blob: f97e564bdf118f140e66fb8ad1978d20abf2c939 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * linux/mm/slab.c
3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
5 *
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
7 *
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
10 *
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
13 *
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
21 *
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
27 *
28 * This means, that your constructor is used only for newly allocated
Simon Arlott183ff222007-10-20 01:27:18 +020029 * slabs and you must pass objects with the same initializations to
Linus Torvalds1da177e2005-04-16 15:20:36 -070030 * kmem_cache_free.
31 *
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
35 *
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
38 * partial slabs
39 * empty slabs with no allocated objects
40 *
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
43 *
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
46 *
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
52 *
Andrew Mortona737b3e2006-03-22 00:08:11 -080053 * The c_cpuarray may not be read with enabled local interrupts -
Linus Torvalds1da177e2005-04-16 15:20:36 -070054 * it's changed with a smp_call_function().
55 *
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
Pekka Enberg343e0d72006-02-01 03:05:50 -080058 * Several members in struct kmem_cache and struct slab never change, they
Linus Torvalds1da177e2005-04-16 15:20:36 -070059 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
63 *
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
66 * his patch.
67 *
68 * Further notes from the original documentation:
69 *
70 * 11 April '97. Started multi-threading - markhe
Ingo Molnarfc0abb12006-01-18 17:42:33 -080071 * The global cache-chain is protected by the mutex 'cache_chain_mutex'.
Linus Torvalds1da177e2005-04-16 15:20:36 -070072 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
75 *
76 * At present, each engine can be growing a cache. This should be blocked.
77 *
Christoph Lametere498be72005-09-09 13:03:32 -070078 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
83 *
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
Linus Torvalds1da177e2005-04-16 15:20:36 -070087 */
88
Linus Torvalds1da177e2005-04-16 15:20:36 -070089#include <linux/slab.h>
90#include <linux/mm.h>
Randy Dunlapc9cf5522006-06-27 02:53:52 -070091#include <linux/poison.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070092#include <linux/swap.h>
93#include <linux/cache.h>
94#include <linux/interrupt.h>
95#include <linux/init.h>
96#include <linux/compiler.h>
Paul Jackson101a5002006-03-24 03:16:07 -080097#include <linux/cpuset.h>
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +040098#include <linux/proc_fs.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070099#include <linux/seq_file.h>
100#include <linux/notifier.h>
101#include <linux/kallsyms.h>
102#include <linux/cpu.h>
103#include <linux/sysctl.h>
104#include <linux/module.h>
105#include <linux/rcupdate.h>
Paulo Marques543537b2005-06-23 00:09:02 -0700106#include <linux/string.h>
Andrew Morton138ae662006-12-06 20:36:41 -0800107#include <linux/uaccess.h>
Christoph Lametere498be72005-09-09 13:03:32 -0700108#include <linux/nodemask.h>
Christoph Lameterdc85da12006-01-18 17:42:36 -0800109#include <linux/mempolicy.h>
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800110#include <linux/mutex.h>
Akinobu Mita8a8b6502006-12-08 02:39:44 -0800111#include <linux/fault-inject.h>
Ingo Molnare7eebaf2006-06-27 02:54:55 -0700112#include <linux/rtmutex.h>
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800113#include <linux/reciprocal_div.h>
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -0700114#include <linux/debugobjects.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700115
Linus Torvalds1da177e2005-04-16 15:20:36 -0700116#include <asm/cacheflush.h>
117#include <asm/tlbflush.h>
118#include <asm/page.h>
119
120/*
Christoph Lameter50953fe2007-05-06 14:50:16 -0700121 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700122 * 0 for faster, smaller code (especially in the critical paths).
123 *
124 * STATS - 1 to collect stats for /proc/slabinfo.
125 * 0 for faster, smaller code (especially in the critical paths).
126 *
127 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
128 */
129
130#ifdef CONFIG_DEBUG_SLAB
131#define DEBUG 1
132#define STATS 1
133#define FORCED_DEBUG 1
134#else
135#define DEBUG 0
136#define STATS 0
137#define FORCED_DEBUG 0
138#endif
139
Linus Torvalds1da177e2005-04-16 15:20:36 -0700140/* Shouldn't this be in a header file somewhere? */
141#define BYTES_PER_WORD sizeof(void *)
David Woodhouse87a927c2007-07-04 21:26:44 -0400142#define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700143
Linus Torvalds1da177e2005-04-16 15:20:36 -0700144#ifndef ARCH_KMALLOC_MINALIGN
145/*
146 * Enforce a minimum alignment for the kmalloc caches.
147 * Usually, the kmalloc caches are cache_line_size() aligned, except when
148 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
149 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
David Woodhouseb46b8f12007-05-08 00:22:59 -0700150 * alignment larger than the alignment of a 64-bit integer.
151 * ARCH_KMALLOC_MINALIGN allows that.
152 * Note that increasing this value may disable some debug features.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700153 */
David Woodhouseb46b8f12007-05-08 00:22:59 -0700154#define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700155#endif
156
157#ifndef ARCH_SLAB_MINALIGN
158/*
159 * Enforce a minimum alignment for all caches.
160 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
161 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
162 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
163 * some debug features.
164 */
165#define ARCH_SLAB_MINALIGN 0
166#endif
167
168#ifndef ARCH_KMALLOC_FLAGS
169#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
170#endif
171
172/* Legal flag mask for kmem_cache_create(). */
173#if DEBUG
Christoph Lameter50953fe2007-05-06 14:50:16 -0700174# define CREATE_MASK (SLAB_RED_ZONE | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700175 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
Christoph Lameterac2b8982006-03-22 00:08:15 -0800176 SLAB_CACHE_DMA | \
Christoph Lameter5af60832007-05-06 14:49:56 -0700177 SLAB_STORE_USER | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700178 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -0700179 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
180 SLAB_DEBUG_OBJECTS)
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 | \
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -0700185 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
186 SLAB_DEBUG_OBJECTS)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700187#endif
188
189/*
190 * kmem_bufctl_t:
191 *
192 * Bufctl's are used for linking objs within a slab
193 * linked offsets.
194 *
195 * This implementation relies on "struct page" for locating the cache &
196 * slab an object belongs to.
197 * This allows the bufctl structure to be small (one int), but limits
198 * the number of objects a slab (not a cache) can contain when off-slab
199 * bufctls are used. The limit is the size of the largest general cache
200 * that does not use off-slab slabs.
201 * For 32bit archs with 4 kB pages, is this 56.
202 * This is not serious, as it is only for large objects, when it is unwise
203 * to have too many per slab.
204 * Note: This limit can be raised by introducing a general cache whose size
205 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
206 */
207
Kyle Moffettfa5b08d2005-09-03 15:55:03 -0700208typedef unsigned int kmem_bufctl_t;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700209#define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
210#define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
Al Viro871751e2006-03-25 03:06:39 -0800211#define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
212#define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700213
Linus Torvalds1da177e2005-04-16 15:20:36 -0700214/*
215 * struct slab
216 *
217 * Manages the objs in a slab. Placed either at the beginning of mem allocated
218 * for a slab, or allocated from an general cache.
219 * Slabs are chained into three list: fully used, partial, fully free slabs.
220 */
221struct slab {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800222 struct list_head list;
223 unsigned long colouroff;
224 void *s_mem; /* including colour offset */
225 unsigned int inuse; /* num of objs active in slab */
226 kmem_bufctl_t free;
227 unsigned short nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700228};
229
230/*
231 * struct slab_rcu
232 *
233 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
234 * arrange for kmem_freepages to be called via RCU. This is useful if
235 * we need to approach a kernel structure obliquely, from its address
236 * obtained without the usual locking. We can lock the structure to
237 * stabilize it and check it's still at the given address, only if we
238 * can be sure that the memory has not been meanwhile reused for some
239 * other kind of object (which our subsystem's lock might corrupt).
240 *
241 * rcu_read_lock before reading the address, then rcu_read_unlock after
242 * taking the spinlock within the structure expected at that address.
243 *
244 * We assume struct slab_rcu can overlay struct slab when destroying.
245 */
246struct slab_rcu {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800247 struct rcu_head head;
Pekka Enberg343e0d72006-02-01 03:05:50 -0800248 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800249 void *addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700250};
251
252/*
253 * struct array_cache
254 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700255 * Purpose:
256 * - LIFO ordering, to hand out cache-warm objects from _alloc
257 * - reduce the number of linked list operations
258 * - reduce spinlock operations
259 *
260 * The limit is stored in the per-cpu structure to reduce the data cache
261 * footprint.
262 *
263 */
264struct array_cache {
265 unsigned int avail;
266 unsigned int limit;
267 unsigned int batchcount;
268 unsigned int touched;
Christoph Lametere498be72005-09-09 13:03:32 -0700269 spinlock_t lock;
Robert P. J. Daybda5b652007-10-16 23:30:05 -0700270 void *entry[]; /*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800271 * Must have this definition in here for the proper
272 * alignment of array_cache. Also simplifies accessing
273 * the entries.
Andrew Mortona737b3e2006-03-22 00:08:11 -0800274 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700275};
276
Andrew Mortona737b3e2006-03-22 00:08:11 -0800277/*
278 * bootstrap: The caches do not work without cpuarrays anymore, but the
279 * cpuarrays are allocated from the generic caches...
Linus Torvalds1da177e2005-04-16 15:20:36 -0700280 */
281#define BOOT_CPUCACHE_ENTRIES 1
282struct arraycache_init {
283 struct array_cache cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800284 void *entries[BOOT_CPUCACHE_ENTRIES];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700285};
286
287/*
Christoph Lametere498be72005-09-09 13:03:32 -0700288 * The slab lists for all objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700289 */
290struct kmem_list3 {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800291 struct list_head slabs_partial; /* partial list first, better asm code */
292 struct list_head slabs_full;
293 struct list_head slabs_free;
294 unsigned long free_objects;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800295 unsigned int free_limit;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800296 unsigned int colour_next; /* Per-node cache coloring */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800297 spinlock_t list_lock;
298 struct array_cache *shared; /* shared per node */
299 struct array_cache **alien; /* on other nodes */
Christoph Lameter35386e32006-03-22 00:09:05 -0800300 unsigned long next_reap; /* updated without locking */
301 int free_touched; /* updated without locking */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700302};
303
Christoph Lametere498be72005-09-09 13:03:32 -0700304/*
305 * Need this for bootstrapping a per node allocator.
306 */
Pekka Enberg556a1692008-01-25 08:20:51 +0200307#define NUM_INIT_LISTS (3 * MAX_NUMNODES)
Christoph Lametere498be72005-09-09 13:03:32 -0700308struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
309#define CACHE_CACHE 0
Pekka Enberg556a1692008-01-25 08:20:51 +0200310#define SIZE_AC MAX_NUMNODES
311#define SIZE_L3 (2 * MAX_NUMNODES)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700312
Christoph Lametered11d9e2006-06-30 01:55:45 -0700313static int drain_freelist(struct kmem_cache *cache,
314 struct kmem_list3 *l3, int tofree);
315static void free_block(struct kmem_cache *cachep, void **objpp, int len,
316 int node);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -0700317static int enable_cpucache(struct kmem_cache *cachep);
David Howells65f27f32006-11-22 14:55:48 +0000318static void cache_reap(struct work_struct *unused);
Christoph Lametered11d9e2006-06-30 01:55:45 -0700319
Christoph Lametere498be72005-09-09 13:03:32 -0700320/*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800321 * This function must be completely optimized away if a constant is passed to
322 * it. Mostly the same as what is in linux/slab.h except it returns an index.
Christoph Lametere498be72005-09-09 13:03:32 -0700323 */
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700324static __always_inline int index_of(const size_t size)
Christoph Lametere498be72005-09-09 13:03:32 -0700325{
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800326 extern void __bad_size(void);
327
Christoph Lametere498be72005-09-09 13:03:32 -0700328 if (__builtin_constant_p(size)) {
329 int i = 0;
330
331#define CACHE(x) \
332 if (size <=x) \
333 return i; \
334 else \
335 i++;
Joe Perches1c61fc42008-03-05 13:58:17 -0800336#include <linux/kmalloc_sizes.h>
Christoph Lametere498be72005-09-09 13:03:32 -0700337#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 */
Alexey Dobriyan51cc5062008-07-25 19:45:34 -0700410 void (*ctor)(void *obj);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700411
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800412/* 5) cache creation/removal */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800413 const char *name;
414 struct list_head next;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700415
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800416/* 6) statistics */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700417#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800418 unsigned long num_active;
419 unsigned long num_allocations;
420 unsigned long high_mark;
421 unsigned long grown;
422 unsigned long reaped;
423 unsigned long errors;
424 unsigned long max_freeable;
425 unsigned long node_allocs;
426 unsigned long node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700427 unsigned long node_overflow;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800428 atomic_t allochit;
429 atomic_t allocmiss;
430 atomic_t freehit;
431 atomic_t freemiss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700432#endif
433#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800434 /*
435 * If debugging is enabled, then the allocator can add additional
436 * fields and/or padding to every object. buffer_size contains the total
437 * object size including these internal fields, the following two
438 * variables contain the offset to the user object and its size.
439 */
440 int obj_offset;
441 int obj_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700442#endif
Eric Dumazet8da34302007-05-06 14:49:29 -0700443 /*
444 * We put nodelists[] at the end of kmem_cache, because we want to size
445 * this array to nr_node_ids slots instead of MAX_NUMNODES
446 * (see kmem_cache_init())
447 * We still use [MAX_NUMNODES] and not [1] or [0] because cache_cache
448 * is statically defined, so we reserve the max number of nodes.
449 */
450 struct kmem_list3 *nodelists[MAX_NUMNODES];
451 /*
452 * Do not add fields after nodelists[]
453 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700454};
455
456#define CFLGS_OFF_SLAB (0x80000000UL)
457#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
458
459#define BATCHREFILL_LIMIT 16
Andrew Mortona737b3e2006-03-22 00:08:11 -0800460/*
461 * Optimization question: fewer reaps means less probability for unnessary
462 * cpucache drain/refill cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700463 *
Adrian Bunkdc6f3f22005-11-08 16:44:08 +0100464 * OTOH the cpuarrays can contain lots of objects,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700465 * which could lock up otherwise freeable slabs.
466 */
467#define REAPTIMEOUT_CPUC (2*HZ)
468#define REAPTIMEOUT_LIST3 (4*HZ)
469
470#if STATS
471#define STATS_INC_ACTIVE(x) ((x)->num_active++)
472#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
473#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
474#define STATS_INC_GROWN(x) ((x)->grown++)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700475#define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
Andrew Mortona737b3e2006-03-22 00:08:11 -0800476#define STATS_SET_HIGH(x) \
477 do { \
478 if ((x)->num_active > (x)->high_mark) \
479 (x)->high_mark = (x)->num_active; \
480 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700481#define STATS_INC_ERR(x) ((x)->errors++)
482#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
Christoph Lametere498be72005-09-09 13:03:32 -0700483#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700484#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800485#define STATS_SET_FREEABLE(x, i) \
486 do { \
487 if ((x)->max_freeable < i) \
488 (x)->max_freeable = i; \
489 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700490#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
491#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
492#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
493#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
494#else
495#define STATS_INC_ACTIVE(x) do { } while (0)
496#define STATS_DEC_ACTIVE(x) do { } while (0)
497#define STATS_INC_ALLOCED(x) do { } while (0)
498#define STATS_INC_GROWN(x) do { } while (0)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700499#define STATS_ADD_REAPED(x,y) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700500#define STATS_SET_HIGH(x) do { } while (0)
501#define STATS_INC_ERR(x) do { } while (0)
502#define STATS_INC_NODEALLOCS(x) do { } while (0)
Christoph Lametere498be72005-09-09 13:03:32 -0700503#define STATS_INC_NODEFREES(x) do { } while (0)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700504#define STATS_INC_ACOVERFLOW(x) do { } while (0)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800505#define STATS_SET_FREEABLE(x, i) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700506#define STATS_INC_ALLOCHIT(x) do { } while (0)
507#define STATS_INC_ALLOCMISS(x) do { } while (0)
508#define STATS_INC_FREEHIT(x) do { } while (0)
509#define STATS_INC_FREEMISS(x) do { } while (0)
510#endif
511
512#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -0700513
Andrew Mortona737b3e2006-03-22 00:08:11 -0800514/*
515 * memory layout of objects:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700516 * 0 : objp
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800517 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
Linus Torvalds1da177e2005-04-16 15:20:36 -0700518 * the end of an object is aligned with the end of the real
519 * allocation. Catches writes behind the end of the allocation.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800520 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700521 * redzone word.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800522 * cachep->obj_offset: The real object.
