongs to.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700600 */
Pekka Enberg065d41c2005-11-13 16:06:46 -0800601static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
602{
603 page->lru.next = (struct list_head *)cache;
604}
605
606static inline struct kmem_cache *page_get_cache(struct page *page)
607{
Christoph Lameterd85f3382007-05-06 14:49:39 -0700608 page = compound_head(page);
Pekka Enbergddc2e812006-06-23 02:03:40 -0700609 BUG_ON(!PageSlab(page));
Pekka Enberg065d41c2005-11-13 16:06:46 -0800610 return (struct kmem_cache *)page->lru.next;
611}
612
613static inline void page_set_slab(struct page *page, struct slab *slab)
614{
615 page->lru.prev = (struct list_head *)slab;
616}
617
618static inline struct slab *page_get_slab(struct page *page)
619{
Pekka Enbergddc2e812006-06-23 02:03:40 -0700620 BUG_ON(!PageSlab(page));
Pekka Enberg065d41c2005-11-13 16:06:46 -0800621 return (struct slab *)page->lru.prev;
622}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700623
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800624static inline struct kmem_cache *virt_to_cache(const void *obj)
625{
Christoph Lameterb49af682007-05-06 14:49:41 -0700626 struct page *page = virt_to_head_page(obj);
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800627 return page_get_cache(page);
628}
629
630static inline struct slab *virt_to_slab(const void *obj)
631{
Christoph Lameterb49af682007-05-06 14:49:41 -0700632 struct page *page = virt_to_head_page(obj);
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800633 return page_get_slab(page);
634}
635
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800636static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
637 unsigned int idx)
638{
639 return slab->s_mem + cache->buffer_size * idx;
640}
641
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800642/*
643 * We want to avoid an expensive divide : (offset / cache->buffer_size)
644 * Using the fact that buffer_size is a constant for a particular cache,
645 * we can replace (offset / cache->buffer_size) by
646 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
647 */
648static inline unsigned int obj_to_index(const struct kmem_cache *cache,
649 const struct slab *slab, void *obj)
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800650{
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800651 u32 offset = (obj - slab->s_mem);
652 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800653}
654
Andrew Mortona737b3e2006-03-22 00:08:11 -0800655/*
656 * These are the default caches for kmalloc. Custom caches can have other sizes.
657 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700658struct cache_sizes malloc_sizes[] = {
659#define CACHE(x) { .cs_size = (x) },
660#include <linux/kmalloc_sizes.h>
661 CACHE(ULONG_MAX)
662#undef CACHE
663};
664EXPORT_SYMBOL(malloc_sizes);
665
666/* Must match cache_sizes above. Out of line to keep cache footprint low. */
667struct cache_names {
668 char *name;
669 char *name_dma;
670};
671
672static struct cache_names __initdata cache_names[] = {
673#define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
674#include <linux/kmalloc_sizes.h>
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800675 {NULL,}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700676#undef CACHE
677};
678
679static struct arraycache_init initarray_cache __initdata =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800680 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700681static struct arraycache_init initarray_generic =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800682 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700683
684/* internal cache of cache description objs */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800685static struct kmem_cache cache_cache = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800686 .batchcount = 1,
687 .limit = BOOT_CPUCACHE_ENTRIES,
688 .shared = 1,
Pekka Enberg343e0d72006-02-01 03:05:50 -0800689 .buffer_size = sizeof(struct kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800690 .name = "kmem_cache",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700691};
692
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700693#define BAD_ALIEN_MAGIC 0x01020304ul
694
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200695#ifdef CONFIG_LOCKDEP
696
697/*
698 * Slab sometimes uses the kmalloc slabs to store the slab headers
699 * for other slabs "off slab".
700 * The locking for this is tricky in that it nests within the locks
701 * of all other slabs in a few places; to deal with this special
702 * locking we put on-slab caches into a separate lock-class.
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700703 *
704 * We set lock class for alien array caches which are up during init.
705 * The lock annotation will be lost if all cpus of a node goes down and
706 * then comes back up during hotplug
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200707 */
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700708static struct lock_class_key on_slab_l3_key;
709static struct lock_class_key on_slab_alc_key;
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200710
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700711static inline void init_lock_keys(void)
712
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200713{
714 int q;
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700715 struct cache_sizes *s = malloc_sizes;
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200716
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700717 while (s->cs_size != ULONG_MAX) {
718 for_each_node(q) {
719 struct array_cache **alc;
720 int r;
721 struct kmem_list3 *l3 = s->cs_cachep->nodelists[q];
722 if (!l3 || OFF_SLAB(s->cs_cachep))
723 continue;
724 lockdep_set_class(&l3->list_lock, &on_slab_l3_key);
725 alc = l3->alien;
726 /*
727 * FIXME: This check for BAD_ALIEN_MAGIC
728 * should go away when common slab code is taught to
729 * work even without alien caches.
730 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
731 * for alloc_alien_cache,
732 */
733 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
734 continue;
735 for_each_node(r) {
736 if (alc[r])
737 lockdep_set_class(&alc[r]->lock,
738 &on_slab_alc_key);
739 }
740 }
741 s++;
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200742 }
743}
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200744#else
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700745static inline void init_lock_keys(void)
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200746{
747}
748#endif
749
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -0800750/*
751 * 1. Guard access to the cache-chain.
752 * 2. Protect sanity of cpu_online_map against cpu hotplug events
753 */
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800754static DEFINE_MUTEX(cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700755static struct list_head cache_chain;
756
757/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700758 * chicken and egg problem: delay the per-cpu array allocation
759 * until the general caches are up.
760 */
761static enum {
762 NONE,
Christoph Lametere498be72005-09-09 13:03:32 -0700763 PARTIAL_AC,
764 PARTIAL_L3,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700765 FULL
766} g_cpucache_up;
767
Mike Kravetz39d24e62006-05-15 09:44:13 -0700768/*
769 * used by boot code to determine if it can use slab based allocator
770 */
771int slab_is_available(void)
772{
773 return g_cpucache_up == FULL;
774}
775
David Howells52bad642006-11-22 14:54:01 +0000776static DEFINE_PER_CPU(struct delayed_work, reap_work);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700777
Pekka Enberg343e0d72006-02-01 03:05:50 -0800778static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700779{
780 return cachep->array[smp_processor_id()];
781}
782
Andrew Mortona737b3e2006-03-22 00:08:11 -0800783static inline struct kmem_cache *__find_general_cachep(size_t size,
784 gfp_t gfpflags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700785{
786 struct cache_sizes *csizep = malloc_sizes;
787
788#if DEBUG
789 /* This happens if someone tries to call
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800790 * kmem_cache_create(), or __kmalloc(), before
791 * the generic caches are initialized.
792 */
Alok Katariac7e43c72005-09-14 12:17:53 -0700793 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700794#endif
795 while (size > csizep->cs_size)
796 csizep++;
797
798 /*
Martin Hicks0abf40c2005-09-03 15:54:54 -0700799 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
Linus Torvalds1da177e2005-04-16 15:20:36 -0700800 * has cs_{dma,}cachep==NULL. Thus no special case
801 * for large kmalloc calls required.
802 */
Christoph Lameter4b51d662007-02-10 01:43:10 -0800803#ifdef CONFIG_ZONE_DMA
Linus Torvalds1da177e2005-04-16 15:20:36 -0700804 if (unlikely(gfpflags & GFP_DMA))
805 return csizep->cs_dmacachep;
Christoph Lameter4b51d662007-02-10 01:43:10 -0800806#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700807 return csizep->cs_cachep;
808}
809
Adrian Bunkb2213852006-09-25 23:31:02 -0700810static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700811{
812 return __find_general_cachep(size, gfpflags);
813}
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700814
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800815static size_t slab_mgmt_size(size_t nr_objs, size_t align)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700816{
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800817 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
818}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700819
Andrew Mortona737b3e2006-03-22 00:08:11 -0800820/*
821 * Calculate the number of objects and left-over bytes for a given buffer size.
822 */
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800823static void cache_estimate(unsigned long gfporder, size_t buffer_size,
824 size_t align, int flags, size_t *left_over,
825 unsigned int *num)
826{
827 int nr_objs;
828 size_t mgmt_size;
829 size_t slab_size = PAGE_SIZE << gfporder;
830
831 /*
832 * The slab management structure can be either off the slab or
833 * on it. For the latter case, the memory allocated for a
834 * slab is used for:
835 *
836 * - The struct slab
837 * - One kmem_bufctl_t for each object
838 * - Padding to respect alignment of @align
839 * - @buffer_size bytes for each object
840 *
841 * If the slab management structure is off the slab, then the
842 * alignment will already be calculated into the size. Because
843 * the slabs are all pages aligned, the objects will be at the
844 * correct alignment when allocated.
845 */
846 if (flags & CFLGS_OFF_SLAB) {
847 mgmt_size = 0;
848 nr_objs = slab_size / buffer_size;
849
850 if (nr_objs > SLAB_LIMIT)
851 nr_objs = SLAB_LIMIT;
852 } else {
853 /*
854 * Ignore padding for the initial guess. The padding
855 * is at most @align-1 bytes, and @buffer_size is at
856 * least @align. In the worst case, this result will
857 * be one greater than the number of objects that fit
858 * into the memory allocation when taking the padding
859 * into account.
860 */
861 nr_objs = (slab_size - sizeof(struct slab)) /
862 (buffer_size + sizeof(kmem_bufctl_t));
863
864 /*
865 * This calculated number will be either the right
866 * amount, or one greater than what we want.
867 */
868 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
869 > slab_size)
870 nr_objs--;
871
872 if (nr_objs > SLAB_LIMIT)
873 nr_objs = SLAB_LIMIT;
874
875 mgmt_size = slab_mgmt_size(nr_objs, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700876 }
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800877 *num = nr_objs;
878 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700879}
880
881#define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
882
Andrew Mortona737b3e2006-03-22 00:08:11 -0800883static void __slab_error(const char *function, struct kmem_cache *cachep,
884 char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700885{
886 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800887 function, cachep->name, msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700888 dump_stack();
889}
890
Paul Menage3395ee02006-12-06 20:32:16 -0800891/*
892 * By default on NUMA we use alien caches to stage the freeing of
893 * objects allocated from other nodes. This causes massive memory
894 * inefficiencies when using fake NUMA setup to split memory into a
895 * large number of small nodes, so it can be disabled on the command
896 * line
897 */
898
899static int use_alien_caches __read_mostly = 1;
900static int __init noaliencache_setup(char *s)
901{
902 use_alien_caches = 0;
903 return 1;
904}
905__setup("noaliencache", noaliencache_setup);
906
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800907#ifdef CONFIG_NUMA
908/*
909 * Special reaping functions for NUMA systems called from cache_reap().
910 * These take care of doing round robin flushing of alien caches (containing
911 * objects freed on different nodes from which they were allocated) and the
912 * flushing of remote pcps by calling drain_node_pages.
913 */
914static DEFINE_PER_CPU(unsigned long, reap_node);
915
916static void init_reap_node(int cpu)
917{
918 int node;
919
920 node = next_node(cpu_to_node(cpu), node_online_map);
921 if (node == MAX_NUMNODES)
Paul Jackson442295c2006-03-22 00:09:11 -0800922 node = first_node(node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800923
Daniel Yeisley7f6b8872006-11-02 22:07:14 -0800924 per_cpu(reap_node, cpu) = node;
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800925}
926
927static void next_reap_node(void)
928{
929 int node = __get_cpu_var(reap_node);
930
931 /*
932 * Also drain per cpu pages on remote zones
933 */
934 if (node != numa_node_id())
935 drain_node_pages(node);
936
937 node = next_node(node, node_online_map);
938 if (unlikely(node >= MAX_NUMNODES))
939 node = first_node(node_online_map);
940 __get_cpu_var(reap_node) = node;
941}
942
943#else
944#define init_reap_node(cpu) do { } while (0)
945#define next_reap_node(void) do { } while (0)
946#endif
947
Linus Torvalds1da177e2005-04-16 15:20:36 -0700948/*
949 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
950 * via the workqueue/eventd.
951 * Add the CPU number into the expiration time to minimize the possibility of
952 * the CPUs getting into lockstep and contending for the global cache chain
953 * lock.
954 */
955static void __devinit start_cpu_timer(int cpu)
956{
David Howells52bad642006-11-22 14:54:01 +0000957 struct delayed_work *reap_work = &per_cpu(reap_work, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700958
959 /*
960 * When this gets called from do_initcalls via cpucache_init(),
961 * init_workqueues() has already run, so keventd will be setup
962 * at that time.
963 */
David Howells52bad642006-11-22 14:54:01 +0000964 if (keventd_up() && reap_work->work.func == NULL) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800965 init_reap_node(cpu);
David Howells65f27f32006-11-22 14:55:48 +0000966 INIT_DELAYED_WORK(reap_work, cache_reap);
Arjan van de Ven2b284212006-12-10 02:21:28 -0800967 schedule_delayed_work_on(cpu, reap_work,
968 __round_jiffies_relative(HZ, cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700969 }
970}
971
Christoph Lametere498be72005-09-09 13:03:32 -0700972static struct array_cache *alloc_arraycache(int node, int entries,
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800973 int batchcount)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700974{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800975 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700976 struct array_cache *nc = NULL;
977
Christoph Lametere498be72005-09-09 13:03:32 -0700978 nc = kmalloc_node(memsize, GFP_KERNEL, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700979 if (nc) {
980 nc->avail = 0;
981 nc->limit = entries;
982 nc->batchcount = batchcount;
983 nc->touched = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700984 spin_lock_init(&nc->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700985 }
986 return nc;
987}
988
Christoph Lameter3ded1752006-03-25 03:06:44 -0800989/*
990 * Transfer objects in one arraycache to another.
991 * Locking must be handled by the caller.
992 *
993 * Return the number of entries transferred.
994 */
995static int transfer_objects(struct array_cache *to,
996 struct array_cache *from, unsigned int max)
997{
998 /* Figure out how many entries to transfer */
999 int nr = min(min(from->avail, max), to->limit - to->avail);
1000
1001 if (!nr)
1002 return 0;
1003
1004 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
1005 sizeof(void *) *nr);
1006
1007 from->avail -= nr;
1008 to->avail += nr;
1009 to->touched = 1;
1010 return nr;
1011}
1012
Christoph Lameter765c4502006-09-27 01:50:08 -07001013#ifndef CONFIG_NUMA
1014
1015#define drain_alien_cache(cachep, alien) do { } while (0)
1016#define reap_alien(cachep, l3) do { } while (0)
1017
1018static inline struct array_cache **alloc_alien_cache(int node, int limit)
1019{
1020 return (struct array_cache **)BAD_ALIEN_MAGIC;
1021}
1022
1023static inline void free_alien_cache(struct array_cache **ac_ptr)
1024{
1025}
1026
1027static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1028{
1029 return 0;
1030}
1031
1032static inline void *alternate_node_alloc(struct kmem_cache *cachep,
1033 gfp_t flags)
1034{
1035 return NULL;
1036}
1037
Christoph Hellwig8b98c162006-12-06 20:32:30 -08001038static inline void *____cache_alloc_node(struct kmem_cache *cachep,
Christoph Lameter765c4502006-09-27 01:50:08 -07001039 gfp_t flags, int nodeid)
1040{
1041 return NULL;
1042}
1043
1044#else /* CONFIG_NUMA */
1045
Christoph Hellwig8b98c162006-12-06 20:32:30 -08001046static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
Paul Jacksonc61afb12006-03-24 03:16:08 -08001047static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
Christoph Lameterdc85da12006-01-18 17:42:36 -08001048
Pekka Enberg5295a742006-02-01 03:05:48 -08001049static struct array_cache **alloc_alien_cache(int node, int limit)
Christoph Lametere498be72005-09-09 13:03:32 -07001050{
1051 struct array_cache **ac_ptr;
Christoph Lameter8ef82862007-02-20 13:57:52 -08001052 int memsize = sizeof(void *) * nr_node_ids;
Christoph Lametere498be72005-09-09 13:03:32 -07001053 int i;
1054
1055 if (limit > 1)
1056 limit = 12;
1057 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
1058 if (ac_ptr) {
1059 for_each_node(i) {
1060 if (i == node || !node_online(i)) {
1061 ac_ptr[i] = NULL;
1062 continue;
1063 }
1064 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
1065 if (!ac_ptr[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001066 for (i--; i <= 0; i--)
Christoph Lametere498be72005-09-09 13:03:32 -07001067 kfree(ac_ptr[i]);
1068 kfree(ac_ptr);
1069 return NULL;
1070 }
1071 }
1072 }
1073 return ac_ptr;
1074}
1075
Pekka Enberg5295a742006-02-01 03:05:48 -08001076static void free_alien_cache(struct array_cache **ac_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -07001077{
1078 int i;
1079
1080 if (!ac_ptr)
1081 return;
Christoph Lametere498be72005-09-09 13:03:32 -07001082 for_each_node(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001083 kfree(ac_ptr[i]);
Christoph Lametere498be72005-09-09 13:03:32 -07001084 kfree(ac_ptr);
1085}
1086
Pekka Enberg343e0d72006-02-01 03:05:50 -08001087static void __drain_alien_cache(struct kmem_cache *cachep,
Pekka Enberg5295a742006-02-01 03:05:48 -08001088 struct array_cache *ac, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07001089{
1090 struct kmem_list3 *rl3 = cachep->nodelists[node];
1091
1092 if (ac->avail) {
1093 spin_lock(&rl3->list_lock);
Christoph Lametere00946f2006-03-25 03:06:45 -08001094 /*
1095 * Stuff objects into the remote nodes shared array first.
1096 * That way we could avoid the overhead of putting the objects
1097 * into the free lists and getting them back later.
1098 */
shin, jacob693f7d32006-04-28 10:54:37 -05001099 if (rl3->shared)
1100 transfer_objects(rl3->shared, ac, ac->limit);
Christoph Lametere00946f2006-03-25 03:06:45 -08001101
Christoph Lameterff694162005-09-22 21:44:02 -07001102 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001103 ac->avail = 0;
1104 spin_unlock(&rl3->list_lock);
1105 }
1106}
1107
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001108/*
1109 * Called from cache_reap() to regularly drain alien caches round robin.
1110 */
1111static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1112{
1113 int node = __get_cpu_var(reap_node);
1114
1115 if (l3->alien) {
1116 struct array_cache *ac = l3->alien[node];
Christoph Lametere00946f2006-03-25 03:06:45 -08001117
1118 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001119 __drain_alien_cache(cachep, ac, node);
1120 spin_unlock_irq(&ac->lock);
1121 }
1122 }
1123}
1124
Andrew Mortona737b3e2006-03-22 00:08:11 -08001125static void drain_alien_cache(struct kmem_cache *cachep,
1126 struct array_cache **alien)
Christoph Lametere498be72005-09-09 13:03:32 -07001127{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001128 int i = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07001129 struct array_cache *ac;
1130 unsigned long flags;
1131
1132 for_each_online_node(i) {
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001133 ac = alien[i];
Christoph Lametere498be72005-09-09 13:03:32 -07001134 if (ac) {
1135 spin_lock_irqsave(&ac->lock, flags);
1136 __drain_alien_cache(cachep, ac, i);
1137 spin_unlock_irqrestore(&ac->lock, flags);
1138 }
1139 }
1140}
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001141
Ingo Molnar873623d2006-07-13 14:44:38 +02001142static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001143{
1144 struct slab *slabp = virt_to_slab(objp);
1145 int nodeid = slabp->nodeid;
1146 struct kmem_list3 *l3;
1147 struct array_cache *alien = NULL;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001148 int node;
1149
1150 node = numa_node_id();
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001151
1152 /*
1153 * Make sure we are not freeing a object from another node to the array
1154 * cache on this cpu.
1155 */
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001156 if (likely(slabp->nodeid == node))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001157 return 0;
1158
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001159 l3 = cachep->nodelists[node];
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001160 STATS_INC_NODEFREES(cachep);
1161 if (l3->alien && l3->alien[nodeid]) {
1162 alien = l3->alien[nodeid];
Ingo Molnar873623d2006-07-13 14:44:38 +02001163 spin_lock(&alien->lock);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001164 if (unlikely(alien->avail == alien->limit)) {
1165 STATS_INC_ACOVERFLOW(cachep);
1166 __drain_alien_cache(cachep, alien, nodeid);
1167 }
1168 alien->entry[alien->avail++] = objp;
1169 spin_unlock(&alien->lock);
1170 } else {
1171 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1172 free_block(cachep, &objp, 1, nodeid);
1173 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1174 }
1175 return 1;
1176}
Christoph Lametere498be72005-09-09 13:03:32 -07001177#endif
1178
Chandra Seetharaman8c78f302006-07-30 03:03:35 -07001179static int __cpuinit cpuup_callback(struct notifier_block *nfb,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001180 unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001181{
1182 long cpu = (long)hcpu;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001183 struct kmem_cache *cachep;
Christoph Lametere498be72005-09-09 13:03:32 -07001184 struct kmem_list3 *l3 = NULL;
1185 int node = cpu_to_node(cpu);
1186 int memsize = sizeof(struct kmem_list3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001187
1188 switch (action) {
Heiko Carstens38c3bd92007-05-09 02:34:05 -07001189 case CPU_LOCK_ACQUIRE:
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001190 mutex_lock(&cache_chain_mutex);
Heiko Carstens38c3bd92007-05-09 02:34:05 -07001191 break;
1192 case CPU_UP_PREPARE:
Andrew Mortona737b3e2006-03-22 00:08:11 -08001193 /*
1194 * We need to do this right in the beginning since
Christoph Lametere498be72005-09-09 13:03:32 -07001195 * alloc_arraycache's are going to use this list.
1196 * kmalloc_node allows us to add the slab to the right
1197 * kmem_list3 and not this cpu's kmem_list3
1198 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001199
Christoph Lametere498be72005-09-09 13:03:32 -07001200 list_for_each_entry(cachep, &cache_chain, next) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08001201 /*
1202 * Set up the size64 kmemlist for cpu before we can
Christoph Lametere498be72005-09-09 13:03:32 -07001203 * begin anything. Make sure some other cpu on this
1204 * node has not already allocated this
1205 */
1206 if (!cachep->nodelists[node]) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08001207 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1208 if (!l3)
Christoph Lametere498be72005-09-09 13:03:32 -07001209 goto bad;
1210 kmem_list3_init(l3);
1211 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001212 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -07001213
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001214 /*
1215 * The l3s don't come and go as CPUs come and
1216 * go. cache_chain_mutex is sufficient
1217 * protection here.
1218 */
Christoph Lametere498be72005-09-09 13:03:32 -07001219 cachep->nodelists[node] = l3;
1220 }
1221
1222 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1223 cachep->nodelists[node]->free_limit =
Andrew Mortona737b3e2006-03-22 00:08:11 -08001224 (1 + nr_cpus_node(node)) *
1225 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07001226 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1227 }
1228
Andrew Mortona737b3e2006-03-22 00:08:11 -08001229 /*
1230 * Now we can go ahead with allocating the shared arrays and
1231 * array caches
1232 */
Christoph Lametere498be72005-09-09 13:03:32 -07001233 list_for_each_entry(cachep, &cache_chain, next) {
Tobias Klausercd105df2006-01-08 01:00:59 -08001234 struct array_cache *nc;
Eric Dumazet63109842007-05-06 14:49:28 -07001235 struct array_cache *shared = NULL;
Paul Menage3395ee02006-12-06 20:32:16 -08001236 struct array_cache **alien = NULL;
Tobias Klausercd105df2006-01-08 01:00:59 -08001237
Christoph Lametere498be72005-09-09 13:03:32 -07001238 nc = alloc_arraycache(node, cachep->limit,
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001239 cachep->batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001240 if (!nc)
1241 goto bad;
Eric Dumazet63109842007-05-06 14:49:28 -07001242 if (cachep->shared) {
1243 shared = alloc_arraycache(node,
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001244 cachep->shared * cachep->batchcount,
1245 0xbaadf00d);
Eric Dumazet63109842007-05-06 14:49:28 -07001246 if (!shared)
1247 goto bad;
1248 }
Paul Menage3395ee02006-12-06 20:32:16 -08001249 if (use_alien_caches) {
1250 alien = alloc_alien_cache(node, cachep->limit);
1251 if (!alien)
1252 goto bad;
1253 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001254 cachep->array[cpu] = nc;
Christoph Lametere498be72005-09-09 13:03:32 -07001255 l3 = cachep->nodelists[node];
1256 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07001257
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001258 spin_lock_irq(&l3->list_lock);
1259 if (!l3->shared) {
1260 /*
1261 * We are serialised from CPU_DEAD or
1262 * CPU_UP_CANCELLED by the cpucontrol lock
1263 */
1264 l3->shared = shared;
1265 shared = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07001266 }
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001267#ifdef CONFIG_NUMA
1268 if (!l3->alien) {
1269 l3->alien = alien;
1270 alien = NULL;
1271 }
1272#endif
1273 spin_unlock_irq(&l3->list_lock);
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001274 kfree(shared);
1275 free_alien_cache(alien);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001276 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001277 break;
1278 case CPU_ONLINE:
1279 start_cpu_timer(cpu);
1280 break;
1281#ifdef CONFIG_HOTPLUG_CPU
1282 case CPU_DEAD:
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001283 /*
1284 * Even if all the cpus of a node are down, we don't free the
1285 * kmem_list3 of any cache. This to avoid a race between
1286 * cpu_down, and a kmalloc allocation from another cpu for
1287 * memory from the node of the cpu going down. The list3
1288 * structure is usually allocated from kmem_cache_create() and
1289 * gets destroyed at kmem_cache_destroy().
1290 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001291 /* fall thru */
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08001292#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001293 case CPU_UP_CANCELED:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001294 list_for_each_entry(cachep, &cache_chain, next) {
1295 struct array_cache *nc;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001296 struct array_cache *shared;
1297 struct array_cache **alien;
Christoph Lametere498be72005-09-09 13:03:32 -07001298 cpumask_t mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001299
Christoph Lametere498be72005-09-09 13:03:32 -07001300 mask = node_to_cpumask(node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001301 /* cpu is dead; no one can alloc from it. */
1302 nc = cachep->array[cpu];
1303 cachep->array[cpu] = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07001304 l3 = cachep->nodelists[node];
1305
1306 if (!l3)
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001307 goto free_array_cache;
Christoph Lametere498be72005-09-09 13:03:32 -07001308
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08001309 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07001310
1311 /* Free limit for this kmem_list3 */
1312 l3->free_limit -= cachep->batchcount;
1313 if (nc)
Christoph Lameterff694162005-09-22 21:44:02 -07001314 free_block(cachep, nc->entry, nc->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001315
1316 if (!cpus_empty(mask)) {
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08001317 spin_unlock_irq(&l3->list_lock);
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001318 goto free_array_cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001319 }
Christoph Lametere498be72005-09-09 13:03:32 -07001320
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001321 shared = l3->shared;
1322 if (shared) {
Eric Dumazet63109842007-05-06 14:49:28 -07001323 free_block(cachep, shared->entry,
1324 shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001325 l3->shared = NULL;
1326 }
Christoph Lametere498be72005-09-09 13:03:32 -07001327
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001328 alien = l3->alien;
1329 l3->alien = NULL;
1330
1331 spin_unlock_irq(&l3->list_lock);
1332
1333 kfree(shared);
1334 if (alien) {
1335 drain_alien_cache(cachep, alien);
1336 free_alien_cache(alien);
Christoph Lametere498be72005-09-09 13:03:32 -07001337 }
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001338free_array_cache:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001339 kfree(nc);
1340 }
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001341 /*
1342 * In the previous loop, all the objects were freed to
1343 * the respective cache's slabs, now we can go ahead and
1344 * shrink each nodelist to its limit.
1345 */
1346 list_for_each_entry(cachep, &cache_chain, next) {
1347 l3 = cachep->nodelists[node];
1348 if (!l3)
1349 continue;
Christoph Lametered11d9e2006-06-30 01:55:45 -07001350 drain_freelist(cachep, l3, l3->free_objects);
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001351 }
Heiko Carstens38c3bd92007-05-09 02:34:05 -07001352 break;
1353 case CPU_LOCK_RELEASE:
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001354 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001355 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001356 }
1357 return NOTIFY_OK;
Andrew Mortona737b3e2006-03-22 00:08:11 -08001358bad:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001359 return NOTIFY_BAD;
1360}
1361
Chandra Seetharaman74b85f32006-06-27 02:54:09 -07001362static struct notifier_block __cpuinitdata cpucache_notifier = {
1363 &cpuup_callback, NULL, 0
1364};
Linus Torvalds1da177e2005-04-16 15:20:36 -07001365
Christoph Lametere498be72005-09-09 13:03:32 -07001366/*
1367 * swap the static kmem_list3 with kmalloced memory
1368 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001369static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1370 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001371{
1372 struct kmem_list3 *ptr;
1373
Christoph Lametere498be72005-09-09 13:03:32 -07001374 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1375 BUG_ON(!ptr);
1376
1377 local_irq_disable();
1378 memcpy(ptr, list, sizeof(struct kmem_list3));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001379 /*
1380 * Do not assume that spinlocks can be initialized via memcpy:
1381 */
1382 spin_lock_init(&ptr->list_lock);
1383
Christoph Lametere498be72005-09-09 13:03:32 -07001384 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1385 cachep->nodelists[nodeid] = ptr;
1386 local_irq_enable();
1387}
1388
Andrew Mortona737b3e2006-03-22 00:08:11 -08001389/*
1390 * Initialisation. Called after the page allocator have been initialised and
1391 * before smp_init().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001392 */
1393void __init kmem_cache_init(void)
1394{
1395 size_t left_over;
1396 struct cache_sizes *sizes;
1397 struct cache_names *names;
Christoph Lametere498be72005-09-09 13:03:32 -07001398 int i;
Jack Steiner07ed76b2006-03-07 21:55:46 -08001399 int order;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001400 int node;
Christoph Lametere498be72005-09-09 13:03:32 -07001401
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001402 if (num_possible_nodes() == 1)
1403 use_alien_caches = 0;
1404
Christoph Lametere498be72005-09-09 13:03:32 -07001405 for (i = 0; i < NUM_INIT_LISTS; i++) {
1406 kmem_list3_init(&initkmem_list3[i]);
1407 if (i < MAX_NUMNODES)
1408 cache_cache.nodelists[i] = NULL;
1409 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001410
1411 /*
1412 * Fragmentation resistance on low memory - only use bigger
1413 * page orders on machines with more than 32MB of memory.
1414 */
1415 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1416 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1417
Linus Torvalds1da177e2005-04-16 15:20:36 -07001418 /* Bootstrap is tricky, because several objects are allocated
1419 * from caches that do not exist yet:
Andrew Mortona737b3e2006-03-22 00:08:11 -08001420 * 1) initialize the cache_cache cache: it contains the struct
1421 * kmem_cache structures of all caches, except cache_cache itself:
1422 * cache_cache is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001423 * Initially an __init data area is used for the head array and the
1424 * kmem_list3 structures, it's replaced with a kmalloc allocated
1425 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001426 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001427 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001428 * An __init data area is used for the head array.
1429 * 3) Create the remaining kmalloc caches, with minimally sized
1430 * head arrays.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001431 * 4) Replace the __init data head arrays for cache_cache and the first
1432 * kmalloc cache with kmalloc allocated arrays.
Christoph Lametere498be72005-09-09 13:03:32 -07001433 * 5) Replace the __init data for kmem_list3 for cache_cache and
1434 * the other cache's with kmalloc allocated memory.
1435 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001436 */
1437
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001438 node = numa_node_id();
1439
Linus Torvalds1da177e2005-04-16 15:20:36 -07001440 /* 1) create the cache_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001441 INIT_LIST_HEAD(&cache_chain);
1442 list_add(&cache_cache.next, &cache_chain);
1443 cache_cache.colour_off = cache_line_size();
1444 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001445 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001446
Eric Dumazet8da34302007-05-06 14:49:29 -07001447 /*
1448 * struct kmem_cache size depends on nr_node_ids, which
1449 * can be less than MAX_NUMNODES.
1450 */
1451 cache_cache.buffer_size = offsetof(struct kmem_cache, nodelists) +
1452 nr_node_ids * sizeof(struct kmem_list3 *);
1453#if DEBUG
1454 cache_cache.obj_size = cache_cache.buffer_size;
1455#endif
Andrew Mortona737b3e2006-03-22 00:08:11 -08001456 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1457 cache_line_size());
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08001458 cache_cache.reciprocal_buffer_size =
1459 reciprocal_value(cache_cache.buffer_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001460
Jack Steiner07ed76b2006-03-07 21:55:46 -08001461 for (order = 0; order < MAX_ORDER; order++) {
1462 cache_estimate(order, cache_cache.buffer_size,
1463 cache_line_size(), 0, &left_over, &cache_cache.num);
1464 if (cache_cache.num)
1465 break;
1466 }
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02001467 BUG_ON(!cache_cache.num);
Jack Steiner07ed76b2006-03-07 21:55:46 -08001468 cache_cache.gfporder = order;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001469 cache_cache.colour = left_over / cache_cache.colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001470 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1471 sizeof(struct slab), cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001472
1473 /* 2+3) create the kmalloc caches */
1474 sizes = malloc_sizes;
1475 names = cache_names;
1476
Andrew Mortona737b3e2006-03-22 00:08:11 -08001477 /*
1478 * Initialize the caches that provide memory for the array cache and the
1479 * kmem_list3 structures first. Without this, further allocations will
1480 * bug.
Christoph Lametere498be72005-09-09 13:03:32 -07001481 */
1482
1483 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001484 sizes[INDEX_AC].cs_size,
1485 ARCH_KMALLOC_MINALIGN,
1486 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1487 NULL, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07001488
Andrew Mortona737b3e2006-03-22 00:08:11 -08001489 if (INDEX_AC != INDEX_L3) {
Christoph Lametere498be72005-09-09 13:03:32 -07001490 sizes[INDEX_L3].cs_cachep =
Andrew Mortona737b3e2006-03-22 00:08:11 -08001491 kmem_cache_create(names[INDEX_L3].name,
1492 sizes[INDEX_L3].cs_size,
1493 ARCH_KMALLOC_MINALIGN,
1494 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1495 NULL, NULL);
1496 }
Christoph Lametere498be72005-09-09 13:03:32 -07001497
Ingo Molnare0a42722006-06-23 02:03:46 -07001498 slab_early_init = 0;
1499
Linus Torvalds1da177e2005-04-16 15:20:36 -07001500 while (sizes->cs_size != ULONG_MAX) {
Christoph Lametere498be72005-09-09 13:03:32 -07001501 /*
1502 * For performance, all the general caches are L1 aligned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001503 * This should be particularly beneficial on SMP boxes, as it
1504 * eliminates "false sharing".
1505 * Note for systems short on memory removing the alignment will
Christoph Lametere498be72005-09-09 13:03:32 -07001506 * allow tighter packing of the smaller caches.