523 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
Andrew Mortona737b3e2006-03-22 00:08:11 -0800524 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
525 * [BYTES_PER_WORD long]
Linus Torvalds1da177e2005-04-16 15:20:36 -0700526 */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800527static int obj_offset(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700528{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800529 return cachep->obj_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700530}
531
Pekka Enberg343e0d72006-02-01 03:05:50 -0800532static int obj_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700533{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800534 return cachep->obj_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700535}
536
David Woodhouseb46b8f12007-05-08 00:22:59 -0700537static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700538{
539 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
David Woodhouseb46b8f12007-05-08 00:22:59 -0700540 return (unsigned long long*) (objp + obj_offset(cachep) -
541 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700542}
543
David Woodhouseb46b8f12007-05-08 00:22:59 -0700544static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700545{
546 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
547 if (cachep->flags & SLAB_STORE_USER)
David Woodhouseb46b8f12007-05-08 00:22:59 -0700548 return (unsigned long long *)(objp + cachep->buffer_size -
549 sizeof(unsigned long long) -
David Woodhouse87a927c2007-07-04 21:26:44 -0400550 REDZONE_ALIGN);
David Woodhouseb46b8f12007-05-08 00:22:59 -0700551 return (unsigned long long *) (objp + cachep->buffer_size -
552 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700553}
554
Pekka Enberg343e0d72006-02-01 03:05:50 -0800555static void **dbg_userword(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700556{
557 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800558 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700559}
560
561#else
562
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800563#define obj_offset(x) 0
564#define obj_size(cachep) (cachep->buffer_size)
David Woodhouseb46b8f12007-05-08 00:22:59 -0700565#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
566#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
Linus Torvalds1da177e2005-04-16 15:20:36 -0700567#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
568
569#endif
570
571/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700572 * Do not go above this order unless 0 objects fit into the slab.
573 */
574#define BREAK_GFP_ORDER_HI 1
575#define BREAK_GFP_ORDER_LO 0
576static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
577
Andrew Mortona737b3e2006-03-22 00:08:11 -0800578/*
579 * Functions for storing/retrieving the cachep and or slab from the page
580 * allocator. These are used to find the slab an obj belongs to. With kfree(),
581 * these are used to find the cache which an obj belongs to.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700582 */
Pekka Enberg065d41c2005-11-13 16:06:46 -0800583static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
584{
585 page->lru.next = (struct list_head *)cache;
586}
587
588static inline struct kmem_cache *page_get_cache(struct page *page)
589{
Christoph Lameterd85f3382007-05-06 14:49:39 -0700590 page = compound_head(page);
Pekka Enbergddc2e812006-06-23 02:03:40 -0700591 BUG_ON(!PageSlab(page));
Pekka Enberg065d41c2005-11-13 16:06:46 -0800592 return (struct kmem_cache *)page->lru.next;
593}
594
595static inline void page_set_slab(struct page *page, struct slab *slab)
596{
597 page->lru.prev = (struct list_head *)slab;
598}
599
600static inline struct slab *page_get_slab(struct page *page)
601{
Pekka Enbergddc2e812006-06-23 02:03:40 -0700602 BUG_ON(!PageSlab(page));
Pekka Enberg065d41c2005-11-13 16:06:46 -0800603 return (struct slab *)page->lru.prev;
604}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700605
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800606static inline struct kmem_cache *virt_to_cache(const void *obj)
607{
Christoph Lameterb49af682007-05-06 14:49:41 -0700608 struct page *page = virt_to_head_page(obj);
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800609 return page_get_cache(page);
610}
611
612static inline struct slab *virt_to_slab(const void *obj)
613{
Christoph Lameterb49af682007-05-06 14:49:41 -0700614 struct page *page = virt_to_head_page(obj);
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800615 return page_get_slab(page);
616}
617
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800618static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
619 unsigned int idx)
620{
621 return slab->s_mem + cache->buffer_size * idx;
622}
623
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800624/*
625 * We want to avoid an expensive divide : (offset / cache->buffer_size)
626 * Using the fact that buffer_size is a constant for a particular cache,
627 * we can replace (offset / cache->buffer_size) by
628 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
629 */
630static inline unsigned int obj_to_index(const struct kmem_cache *cache,
631 const struct slab *slab, void *obj)
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800632{
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800633 u32 offset = (obj - slab->s_mem);
634 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800635}
636
Andrew Mortona737b3e2006-03-22 00:08:11 -0800637/*
638 * These are the default caches for kmalloc. Custom caches can have other sizes.
639 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700640struct cache_sizes malloc_sizes[] = {
641#define CACHE(x) { .cs_size = (x) },
642#include <linux/kmalloc_sizes.h>
643 CACHE(ULONG_MAX)
644#undef CACHE
645};
646EXPORT_SYMBOL(malloc_sizes);
647
648/* Must match cache_sizes above. Out of line to keep cache footprint low. */
649struct cache_names {
650 char *name;
651 char *name_dma;
652};
653
654static struct cache_names __initdata cache_names[] = {
655#define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
656#include <linux/kmalloc_sizes.h>
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800657 {NULL,}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700658#undef CACHE
659};
660
661static struct arraycache_init initarray_cache __initdata =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800662 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700663static struct arraycache_init initarray_generic =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800664 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700665
666/* internal cache of cache description objs */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800667static struct kmem_cache cache_cache = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800668 .batchcount = 1,
669 .limit = BOOT_CPUCACHE_ENTRIES,
670 .shared = 1,
Pekka Enberg343e0d72006-02-01 03:05:50 -0800671 .buffer_size = sizeof(struct kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800672 .name = "kmem_cache",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700673};
674
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700675#define BAD_ALIEN_MAGIC 0x01020304ul
676
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200677#ifdef CONFIG_LOCKDEP
678
679/*
680 * Slab sometimes uses the kmalloc slabs to store the slab headers
681 * for other slabs "off slab".
682 * The locking for this is tricky in that it nests within the locks
683 * of all other slabs in a few places; to deal with this special
684 * locking we put on-slab caches into a separate lock-class.
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700685 *
686 * We set lock class for alien array caches which are up during init.
687 * The lock annotation will be lost if all cpus of a node goes down and
688 * then comes back up during hotplug
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200689 */
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700690static struct lock_class_key on_slab_l3_key;
691static struct lock_class_key on_slab_alc_key;
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200692
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700693static inline void init_lock_keys(void)
694
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200695{
696 int q;
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700697 struct cache_sizes *s = malloc_sizes;
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200698
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700699 while (s->cs_size != ULONG_MAX) {
700 for_each_node(q) {
701 struct array_cache **alc;
702 int r;
703 struct kmem_list3 *l3 = s->cs_cachep->nodelists[q];
704 if (!l3 || OFF_SLAB(s->cs_cachep))
705 continue;
706 lockdep_set_class(&l3->list_lock, &on_slab_l3_key);
707 alc = l3->alien;
708 /*
709 * FIXME: This check for BAD_ALIEN_MAGIC
710 * should go away when common slab code is taught to
711 * work even without alien caches.
712 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
713 * for alloc_alien_cache,
714 */
715 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
716 continue;
717 for_each_node(r) {
718 if (alc[r])
719 lockdep_set_class(&alc[r]->lock,
720 &on_slab_alc_key);
721 }
722 }
723 s++;
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200724 }
725}
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200726#else
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700727static inline void init_lock_keys(void)
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200728{
729}
730#endif
731
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -0800732/*
Gautham R Shenoy95402b32008-01-25 21:08:02 +0100733 * Guard access to the cache-chain.
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -0800734 */
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800735static DEFINE_MUTEX(cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700736static struct list_head cache_chain;
737
738/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700739 * chicken and egg problem: delay the per-cpu array allocation
740 * until the general caches are up.
741 */
742static enum {
743 NONE,
Christoph Lametere498be72005-09-09 13:03:32 -0700744 PARTIAL_AC,
745 PARTIAL_L3,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700746 FULL
747} g_cpucache_up;
748
Mike Kravetz39d24e62006-05-15 09:44:13 -0700749/*
750 * used by boot code to determine if it can use slab based allocator
751 */
752int slab_is_available(void)
753{
754 return g_cpucache_up == FULL;
755}
756
David Howells52bad642006-11-22 14:54:01 +0000757static DEFINE_PER_CPU(struct delayed_work, reap_work);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700758
Pekka Enberg343e0d72006-02-01 03:05:50 -0800759static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700760{
761 return cachep->array[smp_processor_id()];
762}
763
Andrew Mortona737b3e2006-03-22 00:08:11 -0800764static inline struct kmem_cache *__find_general_cachep(size_t size,
765 gfp_t gfpflags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700766{
767 struct cache_sizes *csizep = malloc_sizes;
768
769#if DEBUG
770 /* This happens if someone tries to call
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800771 * kmem_cache_create(), or __kmalloc(), before
772 * the generic caches are initialized.
773 */
Alok Katariac7e43c72005-09-14 12:17:53 -0700774 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700775#endif
Christoph Lameter6cb8f912007-07-17 04:03:22 -0700776 if (!size)
777 return ZERO_SIZE_PTR;
778
Linus Torvalds1da177e2005-04-16 15:20:36 -0700779 while (size > csizep->cs_size)
780 csizep++;
781
782 /*
Martin Hicks0abf40c2005-09-03 15:54:54 -0700783 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
Linus Torvalds1da177e2005-04-16 15:20:36 -0700784 * has cs_{dma,}cachep==NULL. Thus no special case
785 * for large kmalloc calls required.
786 */
Christoph Lameter4b51d662007-02-10 01:43:10 -0800787#ifdef CONFIG_ZONE_DMA
Linus Torvalds1da177e2005-04-16 15:20:36 -0700788 if (unlikely(gfpflags & GFP_DMA))
789 return csizep->cs_dmacachep;
Christoph Lameter4b51d662007-02-10 01:43:10 -0800790#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700791 return csizep->cs_cachep;
792}
793
Adrian Bunkb2213852006-09-25 23:31:02 -0700794static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700795{
796 return __find_general_cachep(size, gfpflags);
797}
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700798
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800799static size_t slab_mgmt_size(size_t nr_objs, size_t align)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700800{
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800801 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
802}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700803
Andrew Mortona737b3e2006-03-22 00:08:11 -0800804/*
805 * Calculate the number of objects and left-over bytes for a given buffer size.
806 */
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800807static void cache_estimate(unsigned long gfporder, size_t buffer_size,
808 size_t align, int flags, size_t *left_over,
809 unsigned int *num)
810{
811 int nr_objs;
812 size_t mgmt_size;
813 size_t slab_size = PAGE_SIZE << gfporder;
814
815 /*
816 * The slab management structure can be either off the slab or
817 * on it. For the latter case, the memory allocated for a
818 * slab is used for:
819 *
820 * - The struct slab
821 * - One kmem_bufctl_t for each object
822 * - Padding to respect alignment of @align
823 * - @buffer_size bytes for each object
824 *
825 * If the slab management structure is off the slab, then the
826 * alignment will already be calculated into the size. Because
827 * the slabs are all pages aligned, the objects will be at the
828 * correct alignment when allocated.
829 */
830 if (flags & CFLGS_OFF_SLAB) {
831 mgmt_size = 0;
832 nr_objs = slab_size / buffer_size;
833
834 if (nr_objs > SLAB_LIMIT)
835 nr_objs = SLAB_LIMIT;
836 } else {
837 /*
838 * Ignore padding for the initial guess. The padding
839 * is at most @align-1 bytes, and @buffer_size is at
840 * least @align. In the worst case, this result will
841 * be one greater than the number of objects that fit
842 * into the memory allocation when taking the padding
843 * into account.
844 */
845 nr_objs = (slab_size - sizeof(struct slab)) /
846 (buffer_size + sizeof(kmem_bufctl_t));
847
848 /*
849 * This calculated number will be either the right
850 * amount, or one greater than what we want.
851 */
852 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
853 > slab_size)
854 nr_objs--;
855
856 if (nr_objs > SLAB_LIMIT)
857 nr_objs = SLAB_LIMIT;
858
859 mgmt_size = slab_mgmt_size(nr_objs, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700860 }
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800861 *num = nr_objs;
862 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700863}
864
Harvey Harrisond40cee22008-04-30 00:55:07 -0700865#define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700866
Andrew Mortona737b3e2006-03-22 00:08:11 -0800867static void __slab_error(const char *function, struct kmem_cache *cachep,
868 char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700869{
870 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800871 function, cachep->name, msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700872 dump_stack();
873}
874
Paul Menage3395ee02006-12-06 20:32:16 -0800875/*
876 * By default on NUMA we use alien caches to stage the freeing of
877 * objects allocated from other nodes. This causes massive memory
878 * inefficiencies when using fake NUMA setup to split memory into a
879 * large number of small nodes, so it can be disabled on the command
880 * line
881 */
882
883static int use_alien_caches __read_mostly = 1;
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -0700884static int numa_platform __read_mostly = 1;
Paul Menage3395ee02006-12-06 20:32:16 -0800885static int __init noaliencache_setup(char *s)
886{
887 use_alien_caches = 0;
888 return 1;
889}
890__setup("noaliencache", noaliencache_setup);
891
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800892#ifdef CONFIG_NUMA
893/*
894 * Special reaping functions for NUMA systems called from cache_reap().
895 * These take care of doing round robin flushing of alien caches (containing
896 * objects freed on different nodes from which they were allocated) and the
897 * flushing of remote pcps by calling drain_node_pages.
898 */
899static DEFINE_PER_CPU(unsigned long, reap_node);
900
901static void init_reap_node(int cpu)
902{
903 int node;
904
905 node = next_node(cpu_to_node(cpu), node_online_map);
906 if (node == MAX_NUMNODES)
Paul Jackson442295c2006-03-22 00:09:11 -0800907 node = first_node(node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800908
Daniel Yeisley7f6b8872006-11-02 22:07:14 -0800909 per_cpu(reap_node, cpu) = node;
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800910}
911
912static void next_reap_node(void)
913{
914 int node = __get_cpu_var(reap_node);
915
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800916 node = next_node(node, node_online_map);
917 if (unlikely(node >= MAX_NUMNODES))
918 node = first_node(node_online_map);
919 __get_cpu_var(reap_node) = node;
920}
921
922#else
923#define init_reap_node(cpu) do { } while (0)
924#define next_reap_node(void) do { } while (0)
925#endif
926
Linus Torvalds1da177e2005-04-16 15:20:36 -0700927/*
928 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
929 * via the workqueue/eventd.
930 * Add the CPU number into the expiration time to minimize the possibility of
931 * the CPUs getting into lockstep and contending for the global cache chain
932 * lock.
933 */
Adrian Bunk897e6792007-07-15 23:38:20 -0700934static void __cpuinit start_cpu_timer(int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700935{
David Howells52bad642006-11-22 14:54:01 +0000936 struct delayed_work *reap_work = &per_cpu(reap_work, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700937
938 /*
939 * When this gets called from do_initcalls via cpucache_init(),
940 * init_workqueues() has already run, so keventd will be setup
941 * at that time.
942 */
David Howells52bad642006-11-22 14:54:01 +0000943 if (keventd_up() && reap_work->work.func == NULL) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800944 init_reap_node(cpu);
David Howells65f27f32006-11-22 14:55:48 +0000945 INIT_DELAYED_WORK(reap_work, cache_reap);
Arjan van de Ven2b284212006-12-10 02:21:28 -0800946 schedule_delayed_work_on(cpu, reap_work,
947 __round_jiffies_relative(HZ, cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700948 }
949}
950
Christoph Lametere498be72005-09-09 13:03:32 -0700951static struct array_cache *alloc_arraycache(int node, int entries,
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800952 int batchcount)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700953{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800954 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700955 struct array_cache *nc = NULL;
956
Christoph Lametere498be72005-09-09 13:03:32 -0700957 nc = kmalloc_node(memsize, GFP_KERNEL, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700958 if (nc) {
959 nc->avail = 0;
960 nc->limit = entries;
961 nc->batchcount = batchcount;
962 nc->touched = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700963 spin_lock_init(&nc->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700964 }
965 return nc;
966}
967
Christoph Lameter3ded1752006-03-25 03:06:44 -0800968/*
969 * Transfer objects in one arraycache to another.
970 * Locking must be handled by the caller.
971 *
972 * Return the number of entries transferred.
973 */
974static int transfer_objects(struct array_cache *to,
975 struct array_cache *from, unsigned int max)
976{
977 /* Figure out how many entries to transfer */
978 int nr = min(min(from->avail, max), to->limit - to->avail);
979
980 if (!nr)
981 return 0;
982
983 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
984 sizeof(void *) *nr);
985
986 from->avail -= nr;
987 to->avail += nr;
988 to->touched = 1;
989 return nr;
990}
991
Christoph Lameter765c4502006-09-27 01:50:08 -0700992#ifndef CONFIG_NUMA
993
994#define drain_alien_cache(cachep, alien) do { } while (0)
995#define reap_alien(cachep, l3) do { } while (0)
996
997static inline struct array_cache **alloc_alien_cache(int node, int limit)
998{
999 return (struct array_cache **)BAD_ALIEN_MAGIC;
1000}
1001
1002static inline void free_alien_cache(struct array_cache **ac_ptr)
1003{
1004}
1005
1006static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1007{
1008 return 0;
1009}
1010
1011static inline void *alternate_node_alloc(struct kmem_cache *cachep,
1012 gfp_t flags)
1013{
1014 return NULL;
1015}
1016
Christoph Hellwig8b98c162006-12-06 20:32:30 -08001017static inline void *____cache_alloc_node(struct kmem_cache *cachep,
Christoph Lameter765c4502006-09-27 01:50:08 -07001018 gfp_t flags, int nodeid)
1019{
1020 return NULL;
1021}
1022
1023#else /* CONFIG_NUMA */
1024
Christoph Hellwig8b98c162006-12-06 20:32:30 -08001025static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
Paul Jacksonc61afb12006-03-24 03:16:08 -08001026static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
Christoph Lameterdc85da12006-01-18 17:42:36 -08001027
Pekka Enberg5295a742006-02-01 03:05:48 -08001028static struct array_cache **alloc_alien_cache(int node, int limit)
Christoph Lametere498be72005-09-09 13:03:32 -07001029{
1030 struct array_cache **ac_ptr;
Christoph Lameter8ef82862007-02-20 13:57:52 -08001031 int memsize = sizeof(void *) * nr_node_ids;
Christoph Lametere498be72005-09-09 13:03:32 -07001032 int i;
1033
1034 if (limit > 1)
1035 limit = 12;
1036 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
1037 if (ac_ptr) {
1038 for_each_node(i) {
1039 if (i == node || !node_online(i)) {
1040 ac_ptr[i] = NULL;
1041 continue;
1042 }
1043 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
1044 if (!ac_ptr[i]) {
Akinobu Mitacc550de2007-11-14 16:58:35 -08001045 for (i--; i >= 0; i--)
Christoph Lametere498be72005-09-09 13:03:32 -07001046 kfree(ac_ptr[i]);
1047 kfree(ac_ptr);
1048 return NULL;
1049 }
1050 }
1051 }
1052 return ac_ptr;
1053}
1054
Pekka Enberg5295a742006-02-01 03:05:48 -08001055static void free_alien_cache(struct array_cache **ac_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -07001056{
1057 int i;
1058
1059 if (!ac_ptr)
1060 return;
Christoph Lametere498be72005-09-09 13:03:32 -07001061 for_each_node(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001062 kfree(ac_ptr[i]);
Christoph Lametere498be72005-09-09 13:03:32 -07001063 kfree(ac_ptr);
1064}
1065
Pekka Enberg343e0d72006-02-01 03:05:50 -08001066static void __drain_alien_cache(struct kmem_cache *cachep,
Pekka Enberg5295a742006-02-01 03:05:48 -08001067 struct array_cache *ac, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07001068{
1069 struct kmem_list3 *rl3 = cachep->nodelists[node];
1070
1071 if (ac->avail) {
1072 spin_lock(&rl3->list_lock);
Christoph Lametere00946f2006-03-25 03:06:45 -08001073 /*
1074 * Stuff objects into the remote nodes shared array first.
1075 * That way we could avoid the overhead of putting the objects
1076 * into the free lists and getting them back later.
1077 */
shin, jacob693f7d32006-04-28 10:54:37 -05001078 if (rl3->shared)
1079 transfer_objects(rl3->shared, ac, ac->limit);
Christoph Lametere00946f2006-03-25 03:06:45 -08001080
Christoph Lameterff694162005-09-22 21:44:02 -07001081 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001082 ac->avail = 0;
1083 spin_unlock(&rl3->list_lock);
1084 }
1085}
1086
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001087/*
1088 * Called from cache_reap() to regularly drain alien caches round robin.
1089 */
1090static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1091{
1092 int node = __get_cpu_var(reap_node);
1093
1094 if (l3->alien) {
1095 struct array_cache *ac = l3->alien[node];
Christoph Lametere00946f2006-03-25 03:06:45 -08001096
1097 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001098 __drain_alien_cache(cachep, ac, node);
1099 spin_unlock_irq(&ac->lock);
1100 }
1101 }
1102}
1103
Andrew Mortona737b3e2006-03-22 00:08:11 -08001104static void drain_alien_cache(struct kmem_cache *cachep,
1105 struct array_cache **alien)
Christoph Lametere498be72005-09-09 13:03:32 -07001106{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001107 int i = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07001108 struct array_cache *ac;
1109 unsigned long flags;
1110
1111 for_each_online_node(i) {
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001112 ac = alien[i];
Christoph Lametere498be72005-09-09 13:03:32 -07001113 if (ac) {
1114 spin_lock_irqsave(&ac->lock, flags);
1115 __drain_alien_cache(cachep, ac, i);
1116 spin_unlock_irqrestore(&ac->lock, flags);
1117 }
1118 }
1119}
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001120
Ingo Molnar873623d2006-07-13 14:44:38 +02001121static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001122{
1123 struct slab *slabp = virt_to_slab(objp);
1124 int nodeid = slabp->nodeid;
1125 struct kmem_list3 *l3;
1126 struct array_cache *alien = NULL;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001127 int node;
1128
1129 node = numa_node_id();
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001130
1131 /*
1132 * Make sure we are not freeing a object from another node to the array
1133 * cache on this cpu.
1134 */
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001135 if (likely(slabp->nodeid == node))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001136 return 0;
1137
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001138 l3 = cachep->nodelists[node];
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001139 STATS_INC_NODEFREES(cachep);
1140 if (l3->alien && l3->alien[nodeid]) {
1141 alien = l3->alien[nodeid];
Ingo Molnar873623d2006-07-13 14:44:38 +02001142 spin_lock(&alien->lock);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001143 if (unlikely(alien->avail == alien->limit)) {
1144 STATS_INC_ACOVERFLOW(cachep);
1145 __drain_alien_cache(cachep, alien, nodeid);
1146 }
1147 alien->entry[alien->avail++] = objp;
1148 spin_unlock(&alien->lock);
1149 } else {
1150 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1151 free_block(cachep, &objp, 1, nodeid);
1152 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1153 }
1154 return 1;
1155}
Christoph Lametere498be72005-09-09 13:03:32 -07001156#endif
1157
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001158static void __cpuinit cpuup_canceled(long cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001159{
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001160 struct kmem_cache *cachep;
1161 struct kmem_list3 *l3 = NULL;
1162 int node = cpu_to_node(cpu);
Mike Travisc5f59f02008-04-04 18:11:10 -07001163 node_to_cpumask_ptr(mask, node);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001164
1165 list_for_each_entry(cachep, &cache_chain, next) {
1166 struct array_cache *nc;
1167 struct array_cache *shared;
1168 struct array_cache **alien;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001169
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001170 /* cpu is dead; no one can alloc from it. */
1171 nc = cachep->array[cpu];
1172 cachep->array[cpu] = NULL;
1173 l3 = cachep->nodelists[node];
1174
1175 if (!l3)
1176 goto free_array_cache;
1177
1178 spin_lock_irq(&l3->list_lock);
1179
1180 /* Free limit for this kmem_list3 */
1181 l3->free_limit -= cachep->batchcount;
1182 if (nc)
1183 free_block(cachep, nc->entry, nc->avail, node);
1184
Mike Travisc5f59f02008-04-04 18:11:10 -07001185 if (!cpus_empty(*mask)) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001186 spin_unlock_irq(&l3->list_lock);
1187 goto free_array_cache;
1188 }
1189
1190 shared = l3->shared;
1191 if (shared) {
1192 free_block(cachep, shared->entry,
1193 shared->avail, node);
1194 l3->shared = NULL;
1195 }
1196
1197 alien = l3->alien;
1198 l3->alien = NULL;
1199
1200 spin_unlock_irq(&l3->list_lock);
1201
1202 kfree(shared);
1203 if (alien) {
1204 drain_alien_cache(cachep, alien);
1205 free_alien_cache(alien);
1206 }
1207free_array_cache:
1208 kfree(nc);
1209 }
1210 /*
1211 * In the previous loop, all the objects were freed to
1212 * the respective cache's slabs, now we can go ahead and
1213 * shrink each nodelist to its limit.
1214 */
1215 list_for_each_entry(cachep, &cache_chain, next) {
1216 l3 = cachep->nodelists[node];
1217 if (!l3)
1218 continue;
1219 drain_freelist(cachep, l3, l3->free_objects);
1220 }
1221}
1222
1223static int __cpuinit cpuup_prepare(long cpu)
1224{
Pekka Enberg343e0d72006-02-01 03:05:50 -08001225 struct kmem_cache *cachep;
Christoph Lametere498be72005-09-09 13:03:32 -07001226 struct kmem_list3 *l3 = NULL;
1227 int node = cpu_to_node(cpu);
David Howellsea02e3d2007-07-19 01:49:09 -07001228 const int memsize = sizeof(struct kmem_list3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001229
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001230 /*
1231 * We need to do this right in the beginning since
1232 * alloc_arraycache's are going to use this list.
1233 * kmalloc_node allows us to add the slab to the right
1234 * kmem_list3 and not this cpu's kmem_list3
1235 */
1236
1237 list_for_each_entry(cachep, &cache_chain, next) {
1238 /*
1239 * Set up the size64 kmemlist for cpu before we can
1240 * begin anything. Make sure some other cpu on this
1241 * node has not already allocated this
1242 */
1243 if (!cachep->nodelists[node]) {
1244 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1245 if (!l3)
1246 goto bad;
1247 kmem_list3_init(l3);
1248 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1249 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1250
1251 /*
1252 * The l3s don't come and go as CPUs come and
1253 * go. cache_chain_mutex is sufficient
1254 * protection here.
1255 */
1256 cachep->nodelists[node] = l3;
1257 }
1258
1259 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1260 cachep->nodelists[node]->free_limit =
1261 (1 + nr_cpus_node(node)) *
1262 cachep->batchcount + cachep->num;
1263 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1264 }
1265
1266 /*
1267 * Now we can go ahead with allocating the shared arrays and
1268 * array caches
1269 */
1270 list_for_each_entry(cachep, &cache_chain, next) {
1271 struct array_cache *nc;
1272 struct array_cache *shared = NULL;
1273 struct array_cache **alien = NULL;
1274
1275 nc = alloc_arraycache(node, cachep->limit,
1276 cachep->batchcount);
1277 if (!nc)
1278 goto bad;
1279 if (cachep->shared) {
1280 shared = alloc_arraycache(node,
1281 cachep->shared * cachep->batchcount,
1282 0xbaadf00d);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001283 if (!shared) {
1284 kfree(nc);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001285 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001286 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001287 }
1288 if (use_alien_caches) {
1289 alien = alloc_alien_cache(node, cachep->limit);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001290 if (!alien) {
1291 kfree(shared);
1292 kfree(nc);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001293 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001294 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001295 }
1296 cachep->array[cpu] = nc;
1297 l3 = cachep->nodelists[node];
1298 BUG_ON(!l3);
1299
1300 spin_lock_irq(&l3->list_lock);
1301 if (!l3->shared) {
1302 /*
1303 * We are serialised from CPU_DEAD or
1304 * CPU_UP_CANCELLED by the cpucontrol lock
1305 */
1306 l3->shared = shared;
1307 shared = NULL;
1308 }
1309#ifdef CONFIG_NUMA
1310 if (!l3->alien) {
1311 l3->alien = alien;
1312 alien = NULL;
1313 }
1314#endif
1315 spin_unlock_irq(&l3->list_lock);
1316 kfree(shared);
1317 free_alien_cache(alien);
1318 }
1319 return 0;
1320bad:
Akinobu Mita12d00f62007-10-18 03:05:11 -07001321 cpuup_canceled(cpu);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001322 return -ENOMEM;
1323}
1324
1325static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1326 unsigned long action, void *hcpu)
1327{
1328 long cpu = (long)hcpu;
1329 int err = 0;
1330
Linus Torvalds1da177e2005-04-16 15:20:36 -07001331 switch (action) {
Heiko Carstens38c3bd92007-05-09 02:34:05 -07001332 case CPU_UP_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001333 case CPU_UP_PREPARE_FROZEN:
Gautham R Shenoy95402b32008-01-25 21:08:02 +01001334 mutex_lock(&cache_chain_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001335 err = cpuup_prepare(cpu);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01001336 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001337 break;
1338 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001339 case CPU_ONLINE_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001340 start_cpu_timer(cpu);
1341 break;
1342#ifdef CONFIG_HOTPLUG_CPU
Christoph Lameter5830c592007-05-09 02:34:22 -07001343 case CPU_DOWN_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001344 case CPU_DOWN_PREPARE_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001345 /*
1346 * Shutdown cache reaper. Note that the cache_chain_mutex is
1347 * held so that if cache_reap() is invoked it cannot do
1348 * anything expensive but will only modify reap_work
1349 * and reschedule the timer.
1350 */
1351 cancel_rearming_delayed_work(&per_cpu(reap_work, cpu));
1352 /* Now the cache_reaper is guaranteed to be not running. */
1353 per_cpu(reap_work, cpu).work.func = NULL;
1354 break;
1355 case CPU_DOWN_FAILED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001356 case CPU_DOWN_FAILED_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001357 start_cpu_timer(cpu);
1358 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001359 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001360 case CPU_DEAD_FROZEN:
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001361 /*
1362 * Even if all the cpus of a node are down, we don't free the
1363 * kmem_list3 of any cache. This to avoid a race between
1364 * cpu_down, and a kmalloc allocation from another cpu for
1365 * memory from the node of the cpu going down. The list3
1366 * structure is usually allocated from kmem_cache_create() and
1367 * gets destroyed at kmem_cache_destroy().