1507 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001508 if (!sizes->cs_cachep) {
Christoph Lametere498be72005-09-09 13:03:32 -07001509 sizes->cs_cachep = kmem_cache_create(names->name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001510 sizes->cs_size,
1511 ARCH_KMALLOC_MINALIGN,
1512 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1513 NULL, NULL);
1514 }
Christoph Lameter4b51d662007-02-10 01:43:10 -08001515#ifdef CONFIG_ZONE_DMA
1516 sizes->cs_dmacachep = kmem_cache_create(
1517 names->name_dma,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001518 sizes->cs_size,
1519 ARCH_KMALLOC_MINALIGN,
1520 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1521 SLAB_PANIC,
1522 NULL, NULL);
Christoph Lameter4b51d662007-02-10 01:43:10 -08001523#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001524 sizes++;
1525 names++;
1526 }
1527 /* 4) Replace the bootstrap head arrays */
1528 {
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001529 struct array_cache *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001530
Linus Torvalds1da177e2005-04-16 15:20:36 -07001531 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001532
Linus Torvalds1da177e2005-04-16 15:20:36 -07001533 local_irq_disable();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001534 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1535 memcpy(ptr, cpu_cache_get(&cache_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001536 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001537 /*
1538 * Do not assume that spinlocks can be initialized via memcpy:
1539 */
1540 spin_lock_init(&ptr->lock);
1541
Linus Torvalds1da177e2005-04-16 15:20:36 -07001542 cache_cache.array[smp_processor_id()] = ptr;
1543 local_irq_enable();
Christoph Lametere498be72005-09-09 13:03:32 -07001544
Linus Torvalds1da177e2005-04-16 15:20:36 -07001545 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001546
Linus Torvalds1da177e2005-04-16 15:20:36 -07001547 local_irq_disable();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001548 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001549 != &initarray_generic.cache);
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001550 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001551 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001552 /*
1553 * Do not assume that spinlocks can be initialized via memcpy:
1554 */
1555 spin_lock_init(&ptr->lock);
1556
Christoph Lametere498be72005-09-09 13:03:32 -07001557 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001558 ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001559 local_irq_enable();
1560 }
Christoph Lametere498be72005-09-09 13:03:32 -07001561 /* 5) Replace the bootstrap kmem_list3's */
1562 {
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001563 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001564
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001565 /* Replace the static kmem_list3 structures for the boot cpu */
1566 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE], node);
1567
1568 for_each_online_node(nid) {
Christoph Lametere498be72005-09-09 13:03:32 -07001569 init_list(malloc_sizes[INDEX_AC].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001570 &initkmem_list3[SIZE_AC + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001571
1572 if (INDEX_AC != INDEX_L3) {
1573 init_list(malloc_sizes[INDEX_L3].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001574 &initkmem_list3[SIZE_L3 + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001575 }
1576 }
1577 }
1578
1579 /* 6) resize the head arrays to their final sizes */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001580 {
Pekka Enberg343e0d72006-02-01 03:05:50 -08001581 struct kmem_cache *cachep;
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001582 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001583 list_for_each_entry(cachep, &cache_chain, next)
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07001584 if (enable_cpucache(cachep))
1585 BUG();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001586 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001587 }
1588
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001589 /* Annotate slab for lockdep -- annotate the malloc caches */
1590 init_lock_keys();
1591
1592
Linus Torvalds1da177e2005-04-16 15:20:36 -07001593 /* Done! */
1594 g_cpucache_up = FULL;
1595
Andrew Mortona737b3e2006-03-22 00:08:11 -08001596 /*
1597 * Register a cpu startup notifier callback that initializes
1598 * cpu_cache_get for all new cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07001599 */
1600 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001601
Andrew Mortona737b3e2006-03-22 00:08:11 -08001602 /*
1603 * The reap timers are started later, with a module init call: That part
1604 * of the kernel is not yet operational.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001605 */
1606}
1607
1608static int __init cpucache_init(void)
1609{
1610 int cpu;
1611
Andrew Mortona737b3e2006-03-22 00:08:11 -08001612 /*
1613 * Register the timers that return unneeded pages to the page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -07001614 */
Christoph Lametere498be72005-09-09 13:03:32 -07001615 for_each_online_cpu(cpu)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001616 start_cpu_timer(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001617 return 0;
1618}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001619__initcall(cpucache_init);
1620
1621/*
1622 * Interface to system's page allocator. No need to hold the cache-lock.
1623 *
1624 * If we requested dmaable memory, we will get it. Even if we
1625 * did not request dmaable memory, we might get it, but that
1626 * would be relatively rare and ignorable.
1627 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001628static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001629{
1630 struct page *page;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001631 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001632 int i;
1633
Luke Yangd6fef9d2006-04-10 22:52:56 -07001634#ifndef CONFIG_MMU
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001635 /*
1636 * Nommu uses slab's for process anonymous memory allocations, and thus
1637 * requires __GFP_COMP to properly refcount higher order allocations
Luke Yangd6fef9d2006-04-10 22:52:56 -07001638 */
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001639 flags |= __GFP_COMP;
Luke Yangd6fef9d2006-04-10 22:52:56 -07001640#endif
Christoph Lameter765c4502006-09-27 01:50:08 -07001641
Christoph Lameter3c517a62006-12-06 20:33:29 -08001642 flags |= cachep->gfpflags;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001643
1644 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001645 if (!page)
1646 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001647
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001648 nr_pages = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001649 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
Christoph Lameter972d1a72006-09-25 23:31:51 -07001650 add_zone_page_state(page_zone(page),
1651 NR_SLAB_RECLAIMABLE, nr_pages);
1652 else
1653 add_zone_page_state(page_zone(page),
1654 NR_SLAB_UNRECLAIMABLE, nr_pages);
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001655 for (i = 0; i < nr_pages; i++)
1656 __SetPageSlab(page + i);
1657 return page_address(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001658}
1659
1660/*
1661 * Interface to system's page release.
1662 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001663static void kmem_freepages(struct kmem_cache *cachep, void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001664{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001665 unsigned long i = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001666 struct page *page = virt_to_page(addr);
1667 const unsigned long nr_freed = i;
1668
Christoph Lameter972d1a72006-09-25 23:31:51 -07001669 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1670 sub_zone_page_state(page_zone(page),
1671 NR_SLAB_RECLAIMABLE, nr_freed);
1672 else
1673 sub_zone_page_state(page_zone(page),
1674 NR_SLAB_UNRECLAIMABLE, nr_freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001675 while (i--) {
Nick Pigginf205b2f2006-03-22 00:08:02 -08001676 BUG_ON(!PageSlab(page));
1677 __ClearPageSlab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001678 page++;
1679 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001680 if (current->reclaim_state)
1681 current->reclaim_state->reclaimed_slab += nr_freed;
1682 free_pages((unsigned long)addr, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001683}
1684
1685static void kmem_rcu_free(struct rcu_head *head)
1686{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001687 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001688 struct kmem_cache *cachep = slab_rcu->cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001689
1690 kmem_freepages(cachep, slab_rcu->addr);
1691 if (OFF_SLAB(cachep))
1692 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1693}
1694
1695#if DEBUG
1696
1697#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001698static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001699 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001700{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001701 int size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001702
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001703 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001704
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001705 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001706 return;
1707
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001708 *addr++ = 0x12345678;
1709 *addr++ = caller;
1710 *addr++ = smp_processor_id();
1711 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001712 {
1713 unsigned long *sptr = &caller;
1714 unsigned long svalue;
1715
1716 while (!kstack_end(sptr)) {
1717 svalue = *sptr++;
1718 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001719 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001720 size -= sizeof(unsigned long);
1721 if (size <= sizeof(unsigned long))
1722 break;
1723 }
1724 }
1725
1726 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001727 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001728}
1729#endif
1730
Pekka Enberg343e0d72006-02-01 03:05:50 -08001731static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001732{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001733 int size = obj_size(cachep);
1734 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001735
1736 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001737 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001738}
1739
1740static void dump_line(char *data, int offset, int limit)
1741{
1742 int i;
Dave Jonesaa83aa42006-09-29 01:59:51 -07001743 unsigned char error = 0;
1744 int bad_count = 0;
1745
Linus Torvalds1da177e2005-04-16 15:20:36 -07001746 printk(KERN_ERR "%03x:", offset);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001747 for (i = 0; i < limit; i++) {
1748 if (data[offset + i] != POISON_FREE) {
1749 error = data[offset + i];
1750 bad_count++;
1751 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001752 printk(" %02x", (unsigned char)data[offset + i]);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001753 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001754 printk("\n");
Dave Jonesaa83aa42006-09-29 01:59:51 -07001755
1756 if (bad_count == 1) {
1757 error ^= POISON_FREE;
1758 if (!(error & (error - 1))) {
1759 printk(KERN_ERR "Single bit error detected. Probably "
1760 "bad RAM.\n");
1761#ifdef CONFIG_X86
1762 printk(KERN_ERR "Run memtest86+ or a similar memory "
1763 "test tool.\n");
1764#else
1765 printk(KERN_ERR "Run a memory test tool.\n");
1766#endif
1767 }
1768 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001769}
1770#endif
1771
1772#if DEBUG
1773
Pekka Enberg343e0d72006-02-01 03:05:50 -08001774static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001775{
1776 int i, size;
1777 char *realobj;
1778
1779 if (cachep->flags & SLAB_RED_ZONE) {
David Woodhouseb46b8f12007-05-08 00:22:59 -07001780 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001781 *dbg_redzone1(cachep, objp),
1782 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001783 }
1784
1785 if (cachep->flags & SLAB_STORE_USER) {
1786 printk(KERN_ERR "Last user: [<%p>]",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001787 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001788 print_symbol("(%s)",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001789 (unsigned long)*dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001790 printk("\n");
1791 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001792 realobj = (char *)objp + obj_offset(cachep);
1793 size = obj_size(cachep);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001794 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001795 int limit;
1796 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001797 if (i + limit > size)
1798 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001799 dump_line(realobj, i, limit);
1800 }
1801}
1802
Pekka Enberg343e0d72006-02-01 03:05:50 -08001803static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001804{
1805 char *realobj;
1806 int size, i;
1807 int lines = 0;
1808
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001809 realobj = (char *)objp + obj_offset(cachep);
1810 size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001811
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001812 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001813 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001814 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001815 exp = POISON_END;
1816 if (realobj[i] != exp) {
1817 int limit;
1818 /* Mismatch ! */
1819 /* Print header */
1820 if (lines == 0) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001821 printk(KERN_ERR
David Howellse94a40c2007-04-02 23:46:28 +01001822 "Slab corruption: %s start=%p, len=%d\n",
1823 cachep->name, realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001824 print_objinfo(cachep, objp, 0);
1825 }
1826 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001827 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001828 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001829 if (i + limit > size)
1830 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001831 dump_line(realobj, i, limit);
1832 i += 16;
1833 lines++;
1834 /* Limit to 5 lines */
1835 if (lines > 5)
1836 break;
1837 }
1838 }
1839 if (lines != 0) {
1840 /* Print some data about the neighboring objects, if they
1841 * exist:
1842 */
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08001843 struct slab *slabp = virt_to_slab(objp);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001844 unsigned int objnr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001845
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001846 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001847 if (objnr) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001848 objp = index_to_obj(cachep, slabp, objnr - 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001849 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001850 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001851 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001852 print_objinfo(cachep, objp, 2);
1853 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001854 if (objnr + 1 < cachep->num) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001855 objp = index_to_obj(cachep, slabp, objnr + 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001856 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001857 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001858 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001859 print_objinfo(cachep, objp, 2);
1860 }
1861 }
1862}
1863#endif
1864
Linus Torvalds1da177e2005-04-16 15:20:36 -07001865#if DEBUG
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001866/**
Randy Dunlap911851e2006-03-22 00:08:14 -08001867 * slab_destroy_objs - destroy a slab and its objects
1868 * @cachep: cache pointer being destroyed
1869 * @slabp: slab pointer being destroyed
1870 *
1871 * Call the registered destructor for each object in a slab that is being
1872 * destroyed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001873 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001874static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001875{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001876 int i;
1877 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001878 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001879
1880 if (cachep->flags & SLAB_POISON) {
1881#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08001882 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1883 OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001884 kernel_map_pages(virt_to_page(objp),
Andrew Mortona737b3e2006-03-22 00:08:11 -08001885 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001886 else
1887 check_poison_obj(cachep, objp);
1888#else
1889 check_poison_obj(cachep, objp);
1890#endif
1891 }
1892 if (cachep->flags & SLAB_RED_ZONE) {
1893 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1894 slab_error(cachep, "start of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001895 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001896 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1897 slab_error(cachep, "end of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001898 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001899 }
1900 if (cachep->dtor && !(cachep->flags & SLAB_POISON))
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001901 (cachep->dtor) (objp + obj_offset(cachep), cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001902 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001903}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001904#else
Pekka Enberg343e0d72006-02-01 03:05:50 -08001905static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001906{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001907 if (cachep->dtor) {
1908 int i;
1909 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001910 void *objp = index_to_obj(cachep, slabp, i);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001911 (cachep->dtor) (objp, cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001912 }
1913 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001914}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001915#endif
1916
Randy Dunlap911851e2006-03-22 00:08:14 -08001917/**
1918 * slab_destroy - destroy and release all objects in a slab
1919 * @cachep: cache pointer being destroyed
1920 * @slabp: slab pointer being destroyed
1921 *
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001922 * Destroy all the objs in a slab, and release the mem back to the system.
Andrew Mortona737b3e2006-03-22 00:08:11 -08001923 * Before calling the slab must have been unlinked from the cache. The
1924 * cache-lock is not held/needed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001925 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001926static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001927{
1928 void *addr = slabp->s_mem - slabp->colouroff;
1929
1930 slab_destroy_objs(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001931 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1932 struct slab_rcu *slab_rcu;
1933
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001934 slab_rcu = (struct slab_rcu *)slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001935 slab_rcu->cachep = cachep;
1936 slab_rcu->addr = addr;
1937 call_rcu(&slab_rcu->head, kmem_rcu_free);
1938 } else {
1939 kmem_freepages(cachep, addr);
Ingo Molnar873623d2006-07-13 14:44:38 +02001940 if (OFF_SLAB(cachep))
1941 kmem_cache_free(cachep->slabp_cache, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001942 }
1943}
1944
Andrew Mortona737b3e2006-03-22 00:08:11 -08001945/*
1946 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1947 * size of kmem_list3.
1948 */
Andrew Mortona3a02be2007-05-06 14:49:31 -07001949static void __init set_up_list3s(struct kmem_cache *cachep, int index)
Christoph Lametere498be72005-09-09 13:03:32 -07001950{
1951 int node;
1952
1953 for_each_online_node(node) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001954 cachep->nodelists[node] = &initkmem_list3[index + node];
Christoph Lametere498be72005-09-09 13:03:32 -07001955 cachep->nodelists[node]->next_reap = jiffies +
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001956 REAPTIMEOUT_LIST3 +
1957 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -07001958 }
1959}
1960
Christoph Lameter117f6eb2006-09-25 23:31:37 -07001961static void __kmem_cache_destroy(struct kmem_cache *cachep)
1962{
1963 int i;
1964 struct kmem_list3 *l3;
1965
1966 for_each_online_cpu(i)
1967 kfree(cachep->array[i]);
1968
1969 /* NUMA: free the list3 structures */
1970 for_each_online_node(i) {
1971 l3 = cachep->nodelists[i];
1972 if (l3) {
1973 kfree(l3->shared);
1974 free_alien_cache(l3->alien);
1975 kfree(l3);
1976 }
1977 }
1978 kmem_cache_free(&cache_cache, cachep);
1979}
1980
1981
Linus Torvalds1da177e2005-04-16 15:20:36 -07001982/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08001983 * calculate_slab_order - calculate size (page order) of slabs
1984 * @cachep: pointer to the cache that is being created
1985 * @size: size of objects to be created in this cache.
1986 * @align: required alignment for the objects.
1987 * @flags: slab allocation flags
1988 *
1989 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001990 *
1991 * This could be made much more intelligent. For now, try to avoid using
1992 * high order pages for slabs. When the gfp() functions are more friendly
1993 * towards high-order requests, this should be changed.
1994 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001995static size_t calculate_slab_order(struct kmem_cache *cachep,
Randy Dunlapee13d782006-02-01 03:05:53 -08001996 size_t size, size_t align, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001997{
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001998 unsigned long offslab_limit;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001999 size_t left_over = 0;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002000 int gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002001
Andrew Mortona737b3e2006-03-22 00:08:11 -08002002 for (gfporder = 0; gfporder <= MAX_GFP_ORDER; gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002003 unsigned int num;
2004 size_t remainder;
2005
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002006 cache_estimate(gfporder, size, align, flags, &remainder, &num);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002007 if (!num)
2008 continue;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002009
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002010 if (flags & CFLGS_OFF_SLAB) {
2011 /*
2012 * Max number of objs-per-slab for caches which
2013 * use off-slab slabs. Needed to avoid a possible
2014 * looping condition in cache_grow().
2015 */
2016 offslab_limit = size - sizeof(struct slab);
2017 offslab_limit /= sizeof(kmem_bufctl_t);
2018
2019 if (num > offslab_limit)
2020 break;
2021 }
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002022
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002023 /* Found something acceptable - save it away */
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002024 cachep->num = num;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002025 cachep->gfporder = gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002026 left_over = remainder;
2027
2028 /*
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002029 * A VFS-reclaimable slab tends to have most allocations
2030 * as GFP_NOFS and we really don't want to have to be allocating
2031 * higher-order pages when we are unable to shrink dcache.
2032 */
2033 if (flags & SLAB_RECLAIM_ACCOUNT)
2034 break;
2035
2036 /*
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002037 * Large number of objects is good, but very large slabs are
2038 * currently bad for the gfp()s.
2039 */
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002040 if (gfporder >= slab_break_gfp_order)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002041 break;
2042
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002043 /*
2044 * Acceptable internal fragmentation?
2045 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002046 if (left_over * 8 <= (PAGE_SIZE << gfporder))
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002047 break;
2048 }
2049 return left_over;
2050}
2051
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002052static int setup_cpu_cache(struct kmem_cache *cachep)
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002053{
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002054 if (g_cpucache_up == FULL)
2055 return enable_cpucache(cachep);
2056
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002057 if (g_cpucache_up == NONE) {
2058 /*
2059 * Note: the first kmem_cache_create must create the cache
2060 * that's used by kmalloc(24), otherwise the creation of
2061 * further caches will BUG().
2062 */
2063 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2064
2065 /*
2066 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2067 * the first cache, then we need to set up all its list3s,
2068 * otherwise the creation of further caches will BUG().