1368 */
Simon Arlott183ff222007-10-20 01:27:18 +02001369 /* fall through */
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08001370#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001371 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001372 case CPU_UP_CANCELED_FROZEN:
Gautham R Shenoy95402b32008-01-25 21:08:02 +01001373 mutex_lock(&cache_chain_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001374 cpuup_canceled(cpu);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001375 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001376 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001377 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001378 return err ? NOTIFY_BAD : NOTIFY_OK;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001379}
1380
Chandra Seetharaman74b85f32006-06-27 02:54:09 -07001381static struct notifier_block __cpuinitdata cpucache_notifier = {
1382 &cpuup_callback, NULL, 0
1383};
Linus Torvalds1da177e2005-04-16 15:20:36 -07001384
Christoph Lametere498be72005-09-09 13:03:32 -07001385/*
1386 * swap the static kmem_list3 with kmalloced memory
1387 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001388static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1389 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001390{
1391 struct kmem_list3 *ptr;
1392
Christoph Lametere498be72005-09-09 13:03:32 -07001393 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1394 BUG_ON(!ptr);
1395
1396 local_irq_disable();
1397 memcpy(ptr, list, sizeof(struct kmem_list3));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001398 /*
1399 * Do not assume that spinlocks can be initialized via memcpy:
1400 */
1401 spin_lock_init(&ptr->list_lock);
1402
Christoph Lametere498be72005-09-09 13:03:32 -07001403 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1404 cachep->nodelists[nodeid] = ptr;
1405 local_irq_enable();
1406}
1407
Andrew Mortona737b3e2006-03-22 00:08:11 -08001408/*
Pekka Enberg556a1692008-01-25 08:20:51 +02001409 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1410 * size of kmem_list3.
1411 */
1412static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1413{
1414 int node;
1415
1416 for_each_online_node(node) {
1417 cachep->nodelists[node] = &initkmem_list3[index + node];
1418 cachep->nodelists[node]->next_reap = jiffies +
1419 REAPTIMEOUT_LIST3 +
1420 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1421 }
1422}
1423
1424/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08001425 * Initialisation. Called after the page allocator have been initialised and
1426 * before smp_init().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001427 */
1428void __init kmem_cache_init(void)
1429{
1430 size_t left_over;
1431 struct cache_sizes *sizes;
1432 struct cache_names *names;
Christoph Lametere498be72005-09-09 13:03:32 -07001433 int i;
Jack Steiner07ed76b2006-03-07 21:55:46 -08001434 int order;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001435 int node;
Christoph Lametere498be72005-09-09 13:03:32 -07001436
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -07001437 if (num_possible_nodes() == 1) {
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001438 use_alien_caches = 0;
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -07001439 numa_platform = 0;
1440 }
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001441
Christoph Lametere498be72005-09-09 13:03:32 -07001442 for (i = 0; i < NUM_INIT_LISTS; i++) {
1443 kmem_list3_init(&initkmem_list3[i]);
1444 if (i < MAX_NUMNODES)
1445 cache_cache.nodelists[i] = NULL;
1446 }
Pekka Enberg556a1692008-01-25 08:20:51 +02001447 set_up_list3s(&cache_cache, CACHE_CACHE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001448
1449 /*
1450 * Fragmentation resistance on low memory - only use bigger
1451 * page orders on machines with more than 32MB of memory.
1452 */
1453 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1454 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1455
Linus Torvalds1da177e2005-04-16 15:20:36 -07001456 /* Bootstrap is tricky, because several objects are allocated
1457 * from caches that do not exist yet:
Andrew Mortona737b3e2006-03-22 00:08:11 -08001458 * 1) initialize the cache_cache cache: it contains the struct
1459 * kmem_cache structures of all caches, except cache_cache itself:
1460 * cache_cache is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001461 * Initially an __init data area is used for the head array and the
1462 * kmem_list3 structures, it's replaced with a kmalloc allocated
1463 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001464 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001465 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001466 * An __init data area is used for the head array.
1467 * 3) Create the remaining kmalloc caches, with minimally sized
1468 * head arrays.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001469 * 4) Replace the __init data head arrays for cache_cache and the first
1470 * kmalloc cache with kmalloc allocated arrays.
Christoph Lametere498be72005-09-09 13:03:32 -07001471 * 5) Replace the __init data for kmem_list3 for cache_cache and
1472 * the other cache's with kmalloc allocated memory.
1473 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001474 */
1475
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001476 node = numa_node_id();
1477
Linus Torvalds1da177e2005-04-16 15:20:36 -07001478 /* 1) create the cache_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001479 INIT_LIST_HEAD(&cache_chain);
1480 list_add(&cache_cache.next, &cache_chain);
1481 cache_cache.colour_off = cache_line_size();
1482 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
Daniel Yeisleyec1f5ee2008-03-25 23:59:08 +02001483 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001484
Eric Dumazet8da34302007-05-06 14:49:29 -07001485 /*
1486 * struct kmem_cache size depends on nr_node_ids, which
1487 * can be less than MAX_NUMNODES.
1488 */
1489 cache_cache.buffer_size = offsetof(struct kmem_cache, nodelists) +
1490 nr_node_ids * sizeof(struct kmem_list3 *);
1491#if DEBUG
1492 cache_cache.obj_size = cache_cache.buffer_size;
1493#endif
Andrew Mortona737b3e2006-03-22 00:08:11 -08001494 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1495 cache_line_size());
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08001496 cache_cache.reciprocal_buffer_size =
1497 reciprocal_value(cache_cache.buffer_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001498
Jack Steiner07ed76b2006-03-07 21:55:46 -08001499 for (order = 0; order < MAX_ORDER; order++) {
1500 cache_estimate(order, cache_cache.buffer_size,
1501 cache_line_size(), 0, &left_over, &cache_cache.num);
1502 if (cache_cache.num)
1503 break;
1504 }
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02001505 BUG_ON(!cache_cache.num);
Jack Steiner07ed76b2006-03-07 21:55:46 -08001506 cache_cache.gfporder = order;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001507 cache_cache.colour = left_over / cache_cache.colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001508 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1509 sizeof(struct slab), cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001510
1511 /* 2+3) create the kmalloc caches */
1512 sizes = malloc_sizes;
1513 names = cache_names;
1514
Andrew Mortona737b3e2006-03-22 00:08:11 -08001515 /*
1516 * Initialize the caches that provide memory for the array cache and the
1517 * kmem_list3 structures first. Without this, further allocations will
1518 * bug.
Christoph Lametere498be72005-09-09 13:03:32 -07001519 */
1520
1521 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001522 sizes[INDEX_AC].cs_size,
1523 ARCH_KMALLOC_MINALIGN,
1524 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001525 NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07001526
Andrew Mortona737b3e2006-03-22 00:08:11 -08001527 if (INDEX_AC != INDEX_L3) {
Christoph Lametere498be72005-09-09 13:03:32 -07001528 sizes[INDEX_L3].cs_cachep =
Andrew Mortona737b3e2006-03-22 00:08:11 -08001529 kmem_cache_create(names[INDEX_L3].name,
1530 sizes[INDEX_L3].cs_size,
1531 ARCH_KMALLOC_MINALIGN,
1532 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001533 NULL);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001534 }
Christoph Lametere498be72005-09-09 13:03:32 -07001535
Ingo Molnare0a42722006-06-23 02:03:46 -07001536 slab_early_init = 0;
1537
Linus Torvalds1da177e2005-04-16 15:20:36 -07001538 while (sizes->cs_size != ULONG_MAX) {
Christoph Lametere498be72005-09-09 13:03:32 -07001539 /*
1540 * For performance, all the general caches are L1 aligned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001541 * This should be particularly beneficial on SMP boxes, as it
1542 * eliminates "false sharing".
1543 * Note for systems short on memory removing the alignment will
Christoph Lametere498be72005-09-09 13:03:32 -07001544 * allow tighter packing of the smaller caches.
1545 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001546 if (!sizes->cs_cachep) {
Christoph Lametere498be72005-09-09 13:03:32 -07001547 sizes->cs_cachep = kmem_cache_create(names->name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001548 sizes->cs_size,
1549 ARCH_KMALLOC_MINALIGN,
1550 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001551 NULL);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001552 }
Christoph Lameter4b51d662007-02-10 01:43:10 -08001553#ifdef CONFIG_ZONE_DMA
1554 sizes->cs_dmacachep = kmem_cache_create(
1555 names->name_dma,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001556 sizes->cs_size,
1557 ARCH_KMALLOC_MINALIGN,
1558 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1559 SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001560 NULL);
Christoph Lameter4b51d662007-02-10 01:43:10 -08001561#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001562 sizes++;
1563 names++;
1564 }
1565 /* 4) Replace the bootstrap head arrays */
1566 {
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001567 struct array_cache *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001568
Linus Torvalds1da177e2005-04-16 15:20:36 -07001569 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001570
Linus Torvalds1da177e2005-04-16 15:20:36 -07001571 local_irq_disable();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001572 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1573 memcpy(ptr, cpu_cache_get(&cache_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001574 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001575 /*
1576 * Do not assume that spinlocks can be initialized via memcpy:
1577 */
1578 spin_lock_init(&ptr->lock);
1579
Linus Torvalds1da177e2005-04-16 15:20:36 -07001580 cache_cache.array[smp_processor_id()] = ptr;
1581 local_irq_enable();
Christoph Lametere498be72005-09-09 13:03:32 -07001582
Linus Torvalds1da177e2005-04-16 15:20:36 -07001583 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001584
Linus Torvalds1da177e2005-04-16 15:20:36 -07001585 local_irq_disable();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001586 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001587 != &initarray_generic.cache);
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001588 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001589 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001590 /*
1591 * Do not assume that spinlocks can be initialized via memcpy:
1592 */
1593 spin_lock_init(&ptr->lock);
1594
Christoph Lametere498be72005-09-09 13:03:32 -07001595 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001596 ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001597 local_irq_enable();
1598 }
Christoph Lametere498be72005-09-09 13:03:32 -07001599 /* 5) Replace the bootstrap kmem_list3's */
1600 {
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001601 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001602
Mel Gorman9c09a952008-01-24 05:49:54 -08001603 for_each_online_node(nid) {
Daniel Yeisleyec1f5ee2008-03-25 23:59:08 +02001604 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid);
Pekka Enberg556a1692008-01-25 08:20:51 +02001605
Christoph Lametere498be72005-09-09 13:03:32 -07001606 init_list(malloc_sizes[INDEX_AC].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001607 &initkmem_list3[SIZE_AC + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001608
1609 if (INDEX_AC != INDEX_L3) {
1610 init_list(malloc_sizes[INDEX_L3].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001611 &initkmem_list3[SIZE_L3 + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001612 }
1613 }
1614 }
1615
1616 /* 6) resize the head arrays to their final sizes */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001617 {
Pekka Enberg343e0d72006-02-01 03:05:50 -08001618 struct kmem_cache *cachep;
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001619 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001620 list_for_each_entry(cachep, &cache_chain, next)
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07001621 if (enable_cpucache(cachep))
1622 BUG();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001623 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001624 }
1625
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001626 /* Annotate slab for lockdep -- annotate the malloc caches */
1627 init_lock_keys();
1628
1629
Linus Torvalds1da177e2005-04-16 15:20:36 -07001630 /* Done! */
1631 g_cpucache_up = FULL;
1632
Andrew Mortona737b3e2006-03-22 00:08:11 -08001633 /*
1634 * Register a cpu startup notifier callback that initializes
1635 * cpu_cache_get for all new cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07001636 */
1637 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001638
Andrew Mortona737b3e2006-03-22 00:08:11 -08001639 /*
1640 * The reap timers are started later, with a module init call: That part
1641 * of the kernel is not yet operational.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001642 */
1643}
1644
1645static int __init cpucache_init(void)
1646{
1647 int cpu;
1648
Andrew Mortona737b3e2006-03-22 00:08:11 -08001649 /*
1650 * Register the timers that return unneeded pages to the page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -07001651 */
Christoph Lametere498be72005-09-09 13:03:32 -07001652 for_each_online_cpu(cpu)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001653 start_cpu_timer(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001654 return 0;
1655}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001656__initcall(cpucache_init);
1657
1658/*
1659 * Interface to system's page allocator. No need to hold the cache-lock.
1660 *
1661 * If we requested dmaable memory, we will get it. Even if we
1662 * did not request dmaable memory, we might get it, but that
1663 * would be relatively rare and ignorable.
1664 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001665static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001666{
1667 struct page *page;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001668 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001669 int i;
1670
Luke Yangd6fef9d2006-04-10 22:52:56 -07001671#ifndef CONFIG_MMU
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001672 /*
1673 * Nommu uses slab's for process anonymous memory allocations, and thus
1674 * requires __GFP_COMP to properly refcount higher order allocations
Luke Yangd6fef9d2006-04-10 22:52:56 -07001675 */
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001676 flags |= __GFP_COMP;
Luke Yangd6fef9d2006-04-10 22:52:56 -07001677#endif
Christoph Lameter765c4502006-09-27 01:50:08 -07001678
Christoph Lameter3c517a62006-12-06 20:33:29 -08001679 flags |= cachep->gfpflags;
Mel Gormane12ba742007-10-16 01:25:52 -07001680 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1681 flags |= __GFP_RECLAIMABLE;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001682
1683 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001684 if (!page)
1685 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001686
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001687 nr_pages = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001688 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
Christoph Lameter972d1a72006-09-25 23:31:51 -07001689 add_zone_page_state(page_zone(page),
1690 NR_SLAB_RECLAIMABLE, nr_pages);
1691 else
1692 add_zone_page_state(page_zone(page),
1693 NR_SLAB_UNRECLAIMABLE, nr_pages);
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001694 for (i = 0; i < nr_pages; i++)
1695 __SetPageSlab(page + i);
1696 return page_address(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001697}
1698
1699/*
1700 * Interface to system's page release.
1701 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001702static void kmem_freepages(struct kmem_cache *cachep, void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001703{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001704 unsigned long i = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001705 struct page *page = virt_to_page(addr);
1706 const unsigned long nr_freed = i;
1707
Christoph Lameter972d1a72006-09-25 23:31:51 -07001708 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1709 sub_zone_page_state(page_zone(page),
1710 NR_SLAB_RECLAIMABLE, nr_freed);
1711 else
1712 sub_zone_page_state(page_zone(page),
1713 NR_SLAB_UNRECLAIMABLE, nr_freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001714 while (i--) {
Nick Pigginf205b2f2006-03-22 00:08:02 -08001715 BUG_ON(!PageSlab(page));
1716 __ClearPageSlab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001717 page++;
1718 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001719 if (current->reclaim_state)
1720 current->reclaim_state->reclaimed_slab += nr_freed;
1721 free_pages((unsigned long)addr, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001722}
1723
1724static void kmem_rcu_free(struct rcu_head *head)
1725{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001726 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001727 struct kmem_cache *cachep = slab_rcu->cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001728
1729 kmem_freepages(cachep, slab_rcu->addr);
1730 if (OFF_SLAB(cachep))
1731 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1732}
1733
1734#if DEBUG
1735
1736#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001737static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001738 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001739{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001740 int size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001741
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001742 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001743
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001744 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001745 return;
1746
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001747 *addr++ = 0x12345678;
1748 *addr++ = caller;
1749 *addr++ = smp_processor_id();
1750 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001751 {
1752 unsigned long *sptr = &caller;
1753 unsigned long svalue;
1754
1755 while (!kstack_end(sptr)) {
1756 svalue = *sptr++;
1757 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001758 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001759 size -= sizeof(unsigned long);
1760 if (size <= sizeof(unsigned long))
1761 break;
1762 }
1763 }
1764
1765 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001766 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001767}
1768#endif
1769
Pekka Enberg343e0d72006-02-01 03:05:50 -08001770static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001771{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001772 int size = obj_size(cachep);
1773 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001774
1775 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001776 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001777}
1778
1779static void dump_line(char *data, int offset, int limit)
1780{
1781 int i;
Dave Jonesaa83aa42006-09-29 01:59:51 -07001782 unsigned char error = 0;
1783 int bad_count = 0;
1784
Linus Torvalds1da177e2005-04-16 15:20:36 -07001785 printk(KERN_ERR "%03x:", offset);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001786 for (i = 0; i < limit; i++) {
1787 if (data[offset + i] != POISON_FREE) {
1788 error = data[offset + i];
1789 bad_count++;
1790 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001791 printk(" %02x", (unsigned char)data[offset + i]);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001792 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001793 printk("\n");
Dave Jonesaa83aa42006-09-29 01:59:51 -07001794
1795 if (bad_count == 1) {
1796 error ^= POISON_FREE;
1797 if (!(error & (error - 1))) {
1798 printk(KERN_ERR "Single bit error detected. Probably "
1799 "bad RAM.\n");
1800#ifdef CONFIG_X86
1801 printk(KERN_ERR "Run memtest86+ or a similar memory "
1802 "test tool.\n");
1803#else
1804 printk(KERN_ERR "Run a memory test tool.\n");
1805#endif
1806 }
1807 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001808}
1809#endif
1810
1811#if DEBUG
1812
Pekka Enberg343e0d72006-02-01 03:05:50 -08001813static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001814{
1815 int i, size;
1816 char *realobj;
1817
1818 if (cachep->flags & SLAB_RED_ZONE) {
David Woodhouseb46b8f12007-05-08 00:22:59 -07001819 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001820 *dbg_redzone1(cachep, objp),
1821 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001822 }
1823
1824 if (cachep->flags & SLAB_STORE_USER) {
1825 printk(KERN_ERR "Last user: [<%p>]",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001826 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001827 print_symbol("(%s)",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001828 (unsigned long)*dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001829 printk("\n");
1830 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001831 realobj = (char *)objp + obj_offset(cachep);
1832 size = obj_size(cachep);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001833 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001834 int limit;
1835 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001836 if (i + limit > size)
1837 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001838 dump_line(realobj, i, limit);
1839 }
1840}
1841
Pekka Enberg343e0d72006-02-01 03:05:50 -08001842static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001843{
1844 char *realobj;
1845 int size, i;
1846 int lines = 0;
1847
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001848 realobj = (char *)objp + obj_offset(cachep);
1849 size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001850
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001851 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001852 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001853 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001854 exp = POISON_END;
1855 if (realobj[i] != exp) {
1856 int limit;
1857 /* Mismatch ! */
1858 /* Print header */
1859 if (lines == 0) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001860 printk(KERN_ERR
David Howellse94a40c2007-04-02 23:46:28 +01001861 "Slab corruption: %s start=%p, len=%d\n",
1862 cachep->name, realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001863 print_objinfo(cachep, objp, 0);
1864 }
1865 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001866 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001867 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001868 if (i + limit > size)
1869 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001870 dump_line(realobj, i, limit);
1871 i += 16;
1872 lines++;
1873 /* Limit to 5 lines */
1874 if (lines > 5)
1875 break;
1876 }
1877 }
1878 if (lines != 0) {
1879 /* Print some data about the neighboring objects, if they
1880 * exist:
1881 */
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08001882 struct slab *slabp = virt_to_slab(objp);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001883 unsigned int objnr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001884
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001885 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001886 if (objnr) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001887 objp = index_to_obj(cachep, slabp, objnr - 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001888 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001889 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001890 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001891 print_objinfo(cachep, objp, 2);
1892 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001893 if (objnr + 1 < cachep->num) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001894 objp = index_to_obj(cachep, slabp, objnr + 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001895 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001896 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001897 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001898 print_objinfo(cachep, objp, 2);
1899 }
1900 }
1901}
1902#endif
1903
Linus Torvalds1da177e2005-04-16 15:20:36 -07001904#if DEBUG
Rabin Vincente79aec22008-07-04 00:40:32 +05301905static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001906{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001907 int i;
1908 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001909 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001910
1911 if (cachep->flags & SLAB_POISON) {
1912#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08001913 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1914 OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001915 kernel_map_pages(virt_to_page(objp),
Andrew Mortona737b3e2006-03-22 00:08:11 -08001916 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001917 else
1918 check_poison_obj(cachep, objp);
1919#else
1920 check_poison_obj(cachep, objp);
1921#endif
1922 }
1923 if (cachep->flags & SLAB_RED_ZONE) {
1924 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1925 slab_error(cachep, "start of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001926 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001927 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1928 slab_error(cachep, "end of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001929 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001930 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001931 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001932}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001933#else
Rabin Vincente79aec22008-07-04 00:40:32 +05301934static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001935{
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001936}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001937#endif
1938
Randy Dunlap911851e2006-03-22 00:08:14 -08001939/**
1940 * slab_destroy - destroy and release all objects in a slab
1941 * @cachep: cache pointer being destroyed
1942 * @slabp: slab pointer being destroyed
1943 *
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001944 * Destroy all the objs in a slab, and release the mem back to the system.
Andrew Mortona737b3e2006-03-22 00:08:11 -08001945 * Before calling the slab must have been unlinked from the cache. The
1946 * cache-lock is not held/needed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001947 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001948static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001949{
1950 void *addr = slabp->s_mem - slabp->colouroff;
1951
Rabin Vincente79aec22008-07-04 00:40:32 +05301952 slab_destroy_debugcheck(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001953 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1954 struct slab_rcu *slab_rcu;
1955
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001956 slab_rcu = (struct slab_rcu *)slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001957 slab_rcu->cachep = cachep;
1958 slab_rcu->addr = addr;
1959 call_rcu(&slab_rcu->head, kmem_rcu_free);
1960 } else {
1961 kmem_freepages(cachep, addr);
Ingo Molnar873623d2006-07-13 14:44:38 +02001962 if (OFF_SLAB(cachep))
1963 kmem_cache_free(cachep->slabp_cache, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001964 }
1965}
1966
Christoph Lameter117f6eb2006-09-25 23:31:37 -07001967static void __kmem_cache_destroy(struct kmem_cache *cachep)
1968{
1969 int i;
1970 struct kmem_list3 *l3;
1971
1972 for_each_online_cpu(i)
1973 kfree(cachep->array[i]);
1974
1975 /* NUMA: free the list3 structures */
1976 for_each_online_node(i) {
1977 l3 = cachep->nodelists[i];
1978 if (l3) {
1979 kfree(l3->shared);
1980 free_alien_cache(l3->alien);
1981 kfree(l3);
1982 }
1983 }
1984 kmem_cache_free(&cache_cache, cachep);
1985}
1986
1987
Linus Torvalds1da177e2005-04-16 15:20:36 -07001988/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08001989 * calculate_slab_order - calculate size (page order) of slabs
1990 * @cachep: pointer to the cache that is being created
1991 * @size: size of objects to be created in this cache.
1992 * @align: required alignment for the objects.
1993 * @flags: slab allocation flags
1994 *
1995 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001996 *
1997 * This could be made much more intelligent. For now, try to avoid using
1998 * high order pages for slabs. When the gfp() functions are more friendly
1999 * towards high-order requests, this should be changed.
2000 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002001static size_t calculate_slab_order(struct kmem_cache *cachep,
Randy Dunlapee13d782006-02-01 03:05:53 -08002002 size_t size, size_t align, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002003{
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002004 unsigned long offslab_limit;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002005 size_t left_over = 0;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002006 int gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002007
Christoph Lameter0aa817f2007-05-16 22:11:01 -07002008 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002009 unsigned int num;
2010 size_t remainder;
2011
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002012 cache_estimate(gfporder, size, align, flags, &remainder, &num);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002013 if (!num)
2014 continue;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002015
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002016 if (flags & CFLGS_OFF_SLAB) {
2017 /*
2018 * Max number of objs-per-slab for caches which
2019 * use off-slab slabs. Needed to avoid a possible
2020 * looping condition in cache_grow().
2021 */
2022 offslab_limit = size - sizeof(struct slab);
2023 offslab_limit /= sizeof(kmem_bufctl_t);
2024
2025 if (num > offslab_limit)
2026 break;
2027 }
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002028
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002029 /* Found something acceptable - save it away */
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002030 cachep->num = num;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002031 cachep->gfporder = gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002032 left_over = remainder;
2033
2034 /*
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002035 * A VFS-reclaimable slab tends to have most allocations
2036 * as GFP_NOFS and we really don't want to have to be allocating
2037 * higher-order pages when we are unable to shrink dcache.
2038 */
2039 if (flags & SLAB_RECLAIM_ACCOUNT)
2040 break;
2041
2042 /*
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002043 * Large number of objects is good, but very large slabs are
2044 * currently bad for the gfp()s.
2045 */
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002046 if (gfporder >= slab_break_gfp_order)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002047 break;
2048
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002049 /*
2050 * Acceptable internal fragmentation?
2051 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002052 if (left_over * 8 <= (PAGE_SIZE << gfporder))
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002053 break;
2054 }
2055 return left_over;
2056}
2057
Sam Ravnborg38bdc322007-05-17 23:48:19 +02002058static int __init_refok setup_cpu_cache(struct kmem_cache *cachep)
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002059{
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002060 if (g_cpucache_up == FULL)
2061 return enable_cpucache(cachep);
2062
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002063 if (g_cpucache_up == NONE) {
2064 /*
2065 * Note: the first kmem_cache_create must create the cache
2066 * that's used by kmalloc(24), otherwise the creation of
2067 * further caches will BUG().
2068 */
2069 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2070
2071 /*
2072 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2073 * the first cache, then we need to set up all its list3s,
2074 * otherwise the creation of further caches will BUG().
2075 */
2076 set_up_list3s(cachep, SIZE_AC);
2077 if (INDEX_AC == INDEX_L3)
2078 g_cpucache_up = PARTIAL_L3;
2079 else
2080 g_cpucache_up = PARTIAL_AC;
2081 } else {
2082 cachep->array[smp_processor_id()] =
2083 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
2084
2085 if (g_cpucache_up == PARTIAL_AC) {
2086 set_up_list3s(cachep, SIZE_L3);
2087 g_cpucache_up = PARTIAL_L3;
2088 } else {
2089 int node;
Pekka Enberg556a1692008-01-25 08:20:51 +02002090 for_each_online_node(node) {
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002091 cachep->nodelists[node] =
2092 kmalloc_node(sizeof(struct kmem_list3),
2093 GFP_KERNEL, node);
2094 BUG_ON(!cachep->nodelists[node]);
2095 kmem_list3_init(cachep->nodelists[node]);
2096 }
2097 }
2098 }
2099 cachep->nodelists[numa_node_id()]->next_reap =
2100 jiffies + REAPTIMEOUT_LIST3 +
2101 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2102
2103 cpu_cache_get(cachep)->avail = 0;
2104 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2105 cpu_cache_get(cachep)->batchcount = 1;
2106 cpu_cache_get(cachep)->touched = 0;
2107 cachep->batchcount = 1;
2108 cachep->limit = BOOT_CPUCACHE_ENTRIES;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002109 return 0;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002110}
2111
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002112/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07002113 * kmem_cache_create - Create a cache.
2114 * @name: A string which is used in /proc/slabinfo to identify this cache.
2115 * @size: The size of objects to be created in this cache.
2116 * @align: The required alignment for the objects.
2117 * @flags: SLAB flags
2118 * @ctor: A constructor for the objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002119 *
2120 * Returns a ptr to the cache on success, NULL on failure.
2121 * Cannot be called within a int, but can be interrupted.
Paul Mundt20c2df82007-07-20 10:11:58 +09002122 * The @ctor is run when new pages are allocated by the cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002123 *
2124 * @name must be valid until the cache is destroyed. This implies that
Andrew Mortona737b3e2006-03-22 00:08:11 -08002125 * the module calling this has to destroy the cache before getting unloaded.
Catalin Marinas249da162008-11-21 12:56:22 +00002126 * Note that kmem_cache_name() is not guaranteed to return the same pointer,
2127 * therefore applications must manage it themselves.
Andrew Mortona737b3e2006-03-22 00:08:11 -08002128 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002129 * The flags are
2130 *
2131 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2132 * to catch references to uninitialised memory.
2133 *
2134 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2135 * for buffer overruns.
2136 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002137 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2138 * cacheline. This can be beneficial if you're counting cycles as closely
2139 * as davem.
2140 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002141struct kmem_cache *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002142kmem_cache_create (const char *name, size_t size, size_t align,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002143 unsigned long flags, void (*ctor)(void *))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002144{
2145 size_t left_over, slab_size, ralign;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07002146 struct kmem_cache *cachep = NULL, *pc;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002147
2148 /*
2149 * Sanity checks... these are all serious usage bugs.
2150 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002151 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
Paul Mundt20c2df82007-07-20 10:11:58 +09002152 size > KMALLOC_MAX_SIZE) {
Harvey Harrisond40cee22008-04-30 00:55:07 -07002153 printk(KERN_ERR "%s: Early error in slab %s\n", __func__,
Andrew Mortona737b3e2006-03-22 00:08:11 -08002154 name);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002155 BUG();
2156 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002157
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002158 /*
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002159 * We use cache_chain_mutex to ensure a consistent view of
2160 * cpu_online_map as well. Please see cpuup_callback
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002161 */
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002162 get_online_cpus();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002163 mutex_lock(&cache_chain_mutex);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002164
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07002165 list_for_each_entry(pc, &cache_chain, next) {
Andrew Morton4f12bb42005-11-07 00:58:00 -08002166 char tmp;
2167 int res;
2168
2169 /*
2170 * This happens when the module gets unloaded and doesn't
2171 * destroy its slab cache and no-one else reuses the vmalloc
2172 * area of the module. Print a warning.
2173 */
Andrew Morton138ae662006-12-06 20:36:41 -08002174 res = probe_kernel_address(pc->name, tmp);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002175 if (res) {
matzeb4169522007-05-06 14:49:52 -07002176 printk(KERN_ERR
2177 "SLAB: cache with size %d has lost its name\n",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002178 pc->buffer_size);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002179 continue;
2180 }
2181
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002182 if (!strcmp(pc->name, name)) {
matzeb4169522007-05-06 14:49:52 -07002183 printk(KERN_ERR
2184 "kmem_cache_create: duplicate cache %s\n", name);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002185 dump_stack();
2186 goto oops;
2187 }
2188 }
2189
Linus Torvalds1da177e2005-04-16 15:20:36 -07002190#if DEBUG
2191 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002192#if FORCED_DEBUG
2193 /*
2194 * Enable redzoning and last user accounting, except for caches with
2195 * large objects, if the increased size would increase the object size
2196 * above the next power of two: caches with object sizes just above a
2197 * power of two have a significant amount of internal fragmentation.
2198 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002199 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2200 2 * sizeof(unsigned long long)))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002201 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002202 if (!(flags & SLAB_DESTROY_BY_RCU))
2203 flags |= SLAB_POISON;
2204#endif
2205 if (flags & SLAB_DESTROY_BY_RCU)
2206 BUG_ON(flags & SLAB_POISON);
2207#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002208 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002209 * Always checks flags, a caller might be expecting debug support which
2210 * isn't available.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002211 */
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002212 BUG_ON(flags & ~CREATE_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002213
Andrew Mortona737b3e2006-03-22 00:08:11 -08002214 /*
2215 * Check that size is in terms of words. This is needed to avoid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002216 * unaligned accesses for some archs when redzoning is used, and makes
2217 * sure any on-slab bufctl's are also correctly aligned.
2218 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002219 if (size & (BYTES_PER_WORD - 1)) {
2220 size += (BYTES_PER_WORD - 1);
2221 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002222 }
2223
Andrew Mortona737b3e2006-03-22 00:08:11 -08002224 /* calculate the final buffer alignment: */
2225
Linus Torvalds1da177e2005-04-16 15:20:36 -07002226 /* 1) arch recommendation: can be overridden for debug */
2227 if (flags & SLAB_HWCACHE_ALIGN) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002228 /*
2229 * Default alignment: as specified by the arch code. Except if
2230 * an object is really small, then squeeze multiple objects into
2231 * one cacheline.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002232 */
2233 ralign = cache_line_size();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002234 while (size <= ralign / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002235 ralign /= 2;
2236 } else {
2237 ralign = BYTES_PER_WORD;
2238 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002239
2240 /*
David Woodhouse87a927c2007-07-04 21:26:44 -04002241 * Redzoning and user store require word alignment or possibly larger.
2242 * Note this will be overridden by architecture or caller mandated
2243 * alignment if either is greater than BYTES_PER_WORD.