2069 */
2070 set_up_list3s(cachep, SIZE_AC);
2071 if (INDEX_AC == INDEX_L3)
2072 g_cpucache_up = PARTIAL_L3;
2073 else
2074 g_cpucache_up = PARTIAL_AC;
2075 } else {
2076 cachep->array[smp_processor_id()] =
2077 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
2078
2079 if (g_cpucache_up == PARTIAL_AC) {
2080 set_up_list3s(cachep, SIZE_L3);
2081 g_cpucache_up = PARTIAL_L3;
2082 } else {
2083 int node;
2084 for_each_online_node(node) {
2085 cachep->nodelists[node] =
2086 kmalloc_node(sizeof(struct kmem_list3),
2087 GFP_KERNEL, node);
2088 BUG_ON(!cachep->nodelists[node]);
2089 kmem_list3_init(cachep->nodelists[node]);
2090 }
2091 }
2092 }
2093 cachep->nodelists[numa_node_id()]->next_reap =
2094 jiffies + REAPTIMEOUT_LIST3 +
2095 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2096
2097 cpu_cache_get(cachep)->avail = 0;
2098 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2099 cpu_cache_get(cachep)->batchcount = 1;
2100 cpu_cache_get(cachep)->touched = 0;
2101 cachep->batchcount = 1;
2102 cachep->limit = BOOT_CPUCACHE_ENTRIES;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002103 return 0;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002104}
2105
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002106/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07002107 * kmem_cache_create - Create a cache.
2108 * @name: A string which is used in /proc/slabinfo to identify this cache.
2109 * @size: The size of objects to be created in this cache.
2110 * @align: The required alignment for the objects.
2111 * @flags: SLAB flags
2112 * @ctor: A constructor for the objects.
2113 * @dtor: A destructor for the objects.
2114 *
2115 * Returns a ptr to the cache on success, NULL on failure.
2116 * Cannot be called within a int, but can be interrupted.
2117 * The @ctor is run when new pages are allocated by the cache
2118 * and the @dtor is run before the pages are handed back.
2119 *
2120 * @name must be valid until the cache is destroyed. This implies that
Andrew Mortona737b3e2006-03-22 00:08:11 -08002121 * the module calling this has to destroy the cache before getting unloaded.
2122 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002123 * The flags are
2124 *
2125 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2126 * to catch references to uninitialised memory.
2127 *
2128 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2129 * for buffer overruns.
2130 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002131 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2132 * cacheline. This can be beneficial if you're counting cycles as closely
2133 * as davem.
2134 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002135struct kmem_cache *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002136kmem_cache_create (const char *name, size_t size, size_t align,
Andrew Mortona737b3e2006-03-22 00:08:11 -08002137 unsigned long flags,
2138 void (*ctor)(void*, struct kmem_cache *, unsigned long),
Pekka Enberg343e0d72006-02-01 03:05:50 -08002139 void (*dtor)(void*, struct kmem_cache *, unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002140{
2141 size_t left_over, slab_size, ralign;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07002142 struct kmem_cache *cachep = NULL, *pc;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002143
2144 /*
2145 * Sanity checks... these are all serious usage bugs.
2146 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002147 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002148 (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002149 printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__,
2150 name);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002151 BUG();
2152 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002153
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002154 /*
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002155 * We use cache_chain_mutex to ensure a consistent view of
2156 * cpu_online_map as well. Please see cpuup_callback
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002157 */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002158 mutex_lock(&cache_chain_mutex);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002159
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07002160 list_for_each_entry(pc, &cache_chain, next) {
Andrew Morton4f12bb42005-11-07 00:58:00 -08002161 char tmp;
2162 int res;
2163
2164 /*
2165 * This happens when the module gets unloaded and doesn't
2166 * destroy its slab cache and no-one else reuses the vmalloc
2167 * area of the module. Print a warning.
2168 */
Andrew Morton138ae662006-12-06 20:36:41 -08002169 res = probe_kernel_address(pc->name, tmp);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002170 if (res) {
matzeb4169522007-05-06 14:49:52 -07002171 printk(KERN_ERR
2172 "SLAB: cache with size %d has lost its name\n",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002173 pc->buffer_size);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002174 continue;
2175 }
2176
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002177 if (!strcmp(pc->name, name)) {
matzeb4169522007-05-06 14:49:52 -07002178 printk(KERN_ERR
2179 "kmem_cache_create: duplicate cache %s\n", name);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002180 dump_stack();
2181 goto oops;
2182 }
2183 }
2184
Linus Torvalds1da177e2005-04-16 15:20:36 -07002185#if DEBUG
2186 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002187#if FORCED_DEBUG
2188 /*
2189 * Enable redzoning and last user accounting, except for caches with
2190 * large objects, if the increased size would increase the object size
2191 * above the next power of two: caches with object sizes just above a
2192 * power of two have a significant amount of internal fragmentation.
2193 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002194 if (size < 4096 || fls(size - 1) == fls(size-1 + 3 * BYTES_PER_WORD))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002195 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002196 if (!(flags & SLAB_DESTROY_BY_RCU))
2197 flags |= SLAB_POISON;
2198#endif
2199 if (flags & SLAB_DESTROY_BY_RCU)
2200 BUG_ON(flags & SLAB_POISON);
2201#endif
2202 if (flags & SLAB_DESTROY_BY_RCU)
2203 BUG_ON(dtor);
2204
2205 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002206 * Always checks flags, a caller might be expecting debug support which
2207 * isn't available.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002208 */
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002209 BUG_ON(flags & ~CREATE_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002210
Andrew Mortona737b3e2006-03-22 00:08:11 -08002211 /*
2212 * Check that size is in terms of words. This is needed to avoid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002213 * unaligned accesses for some archs when redzoning is used, and makes
2214 * sure any on-slab bufctl's are also correctly aligned.
2215 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002216 if (size & (BYTES_PER_WORD - 1)) {
2217 size += (BYTES_PER_WORD - 1);
2218 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002219 }
2220
Andrew Mortona737b3e2006-03-22 00:08:11 -08002221 /* calculate the final buffer alignment: */
2222
Linus Torvalds1da177e2005-04-16 15:20:36 -07002223 /* 1) arch recommendation: can be overridden for debug */
2224 if (flags & SLAB_HWCACHE_ALIGN) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002225 /*
2226 * Default alignment: as specified by the arch code. Except if
2227 * an object is really small, then squeeze multiple objects into
2228 * one cacheline.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002229 */
2230 ralign = cache_line_size();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002231 while (size <= ralign / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002232 ralign /= 2;
2233 } else {
2234 ralign = BYTES_PER_WORD;
2235 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002236
2237 /*
2238 * Redzoning and user store require word alignment. Note this will be
2239 * overridden by architecture or caller mandated alignment if either
2240 * is greater than BYTES_PER_WORD.
2241 */
2242 if (flags & SLAB_RED_ZONE || flags & SLAB_STORE_USER)
David Woodhouseb46b8f12007-05-08 00:22:59 -07002243 ralign = __alignof__(unsigned long long);
Pekka Enbergca5f9702006-09-25 23:31:25 -07002244
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002245 /* 2) arch mandated alignment */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002246 if (ralign < ARCH_SLAB_MINALIGN) {
2247 ralign = ARCH_SLAB_MINALIGN;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002248 }
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002249 /* 3) caller mandated alignment */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002250 if (ralign < align) {
2251 ralign = align;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002252 }
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002253 /* disable debug if necessary */
David Woodhouseb46b8f12007-05-08 00:22:59 -07002254 if (ralign > __alignof__(unsigned long long))
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002255 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002256 /*
Pekka Enbergca5f9702006-09-25 23:31:25 -07002257 * 4) Store it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002258 */
2259 align = ralign;
2260
2261 /* Get cache's description obj. */
Christoph Lametere94b1762006-12-06 20:33:17 -08002262 cachep = kmem_cache_zalloc(&cache_cache, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002263 if (!cachep)
Andrew Morton4f12bb42005-11-07 00:58:00 -08002264 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002265
2266#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002267 cachep->obj_size = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002268
Pekka Enbergca5f9702006-09-25 23:31:25 -07002269 /*
2270 * Both debugging options require word-alignment which is calculated
2271 * into align above.
2272 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002273 if (flags & SLAB_RED_ZONE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002274 /* add space for red zone words */
David Woodhouseb46b8f12007-05-08 00:22:59 -07002275 cachep->obj_offset += sizeof(unsigned long long);
2276 size += 2 * sizeof(unsigned long long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002277 }
2278 if (flags & SLAB_STORE_USER) {
Pekka Enbergca5f9702006-09-25 23:31:25 -07002279 /* user store requires one word storage behind the end of
2280 * the real object.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002281 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002282 size += BYTES_PER_WORD;
2283 }
2284#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002285 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002286 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2287 cachep->obj_offset += PAGE_SIZE - size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002288 size = PAGE_SIZE;
2289 }
2290#endif
2291#endif
2292
Ingo Molnare0a42722006-06-23 02:03:46 -07002293 /*
2294 * Determine if the slab management is 'on' or 'off' slab.
2295 * (bootstrapping cannot cope with offslab caches so don't do
2296 * it too early on.)
2297 */
2298 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002299 /*
2300 * Size is large, assume best to place the slab management obj
2301 * off-slab (should allow better packing of objs).
2302 */
2303 flags |= CFLGS_OFF_SLAB;
2304
2305 size = ALIGN(size, align);
2306
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002307 left_over = calculate_slab_order(cachep, size, align, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002308
2309 if (!cachep->num) {
matzeb4169522007-05-06 14:49:52 -07002310 printk(KERN_ERR
2311 "kmem_cache_create: couldn't create cache %s.\n", name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002312 kmem_cache_free(&cache_cache, cachep);
2313 cachep = NULL;
Andrew Morton4f12bb42005-11-07 00:58:00 -08002314 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002315 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002316 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2317 + sizeof(struct slab), align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002318
2319 /*
2320 * If the slab has been placed off-slab, and we have enough space then
2321 * move it on-slab. This is at the expense of any extra colouring.
2322 */
2323 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2324 flags &= ~CFLGS_OFF_SLAB;
2325 left_over -= slab_size;
2326 }
2327
2328 if (flags & CFLGS_OFF_SLAB) {
2329 /* really off slab. No need for manual alignment */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002330 slab_size =
2331 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002332 }
2333
2334 cachep->colour_off = cache_line_size();
2335 /* Offset must be a multiple of the alignment. */
2336 if (cachep->colour_off < align)
2337 cachep->colour_off = align;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002338 cachep->colour = left_over / cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002339 cachep->slab_size = slab_size;
2340 cachep->flags = flags;
2341 cachep->gfpflags = 0;
Christoph Lameter4b51d662007-02-10 01:43:10 -08002342 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002343 cachep->gfpflags |= GFP_DMA;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002344 cachep->buffer_size = size;
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08002345 cachep->reciprocal_buffer_size = reciprocal_value(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002346
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002347 if (flags & CFLGS_OFF_SLAB) {
Victor Fuscob2d55072005-09-10 00:26:36 -07002348 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002349 /*
2350 * This is a possibility for one of the malloc_sizes caches.
2351 * But since we go off slab only for object size greater than
2352 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2353 * this should not happen at all.
2354 * But leave a BUG_ON for some lucky dude.
2355 */
2356 BUG_ON(!cachep->slabp_cache);
2357 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002358 cachep->ctor = ctor;
2359 cachep->dtor = dtor;
2360 cachep->name = name;
2361
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002362 if (setup_cpu_cache(cachep)) {
2363 __kmem_cache_destroy(cachep);
2364 cachep = NULL;
2365 goto oops;
2366 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002367
Linus Torvalds1da177e2005-04-16 15:20:36 -07002368 /* cache setup completed, link it into the list */
2369 list_add(&cachep->next, &cache_chain);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002370oops:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002371 if (!cachep && (flags & SLAB_PANIC))
2372 panic("kmem_cache_create(): failed to create slab `%s'\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002373 name);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002374 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002375 return cachep;
2376}
2377EXPORT_SYMBOL(kmem_cache_create);
2378
2379#if DEBUG
2380static void check_irq_off(void)
2381{
2382 BUG_ON(!irqs_disabled());
2383}
2384
2385static void check_irq_on(void)
2386{
2387 BUG_ON(irqs_disabled());
2388}
2389
Pekka Enberg343e0d72006-02-01 03:05:50 -08002390static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002391{
2392#ifdef CONFIG_SMP
2393 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002394 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002395#endif
2396}
Christoph Lametere498be72005-09-09 13:03:32 -07002397
Pekka Enberg343e0d72006-02-01 03:05:50 -08002398static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002399{
2400#ifdef CONFIG_SMP
2401 check_irq_off();
2402 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2403#endif
2404}
2405
Linus Torvalds1da177e2005-04-16 15:20:36 -07002406#else
2407#define check_irq_off() do { } while(0)
2408#define check_irq_on() do { } while(0)
2409#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002410#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002411#endif
2412
Christoph Lameteraab22072006-03-22 00:09:06 -08002413static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2414 struct array_cache *ac,
2415 int force, int node);
2416
Linus Torvalds1da177e2005-04-16 15:20:36 -07002417static void do_drain(void *arg)
2418{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002419 struct kmem_cache *cachep = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002420 struct array_cache *ac;
Christoph Lameterff694162005-09-22 21:44:02 -07002421 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002422
2423 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002424 ac = cpu_cache_get(cachep);
Christoph Lameterff694162005-09-22 21:44:02 -07002425 spin_lock(&cachep->nodelists[node]->list_lock);
2426 free_block(cachep, ac->entry, ac->avail, node);
2427 spin_unlock(&cachep->nodelists[node]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002428 ac->avail = 0;
2429}
2430
Pekka Enberg343e0d72006-02-01 03:05:50 -08002431static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002432{
Christoph Lametere498be72005-09-09 13:03:32 -07002433 struct kmem_list3 *l3;
2434 int node;
2435
Andrew Mortona07fa392006-03-22 00:08:17 -08002436 on_each_cpu(do_drain, cachep, 1, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002437 check_irq_on();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002438 for_each_online_node(node) {
Christoph Lametere498be72005-09-09 13:03:32 -07002439 l3 = cachep->nodelists[node];
Roland Dreiera4523a82006-05-15 11:41:00 -07002440 if (l3 && l3->alien)
2441 drain_alien_cache(cachep, l3->alien);
2442 }
2443
2444 for_each_online_node(node) {
2445 l3 = cachep->nodelists[node];
2446 if (l3)
Christoph Lameteraab22072006-03-22 00:09:06 -08002447 drain_array(cachep, l3, l3->shared, 1, node);
Christoph Lametere498be72005-09-09 13:03:32 -07002448 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002449}
2450
Christoph Lametered11d9e2006-06-30 01:55:45 -07002451/*
2452 * Remove slabs from the list of free slabs.
2453 * Specify the number of slabs to drain in tofree.
2454 *
2455 * Returns the actual number of slabs released.
2456 */
2457static int drain_freelist(struct kmem_cache *cache,
2458 struct kmem_list3 *l3, int tofree)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002459{
Christoph Lametered11d9e2006-06-30 01:55:45 -07002460 struct list_head *p;
2461 int nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002462 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002463
Christoph Lametered11d9e2006-06-30 01:55:45 -07002464 nr_freed = 0;
2465 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002466
Christoph Lametered11d9e2006-06-30 01:55:45 -07002467 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002468 p = l3->slabs_free.prev;
Christoph Lametered11d9e2006-06-30 01:55:45 -07002469 if (p == &l3->slabs_free) {
2470 spin_unlock_irq(&l3->list_lock);
2471 goto out;
2472 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002473
Christoph Lametered11d9e2006-06-30 01:55:45 -07002474 slabp = list_entry(p, struct slab, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002475#if DEBUG
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002476 BUG_ON(slabp->inuse);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002477#endif
2478 list_del(&slabp->list);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002479 /*
2480 * Safe to drop the lock. The slab is no longer linked
2481 * to the cache.
2482 */
2483 l3->free_objects -= cache->num;
Christoph Lametere498be72005-09-09 13:03:32 -07002484 spin_unlock_irq(&l3->list_lock);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002485 slab_destroy(cache, slabp);
2486 nr_freed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002487 }
Christoph Lametered11d9e2006-06-30 01:55:45 -07002488out:
2489 return nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002490}
2491
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002492/* Called with cache_chain_mutex held to protect against cpu hotplug */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002493static int __cache_shrink(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07002494{
2495 int ret = 0, i = 0;
2496 struct kmem_list3 *l3;
2497
2498 drain_cpu_caches(cachep);
2499
2500 check_irq_on();
2501 for_each_online_node(i) {
2502 l3 = cachep->nodelists[i];
Christoph Lametered11d9e2006-06-30 01:55:45 -07002503 if (!l3)
2504 continue;
2505
2506 drain_freelist(cachep, l3, l3->free_objects);
2507
2508 ret += !list_empty(&l3->slabs_full) ||
2509 !list_empty(&l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07002510 }
2511 return (ret ? 1 : 0);
2512}
2513
Linus Torvalds1da177e2005-04-16 15:20:36 -07002514/**
2515 * kmem_cache_shrink - Shrink a cache.
2516 * @cachep: The cache to shrink.
2517 *
2518 * Releases as many slabs as possible for a cache.
2519 * To help debugging, a zero exit status indicates all slabs were released.
2520 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002521int kmem_cache_shrink(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002522{
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002523 int ret;
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002524 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002525
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002526 mutex_lock(&cache_chain_mutex);
2527 ret = __cache_shrink(cachep);
2528 mutex_unlock(&cache_chain_mutex);
2529 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002530}
2531EXPORT_SYMBOL(kmem_cache_shrink);
2532
2533/**
2534 * kmem_cache_destroy - delete a cache
2535 * @cachep: the cache to destroy
2536 *
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08002537 * Remove a &struct kmem_cache object from the slab cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002538 *
2539 * It is expected this function will be called by a module when it is
2540 * unloaded. This will remove the cache completely, and avoid a duplicate
2541 * cache being allocated each time a module is loaded and unloaded, if the
2542 * module doesn't have persistent in-kernel storage across loads and unloads.
2543 *
2544 * The cache must be empty before calling this function.
2545 *
2546 * The caller must guarantee that noone will allocate memory from the cache
2547 * during the kmem_cache_destroy().
2548 */
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002549void kmem_cache_destroy(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002550{
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002551 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002552
Linus Torvalds1da177e2005-04-16 15:20:36 -07002553 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002554 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002555 /*
2556 * the chain is never empty, cache_cache is never destroyed
2557 */
2558 list_del(&cachep->next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002559 if (__cache_shrink(cachep)) {
2560 slab_error(cachep, "Can't free all objects");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002561 list_add(&cachep->next, &cache_chain);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002562 mutex_unlock(&cache_chain_mutex);
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002563 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002564 }
2565
2566 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
Paul E. McKenneyfbd568a3e2005-05-01 08:59:04 -07002567 synchronize_rcu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002568
Christoph Lameter117f6eb2006-09-25 23:31:37 -07002569 __kmem_cache_destroy(cachep);
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002570 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002571}
2572EXPORT_SYMBOL(kmem_cache_destroy);
2573
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002574/*
2575 * Get the memory for a slab management obj.
2576 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2577 * always come from malloc_sizes caches. The slab descriptor cannot
2578 * come from the same cache which is getting created because,
2579 * when we are searching for an appropriate cache for these
2580 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2581 * If we are creating a malloc_sizes cache here it would not be visible to
2582 * kmem_find_general_cachep till the initialization is complete.
2583 * Hence we cannot have slabp_cache same as the original cache.
2584 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002585static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002586 int colour_off, gfp_t local_flags,
2587 int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002588{
2589 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002590
Linus Torvalds1da177e2005-04-16 15:20:36 -07002591 if (OFF_SLAB(cachep)) {
2592 /* Slab management obj is off-slab. */
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002593 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
Christoph Lameter3c517a62006-12-06 20:33:29 -08002594 local_flags & ~GFP_THISNODE, nodeid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002595 if (!slabp)
2596 return NULL;
2597 } else {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002598 slabp = objp + colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002599 colour_off += cachep->slab_size;
2600 }
2601 slabp->inuse = 0;
2602 slabp->colouroff = colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002603 slabp->s_mem = objp + colour_off;
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002604 slabp->nodeid = nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002605 return slabp;
2606}
2607
2608static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2609{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002610 return (kmem_bufctl_t *) (slabp + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002611}
2612
Pekka Enberg343e0d72006-02-01 03:05:50 -08002613static void cache_init_objs(struct kmem_cache *cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002614 struct slab *slabp, unsigned long ctor_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002615{
2616 int i;
2617
2618 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002619 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002620#if DEBUG
2621 /* need to poison the objs? */
2622 if (cachep->flags & SLAB_POISON)
2623 poison_obj(cachep, objp, POISON_FREE);
2624 if (cachep->flags & SLAB_STORE_USER)
2625 *dbg_userword(cachep, objp) = NULL;
2626
2627 if (cachep->flags & SLAB_RED_ZONE) {
2628 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2629 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2630 }
2631 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002632 * Constructors are not allowed to allocate memory from the same
2633 * cache which they are a constructor for. Otherwise, deadlock.
2634 * They must also be threaded.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002635 */
2636 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002637 cachep->ctor(objp + obj_offset(cachep), cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002638 ctor_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002639
2640 if (cachep->flags & SLAB_RED_ZONE) {
2641 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2642 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002643 " end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002644 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2645 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002646 " start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002647 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08002648 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2649 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002650 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002651 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002652#else
2653 if (cachep->ctor)
2654 cachep->ctor(objp, cachep, ctor_flags);
2655#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002656 slab_bufctl(slabp)[i] = i + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002657 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002658 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002659 slabp->free = 0;
2660}
2661
Pekka Enberg343e0d72006-02-01 03:05:50 -08002662static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002663{
Christoph Lameter4b51d662007-02-10 01:43:10 -08002664 if (CONFIG_ZONE_DMA_FLAG) {
2665 if (flags & GFP_DMA)
2666 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2667 else
2668 BUG_ON(cachep->gfpflags & GFP_DMA);
2669 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002670}
2671
Andrew Mortona737b3e2006-03-22 00:08:11 -08002672static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2673 int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002674{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002675 void *objp = index_to_obj(cachep, slabp, slabp->free);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002676 kmem_bufctl_t next;
2677
2678 slabp->inuse++;
2679 next = slab_bufctl(slabp)[slabp->free];
2680#if DEBUG
2681 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2682 WARN_ON(slabp->nodeid != nodeid);
2683#endif
2684 slabp->free = next;
2685
2686 return objp;
2687}
2688
Andrew Mortona737b3e2006-03-22 00:08:11 -08002689static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2690 void *objp, int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002691{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002692 unsigned int objnr = obj_to_index(cachep, slabp, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002693
2694#if DEBUG
2695 /* Verify that the slab belongs to the intended node */
2696 WARN_ON(slabp->nodeid != nodeid);
2697
Al Viro871751e2006-03-25 03:06:39 -08002698 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
Matthew Dobson78d382d2006-02-01 03:05:47 -08002699 printk(KERN_ERR "slab: double free detected in cache "
Andrew Mortona737b3e2006-03-22 00:08:11 -08002700 "'%s', objp %p\n", cachep->name, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002701 BUG();
2702 }
2703#endif
2704 slab_bufctl(slabp)[objnr] = slabp->free;
2705 slabp->free = objnr;
2706 slabp->inuse--;
2707}
2708
Pekka Enberg47768742006-06-23 02:03:07 -07002709/*
2710 * Map pages beginning at addr to the given cache and slab. This is required
2711 * for the slab allocator to be able to lookup the cache and slab of a
2712 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2713 */
2714static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2715 void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002716{
Pekka Enberg47768742006-06-23 02:03:07 -07002717 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002718 struct page *page;
2719
Pekka Enberg47768742006-06-23 02:03:07 -07002720 page = virt_to_page(addr);
Nick Piggin84097512006-03-22 00:08:34 -08002721
Pekka Enberg47768742006-06-23 02:03:07 -07002722 nr_pages = 1;
Nick Piggin84097512006-03-22 00:08:34 -08002723 if (likely(!PageCompound(page)))
Pekka Enberg47768742006-06-23 02:03:07 -07002724 nr_pages <<= cache->gfporder;
2725
Linus Torvalds1da177e2005-04-16 15:20:36 -07002726 do {
Pekka Enberg47768742006-06-23 02:03:07 -07002727 page_set_cache(page, cache);
2728 page_set_slab(page, slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002729 page++;
Pekka Enberg47768742006-06-23 02:03:07 -07002730 } while (--nr_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002731}
2732
2733/*
2734 * Grow (by 1) the number of slabs within a cache. This is called by
2735 * kmem_cache_alloc() when there are no active objs left in a cache.
2736 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002737static int cache_grow(struct kmem_cache *cachep,
2738 gfp_t flags, int nodeid, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002739{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002740 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002741 size_t offset;
2742 gfp_t local_flags;
2743 unsigned long ctor_flags;
Christoph Lametere498be72005-09-09 13:03:32 -07002744 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002745
Andrew Mortona737b3e2006-03-22 00:08:11 -08002746 /*
2747 * Be lazy and only check for valid flags here, keeping it out of the
2748 * critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002749 */
Christoph Lametercfce6602007-05-06 14:50:17 -07002750 BUG_ON(flags & ~(GFP_DMA | GFP_LEVEL_MASK));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002751
2752 ctor_flags = SLAB_CTOR_CONSTRUCTOR;
Christoph Lametera06d72c2006-12-06 20:33:12 -08002753 local_flags = (flags & GFP_LEVEL_MASK);
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002754 /* Take the l3 list lock to change the colour_next on this node */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002755 check_irq_off();
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002756 l3 = cachep->nodelists[nodeid];
2757 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002758
2759 /* Get colour for the slab, and cal the next value. */
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002760 offset = l3->colour_next;
2761 l3->colour_next++;
2762 if (l3->colour_next >= cachep->colour)
2763 l3->colour_next = 0;
2764 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002765
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002766 offset *= cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002767
2768 if (local_flags & __GFP_WAIT)
2769 local_irq_enable();
2770
2771 /*
2772 * The test for missing atomic flag is performed here, rather than
2773 * the more obvious place, simply to reduce the critical path length
2774 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2775 * will eventually be caught here (where it matters).
2776 */
2777 kmem_flagcheck(cachep, flags);
2778
Andrew Mortona737b3e2006-03-22 00:08:11 -08002779 /*
2780 * Get mem for the objs. Attempt to allocate a physical page from
2781 * 'nodeid'.
Christoph Lametere498be72005-09-09 13:03:32 -07002782 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002783 if (!objp)
2784 objp = kmem_getpages(cachep, flags, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002785 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002786 goto failed;
2787
2788 /* Get slab management. */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002789 slabp = alloc_slabmgmt(cachep, objp, offset,
2790 local_flags & ~GFP_THISNODE, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002791 if (!slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002792 goto opps1;
2793
Christoph Lametere498be72005-09-09 13:03:32 -07002794 slabp->nodeid = nodeid;
Pekka Enberg47768742006-06-23 02:03:07 -07002795 slab_map_pages(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002796
2797 cache_init_objs(cachep, slabp, ctor_flags);
2798
2799 if (local_flags & __GFP_WAIT)
2800 local_irq_disable();
2801 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002802 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002803
2804 /* Make slab active. */
Christoph Lametere498be72005-09-09 13:03:32 -07002805 list_add_tail(&slabp->list, &(l3->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002806 STATS_INC_GROWN(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002807 l3->free_objects += cachep->num;
2808 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002809 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002810opps1:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002811 kmem_freepages(cachep, objp);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002812failed:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002813 if (local_flags & __GFP_WAIT)
2814 local_irq_disable();
2815 return 0;
2816}
2817
2818#if DEBUG
2819
2820/*
2821 * Perform extra freeing checks:
2822 * - detect bad pointers.
2823 * - POISON/RED_ZONE checking
2824 * - destructor calls, for caches with POISON+dtor
2825 */
2826static void kfree_debugcheck(const void *objp)
2827{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002828 if (!virt_addr_valid(objp)) {
2829 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002830 (unsigned long)objp);
2831 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002832 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002833}
2834
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002835static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2836{
David Woodhouseb46b8f12007-05-08 00:22:59 -07002837 unsigned long long redzone1, redzone2;
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002838
2839 redzone1 = *dbg_redzone1(cache, obj);
2840 redzone2 = *dbg_redzone2(cache, obj);
2841
2842 /*
2843 * Redzone is ok.
2844 */
2845 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2846 return;
2847
2848 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2849 slab_error(cache, "double free detected");
2850 else
2851 slab_error(cache, "memory outside object was overwritten");
2852
David Woodhouseb46b8f12007-05-08 00:22:59 -07002853 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002854 obj, redzone1, redzone2);
2855}
2856
Pekka Enberg343e0d72006-02-01 03:05:50 -08002857static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002858 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002859{
2860 struct page *page;
2861 unsigned int objnr;
2862 struct slab *slabp;
2863
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002864 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002865 kfree_debugcheck(objp);
Christoph Lameterb49af682007-05-06 14:49:41 -07002866 page = virt_to_head_page(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002867
Pekka Enberg065d41c2005-11-13 16:06:46 -08002868 slabp = page_get_slab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002869
2870 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002871 verify_redzone_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002872 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2873 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2874 }
2875 if (cachep->flags & SLAB_STORE_USER)
2876 *dbg_userword(cachep, objp) = caller;
2877
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002878 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002879
2880 BUG_ON(objnr >= cachep->num);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002881 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002882
Linus Torvalds1da177e2005-04-16 15:20:36 -07002883 if (cachep->flags & SLAB_POISON && cachep->dtor) {
2884 /* we want to cache poison the object,
2885 * call the destruction callback
2886 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002887 cachep->dtor(objp + obj_offset(cachep), cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002888 }
Al Viro871751e2006-03-25 03:06:39 -08002889#ifdef CONFIG_DEBUG_SLAB_LEAK
2890 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2891#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002892 if (cachep->flags & SLAB_POISON) {
2893#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08002894 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002895 store_stackinfo(cachep, objp, (unsigned long)caller);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002896 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002897 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002898 } else {
2899 poison_obj(cachep, objp, POISON_FREE);
2900 }
2901#else
2902 poison_obj(cachep, objp, POISON_FREE);
2903#endif
2904 }
2905 return objp;
2906}
2907
Pekka Enberg343e0d72006-02-01 03:05:50 -08002908static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002909{
2910 kmem_bufctl_t i;
2911 int entries = 0;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002912
Linus Torvalds1da177e2005-04-16 15:20:36 -07002913 /* Check slab's freelist to see if this obj is there. */
2914 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2915 entries++;
2916 if (entries > cachep->num || i >= cachep->num)
2917 goto bad;
2918 }
2919 if (entries != cachep->num - slabp->inuse) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002920bad:
2921 printk(KERN_ERR "slab: Internal list corruption detected in "
2922 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2923 cachep->name, cachep->num, slabp, slabp->inuse);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002924 for (i = 0;
Linus Torvalds264132b2006-03-06 12:10:07 -08002925 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002926 i++) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002927 if (i % 16 == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002928 printk("\n%03x:", i);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002929 printk(" %02x", ((unsigned char *)slabp)[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002930 }
2931 printk("\n");
2932 BUG();
2933 }
2934}
2935#else
2936#define kfree_debugcheck(x) do { } while(0)
2937#define cache_free_debugcheck(x,objp,z) (objp)
2938#define check_slabp(x,y) do { } while(0)
2939#endif
2940
Pekka Enberg343e0d72006-02-01 03:05:50 -08002941static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002942{
2943 int batchcount;
2944 struct kmem_list3 *l3;
2945 struct array_cache *ac;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002946 int node;
2947
2948 node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002949
2950 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002951 ac = cpu_cache_get(cachep);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002952retry:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002953 batchcount = ac->batchcount;
2954 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002955 /*
2956 * If there was little recent activity on this cache, then
2957 * perform only a partial refill. Otherwise we could generate
2958 * refill bouncing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002959 */
2960 batchcount = BATCHREFILL_LIMIT;
2961 }
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002962 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002963
Christoph Lametere498be72005-09-09 13:03:32 -07002964 BUG_ON(ac->avail > 0 || !l3);
2965 spin_lock(&l3->list_lock);
2966
Christoph Lameter3ded1752006-03-25 03:06:44 -08002967 /* See if we can refill from the shared array */
2968 if (l3->shared && transfer_objects(ac, l3->shared, batchcount))
2969 goto alloc_done;
2970
Linus Torvalds1da177e2005-04-16 15:20:36 -07002971 while (batchcount > 0) {
2972 struct list_head *entry;
2973 struct slab *slabp;
2974 /* Get slab alloc is to come from. */
2975 entry = l3->slabs_partial.next;
2976 if (entry == &l3->slabs_partial) {
2977 l3->free_touched = 1;
2978 entry = l3->slabs_free.next;
2979 if (entry == &l3->slabs_free)
2980 goto must_grow;
2981 }
2982
2983 slabp = list_entry(entry, struct slab, list);
2984 check_slabp(cachep, slabp);
2985 check_spinlock_acquired(cachep);
Pekka Enberg714b81712007-05-06 14:49:03 -07002986
2987 /*
2988 * The slab was either on partial or free list so
2989 * there must be at least one object available for
2990 * allocation.