Pekka Enbergca5f9702006-09-25 23:31:25 -07002244 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002245 if (flags & SLAB_STORE_USER)
2246 ralign = BYTES_PER_WORD;
2247
2248 if (flags & SLAB_RED_ZONE) {
2249 ralign = REDZONE_ALIGN;
2250 /* If redzoning, ensure that the second redzone is suitably
2251 * aligned, by adjusting the object size accordingly. */
2252 size += REDZONE_ALIGN - 1;
2253 size &= ~(REDZONE_ALIGN - 1);
2254 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002255
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002256 /* 2) arch mandated alignment */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002257 if (ralign < ARCH_SLAB_MINALIGN) {
2258 ralign = ARCH_SLAB_MINALIGN;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002259 }
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002260 /* 3) caller mandated alignment */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002261 if (ralign < align) {
2262 ralign = align;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002263 }
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002264 /* disable debug if necessary */
David Woodhouseb46b8f12007-05-08 00:22:59 -07002265 if (ralign > __alignof__(unsigned long long))
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002266 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002267 /*
Pekka Enbergca5f9702006-09-25 23:31:25 -07002268 * 4) Store it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002269 */
2270 align = ralign;
2271
2272 /* Get cache's description obj. */
Christoph Lametere94b1762006-12-06 20:33:17 -08002273 cachep = kmem_cache_zalloc(&cache_cache, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002274 if (!cachep)
Andrew Morton4f12bb42005-11-07 00:58:00 -08002275 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002276
2277#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002278 cachep->obj_size = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002279
Pekka Enbergca5f9702006-09-25 23:31:25 -07002280 /*
2281 * Both debugging options require word-alignment which is calculated
2282 * into align above.
2283 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002284 if (flags & SLAB_RED_ZONE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002285 /* add space for red zone words */
David Woodhouseb46b8f12007-05-08 00:22:59 -07002286 cachep->obj_offset += sizeof(unsigned long long);
2287 size += 2 * sizeof(unsigned long long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002288 }
2289 if (flags & SLAB_STORE_USER) {
Pekka Enbergca5f9702006-09-25 23:31:25 -07002290 /* user store requires one word storage behind the end of
David Woodhouse87a927c2007-07-04 21:26:44 -04002291 * the real object. But if the second red zone needs to be
2292 * aligned to 64 bits, we must allow that much space.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002293 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002294 if (flags & SLAB_RED_ZONE)
2295 size += REDZONE_ALIGN;
2296 else
2297 size += BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002298 }
2299#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002300 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002301 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2302 cachep->obj_offset += PAGE_SIZE - size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002303 size = PAGE_SIZE;
2304 }
2305#endif
2306#endif
2307
Ingo Molnare0a42722006-06-23 02:03:46 -07002308 /*
2309 * Determine if the slab management is 'on' or 'off' slab.
2310 * (bootstrapping cannot cope with offslab caches so don't do
2311 * it too early on.)
2312 */
2313 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002314 /*
2315 * Size is large, assume best to place the slab management obj
2316 * off-slab (should allow better packing of objs).
2317 */
2318 flags |= CFLGS_OFF_SLAB;
2319
2320 size = ALIGN(size, align);
2321
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002322 left_over = calculate_slab_order(cachep, size, align, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002323
2324 if (!cachep->num) {
matzeb4169522007-05-06 14:49:52 -07002325 printk(KERN_ERR
2326 "kmem_cache_create: couldn't create cache %s.\n", name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002327 kmem_cache_free(&cache_cache, cachep);
2328 cachep = NULL;
Andrew Morton4f12bb42005-11-07 00:58:00 -08002329 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002330 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002331 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2332 + sizeof(struct slab), align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002333
2334 /*
2335 * If the slab has been placed off-slab, and we have enough space then
2336 * move it on-slab. This is at the expense of any extra colouring.
2337 */
2338 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2339 flags &= ~CFLGS_OFF_SLAB;
2340 left_over -= slab_size;
2341 }
2342
2343 if (flags & CFLGS_OFF_SLAB) {
2344 /* really off slab. No need for manual alignment */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002345 slab_size =
2346 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002347 }
2348
2349 cachep->colour_off = cache_line_size();
2350 /* Offset must be a multiple of the alignment. */
2351 if (cachep->colour_off < align)
2352 cachep->colour_off = align;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002353 cachep->colour = left_over / cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002354 cachep->slab_size = slab_size;
2355 cachep->flags = flags;
2356 cachep->gfpflags = 0;
Christoph Lameter4b51d662007-02-10 01:43:10 -08002357 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002358 cachep->gfpflags |= GFP_DMA;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002359 cachep->buffer_size = size;
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08002360 cachep->reciprocal_buffer_size = reciprocal_value(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002361
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002362 if (flags & CFLGS_OFF_SLAB) {
Victor Fuscob2d55072005-09-10 00:26:36 -07002363 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002364 /*
2365 * This is a possibility for one of the malloc_sizes caches.
2366 * But since we go off slab only for object size greater than
2367 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2368 * this should not happen at all.
2369 * But leave a BUG_ON for some lucky dude.
2370 */
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002371 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002372 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002373 cachep->ctor = ctor;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002374 cachep->name = name;
2375
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002376 if (setup_cpu_cache(cachep)) {
2377 __kmem_cache_destroy(cachep);
2378 cachep = NULL;
2379 goto oops;
2380 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002381
Linus Torvalds1da177e2005-04-16 15:20:36 -07002382 /* cache setup completed, link it into the list */
2383 list_add(&cachep->next, &cache_chain);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002384oops:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002385 if (!cachep && (flags & SLAB_PANIC))
2386 panic("kmem_cache_create(): failed to create slab `%s'\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002387 name);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002388 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002389 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002390 return cachep;
2391}
2392EXPORT_SYMBOL(kmem_cache_create);
2393
2394#if DEBUG
2395static void check_irq_off(void)
2396{
2397 BUG_ON(!irqs_disabled());
2398}
2399
2400static void check_irq_on(void)
2401{
2402 BUG_ON(irqs_disabled());
2403}
2404
Pekka Enberg343e0d72006-02-01 03:05:50 -08002405static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002406{
2407#ifdef CONFIG_SMP
2408 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002409 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002410#endif
2411}
Christoph Lametere498be72005-09-09 13:03:32 -07002412
Pekka Enberg343e0d72006-02-01 03:05:50 -08002413static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002414{
2415#ifdef CONFIG_SMP
2416 check_irq_off();
2417 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2418#endif
2419}
2420
Linus Torvalds1da177e2005-04-16 15:20:36 -07002421#else
2422#define check_irq_off() do { } while(0)
2423#define check_irq_on() do { } while(0)
2424#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002425#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002426#endif
2427
Christoph Lameteraab22072006-03-22 00:09:06 -08002428static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2429 struct array_cache *ac,
2430 int force, int node);
2431
Linus Torvalds1da177e2005-04-16 15:20:36 -07002432static void do_drain(void *arg)
2433{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002434 struct kmem_cache *cachep = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002435 struct array_cache *ac;
Christoph Lameterff694162005-09-22 21:44:02 -07002436 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002437
2438 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002439 ac = cpu_cache_get(cachep);
Christoph Lameterff694162005-09-22 21:44:02 -07002440 spin_lock(&cachep->nodelists[node]->list_lock);
2441 free_block(cachep, ac->entry, ac->avail, node);
2442 spin_unlock(&cachep->nodelists[node]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002443 ac->avail = 0;
2444}
2445
Pekka Enberg343e0d72006-02-01 03:05:50 -08002446static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002447{
Christoph Lametere498be72005-09-09 13:03:32 -07002448 struct kmem_list3 *l3;
2449 int node;
2450
Jens Axboe15c8b6c2008-05-09 09:39:44 +02002451 on_each_cpu(do_drain, cachep, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002452 check_irq_on();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002453 for_each_online_node(node) {
Christoph Lametere498be72005-09-09 13:03:32 -07002454 l3 = cachep->nodelists[node];
Roland Dreiera4523a82006-05-15 11:41:00 -07002455 if (l3 && l3->alien)
2456 drain_alien_cache(cachep, l3->alien);
2457 }
2458
2459 for_each_online_node(node) {
2460 l3 = cachep->nodelists[node];
2461 if (l3)
Christoph Lameteraab22072006-03-22 00:09:06 -08002462 drain_array(cachep, l3, l3->shared, 1, node);
Christoph Lametere498be72005-09-09 13:03:32 -07002463 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002464}
2465
Christoph Lametered11d9e2006-06-30 01:55:45 -07002466/*
2467 * Remove slabs from the list of free slabs.
2468 * Specify the number of slabs to drain in tofree.
2469 *
2470 * Returns the actual number of slabs released.
2471 */
2472static int drain_freelist(struct kmem_cache *cache,
2473 struct kmem_list3 *l3, int tofree)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002474{
Christoph Lametered11d9e2006-06-30 01:55:45 -07002475 struct list_head *p;
2476 int nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002477 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002478
Christoph Lametered11d9e2006-06-30 01:55:45 -07002479 nr_freed = 0;
2480 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002481
Christoph Lametered11d9e2006-06-30 01:55:45 -07002482 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002483 p = l3->slabs_free.prev;
Christoph Lametered11d9e2006-06-30 01:55:45 -07002484 if (p == &l3->slabs_free) {
2485 spin_unlock_irq(&l3->list_lock);
2486 goto out;
2487 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002488
Christoph Lametered11d9e2006-06-30 01:55:45 -07002489 slabp = list_entry(p, struct slab, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002490#if DEBUG
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002491 BUG_ON(slabp->inuse);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002492#endif
2493 list_del(&slabp->list);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002494 /*
2495 * Safe to drop the lock. The slab is no longer linked
2496 * to the cache.
2497 */
2498 l3->free_objects -= cache->num;
Christoph Lametere498be72005-09-09 13:03:32 -07002499 spin_unlock_irq(&l3->list_lock);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002500 slab_destroy(cache, slabp);
2501 nr_freed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002502 }
Christoph Lametered11d9e2006-06-30 01:55:45 -07002503out:
2504 return nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002505}
2506
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002507/* Called with cache_chain_mutex held to protect against cpu hotplug */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002508static int __cache_shrink(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07002509{
2510 int ret = 0, i = 0;
2511 struct kmem_list3 *l3;
2512
2513 drain_cpu_caches(cachep);
2514
2515 check_irq_on();
2516 for_each_online_node(i) {
2517 l3 = cachep->nodelists[i];
Christoph Lametered11d9e2006-06-30 01:55:45 -07002518 if (!l3)
2519 continue;
2520
2521 drain_freelist(cachep, l3, l3->free_objects);
2522
2523 ret += !list_empty(&l3->slabs_full) ||
2524 !list_empty(&l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07002525 }
2526 return (ret ? 1 : 0);
2527}
2528
Linus Torvalds1da177e2005-04-16 15:20:36 -07002529/**
2530 * kmem_cache_shrink - Shrink a cache.
2531 * @cachep: The cache to shrink.
2532 *
2533 * Releases as many slabs as possible for a cache.
2534 * To help debugging, a zero exit status indicates all slabs were released.
2535 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002536int kmem_cache_shrink(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002537{
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002538 int ret;
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002539 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002540
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002541 get_online_cpus();
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002542 mutex_lock(&cache_chain_mutex);
2543 ret = __cache_shrink(cachep);
2544 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002545 put_online_cpus();
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002546 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002547}
2548EXPORT_SYMBOL(kmem_cache_shrink);
2549
2550/**
2551 * kmem_cache_destroy - delete a cache
2552 * @cachep: the cache to destroy
2553 *
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08002554 * Remove a &struct kmem_cache object from the slab cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002555 *
2556 * It is expected this function will be called by a module when it is
2557 * unloaded. This will remove the cache completely, and avoid a duplicate
2558 * cache being allocated each time a module is loaded and unloaded, if the
2559 * module doesn't have persistent in-kernel storage across loads and unloads.
2560 *
2561 * The cache must be empty before calling this function.
2562 *
2563 * The caller must guarantee that noone will allocate memory from the cache
2564 * during the kmem_cache_destroy().
2565 */
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002566void kmem_cache_destroy(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002567{
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002568 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002569
Linus Torvalds1da177e2005-04-16 15:20:36 -07002570 /* Find the cache in the chain of caches. */
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002571 get_online_cpus();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002572 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002573 /*
2574 * the chain is never empty, cache_cache is never destroyed
2575 */
2576 list_del(&cachep->next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002577 if (__cache_shrink(cachep)) {
2578 slab_error(cachep, "Can't free all objects");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002579 list_add(&cachep->next, &cache_chain);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002580 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002581 put_online_cpus();
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002582 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002583 }
2584
2585 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
Paul E. McKenneyfbd568a3e2005-05-01 08:59:04 -07002586 synchronize_rcu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002587
Christoph Lameter117f6eb2006-09-25 23:31:37 -07002588 __kmem_cache_destroy(cachep);
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002589 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002590 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002591}
2592EXPORT_SYMBOL(kmem_cache_destroy);
2593
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002594/*
2595 * Get the memory for a slab management obj.
2596 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2597 * always come from malloc_sizes caches. The slab descriptor cannot
2598 * come from the same cache which is getting created because,
2599 * when we are searching for an appropriate cache for these
2600 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2601 * If we are creating a malloc_sizes cache here it would not be visible to
2602 * kmem_find_general_cachep till the initialization is complete.
2603 * Hence we cannot have slabp_cache same as the original cache.
2604 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002605static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002606 int colour_off, gfp_t local_flags,
2607 int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002608{
2609 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002610
Linus Torvalds1da177e2005-04-16 15:20:36 -07002611 if (OFF_SLAB(cachep)) {
2612 /* Slab management obj is off-slab. */
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002613 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
Pekka Enberg8759ec52008-11-26 10:01:31 +02002614 local_flags, nodeid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002615 if (!slabp)
2616 return NULL;
2617 } else {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002618 slabp = objp + colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002619 colour_off += cachep->slab_size;
2620 }
2621 slabp->inuse = 0;
2622 slabp->colouroff = colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002623 slabp->s_mem = objp + colour_off;
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002624 slabp->nodeid = nodeid;
Marcin Slusarze51bfd02008-02-10 11:21:54 +01002625 slabp->free = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002626 return slabp;
2627}
2628
2629static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2630{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002631 return (kmem_bufctl_t *) (slabp + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002632}
2633
Pekka Enberg343e0d72006-02-01 03:05:50 -08002634static void cache_init_objs(struct kmem_cache *cachep,
Christoph Lametera35afb82007-05-16 22:10:57 -07002635 struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002636{
2637 int i;
2638
2639 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002640 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002641#if DEBUG
2642 /* need to poison the objs? */
2643 if (cachep->flags & SLAB_POISON)
2644 poison_obj(cachep, objp, POISON_FREE);
2645 if (cachep->flags & SLAB_STORE_USER)
2646 *dbg_userword(cachep, objp) = NULL;
2647
2648 if (cachep->flags & SLAB_RED_ZONE) {
2649 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2650 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2651 }
2652 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002653 * Constructors are not allowed to allocate memory from the same
2654 * cache which they are a constructor for. Otherwise, deadlock.