2991 */
2992 BUG_ON(slabp->inuse < 0 || slabp->inuse >= cachep->num);
2993
Linus Torvalds1da177e2005-04-16 15:20:36 -07002994 while (slabp->inuse < cachep->num && batchcount--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002995 STATS_INC_ALLOCED(cachep);
2996 STATS_INC_ACTIVE(cachep);
2997 STATS_SET_HIGH(cachep);
2998
Matthew Dobson78d382d2006-02-01 03:05:47 -08002999 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07003000 node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003001 }
3002 check_slabp(cachep, slabp);
3003
3004 /* move slabp to correct slabp list: */
3005 list_del(&slabp->list);
3006 if (slabp->free == BUFCTL_END)
3007 list_add(&slabp->list, &l3->slabs_full);
3008 else
3009 list_add(&slabp->list, &l3->slabs_partial);
3010 }
3011
Andrew Mortona737b3e2006-03-22 00:08:11 -08003012must_grow:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003013 l3->free_objects -= ac->avail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003014alloc_done:
Christoph Lametere498be72005-09-09 13:03:32 -07003015 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003016
3017 if (unlikely(!ac->avail)) {
3018 int x;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003019 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07003020
Andrew Mortona737b3e2006-03-22 00:08:11 -08003021 /* cache_grow can reenable interrupts, then ac could change. */
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003022 ac = cpu_cache_get(cachep);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003023 if (!x && ac->avail == 0) /* no objects in sight? abort */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003024 return NULL;
3025
Andrew Mortona737b3e2006-03-22 00:08:11 -08003026 if (!ac->avail) /* objects refilled by interrupt? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003027 goto retry;
3028 }
3029 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07003030 return ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003031}
3032
Andrew Mortona737b3e2006-03-22 00:08:11 -08003033static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3034 gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003035{
3036 might_sleep_if(flags & __GFP_WAIT);
3037#if DEBUG
3038 kmem_flagcheck(cachep, flags);
3039#endif
3040}
3041
3042#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003043static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3044 gfp_t flags, void *objp, void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003045{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003046 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003047 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003048 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003049#ifdef CONFIG_DEBUG_PAGEALLOC
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003050 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003051 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003052 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003053 else
3054 check_poison_obj(cachep, objp);
3055#else
3056 check_poison_obj(cachep, objp);
3057#endif
3058 poison_obj(cachep, objp, POISON_INUSE);
3059 }
3060 if (cachep->flags & SLAB_STORE_USER)
3061 *dbg_userword(cachep, objp) = caller;
3062
3063 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003064 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3065 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3066 slab_error(cachep, "double free, or memory outside"
3067 " object was overwritten");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003068 printk(KERN_ERR
David Woodhouseb46b8f12007-05-08 00:22:59 -07003069 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08003070 objp, *dbg_redzone1(cachep, objp),
3071 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003072 }
3073 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3074 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3075 }
Al Viro871751e2006-03-25 03:06:39 -08003076#ifdef CONFIG_DEBUG_SLAB_LEAK
3077 {
3078 struct slab *slabp;
3079 unsigned objnr;
3080
Christoph Lameterb49af682007-05-06 14:49:41 -07003081 slabp = page_get_slab(virt_to_head_page(objp));
Al Viro871751e2006-03-25 03:06:39 -08003082 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3083 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3084 }
3085#endif
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003086 objp += obj_offset(cachep);
Christoph Lameter4f104932007-05-06 14:50:17 -07003087 if (cachep->ctor && cachep->flags & SLAB_POISON)
3088 cachep->ctor(objp, cachep, SLAB_CTOR_CONSTRUCTOR);
Kevin Hilmana44b56d2006-12-06 20:32:11 -08003089#if ARCH_SLAB_MINALIGN
3090 if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) {
3091 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3092 objp, ARCH_SLAB_MINALIGN);
3093 }
3094#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003095 return objp;
3096}
3097#else
3098#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3099#endif
3100
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003101#ifdef CONFIG_FAILSLAB
3102
3103static struct failslab_attr {
3104
3105 struct fault_attr attr;
3106
3107 u32 ignore_gfp_wait;
3108#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3109 struct dentry *ignore_gfp_wait_file;
3110#endif
3111
3112} failslab = {
3113 .attr = FAULT_ATTR_INITIALIZER,
Don Mullis6b1b60f2006-12-08 02:39:53 -08003114 .ignore_gfp_wait = 1,
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003115};
3116
3117static int __init setup_failslab(char *str)
3118{
3119 return setup_fault_attr(&failslab.attr, str);
3120}
3121__setup("failslab=", setup_failslab);
3122
3123static int should_failslab(struct kmem_cache *cachep, gfp_t flags)
3124{
3125 if (cachep == &cache_cache)
3126 return 0;
3127 if (flags & __GFP_NOFAIL)
3128 return 0;
3129 if (failslab.ignore_gfp_wait && (flags & __GFP_WAIT))
3130 return 0;
3131
3132 return should_fail(&failslab.attr, obj_size(cachep));
3133}
3134
3135#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3136
3137static int __init failslab_debugfs(void)
3138{
3139 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
3140 struct dentry *dir;
3141 int err;
3142
Akinobu Mita824ebef2007-05-06 14:49:58 -07003143 err = init_fault_attr_dentries(&failslab.attr, "failslab");
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003144 if (err)
3145 return err;
3146 dir = failslab.attr.dentries.dir;
3147
3148 failslab.ignore_gfp_wait_file =
3149 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3150 &failslab.ignore_gfp_wait);
3151
3152 if (!failslab.ignore_gfp_wait_file) {
3153 err = -ENOMEM;
3154 debugfs_remove(failslab.ignore_gfp_wait_file);
3155 cleanup_fault_attr_dentries(&failslab.attr);
3156 }
3157
3158 return err;
3159}
3160
3161late_initcall(failslab_debugfs);
3162
3163#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3164
3165#else /* CONFIG_FAILSLAB */
3166
3167static inline int should_failslab(struct kmem_cache *cachep, gfp_t flags)
3168{
3169 return 0;
3170}
3171
3172#endif /* CONFIG_FAILSLAB */
3173
Pekka Enberg343e0d72006-02-01 03:05:50 -08003174static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003175{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003176 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003177 struct array_cache *ac;
3178
Alok N Kataria5c382302005-09-27 21:45:46 -07003179 check_irq_off();
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003180
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003181 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003182 if (likely(ac->avail)) {
3183 STATS_INC_ALLOCHIT(cachep);
3184 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07003185 objp = ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003186 } else {
3187 STATS_INC_ALLOCMISS(cachep);
3188 objp = cache_alloc_refill(cachep, flags);
3189 }
Alok N Kataria5c382302005-09-27 21:45:46 -07003190 return objp;
3191}
3192
Christoph Lametere498be72005-09-09 13:03:32 -07003193#ifdef CONFIG_NUMA
3194/*
Paul Jacksonb2455392006-03-24 03:16:12 -08003195 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
Paul Jacksonc61afb12006-03-24 03:16:08 -08003196 *
3197 * If we are in_interrupt, then process context, including cpusets and
3198 * mempolicy, may not apply and should not be used for allocation policy.
3199 */
3200static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3201{
3202 int nid_alloc, nid_here;
3203
Christoph Lameter765c4502006-09-27 01:50:08 -07003204 if (in_interrupt() || (flags & __GFP_THISNODE))
Paul Jacksonc61afb12006-03-24 03:16:08 -08003205 return NULL;
3206 nid_alloc = nid_here = numa_node_id();
3207 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3208 nid_alloc = cpuset_mem_spread_node();
3209 else if (current->mempolicy)
3210 nid_alloc = slab_node(current->mempolicy);
3211 if (nid_alloc != nid_here)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003212 return ____cache_alloc_node(cachep, flags, nid_alloc);
Paul Jacksonc61afb12006-03-24 03:16:08 -08003213 return NULL;
3214}
3215
3216/*
Christoph Lameter765c4502006-09-27 01:50:08 -07003217 * Fallback function if there was no memory available and no objects on a
Christoph Lameter3c517a62006-12-06 20:33:29 -08003218 * certain node and fall back is permitted. First we scan all the
3219 * available nodelists for available objects. If that fails then we
3220 * perform an allocation without specifying a node. This allows the page
3221 * allocator to do its reclaim / fallback magic. We then insert the
3222 * slab into the proper nodelist and then allocate from it.
Christoph Lameter765c4502006-09-27 01:50:08 -07003223 */
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003224static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
Christoph Lameter765c4502006-09-27 01:50:08 -07003225{
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003226 struct zonelist *zonelist;
3227 gfp_t local_flags;
Christoph Lameter765c4502006-09-27 01:50:08 -07003228 struct zone **z;
3229 void *obj = NULL;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003230 int nid;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003231
3232 if (flags & __GFP_THISNODE)
3233 return NULL;
3234
3235 zonelist = &NODE_DATA(slab_node(current->mempolicy))
3236 ->node_zonelists[gfp_zone(flags)];
3237 local_flags = (flags & GFP_LEVEL_MASK);
Christoph Lameter765c4502006-09-27 01:50:08 -07003238
Christoph Lameter3c517a62006-12-06 20:33:29 -08003239retry:
3240 /*
3241 * Look through allowed nodes for objects available
3242 * from existing per node queues.
3243 */
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003244 for (z = zonelist->zones; *z && !obj; z++) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003245 nid = zone_to_nid(*z);
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003246
Paul Jackson02a0e532006-12-13 00:34:25 -08003247 if (cpuset_zone_allowed_hardwall(*z, flags) &&
Christoph Lameter3c517a62006-12-06 20:33:29 -08003248 cache->nodelists[nid] &&
3249 cache->nodelists[nid]->free_objects)
3250 obj = ____cache_alloc_node(cache,
3251 flags | GFP_THISNODE, nid);
3252 }
3253
Christoph Lametercfce6602007-05-06 14:50:17 -07003254 if (!obj) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003255 /*
3256 * This allocation will be performed within the constraints
3257 * of the current cpuset / memory policy requirements.
3258 * We may trigger various forms of reclaim on the allowed
3259 * set and go into memory reserves if necessary.
3260 */
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003261 if (local_flags & __GFP_WAIT)
3262 local_irq_enable();
3263 kmem_flagcheck(cache, flags);
Christoph Lameter3c517a62006-12-06 20:33:29 -08003264 obj = kmem_getpages(cache, flags, -1);
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003265 if (local_flags & __GFP_WAIT)
3266 local_irq_disable();
Christoph Lameter3c517a62006-12-06 20:33:29 -08003267 if (obj) {
3268 /*
3269 * Insert into the appropriate per node queues
3270 */
3271 nid = page_to_nid(virt_to_page(obj));
3272 if (cache_grow(cache, flags, nid, obj)) {
3273 obj = ____cache_alloc_node(cache,
3274 flags | GFP_THISNODE, nid);
3275 if (!obj)
3276 /*
3277 * Another processor may allocate the
3278 * objects in the slab since we are
3279 * not holding any locks.
3280 */
3281 goto retry;
3282 } else {
Hugh Dickinsb6a60452007-01-05 16:36:36 -08003283 /* cache_grow already freed obj */
Christoph Lameter3c517a62006-12-06 20:33:29 -08003284 obj = NULL;
3285 }
3286 }
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003287 }
Christoph Lameter765c4502006-09-27 01:50:08 -07003288 return obj;
3289}
3290
3291/*
Christoph Lametere498be72005-09-09 13:03:32 -07003292 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07003293 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003294static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003295 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07003296{
3297 struct list_head *entry;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003298 struct slab *slabp;
3299 struct kmem_list3 *l3;
3300 void *obj;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003301 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003302
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003303 l3 = cachep->nodelists[nodeid];
3304 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07003305
Andrew Mortona737b3e2006-03-22 00:08:11 -08003306retry:
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003307 check_irq_off();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003308 spin_lock(&l3->list_lock);
3309 entry = l3->slabs_partial.next;
3310 if (entry == &l3->slabs_partial) {
3311 l3->free_touched = 1;
3312 entry = l3->slabs_free.next;
3313 if (entry == &l3->slabs_free)
3314 goto must_grow;
3315 }
Christoph Lametere498be72005-09-09 13:03:32 -07003316
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003317 slabp = list_entry(entry, struct slab, list);
3318 check_spinlock_acquired_node(cachep, nodeid);
3319 check_slabp(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07003320
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003321 STATS_INC_NODEALLOCS(cachep);
3322 STATS_INC_ACTIVE(cachep);
3323 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003324
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003325 BUG_ON(slabp->inuse == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07003326
Matthew Dobson78d382d2006-02-01 03:05:47 -08003327 obj = slab_get_obj(cachep, slabp, nodeid);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003328 check_slabp(cachep, slabp);
3329 l3->free_objects--;
3330 /* move slabp to correct slabp list: */
3331 list_del(&slabp->list);
Christoph Lametere498be72005-09-09 13:03:32 -07003332
Andrew Mortona737b3e2006-03-22 00:08:11 -08003333 if (slabp->free == BUFCTL_END)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003334 list_add(&slabp->list, &l3->slabs_full);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003335 else
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003336 list_add(&slabp->list, &l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07003337
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003338 spin_unlock(&l3->list_lock);
3339 goto done;
Christoph Lametere498be72005-09-09 13:03:32 -07003340
Andrew Mortona737b3e2006-03-22 00:08:11 -08003341must_grow:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003342 spin_unlock(&l3->list_lock);
Christoph Lameter3c517a62006-12-06 20:33:29 -08003343 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
Christoph Lameter765c4502006-09-27 01:50:08 -07003344 if (x)
3345 goto retry;
Christoph Lametere498be72005-09-09 13:03:32 -07003346
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003347 return fallback_alloc(cachep, flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003348
Andrew Mortona737b3e2006-03-22 00:08:11 -08003349done:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003350 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07003351}
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003352
3353/**
3354 * kmem_cache_alloc_node - Allocate an object on the specified node
3355 * @cachep: The cache to allocate from.
3356 * @flags: See kmalloc().
3357 * @nodeid: node number of the target node.
3358 * @caller: return address of caller, used for debug information
3359 *
3360 * Identical to kmem_cache_alloc but it will allocate memory on the given
3361 * node, which can improve the performance for cpu bound structures.
3362 *
3363 * Fallback to other node is possible if __GFP_THISNODE is not set.
3364 */
3365static __always_inline void *
3366__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3367 void *caller)
3368{
3369 unsigned long save_flags;
3370 void *ptr;
3371
Akinobu Mita824ebef2007-05-06 14:49:58 -07003372 if (should_failslab(cachep, flags))
3373 return NULL;
3374
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003375 cache_alloc_debugcheck_before(cachep, flags);
3376 local_irq_save(save_flags);
3377
3378 if (unlikely(nodeid == -1))
3379 nodeid = numa_node_id();
3380
3381 if (unlikely(!cachep->nodelists[nodeid])) {
3382 /* Node not bootstrapped yet */
3383 ptr = fallback_alloc(cachep, flags);
3384 goto out;
3385 }
3386
3387 if (nodeid == numa_node_id()) {
3388 /*
3389 * Use the locally cached objects if possible.
3390 * However ____cache_alloc does not allow fallback
3391 * to other nodes. It may fail while we still have
3392 * objects on other nodes available.
3393 */
3394 ptr = ____cache_alloc(cachep, flags);
3395 if (ptr)
3396 goto out;
3397 }
3398 /* ___cache_alloc_node can fall back to other nodes */
3399 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3400 out:
3401 local_irq_restore(save_flags);
3402 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3403
3404 return ptr;
3405}
3406
3407static __always_inline void *
3408__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3409{
3410 void *objp;
3411
3412 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3413 objp = alternate_node_alloc(cache, flags);
3414 if (objp)
3415 goto out;
3416 }
3417 objp = ____cache_alloc(cache, flags);
3418
3419 /*
3420 * We may just have run out of memory on the local node.
3421 * ____cache_alloc_node() knows how to locate memory on other nodes
3422 */
3423 if (!objp)
3424 objp = ____cache_alloc_node(cache, flags, numa_node_id());
3425
3426 out:
3427 return objp;
3428}
3429#else
3430
3431static __always_inline void *
3432__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3433{
3434 return ____cache_alloc(cachep, flags);
3435}
3436
3437#endif /* CONFIG_NUMA */
3438
3439static __always_inline void *
3440__cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3441{
3442 unsigned long save_flags;
3443 void *objp;
3444
Akinobu Mita824ebef2007-05-06 14:49:58 -07003445 if (should_failslab(cachep, flags))
3446 return NULL;
3447
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003448 cache_alloc_debugcheck_before(cachep, flags);
3449 local_irq_save(save_flags);
3450 objp = __do_cache_alloc(cachep, flags);
3451 local_irq_restore(save_flags);
3452 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3453 prefetchw(objp);
3454
3455 return objp;
3456}
Christoph Lametere498be72005-09-09 13:03:32 -07003457
3458/*
3459 * Caller needs to acquire correct kmem_list's list_lock
3460 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003461static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003462 int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003463{
3464 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07003465 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003466
3467 for (i = 0; i < nr_objects; i++) {
3468 void *objp = objpp[i];
3469 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003470
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003471 slabp = virt_to_slab(objp);
Christoph Lameterff694162005-09-22 21:44:02 -07003472 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003473 list_del(&slabp->list);
Christoph Lameterff694162005-09-22 21:44:02 -07003474 check_spinlock_acquired_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003475 check_slabp(cachep, slabp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08003476 slab_put_obj(cachep, slabp, objp, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003477 STATS_DEC_ACTIVE(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003478 l3->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003479 check_slabp(cachep, slabp);
3480
3481 /* fixup slab chains */
3482 if (slabp->inuse == 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07003483 if (l3->free_objects > l3->free_limit) {
3484 l3->free_objects -= cachep->num;
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07003485 /* No need to drop any previously held
3486 * lock here, even if we have a off-slab slab
3487 * descriptor it is guaranteed to come from
3488 * a different cache, refer to comments before
3489 * alloc_slabmgmt.
3490 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003491 slab_destroy(cachep, slabp);
3492 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07003493 list_add(&slabp->list, &l3->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003494 }
3495 } else {
3496 /* Unconditionally move a slab to the end of the
3497 * partial list on free - maximum time for the
3498 * other objects to be freed, too.
3499 */
Christoph Lametere498be72005-09-09 13:03:32 -07003500 list_add_tail(&slabp->list, &l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003501 }
3502 }
3503}
3504
Pekka Enberg343e0d72006-02-01 03:05:50 -08003505static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003506{
3507 int batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07003508 struct kmem_list3 *l3;
Christoph Lameterff694162005-09-22 21:44:02 -07003509 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003510
3511 batchcount = ac->batchcount;
3512#if DEBUG
3513 BUG_ON(!batchcount || batchcount > ac->avail);
3514#endif
3515 check_irq_off();
Christoph Lameterff694162005-09-22 21:44:02 -07003516 l3 = cachep->nodelists[node];
Ingo Molnar873623d2006-07-13 14:44:38 +02003517 spin_lock(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003518 if (l3->shared) {
3519 struct array_cache *shared_array = l3->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003520 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003521 if (max) {
3522 if (batchcount > max)
3523 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07003524 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003525 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003526 shared_array->avail += batchcount;
3527 goto free_done;
3528 }
3529 }
3530
Christoph Lameterff694162005-09-22 21:44:02 -07003531 free_block(cachep, ac->entry, batchcount, node);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003532free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003533#if STATS
3534 {
3535 int i = 0;
3536 struct list_head *p;
3537
Christoph Lametere498be72005-09-09 13:03:32 -07003538 p = l3->slabs_free.next;
3539 while (p != &(l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003540 struct slab *slabp;
3541
3542 slabp = list_entry(p, struct slab, list);
3543 BUG_ON(slabp->inuse);
3544
3545 i++;
3546 p = p->next;
3547 }
3548 STATS_SET_FREEABLE(cachep, i);
3549 }
3550#endif
Christoph Lametere498be72005-09-09 13:03:32 -07003551 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003552 ac->avail -= batchcount;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003553 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003554}
3555
3556/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08003557 * Release an obj back to its cache. If the obj has a constructed state, it must
3558 * be in this state _before_ it is released. Called with disabled ints.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003559 */
Ingo Molnar873623d2006-07-13 14:44:38 +02003560static inline void __cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003561{
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003562 struct array_cache *ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003563
3564 check_irq_off();
3565 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3566
Siddha, Suresh B62918a02007-05-02 19:27:18 +02003567 if (use_alien_caches && cache_free_alien(cachep, objp))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07003568 return;
Christoph Lametere498be72005-09-09 13:03:32 -07003569
Linus Torvalds1da177e2005-04-16 15:20:36 -07003570 if (likely(ac->avail < ac->limit)) {
3571 STATS_INC_FREEHIT(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003572 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003573 return;
3574 } else {
3575 STATS_INC_FREEMISS(cachep);
3576 cache_flusharray(cachep, ac);
Christoph Lametere498be72005-09-09 13:03:32 -07003577 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003578 }
3579}
3580
3581/**
3582 * kmem_cache_alloc - Allocate an object
3583 * @cachep: The cache to allocate from.
3584 * @flags: See kmalloc().
3585 *
3586 * Allocate an object from this cache. The flags are only relevant
3587 * if the cache has no available objects.
3588 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003589void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003590{
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003591 return __cache_alloc(cachep, flags, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003592}
3593EXPORT_SYMBOL(kmem_cache_alloc);
3594
3595/**
Rolf Eike Beerb8008b22006-07-30 03:04:04 -07003596 * kmem_cache_zalloc - Allocate an object. The memory is set to zero.
Pekka Enberga8c0f9a2006-03-25 03:06:42 -08003597 * @cache: The cache to allocate from.
3598 * @flags: See kmalloc().
3599 *
3600 * Allocate an object from this cache and set the allocated memory to zero.
3601 * The flags are only relevant if the cache has no available objects.
3602 */
3603void *kmem_cache_zalloc(struct kmem_cache *cache, gfp_t flags)
3604{
3605 void *ret = __cache_alloc(cache, flags, __builtin_return_address(0));
3606 if (ret)
3607 memset(ret, 0, obj_size(cache));
3608 return ret;
3609}
3610EXPORT_SYMBOL(kmem_cache_zalloc);
3611
3612/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003613 * kmem_ptr_validate - check if an untrusted pointer might
3614 * be a slab entry.
3615 * @cachep: the cache we're checking against
3616 * @ptr: pointer to validate
3617 *
3618 * This verifies that the untrusted pointer looks sane:
3619 * it is _not_ a guarantee that the pointer is actually
3620 * part of the slab cache in question, but it at least
3621 * validates that the pointer can be dereferenced and
3622 * looks half-way sane.
3623 *
3624 * Currently only used for dentry validation.
3625 */
Christoph Lameterb7f869a22006-12-22 01:06:44 -08003626int kmem_ptr_validate(struct kmem_cache *cachep, const void *ptr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003627{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003628 unsigned long addr = (unsigned long)ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003629 unsigned long min_addr = PAGE_OFFSET;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003630 unsigned long align_mask = BYTES_PER_WORD - 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003631 unsigned long size = cachep->buffer_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003632 struct page *page;
3633
3634 if (unlikely(addr < min_addr))
3635 goto out;
3636 if (unlikely(addr > (unsigned long)high_memory - size))
3637 goto out;
3638 if (unlikely(addr & align_mask))
3639 goto out;
3640 if (unlikely(!kern_addr_valid(addr)))
3641 goto out;
3642 if (unlikely(!kern_addr_valid(addr + size - 1)))
3643 goto out;
3644 page = virt_to_page(ptr);
3645 if (unlikely(!PageSlab(page)))
3646 goto out;
Pekka Enberg065d41c2005-11-13 16:06:46 -08003647 if (unlikely(page_get_cache(page) != cachep))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003648 goto out;
3649 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003650out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003651 return 0;
3652}
3653
3654#ifdef CONFIG_NUMA
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003655void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3656{
3657 return __cache_alloc_node(cachep, flags, nodeid,
3658 __builtin_return_address(0));
3659}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003660EXPORT_SYMBOL(kmem_cache_alloc_node);
3661
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003662static __always_inline void *
3663__do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003664{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003665 struct kmem_cache *cachep;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003666
3667 cachep = kmem_find_general_cachep(size, flags);
3668 if (unlikely(cachep == NULL))
3669 return NULL;
3670 return kmem_cache_alloc_node(cachep, flags, node);
3671}
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003672
3673#ifdef CONFIG_DEBUG_SLAB
3674void *__kmalloc_node(size_t size, gfp_t flags, int node)
3675{
3676 return __do_kmalloc_node(size, flags, node,
3677 __builtin_return_address(0));
3678}
Christoph Hellwigdbe5e692006-09-25 23:31:36 -07003679EXPORT_SYMBOL(__kmalloc_node);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003680
3681void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3682 int node, void *caller)
3683{
3684 return __do_kmalloc_node(size, flags, node, caller);
3685}
3686EXPORT_SYMBOL(__kmalloc_node_track_caller);
3687#else
3688void *__kmalloc_node(size_t size, gfp_t flags, int node)
3689{
3690 return __do_kmalloc_node(size, flags, node, NULL);
3691}
3692EXPORT_SYMBOL(__kmalloc_node);
3693#endif /* CONFIG_DEBUG_SLAB */
3694#endif /* CONFIG_NUMA */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003695
3696/**
Paul Drynoff800590f2006-06-23 02:03:48 -07003697 * __do_kmalloc - allocate memory
Linus Torvalds1da177e2005-04-16 15:20:36 -07003698 * @size: how many bytes of memory are required.
Paul Drynoff800590f2006-06-23 02:03:48 -07003699 * @flags: the type of memory to allocate (see kmalloc).
Randy Dunlap911851e2006-03-22 00:08:14 -08003700 * @caller: function caller for debug tracking of the caller
Linus Torvalds1da177e2005-04-16 15:20:36 -07003701 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003702static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3703 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003704{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003705 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003706
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003707 /* If you want to save a few bytes .text space: replace
3708 * __ with kmem_.
3709 * Then kmalloc uses the uninlined functions instead of the inline
3710 * functions.
3711 */
3712 cachep = __find_general_cachep(size, flags);
Andrew Mortondbdb9042005-09-23 13:24:10 -07003713 if (unlikely(cachep == NULL))
3714 return NULL;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003715 return __cache_alloc(cachep, flags, caller);
3716}
3717
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003718
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003719#ifdef CONFIG_DEBUG_SLAB
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003720void *__kmalloc(size_t size, gfp_t flags)
3721{
Al Viro871751e2006-03-25 03:06:39 -08003722 return __do_kmalloc(size, flags, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003723}
3724EXPORT_SYMBOL(__kmalloc);
3725
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003726void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller)
3727{
3728 return __do_kmalloc(size, flags, caller);
3729}
3730EXPORT_SYMBOL(__kmalloc_track_caller);
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003731
3732#else
3733void *__kmalloc(size_t size, gfp_t flags)
3734{
3735 return __do_kmalloc(size, flags, NULL);
3736}
3737EXPORT_SYMBOL(__kmalloc);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003738#endif
3739
Linus Torvalds1da177e2005-04-16 15:20:36 -07003740/**
Pekka Enbergfd76bab2007-05-06 14:48:40 -07003741 * krealloc - reallocate memory. The contents will remain unchanged.
Pekka Enbergfd76bab2007-05-06 14:48:40 -07003742 * @p: object to reallocate memory for.
3743 * @new_size: how many bytes of memory are required.
3744 * @flags: the type of memory to allocate.
3745 *
3746 * The contents of the object pointed to are preserved up to the
3747 * lesser of the new and old sizes. If @p is %NULL, krealloc()
3748 * behaves exactly like kmalloc(). If @size is 0 and @p is not a
3749 * %NULL pointer, the object pointed to is freed.
3750 */
3751void *krealloc(const void *p, size_t new_size, gfp_t flags)
3752{
3753 struct kmem_cache *cache, *new_cache;
3754 void *ret;
3755
3756 if (unlikely(!p))
3757 return kmalloc_track_caller(new_size, flags);
3758
3759 if (unlikely(!new_size)) {
3760 kfree(p);
3761 return NULL;
3762 }
3763
3764 cache = virt_to_cache(p);
3765 new_cache = __find_general_cachep(new_size, flags);
3766
3767 /*
3768 * If new size fits in the current cache, bail out.
3769 */
3770 if (likely(cache == new_cache))
3771 return (void *)p;
3772
3773 /*
3774 * We are on the slow-path here so do not use __cache_alloc
3775 * because it bloats kernel text.
3776 */
3777 ret = kmalloc_track_caller(new_size, flags);
3778 if (ret) {
3779 memcpy(ret, p, min(new_size, ksize(p)));
3780 kfree(p);
3781 }
3782 return ret;
3783}
3784EXPORT_SYMBOL(krealloc);
3785
3786/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003787 * kmem_cache_free - Deallocate an object
3788 * @cachep: The cache the allocation was from.
3789 * @objp: The previously allocated object.
3790 *
3791 * Free an object which was previously allocated from this
3792 * cache.
3793 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003794void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003795{
3796 unsigned long flags;
3797
Pekka Enbergddc2e812006-06-23 02:03:40 -07003798 BUG_ON(virt_to_cache(objp) != cachep);
3799
Linus Torvalds1da177e2005-04-16 15:20:36 -07003800 local_irq_save(flags);
Ingo Molnar898552c2007-02-10 01:44:57 -08003801 debug_check_no_locks_freed(objp, obj_size(cachep));
Ingo Molnar873623d2006-07-13 14:44:38 +02003802 __cache_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003803 local_irq_restore(flags);
3804}
3805EXPORT_SYMBOL(kmem_cache_free);
3806
3807/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003808 * kfree - free previously allocated memory
3809 * @objp: pointer returned by kmalloc.
3810 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003811 * If @objp is NULL, no operation is performed.
3812 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003813 * Don't free memory not originally allocated by kmalloc()
3814 * or you will run into trouble.
3815 */
3816void kfree(const void *objp)
3817{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003818 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003819 unsigned long flags;
3820
3821 if (unlikely(!objp))
3822 return;
3823 local_irq_save(flags);
3824 kfree_debugcheck(objp);
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003825 c = virt_to_cache(objp);
Ingo Molnarf9b84042006-06-27 02:54:49 -07003826 debug_check_no_locks_freed(objp, obj_size(c));
Ingo Molnar873623d2006-07-13 14:44:38 +02003827 __cache_free(c, (void *)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003828 local_irq_restore(flags);
3829}
3830EXPORT_SYMBOL(kfree);
3831
Pekka Enberg343e0d72006-02-01 03:05:50 -08003832unsigned int kmem_cache_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003833{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003834 return obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003835}
3836EXPORT_SYMBOL(kmem_cache_size);
3837
Pekka Enberg343e0d72006-02-01 03:05:50 -08003838const char *kmem_cache_name(struct kmem_cache *cachep)
Arnaldo Carvalho de Melo19449722005-06-18 22:46:19 -07003839{
3840 return cachep->name;
3841}
3842EXPORT_SYMBOL_GPL(kmem_cache_name);
3843
Christoph Lametere498be72005-09-09 13:03:32 -07003844/*
Christoph Lameter0718dc22006-03-25 03:06:47 -08003845 * This initializes kmem_list3 or resizes varioius caches for all nodes.
Christoph Lametere498be72005-09-09 13:03:32 -07003846 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003847static int alloc_kmemlist(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07003848{
3849 int node;
3850 struct kmem_list3 *l3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003851 struct array_cache *new_shared;
Paul Menage3395ee02006-12-06 20:32:16 -08003852 struct array_cache **new_alien = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003853
3854 for_each_online_node(node) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003855
Paul Menage3395ee02006-12-06 20:32:16 -08003856 if (use_alien_caches) {
3857 new_alien = alloc_alien_cache(node, cachep->limit);
3858 if (!new_alien)
3859 goto fail;
3860 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003861
Eric Dumazet63109842007-05-06 14:49:28 -07003862 new_shared = NULL;
3863 if (cachep->shared) {
3864 new_shared = alloc_arraycache(node,
Christoph Lameter0718dc22006-03-25 03:06:47 -08003865 cachep->shared*cachep->batchcount,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003866 0xbaadf00d);
Eric Dumazet63109842007-05-06 14:49:28 -07003867 if (!new_shared) {
3868 free_alien_cache(new_alien);
3869 goto fail;
3870 }
Christoph Lameter0718dc22006-03-25 03:06:47 -08003871 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003872
Andrew Mortona737b3e2006-03-22 00:08:11 -08003873 l3 = cachep->nodelists[node];
3874 if (l3) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003875 struct array_cache *shared = l3->shared;
3876
Christoph Lametere498be72005-09-09 13:03:32 -07003877 spin_lock_irq(&l3->list_lock);
3878
Christoph Lametercafeb022006-03-25 03:06:46 -08003879 if (shared)
Christoph Lameter0718dc22006-03-25 03:06:47 -08003880 free_block(cachep, shared->entry,
3881 shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07003882
Christoph Lametercafeb022006-03-25 03:06:46 -08003883 l3->shared = new_shared;
3884 if (!l3->alien) {
Christoph Lametere498be72005-09-09 13:03:32 -07003885 l3->alien = new_alien;
3886 new_alien = NULL;
3887 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003888 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003889 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003890 spin_unlock_irq(&l3->list_lock);
Christoph Lametercafeb022006-03-25 03:06:46 -08003891 kfree(shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003892 free_alien_cache(new_alien);
3893 continue;
3894 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08003895 l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node);
Christoph Lameter0718dc22006-03-25 03:06:47 -08003896 if (!l3) {
3897 free_alien_cache(new_alien);
3898 kfree(new_shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003899 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003900 }
Christoph Lametere498be72005-09-09 13:03:32 -07003901
3902 kmem_list3_init(l3);
3903 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Andrew Mortona737b3e2006-03-22 00:08:11 -08003904 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003905 l3->shared = new_shared;
Christoph Lametere498be72005-09-09 13:03:32 -07003906 l3->alien = new_alien;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003907 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003908 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003909 cachep->nodelists[node] = l3;
3910 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003911 return 0;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003912
Andrew Mortona737b3e2006-03-22 00:08:11 -08003913fail:
Christoph Lameter0718dc22006-03-25 03:06:47 -08003914 if (!cachep->next.next) {
3915 /* Cache is not active yet. Roll back what we did */
3916 node--;
3917 while (node >= 0) {
3918 if (cachep->nodelists[node]) {
3919 l3 = cachep->nodelists[node];
3920
3921 kfree(l3->shared);
3922 free_alien_cache(l3->alien);
3923 kfree(l3);
3924 cachep->nodelists[node] = NULL;
3925 }
3926 node--;
3927 }
3928 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003929 return -ENOMEM;
Christoph Lametere498be72005-09-09 13:03:32 -07003930}
3931
Linus Torvalds1da177e2005-04-16 15:20:36 -07003932struct ccupdate_struct {
Pekka Enberg343e0d72006-02-01 03:05:50 -08003933 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003934 struct array_cache *new[NR_CPUS];
3935};
3936
3937static void do_ccupdate_local(void *info)
3938{
Andrew Mortona737b3e2006-03-22 00:08:11 -08003939 struct ccupdate_struct *new = info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003940 struct array_cache *old;
3941
3942 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003943 old = cpu_cache_get(new->cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003944
Linus Torvalds1da177e2005-04-16 15:20:36 -07003945 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3946 new->new[smp_processor_id()] = old;
3947}
3948
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003949/* Always called with the cache_chain_mutex held */
Andrew Mortona737b3e2006-03-22 00:08:11 -08003950static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3951 int batchcount, int shared)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003952{
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003953 struct ccupdate_struct *new;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003954 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003955
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003956 new = kzalloc(sizeof(*new), GFP_KERNEL);
3957 if (!new)
3958 return -ENOMEM;
3959
Christoph Lametere498be72005-09-09 13:03:32 -07003960 for_each_online_cpu(i) {
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003961 new->new[i] = alloc_arraycache(cpu_to_node(i), limit,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003962 batchcount);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003963 if (!new->new[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003964 for (i--; i >= 0; i--)
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003965 kfree(new->new[i]);
3966 kfree(new);
Christoph Lametere498be72005-09-09 13:03:32 -07003967 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003968 }
3969 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003970 new->cachep = cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003971
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003972 on_each_cpu(do_ccupdate_local, (void *)new, 1, 1);
Christoph Lametere498be72005-09-09 13:03:32 -07003973
Linus Torvalds1da177e2005-04-16 15:20:36 -07003974 check_irq_on();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003975 cachep->batchcount = batchcount;
3976 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07003977 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003978
Christoph Lametere498be72005-09-09 13:03:32 -07003979 for_each_online_cpu(i) {
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003980 struct array_cache *ccold = new->new[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003981 if (!ccold)
3982 continue;
Christoph Lametere498be72005-09-09 13:03:32 -07003983 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -07003984 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
Christoph Lametere498be72005-09-09 13:03:32 -07003985 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003986 kfree(ccold);
3987 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003988 kfree(new);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003989 return alloc_kmemlist(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003990}
3991
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003992/* Called with cache_chain_mutex held always */
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003993static int enable_cpucache(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003994{
3995 int err;
3996 int limit, shared;
3997
Andrew Mortona737b3e2006-03-22 00:08:11 -08003998 /*
3999 * The head array serves three purposes:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004000 * - create a LIFO ordering, i.e. return objects that are cache-warm
4001 * - reduce the number of spinlock operations.
Andrew Mortona737b3e2006-03-22 00:08:11 -08004002 * - reduce the number of linked list operations on the slab and
Linus Torvalds1da177e2005-04-16 15:20:36 -07004003 * bufctl chains: array operations are cheaper.
4004 * The numbers are guessed, we should auto-tune as described by
4005 * Bonwick.
4006 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004007 if (cachep->buffer_size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004008 limit = 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004009 else if (cachep->buffer_size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004010 limit = 8;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004011 else if (cachep->buffer_size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004012 limit = 24;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004013 else if (cachep->buffer_size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004014 limit = 54;
4015 else
4016 limit = 120;
4017
Andrew Mortona737b3e2006-03-22 00:08:11 -08004018 /*
4019 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
Linus Torvalds1da177e2005-04-16 15:20:36 -07004020 * allocation behaviour: Most allocs on one cpu, most free operations
4021 * on another cpu. For these cases, an efficient object passing between
4022 * cpus is necessary. This is provided by a shared array. The array
4023 * replaces Bonwick's magazine layer.