2655 * They must also be threaded.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002656 */
2657 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002658 cachep->ctor(objp + obj_offset(cachep));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002659
2660 if (cachep->flags & SLAB_RED_ZONE) {
2661 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2662 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002663 " end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002664 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2665 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002666 " start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002667 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08002668 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2669 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002670 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002671 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002672#else
2673 if (cachep->ctor)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002674 cachep->ctor(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002675#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002676 slab_bufctl(slabp)[i] = i + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002677 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002678 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002679}
2680
Pekka Enberg343e0d72006-02-01 03:05:50 -08002681static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002682{
Christoph Lameter4b51d662007-02-10 01:43:10 -08002683 if (CONFIG_ZONE_DMA_FLAG) {
2684 if (flags & GFP_DMA)
2685 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2686 else
2687 BUG_ON(cachep->gfpflags & GFP_DMA);
2688 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002689}
2690
Andrew Mortona737b3e2006-03-22 00:08:11 -08002691static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2692 int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002693{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002694 void *objp = index_to_obj(cachep, slabp, slabp->free);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002695 kmem_bufctl_t next;
2696
2697 slabp->inuse++;
2698 next = slab_bufctl(slabp)[slabp->free];
2699#if DEBUG
2700 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2701 WARN_ON(slabp->nodeid != nodeid);
2702#endif
2703 slabp->free = next;
2704
2705 return objp;
2706}
2707
Andrew Mortona737b3e2006-03-22 00:08:11 -08002708static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2709 void *objp, int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002710{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002711 unsigned int objnr = obj_to_index(cachep, slabp, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002712
2713#if DEBUG
2714 /* Verify that the slab belongs to the intended node */
2715 WARN_ON(slabp->nodeid != nodeid);
2716
Al Viro871751e2006-03-25 03:06:39 -08002717 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
Matthew Dobson78d382d2006-02-01 03:05:47 -08002718 printk(KERN_ERR "slab: double free detected in cache "
Andrew Mortona737b3e2006-03-22 00:08:11 -08002719 "'%s', objp %p\n", cachep->name, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002720 BUG();
2721 }
2722#endif
2723 slab_bufctl(slabp)[objnr] = slabp->free;
2724 slabp->free = objnr;
2725 slabp->inuse--;
2726}
2727
Pekka Enberg47768742006-06-23 02:03:07 -07002728/*
2729 * Map pages beginning at addr to the given cache and slab. This is required
2730 * for the slab allocator to be able to lookup the cache and slab of a
2731 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2732 */
2733static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2734 void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002735{
Pekka Enberg47768742006-06-23 02:03:07 -07002736 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002737 struct page *page;
2738
Pekka Enberg47768742006-06-23 02:03:07 -07002739 page = virt_to_page(addr);
Nick Piggin84097512006-03-22 00:08:34 -08002740
Pekka Enberg47768742006-06-23 02:03:07 -07002741 nr_pages = 1;
Nick Piggin84097512006-03-22 00:08:34 -08002742 if (likely(!PageCompound(page)))
Pekka Enberg47768742006-06-23 02:03:07 -07002743 nr_pages <<= cache->gfporder;
2744
Linus Torvalds1da177e2005-04-16 15:20:36 -07002745 do {
Pekka Enberg47768742006-06-23 02:03:07 -07002746 page_set_cache(page, cache);
2747 page_set_slab(page, slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002748 page++;
Pekka Enberg47768742006-06-23 02:03:07 -07002749 } while (--nr_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002750}
2751
2752/*
2753 * Grow (by 1) the number of slabs within a cache. This is called by
2754 * kmem_cache_alloc() when there are no active objs left in a cache.
2755 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002756static int cache_grow(struct kmem_cache *cachep,
2757 gfp_t flags, int nodeid, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002758{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002759 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002760 size_t offset;
2761 gfp_t local_flags;
Christoph Lametere498be72005-09-09 13:03:32 -07002762 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002763
Andrew Mortona737b3e2006-03-22 00:08:11 -08002764 /*
2765 * Be lazy and only check for valid flags here, keeping it out of the
2766 * critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002767 */
Christoph Lameter6cb06222007-10-16 01:25:41 -07002768 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2769 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002770
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002771 /* Take the l3 list lock to change the colour_next on this node */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002772 check_irq_off();
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002773 l3 = cachep->nodelists[nodeid];
2774 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002775
2776 /* Get colour for the slab, and cal the next value. */
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002777 offset = l3->colour_next;
2778 l3->colour_next++;
2779 if (l3->colour_next >= cachep->colour)
2780 l3->colour_next = 0;
2781 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002782
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002783 offset *= cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002784
2785 if (local_flags & __GFP_WAIT)
2786 local_irq_enable();
2787
2788 /*
2789 * The test for missing atomic flag is performed here, rather than
2790 * the more obvious place, simply to reduce the critical path length
2791 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2792 * will eventually be caught here (where it matters).
2793 */
2794 kmem_flagcheck(cachep, flags);
2795
Andrew Mortona737b3e2006-03-22 00:08:11 -08002796 /*
2797 * Get mem for the objs. Attempt to allocate a physical page from
2798 * 'nodeid'.
Christoph Lametere498be72005-09-09 13:03:32 -07002799 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002800 if (!objp)
Andrew Mortonb8c1c5d2007-07-24 12:02:40 -07002801 objp = kmem_getpages(cachep, local_flags, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002802 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002803 goto failed;
2804
2805 /* Get slab management. */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002806 slabp = alloc_slabmgmt(cachep, objp, offset,
Christoph Lameter6cb06222007-10-16 01:25:41 -07002807 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002808 if (!slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002809 goto opps1;
2810
Pekka Enberg47768742006-06-23 02:03:07 -07002811 slab_map_pages(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002812
Christoph Lametera35afb82007-05-16 22:10:57 -07002813 cache_init_objs(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002814
2815 if (local_flags & __GFP_WAIT)
2816 local_irq_disable();
2817 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002818 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002819
2820 /* Make slab active. */
Christoph Lametere498be72005-09-09 13:03:32 -07002821 list_add_tail(&slabp->list, &(l3->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002822 STATS_INC_GROWN(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002823 l3->free_objects += cachep->num;
2824 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002825 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002826opps1:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002827 kmem_freepages(cachep, objp);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002828failed:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002829 if (local_flags & __GFP_WAIT)
2830 local_irq_disable();
2831 return 0;
2832}
2833
2834#if DEBUG
2835
2836/*
2837 * Perform extra freeing checks:
2838 * - detect bad pointers.
2839 * - POISON/RED_ZONE checking
Linus Torvalds1da177e2005-04-16 15:20:36 -07002840 */
2841static void kfree_debugcheck(const void *objp)
2842{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002843 if (!virt_addr_valid(objp)) {
2844 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002845 (unsigned long)objp);
2846 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002847 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002848}
2849
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002850static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2851{
David Woodhouseb46b8f12007-05-08 00:22:59 -07002852 unsigned long long redzone1, redzone2;
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002853
2854 redzone1 = *dbg_redzone1(cache, obj);
2855 redzone2 = *dbg_redzone2(cache, obj);
2856
2857 /*
2858 * Redzone is ok.
2859 */
2860 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2861 return;
2862
2863 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2864 slab_error(cache, "double free detected");
2865 else
2866 slab_error(cache, "memory outside object was overwritten");
2867
David Woodhouseb46b8f12007-05-08 00:22:59 -07002868 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002869 obj, redzone1, redzone2);
2870}
2871
Pekka Enberg343e0d72006-02-01 03:05:50 -08002872static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002873 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002874{
2875 struct page *page;
2876 unsigned int objnr;
2877 struct slab *slabp;
2878
Matthew Wilcox80cbd912007-11-29 12:05:13 -07002879 BUG_ON(virt_to_cache(objp) != cachep);
2880
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002881 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002882 kfree_debugcheck(objp);
Christoph Lameterb49af682007-05-06 14:49:41 -07002883 page = virt_to_head_page(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002884
Pekka Enberg065d41c2005-11-13 16:06:46 -08002885 slabp = page_get_slab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002886
2887 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002888 verify_redzone_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002889 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2890 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2891 }
2892 if (cachep->flags & SLAB_STORE_USER)
2893 *dbg_userword(cachep, objp) = caller;
2894
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002895 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002896
2897 BUG_ON(objnr >= cachep->num);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002898 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002899
Al Viro871751e2006-03-25 03:06:39 -08002900#ifdef CONFIG_DEBUG_SLAB_LEAK
2901 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2902#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002903 if (cachep->flags & SLAB_POISON) {
2904#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08002905 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002906 store_stackinfo(cachep, objp, (unsigned long)caller);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002907 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002908 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002909 } else {
2910 poison_obj(cachep, objp, POISON_FREE);
2911 }
2912#else
2913 poison_obj(cachep, objp, POISON_FREE);
2914#endif
2915 }
2916 return objp;
2917}
2918
Pekka Enberg343e0d72006-02-01 03:05:50 -08002919static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002920{
2921 kmem_bufctl_t i;
2922 int entries = 0;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002923
Linus Torvalds1da177e2005-04-16 15:20:36 -07002924 /* Check slab's freelist to see if this obj is there. */
2925 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2926 entries++;
2927 if (entries > cachep->num || i >= cachep->num)
2928 goto bad;
2929 }
2930 if (entries != cachep->num - slabp->inuse) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002931bad:
2932 printk(KERN_ERR "slab: Internal list corruption detected in "
2933 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2934 cachep->name, cachep->num, slabp, slabp->inuse);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002935 for (i = 0;
Linus Torvalds264132b2006-03-06 12:10:07 -08002936 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002937 i++) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002938 if (i % 16 == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002939 printk("\n%03x:", i);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002940 printk(" %02x", ((unsigned char *)slabp)[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002941 }
2942 printk("\n");
2943 BUG();
2944 }
2945}
2946#else
2947#define kfree_debugcheck(x) do { } while(0)
2948#define cache_free_debugcheck(x,objp,z) (objp)
2949#define check_slabp(x,y) do { } while(0)
2950#endif
2951
Pekka Enberg343e0d72006-02-01 03:05:50 -08002952static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002953{
2954 int batchcount;
2955 struct kmem_list3 *l3;
2956 struct array_cache *ac;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002957 int node;
2958
Andrew Mortona737b3e2006-03-22 00:08:11 -08002959retry:
Joe Korty6d2144d2008-03-05 15:04:59 -08002960 check_irq_off();
2961 node = numa_node_id();
2962 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002963 batchcount = ac->batchcount;
2964 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002965 /*
2966 * If there was little recent activity on this cache, then
2967 * perform only a partial refill. Otherwise we could generate
2968 * refill bouncing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002969 */
2970 batchcount = BATCHREFILL_LIMIT;
2971 }
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002972 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002973
Christoph Lametere498be72005-09-09 13:03:32 -07002974 BUG_ON(ac->avail > 0 || !l3);
2975 spin_lock(&l3->list_lock);
2976
Christoph Lameter3ded1752006-03-25 03:06:44 -08002977 /* See if we can refill from the shared array */
2978 if (l3->shared && transfer_objects(ac, l3->shared, batchcount))
2979 goto alloc_done;
2980
Linus Torvalds1da177e2005-04-16 15:20:36 -07002981 while (batchcount > 0) {
2982 struct list_head *entry;
2983 struct slab *slabp;
2984 /* Get slab alloc is to come from. */
2985 entry = l3->slabs_partial.next;
2986 if (entry == &l3->slabs_partial) {
2987 l3->free_touched = 1;
2988 entry = l3->slabs_free.next;
2989 if (entry == &l3->slabs_free)
2990 goto must_grow;
2991 }
2992
2993 slabp = list_entry(entry, struct slab, list);
2994 check_slabp(cachep, slabp);
2995 check_spinlock_acquired(cachep);
Pekka Enberg714b81712007-05-06 14:49:03 -07002996
2997 /*
2998 * The slab was either on partial or free list so
2999 * there must be at least one object available for
3000 * allocation.
3001 */
roel kluin249b9f32008-10-29 17:18:07 -04003002 BUG_ON(slabp->inuse >= cachep->num);
Pekka Enberg714b81712007-05-06 14:49:03 -07003003
Linus Torvalds1da177e2005-04-16 15:20:36 -07003004 while (slabp->inuse < cachep->num && batchcount--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003005 STATS_INC_ALLOCED(cachep);
3006 STATS_INC_ACTIVE(cachep);
3007 STATS_SET_HIGH(cachep);
3008
Matthew Dobson78d382d2006-02-01 03:05:47 -08003009 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07003010 node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003011 }
3012 check_slabp(cachep, slabp);
3013
3014 /* move slabp to correct slabp list: */
3015 list_del(&slabp->list);
3016 if (slabp->free == BUFCTL_END)
3017 list_add(&slabp->list, &l3->slabs_full);
3018 else
3019 list_add(&slabp->list, &l3->slabs_partial);
3020 }
3021
Andrew Mortona737b3e2006-03-22 00:08:11 -08003022must_grow:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003023 l3->free_objects -= ac->avail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003024alloc_done:
Christoph Lametere498be72005-09-09 13:03:32 -07003025 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003026
3027 if (unlikely(!ac->avail)) {
3028 int x;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003029 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07003030
Andrew Mortona737b3e2006-03-22 00:08:11 -08003031 /* cache_grow can reenable interrupts, then ac could change. */
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003032 ac = cpu_cache_get(cachep);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003033 if (!x && ac->avail == 0) /* no objects in sight? abort */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003034 return NULL;
3035
Andrew Mortona737b3e2006-03-22 00:08:11 -08003036 if (!ac->avail) /* objects refilled by interrupt? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003037 goto retry;
3038 }
3039 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07003040 return ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003041}
3042
Andrew Mortona737b3e2006-03-22 00:08:11 -08003043static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3044 gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003045{
3046 might_sleep_if(flags & __GFP_WAIT);
3047#if DEBUG
3048 kmem_flagcheck(cachep, flags);
3049#endif
3050}
3051
3052#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003053static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3054 gfp_t flags, void *objp, void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003055{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003056 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003057 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003058 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003059#ifdef CONFIG_DEBUG_PAGEALLOC
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003060 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003061 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003062 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003063 else
3064 check_poison_obj(cachep, objp);
3065#else
3066 check_poison_obj(cachep, objp);
3067#endif
3068 poison_obj(cachep, objp, POISON_INUSE);
3069 }
3070 if (cachep->flags & SLAB_STORE_USER)
3071 *dbg_userword(cachep, objp) = caller;
3072
3073 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003074 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3075 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3076 slab_error(cachep, "double free, or memory outside"
3077 " object was overwritten");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003078 printk(KERN_ERR
David Woodhouseb46b8f12007-05-08 00:22:59 -07003079 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08003080 objp, *dbg_redzone1(cachep, objp),
3081 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003082 }
3083 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3084 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3085 }
Al Viro871751e2006-03-25 03:06:39 -08003086#ifdef CONFIG_DEBUG_SLAB_LEAK
3087 {
3088 struct slab *slabp;
3089 unsigned objnr;
3090
Christoph Lameterb49af682007-05-06 14:49:41 -07003091 slabp = page_get_slab(virt_to_head_page(objp));
Al Viro871751e2006-03-25 03:06:39 -08003092 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3093 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3094 }
3095#endif
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003096 objp += obj_offset(cachep);
Christoph Lameter4f104932007-05-06 14:50:17 -07003097 if (cachep->ctor && cachep->flags & SLAB_POISON)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07003098 cachep->ctor(objp);
Kevin Hilmana44b56d2006-12-06 20:32:11 -08003099#if ARCH_SLAB_MINALIGN
3100 if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) {
3101 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3102 objp, ARCH_SLAB_MINALIGN);
3103 }
3104#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003105 return objp;
3106}
3107#else
3108#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3109#endif
3110
Akinobu Mita773ff602008-12-23 19:37:01 +09003111static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003112{
3113 if (cachep == &cache_cache)
Akinobu Mita773ff602008-12-23 19:37:01 +09003114 return false;
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003115
Akinobu Mita773ff602008-12-23 19:37:01 +09003116 return should_failslab(obj_size(cachep), flags);
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003117}
3118
Pekka Enberg343e0d72006-02-01 03:05:50 -08003119static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003120{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003121 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003122 struct array_cache *ac;
3123
Alok N Kataria5c382302005-09-27 21:45:46 -07003124 check_irq_off();
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003125
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003126 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003127 if (likely(ac->avail)) {
3128 STATS_INC_ALLOCHIT(cachep);
3129 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07003130 objp = ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003131 } else {
3132 STATS_INC_ALLOCMISS(cachep);
3133 objp = cache_alloc_refill(cachep, flags);
3134 }
Alok N Kataria5c382302005-09-27 21:45:46 -07003135 return objp;
3136}
3137
Christoph Lametere498be72005-09-09 13:03:32 -07003138#ifdef CONFIG_NUMA
3139/*
Paul Jacksonb2455392006-03-24 03:16:12 -08003140 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
Paul Jacksonc61afb12006-03-24 03:16:08 -08003141 *
3142 * If we are in_interrupt, then process context, including cpusets and
3143 * mempolicy, may not apply and should not be used for allocation policy.