4024 * On uniprocessor, it's functionally equivalent (but less efficient)
4025 * to a larger limit. Thus disabled by default.
4026 */
4027 shared = 0;
Eric Dumazet364fbb22007-05-06 14:49:27 -07004028 if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004029 shared = 8;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004030
4031#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08004032 /*
4033 * With debugging enabled, large batchcount lead to excessively long
4034 * periods with disabled local interrupts. Limit the batchcount
Linus Torvalds1da177e2005-04-16 15:20:36 -07004035 */
4036 if (limit > 32)
4037 limit = 32;
4038#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004039 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004040 if (err)
4041 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004042 cachep->name, -err);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07004043 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004044}
4045
Christoph Lameter1b552532006-03-22 00:09:07 -08004046/*
4047 * Drain an array if it contains any elements taking the l3 lock only if
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004048 * necessary. Note that the l3 listlock also protects the array_cache
4049 * if drain_array() is used on the shared array.
Christoph Lameter1b552532006-03-22 00:09:07 -08004050 */
4051void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
4052 struct array_cache *ac, int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004053{
4054 int tofree;
4055
Christoph Lameter1b552532006-03-22 00:09:07 -08004056 if (!ac || !ac->avail)
4057 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004058 if (ac->touched && !force) {
4059 ac->touched = 0;
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004060 } else {
Christoph Lameter1b552532006-03-22 00:09:07 -08004061 spin_lock_irq(&l3->list_lock);
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004062 if (ac->avail) {
4063 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4064 if (tofree > ac->avail)
4065 tofree = (ac->avail + 1) / 2;
4066 free_block(cachep, ac->entry, tofree, node);
4067 ac->avail -= tofree;
4068 memmove(ac->entry, &(ac->entry[tofree]),
4069 sizeof(void *) * ac->avail);
4070 }
Christoph Lameter1b552532006-03-22 00:09:07 -08004071 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004072 }
4073}
4074
4075/**
4076 * cache_reap - Reclaim memory from caches.
Randy Dunlap05fb6bf2007-02-28 20:12:13 -08004077 * @w: work descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07004078 *
4079 * Called from workqueue/eventd every few seconds.
4080 * Purpose:
4081 * - clear the per-cpu caches for this CPU.
4082 * - return freeable pages to the main free memory pool.
4083 *
Andrew Mortona737b3e2006-03-22 00:08:11 -08004084 * If we cannot acquire the cache chain mutex then just give up - we'll try
4085 * again on the next iteration.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004086 */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004087static void cache_reap(struct work_struct *w)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004088{
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004089 struct kmem_cache *searchp;
Christoph Lametere498be72005-09-09 13:03:32 -07004090 struct kmem_list3 *l3;
Christoph Lameteraab22072006-03-22 00:09:06 -08004091 int node = numa_node_id();
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004092 struct delayed_work *work =
4093 container_of(w, struct delayed_work, work);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004094
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004095 if (!mutex_trylock(&cache_chain_mutex))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004096 /* Give up. Setup the next iteration. */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004097 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004098
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004099 list_for_each_entry(searchp, &cache_chain, next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004100 check_irq_on();
4101
Christoph Lameter35386e32006-03-22 00:09:05 -08004102 /*
4103 * We only take the l3 lock if absolutely necessary and we
4104 * have established with reasonable certainty that
4105 * we can do some work if the lock was obtained.
4106 */
Christoph Lameteraab22072006-03-22 00:09:06 -08004107 l3 = searchp->nodelists[node];
Christoph Lameter35386e32006-03-22 00:09:05 -08004108
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004109 reap_alien(searchp, l3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004110
Christoph Lameteraab22072006-03-22 00:09:06 -08004111 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004112
Christoph Lameter35386e32006-03-22 00:09:05 -08004113 /*
4114 * These are racy checks but it does not matter
4115 * if we skip one check or scan twice.
4116 */
Christoph Lametere498be72005-09-09 13:03:32 -07004117 if (time_after(l3->next_reap, jiffies))
Christoph Lameter35386e32006-03-22 00:09:05 -08004118 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004119
Christoph Lametere498be72005-09-09 13:03:32 -07004120 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004121
Christoph Lameteraab22072006-03-22 00:09:06 -08004122 drain_array(searchp, l3, l3->shared, 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004123
Christoph Lametered11d9e2006-06-30 01:55:45 -07004124 if (l3->free_touched)
Christoph Lametere498be72005-09-09 13:03:32 -07004125 l3->free_touched = 0;
Christoph Lametered11d9e2006-06-30 01:55:45 -07004126 else {
4127 int freed;
4128
4129 freed = drain_freelist(searchp, l3, (l3->free_limit +
4130 5 * searchp->num - 1) / (5 * searchp->num));
4131 STATS_ADD_REAPED(searchp, freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004132 }
Christoph Lameter35386e32006-03-22 00:09:05 -08004133next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004134 cond_resched();
4135 }
4136 check_irq_on();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004137 mutex_unlock(&cache_chain_mutex);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004138 next_reap_node();
Christoph Lameter2244b952006-06-30 01:55:33 -07004139 refresh_cpu_vm_stats(smp_processor_id());
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004140out:
Andrew Mortona737b3e2006-03-22 00:08:11 -08004141 /* Set up the next iteration */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004142 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004143}
4144
4145#ifdef CONFIG_PROC_FS
4146
Pekka Enberg85289f92006-01-08 01:00:36 -08004147static void print_slabinfo_header(struct seq_file *m)
4148{
4149 /*
4150 * Output format version, so at least we can change it
4151 * without _too_ many complaints.
4152 */
4153#if STATS
4154 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4155#else
4156 seq_puts(m, "slabinfo - version: 2.1\n");
4157#endif
4158 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4159 "<objperslab> <pagesperslab>");
4160 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4161 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4162#if STATS
4163 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004164 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
Pekka Enberg85289f92006-01-08 01:00:36 -08004165 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4166#endif
4167 seq_putc(m, '\n');
4168}
4169
Linus Torvalds1da177e2005-04-16 15:20:36 -07004170static void *s_start(struct seq_file *m, loff_t *pos)
4171{
4172 loff_t n = *pos;
4173 struct list_head *p;
4174
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004175 mutex_lock(&cache_chain_mutex);
Pekka Enberg85289f92006-01-08 01:00:36 -08004176 if (!n)
4177 print_slabinfo_header(m);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004178 p = cache_chain.next;
4179 while (n--) {
4180 p = p->next;
4181 if (p == &cache_chain)
4182 return NULL;
4183 }
Pekka Enberg343e0d72006-02-01 03:05:50 -08004184 return list_entry(p, struct kmem_cache, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004185}
4186
4187static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4188{
Pekka Enberg343e0d72006-02-01 03:05:50 -08004189 struct kmem_cache *cachep = p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004190 ++*pos;
Andrew Mortona737b3e2006-03-22 00:08:11 -08004191 return cachep->next.next == &cache_chain ?
4192 NULL : list_entry(cachep->next.next, struct kmem_cache, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004193}
4194
4195static void s_stop(struct seq_file *m, void *p)
4196{
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004197 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004198}
4199
4200static int s_show(struct seq_file *m, void *p)
4201{
Pekka Enberg343e0d72006-02-01 03:05:50 -08004202 struct kmem_cache *cachep = p;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004203 struct slab *slabp;
4204 unsigned long active_objs;
4205 unsigned long num_objs;
4206 unsigned long active_slabs = 0;
4207 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004208 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004209 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07004210 int node;
4211 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004212
Linus Torvalds1da177e2005-04-16 15:20:36 -07004213 active_objs = 0;
4214 num_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004215 for_each_online_node(node) {
4216 l3 = cachep->nodelists[node];
4217 if (!l3)
4218 continue;
4219
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004220 check_irq_on();
4221 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07004222
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004223 list_for_each_entry(slabp, &l3->slabs_full, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004224 if (slabp->inuse != cachep->num && !error)
4225 error = "slabs_full accounting error";
4226 active_objs += cachep->num;
4227 active_slabs++;
4228 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004229 list_for_each_entry(slabp, &l3->slabs_partial, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004230 if (slabp->inuse == cachep->num && !error)
4231 error = "slabs_partial inuse accounting error";
4232 if (!slabp->inuse && !error)
4233 error = "slabs_partial/inuse accounting error";
4234 active_objs += slabp->inuse;
4235 active_slabs++;
4236 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004237 list_for_each_entry(slabp, &l3->slabs_free, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004238 if (slabp->inuse && !error)
4239 error = "slabs_free/inuse accounting error";
4240 num_slabs++;
4241 }
4242 free_objects += l3->free_objects;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08004243 if (l3->shared)
4244 shared_avail += l3->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07004245
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004246 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004247 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004248 num_slabs += active_slabs;
4249 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07004250 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004251 error = "free_objects accounting error";
4252
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004253 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004254 if (error)
4255 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4256
4257 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004258 name, active_objs, num_objs, cachep->buffer_size,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004259 cachep->num, (1 << cachep->gfporder));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004260 seq_printf(m, " : tunables %4u %4u %4u",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004261 cachep->limit, cachep->batchcount, cachep->shared);
Christoph Lametere498be72005-09-09 13:03:32 -07004262 seq_printf(m, " : slabdata %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004263 active_slabs, num_slabs, shared_avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004264#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004265 { /* list3 stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004266 unsigned long high = cachep->high_mark;
4267 unsigned long allocs = cachep->num_allocations;
4268 unsigned long grown = cachep->grown;
4269 unsigned long reaped = cachep->reaped;
4270 unsigned long errors = cachep->errors;
4271 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004272 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07004273 unsigned long node_frees = cachep->node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004274 unsigned long overflows = cachep->node_overflow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004275
Christoph Lametere498be72005-09-09 13:03:32 -07004276 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004277 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown,
Andrew Mortona737b3e2006-03-22 00:08:11 -08004278 reaped, errors, max_freeable, node_allocs,
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004279 node_frees, overflows);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004280 }
4281 /* cpu stats */
4282 {
4283 unsigned long allochit = atomic_read(&cachep->allochit);
4284 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4285 unsigned long freehit = atomic_read(&cachep->freehit);
4286 unsigned long freemiss = atomic_read(&cachep->freemiss);
4287
4288 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004289 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004290 }
4291#endif
4292 seq_putc(m, '\n');
Linus Torvalds1da177e2005-04-16 15:20:36 -07004293 return 0;
4294}
4295
4296/*
4297 * slabinfo_op - iterator that generates /proc/slabinfo
4298 *
4299 * Output layout:
4300 * cache-name
4301 * num-active-objs
4302 * total-objs
4303 * object size
4304 * num-active-slabs
4305 * total-slabs
4306 * num-pages-per-slab
4307 * + further values on SMP and with statistics enabled
4308 */
4309
Helge Deller15ad7cd2006-12-06 20:40:36 -08004310const struct seq_operations slabinfo_op = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004311 .start = s_start,
4312 .next = s_next,
4313 .stop = s_stop,
4314 .show = s_show,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004315};
4316
4317#define MAX_SLABINFO_WRITE 128
4318/**
4319 * slabinfo_write - Tuning for the slab allocator
4320 * @file: unused
4321 * @buffer: user buffer
4322 * @count: data length
4323 * @ppos: unused
4324 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004325ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4326 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004327{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004328 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004329 int limit, batchcount, shared, res;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004330 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004331
Linus Torvalds1da177e2005-04-16 15:20:36 -07004332 if (count > MAX_SLABINFO_WRITE)
4333 return -EINVAL;
4334 if (copy_from_user(&kbuf, buffer, count))
4335 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004336 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07004337
4338 tmp = strchr(kbuf, ' ');
4339 if (!tmp)
4340 return -EINVAL;
4341 *tmp = '\0';
4342 tmp++;
4343 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4344 return -EINVAL;
4345
4346 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004347 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004348 res = -EINVAL;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004349 list_for_each_entry(cachep, &cache_chain, next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004350 if (!strcmp(cachep->name, kbuf)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08004351 if (limit < 1 || batchcount < 1 ||
4352 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07004353 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004354 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07004355 res = do_tune_cpucache(cachep, limit,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004356 batchcount, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004357 }
4358 break;
4359 }
4360 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004361 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004362 if (res >= 0)
4363 res = count;
4364 return res;
4365}
Al Viro871751e2006-03-25 03:06:39 -08004366
4367#ifdef CONFIG_DEBUG_SLAB_LEAK
4368
4369static void *leaks_start(struct seq_file *m, loff_t *pos)
4370{
4371 loff_t n = *pos;
4372 struct list_head *p;
4373
4374 mutex_lock(&cache_chain_mutex);
4375 p = cache_chain.next;
4376 while (n--) {
4377 p = p->next;
4378 if (p == &cache_chain)
4379 return NULL;
4380 }
4381 return list_entry(p, struct kmem_cache, next);
4382}
4383
4384static inline int add_caller(unsigned long *n, unsigned long v)
4385{
4386 unsigned long *p;
4387 int l;
4388 if (!v)
4389 return 1;
4390 l = n[1];
4391 p = n + 2;
4392 while (l) {
4393 int i = l/2;
4394 unsigned long *q = p + 2 * i;
4395 if (*q == v) {
4396 q[1]++;
4397 return 1;
4398 }
4399 if (*q > v) {
4400 l = i;
4401 } else {
4402 p = q + 2;
4403 l -= i + 1;
4404 }
4405 }
4406 if (++n[1] == n[0])
4407 return 0;
4408 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4409 p[0] = v;
4410 p[1] = 1;
4411 return 1;
4412}
4413
4414static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4415{
4416 void *p;
4417 int i;
4418 if (n[0] == n[1])
4419 return;
4420 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4421 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4422 continue;
4423 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4424 return;
4425 }
4426}
4427
4428static void show_symbol(struct seq_file *m, unsigned long address)
4429{
4430#ifdef CONFIG_KALLSYMS
Al Viro871751e2006-03-25 03:06:39 -08004431 unsigned long offset, size;
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004432 char modname[MODULE_NAME_LEN + 1], name[KSYM_NAME_LEN + 1];
Al Viro871751e2006-03-25 03:06:39 -08004433
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004434 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
Al Viro871751e2006-03-25 03:06:39 -08004435 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004436 if (modname[0])
Al Viro871751e2006-03-25 03:06:39 -08004437 seq_printf(m, " [%s]", modname);
4438 return;
4439 }
4440#endif
4441 seq_printf(m, "%p", (void *)address);
4442}
4443
4444static int leaks_show(struct seq_file *m, void *p)
4445{
4446 struct kmem_cache *cachep = p;
Al Viro871751e2006-03-25 03:06:39 -08004447 struct slab *slabp;
4448 struct kmem_list3 *l3;
4449 const char *name;
4450 unsigned long *n = m->private;
4451 int node;
4452 int i;
4453
4454 if (!(cachep->flags & SLAB_STORE_USER))
4455 return 0;
4456 if (!(cachep->flags & SLAB_RED_ZONE))
4457 return 0;
4458
4459 /* OK, we can do it */
4460
4461 n[1] = 0;
4462
4463 for_each_online_node(node) {
4464 l3 = cachep->nodelists[node];
4465 if (!l3)
4466 continue;
4467
4468 check_irq_on();
4469 spin_lock_irq(&l3->list_lock);
4470
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004471 list_for_each_entry(slabp, &l3->slabs_full, list)
Al Viro871751e2006-03-25 03:06:39 -08004472 handle_slab(n, cachep, slabp);
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004473 list_for_each_entry(slabp, &l3->slabs_partial, list)
Al Viro871751e2006-03-25 03:06:39 -08004474 handle_slab(n, cachep, slabp);
Al Viro871751e2006-03-25 03:06:39 -08004475 spin_unlock_irq(&l3->list_lock);
4476 }
4477 name = cachep->name;
4478 if (n[0] == n[1]) {
4479 /* Increase the buffer size */
4480 mutex_unlock(&cache_chain_mutex);
4481 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4482 if (!m->private) {
4483 /* Too bad, we are really out */
4484 m->private = n;
4485 mutex_lock(&cache_chain_mutex);
4486 return -ENOMEM;
4487 }
4488 *(unsigned long *)m->private = n[0] * 2;
4489 kfree(n);
4490 mutex_lock(&cache_chain_mutex);
4491 /* Now make sure this entry will be retried */
4492 m->count = m->size;
4493 return 0;
4494 }
4495 for (i = 0; i < n[1]; i++) {
4496 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4497 show_symbol(m, n[2*i+2]);
4498 seq_putc(m, '\n');
4499 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004500
Al Viro871751e2006-03-25 03:06:39 -08004501 return 0;
4502}
4503
Helge Deller15ad7cd2006-12-06 20:40:36 -08004504const struct seq_operations slabstats_op = {
Al Viro871751e2006-03-25 03:06:39 -08004505 .start = leaks_start,
4506 .next = s_next,
4507 .stop = s_stop,
4508 .show = leaks_show,
4509};
4510#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004511#endif
4512
Manfred Spraul00e145b2005-09-03 15:55:07 -07004513/**
4514 * ksize - get the actual amount of memory allocated for a given object
4515 * @objp: Pointer to the object
4516 *
4517 * kmalloc may internally round up allocations and return more memory
4518 * than requested. ksize() can be used to determine the actual amount of
4519 * memory allocated. The caller may use this additional memory, even though
4520 * a smaller amount of memory was initially specified with the kmalloc call.
4521 * The caller must guarantee that objp points to a valid object previously
4522 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4523 * must not be freed during the duration of the call.
4524 */
Pekka Enbergfd76bab2007-05-06 14:48:40 -07004525size_t ksize(const void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004526{
Manfred Spraul00e145b2005-09-03 15:55:07 -07004527 if (unlikely(objp == NULL))
4528 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004529
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08004530 return obj_size(virt_to_cache(objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004531}