3144 */
3145static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3146{
3147 int nid_alloc, nid_here;
3148
Christoph Lameter765c4502006-09-27 01:50:08 -07003149 if (in_interrupt() || (flags & __GFP_THISNODE))
Paul Jacksonc61afb12006-03-24 03:16:08 -08003150 return NULL;
3151 nid_alloc = nid_here = numa_node_id();
3152 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3153 nid_alloc = cpuset_mem_spread_node();
3154 else if (current->mempolicy)
3155 nid_alloc = slab_node(current->mempolicy);
3156 if (nid_alloc != nid_here)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003157 return ____cache_alloc_node(cachep, flags, nid_alloc);
Paul Jacksonc61afb12006-03-24 03:16:08 -08003158 return NULL;
3159}
3160
3161/*
Christoph Lameter765c4502006-09-27 01:50:08 -07003162 * Fallback function if there was no memory available and no objects on a
Christoph Lameter3c517a62006-12-06 20:33:29 -08003163 * certain node and fall back is permitted. First we scan all the
3164 * available nodelists for available objects. If that fails then we
3165 * perform an allocation without specifying a node. This allows the page
3166 * allocator to do its reclaim / fallback magic. We then insert the
3167 * slab into the proper nodelist and then allocate from it.
Christoph Lameter765c4502006-09-27 01:50:08 -07003168 */
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003169static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
Christoph Lameter765c4502006-09-27 01:50:08 -07003170{
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003171 struct zonelist *zonelist;
3172 gfp_t local_flags;
Mel Gormandd1a2392008-04-28 02:12:17 -07003173 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07003174 struct zone *zone;
3175 enum zone_type high_zoneidx = gfp_zone(flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003176 void *obj = NULL;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003177 int nid;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003178
3179 if (flags & __GFP_THISNODE)
3180 return NULL;
3181
Mel Gorman0e884602008-04-28 02:12:14 -07003182 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
Christoph Lameter6cb06222007-10-16 01:25:41 -07003183 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Christoph Lameter765c4502006-09-27 01:50:08 -07003184
Christoph Lameter3c517a62006-12-06 20:33:29 -08003185retry:
3186 /*
3187 * Look through allowed nodes for objects available
3188 * from existing per node queues.
3189 */
Mel Gorman54a6eb52008-04-28 02:12:16 -07003190 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3191 nid = zone_to_nid(zone);
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003192
Mel Gorman54a6eb52008-04-28 02:12:16 -07003193 if (cpuset_zone_allowed_hardwall(zone, flags) &&
Christoph Lameter3c517a62006-12-06 20:33:29 -08003194 cache->nodelists[nid] &&
Christoph Lameter481c5342008-06-21 16:46:35 -07003195 cache->nodelists[nid]->free_objects) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003196 obj = ____cache_alloc_node(cache,
3197 flags | GFP_THISNODE, nid);
Christoph Lameter481c5342008-06-21 16:46:35 -07003198 if (obj)
3199 break;
3200 }
Christoph Lameter3c517a62006-12-06 20:33:29 -08003201 }
3202
Christoph Lametercfce6602007-05-06 14:50:17 -07003203 if (!obj) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003204 /*
3205 * This allocation will be performed within the constraints
3206 * of the current cpuset / memory policy requirements.
3207 * We may trigger various forms of reclaim on the allowed
3208 * set and go into memory reserves if necessary.
3209 */
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003210 if (local_flags & __GFP_WAIT)
3211 local_irq_enable();
3212 kmem_flagcheck(cache, flags);
Christoph Lameter9ac33b22008-03-04 12:24:22 -08003213 obj = kmem_getpages(cache, local_flags, -1);
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003214 if (local_flags & __GFP_WAIT)
3215 local_irq_disable();
Christoph Lameter3c517a62006-12-06 20:33:29 -08003216 if (obj) {
3217 /*
3218 * Insert into the appropriate per node queues
3219 */
3220 nid = page_to_nid(virt_to_page(obj));
3221 if (cache_grow(cache, flags, nid, obj)) {
3222 obj = ____cache_alloc_node(cache,
3223 flags | GFP_THISNODE, nid);
3224 if (!obj)
3225 /*
3226 * Another processor may allocate the
3227 * objects in the slab since we are
3228 * not holding any locks.
3229 */
3230 goto retry;
3231 } else {
Hugh Dickinsb6a60452007-01-05 16:36:36 -08003232 /* cache_grow already freed obj */
Christoph Lameter3c517a62006-12-06 20:33:29 -08003233 obj = NULL;
3234 }
3235 }
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003236 }
Christoph Lameter765c4502006-09-27 01:50:08 -07003237 return obj;
3238}
3239
3240/*
Christoph Lametere498be72005-09-09 13:03:32 -07003241 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07003242 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003243static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003244 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07003245{
3246 struct list_head *entry;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003247 struct slab *slabp;
3248 struct kmem_list3 *l3;
3249 void *obj;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003250 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003251
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003252 l3 = cachep->nodelists[nodeid];
3253 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07003254
Andrew Mortona737b3e2006-03-22 00:08:11 -08003255retry:
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003256 check_irq_off();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003257 spin_lock(&l3->list_lock);
3258 entry = l3->slabs_partial.next;
3259 if (entry == &l3->slabs_partial) {
3260 l3->free_touched = 1;
3261 entry = l3->slabs_free.next;
3262 if (entry == &l3->slabs_free)
3263 goto must_grow;
3264 }
Christoph Lametere498be72005-09-09 13:03:32 -07003265
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003266 slabp = list_entry(entry, struct slab, list);
3267 check_spinlock_acquired_node(cachep, nodeid);
3268 check_slabp(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07003269
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003270 STATS_INC_NODEALLOCS(cachep);
3271 STATS_INC_ACTIVE(cachep);
3272 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003273
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003274 BUG_ON(slabp->inuse == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07003275
Matthew Dobson78d382d2006-02-01 03:05:47 -08003276 obj = slab_get_obj(cachep, slabp, nodeid);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003277 check_slabp(cachep, slabp);
3278 l3->free_objects--;
3279 /* move slabp to correct slabp list: */
3280 list_del(&slabp->list);
Christoph Lametere498be72005-09-09 13:03:32 -07003281
Andrew Mortona737b3e2006-03-22 00:08:11 -08003282 if (slabp->free == BUFCTL_END)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003283 list_add(&slabp->list, &l3->slabs_full);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003284 else
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003285 list_add(&slabp->list, &l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07003286
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003287 spin_unlock(&l3->list_lock);
3288 goto done;
Christoph Lametere498be72005-09-09 13:03:32 -07003289
Andrew Mortona737b3e2006-03-22 00:08:11 -08003290must_grow:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003291 spin_unlock(&l3->list_lock);
Christoph Lameter3c517a62006-12-06 20:33:29 -08003292 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
Christoph Lameter765c4502006-09-27 01:50:08 -07003293 if (x)
3294 goto retry;
Christoph Lametere498be72005-09-09 13:03:32 -07003295
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003296 return fallback_alloc(cachep, flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003297
Andrew Mortona737b3e2006-03-22 00:08:11 -08003298done:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003299 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07003300}
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003301
3302/**
3303 * kmem_cache_alloc_node - Allocate an object on the specified node
3304 * @cachep: The cache to allocate from.
3305 * @flags: See kmalloc().
3306 * @nodeid: node number of the target node.
3307 * @caller: return address of caller, used for debug information
3308 *
3309 * Identical to kmem_cache_alloc but it will allocate memory on the given
3310 * node, which can improve the performance for cpu bound structures.
3311 *
3312 * Fallback to other node is possible if __GFP_THISNODE is not set.
3313 */
3314static __always_inline void *
3315__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3316 void *caller)
3317{
3318 unsigned long save_flags;
3319 void *ptr;
3320
Akinobu Mita773ff602008-12-23 19:37:01 +09003321 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003322 return NULL;
3323
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003324 cache_alloc_debugcheck_before(cachep, flags);
3325 local_irq_save(save_flags);
3326
3327 if (unlikely(nodeid == -1))
3328 nodeid = numa_node_id();
3329
3330 if (unlikely(!cachep->nodelists[nodeid])) {
3331 /* Node not bootstrapped yet */
3332 ptr = fallback_alloc(cachep, flags);
3333 goto out;
3334 }
3335
3336 if (nodeid == numa_node_id()) {
3337 /*
3338 * Use the locally cached objects if possible.
3339 * However ____cache_alloc does not allow fallback
3340 * to other nodes. It may fail while we still have
3341 * objects on other nodes available.
3342 */
3343 ptr = ____cache_alloc(cachep, flags);
3344 if (ptr)
3345 goto out;
3346 }
3347 /* ___cache_alloc_node can fall back to other nodes */
3348 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3349 out:
3350 local_irq_restore(save_flags);
3351 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3352
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003353 if (unlikely((flags & __GFP_ZERO) && ptr))
3354 memset(ptr, 0, obj_size(cachep));
3355
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003356 return ptr;
3357}
3358
3359static __always_inline void *
3360__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3361{
3362 void *objp;
3363
3364 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3365 objp = alternate_node_alloc(cache, flags);
3366 if (objp)
3367 goto out;
3368 }
3369 objp = ____cache_alloc(cache, flags);
3370
3371 /*
3372 * We may just have run out of memory on the local node.
3373 * ____cache_alloc_node() knows how to locate memory on other nodes
3374 */
3375 if (!objp)
3376 objp = ____cache_alloc_node(cache, flags, numa_node_id());
3377
3378 out:
3379 return objp;
3380}
3381#else
3382
3383static __always_inline void *
3384__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3385{
3386 return ____cache_alloc(cachep, flags);
3387}
3388
3389#endif /* CONFIG_NUMA */
3390
3391static __always_inline void *
3392__cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3393{
3394 unsigned long save_flags;
3395 void *objp;
3396
Akinobu Mita773ff602008-12-23 19:37:01 +09003397 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003398 return NULL;
3399
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003400 cache_alloc_debugcheck_before(cachep, flags);
3401 local_irq_save(save_flags);
3402 objp = __do_cache_alloc(cachep, flags);
3403 local_irq_restore(save_flags);
3404 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3405 prefetchw(objp);
3406
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003407 if (unlikely((flags & __GFP_ZERO) && objp))
3408 memset(objp, 0, obj_size(cachep));
3409
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003410 return objp;
3411}
Christoph Lametere498be72005-09-09 13:03:32 -07003412
3413/*
3414 * Caller needs to acquire correct kmem_list's list_lock
3415 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003416static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003417 int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003418{
3419 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07003420 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003421
3422 for (i = 0; i < nr_objects; i++) {
3423 void *objp = objpp[i];
3424 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003425
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003426 slabp = virt_to_slab(objp);
Christoph Lameterff694162005-09-22 21:44:02 -07003427 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003428 list_del(&slabp->list);
Christoph Lameterff694162005-09-22 21:44:02 -07003429 check_spinlock_acquired_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003430 check_slabp(cachep, slabp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08003431 slab_put_obj(cachep, slabp, objp, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003432 STATS_DEC_ACTIVE(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003433 l3->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003434 check_slabp(cachep, slabp);
3435
3436 /* fixup slab chains */
3437 if (slabp->inuse == 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07003438 if (l3->free_objects > l3->free_limit) {
3439 l3->free_objects -= cachep->num;
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07003440 /* No need to drop any previously held
3441 * lock here, even if we have a off-slab slab
3442 * descriptor it is guaranteed to come from
3443 * a different cache, refer to comments before
3444 * alloc_slabmgmt.
3445 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003446 slab_destroy(cachep, slabp);
3447 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07003448 list_add(&slabp->list, &l3->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003449 }
3450 } else {
3451 /* Unconditionally move a slab to the end of the
3452 * partial list on free - maximum time for the
3453 * other objects to be freed, too.
3454 */
Christoph Lametere498be72005-09-09 13:03:32 -07003455 list_add_tail(&slabp->list, &l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003456 }
3457 }
3458}
3459
Pekka Enberg343e0d72006-02-01 03:05:50 -08003460static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003461{
3462 int batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07003463 struct kmem_list3 *l3;
Christoph Lameterff694162005-09-22 21:44:02 -07003464 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003465
3466 batchcount = ac->batchcount;
3467#if DEBUG
3468 BUG_ON(!batchcount || batchcount > ac->avail);
3469#endif
3470 check_irq_off();
Christoph Lameterff694162005-09-22 21:44:02 -07003471 l3 = cachep->nodelists[node];
Ingo Molnar873623d2006-07-13 14:44:38 +02003472 spin_lock(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003473 if (l3->shared) {
3474 struct array_cache *shared_array = l3->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003475 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003476 if (max) {
3477 if (batchcount > max)
3478 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07003479 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003480 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003481 shared_array->avail += batchcount;
3482 goto free_done;
3483 }
3484 }
3485
Christoph Lameterff694162005-09-22 21:44:02 -07003486 free_block(cachep, ac->entry, batchcount, node);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003487free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003488#if STATS
3489 {
3490 int i = 0;
3491 struct list_head *p;
3492
Christoph Lametere498be72005-09-09 13:03:32 -07003493 p = l3->slabs_free.next;
3494 while (p != &(l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003495 struct slab *slabp;
3496
3497 slabp = list_entry(p, struct slab, list);
3498 BUG_ON(slabp->inuse);
3499
3500 i++;
3501 p = p->next;
3502 }
3503 STATS_SET_FREEABLE(cachep, i);
3504 }
3505#endif
Christoph Lametere498be72005-09-09 13:03:32 -07003506 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003507 ac->avail -= batchcount;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003508 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003509}
3510
3511/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08003512 * Release an obj back to its cache. If the obj has a constructed state, it must
3513 * be in this state _before_ it is released. Called with disabled ints.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003514 */
Ingo Molnar873623d2006-07-13 14:44:38 +02003515static inline void __cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003516{
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003517 struct array_cache *ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003518
3519 check_irq_off();
3520 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3521
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -07003522 /*
3523 * Skip calling cache_free_alien() when the platform is not numa.
3524 * This will avoid cache misses that happen while accessing slabp (which
3525 * is per page memory reference) to get nodeid. Instead use a global
3526 * variable to skip the call, which is mostly likely to be present in
3527 * the cache.
3528 */
3529 if (numa_platform && cache_free_alien(cachep, objp))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07003530 return;
Christoph Lametere498be72005-09-09 13:03:32 -07003531
Linus Torvalds1da177e2005-04-16 15:20:36 -07003532 if (likely(ac->avail < ac->limit)) {
3533 STATS_INC_FREEHIT(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003534 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003535 return;
3536 } else {
3537 STATS_INC_FREEMISS(cachep);
3538 cache_flusharray(cachep, ac);
Christoph Lametere498be72005-09-09 13:03:32 -07003539 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003540 }
3541}
3542
3543/**
3544 * kmem_cache_alloc - Allocate an object
3545 * @cachep: The cache to allocate from.
3546 * @flags: See kmalloc().
3547 *
3548 * Allocate an object from this cache. The flags are only relevant
3549 * if the cache has no available objects.
3550 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003551void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003552{
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003553 return __cache_alloc(cachep, flags, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003554}
3555EXPORT_SYMBOL(kmem_cache_alloc);
3556
3557/**
Randy Dunlap76824862008-03-19 17:00:40 -07003558 * kmem_ptr_validate - check if an untrusted pointer might be a slab entry.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003559 * @cachep: the cache we're checking against
3560 * @ptr: pointer to validate
3561 *
Randy Dunlap76824862008-03-19 17:00:40 -07003562 * This verifies that the untrusted pointer looks sane;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003563 * it is _not_ a guarantee that the pointer is actually
3564 * part of the slab cache in question, but it at least
3565 * validates that the pointer can be dereferenced and
3566 * looks half-way sane.
3567 *
3568 * Currently only used for dentry validation.
3569 */
Christoph Lameterb7f869a22006-12-22 01:06:44 -08003570int kmem_ptr_validate(struct kmem_cache *cachep, const void *ptr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003571{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003572 unsigned long addr = (unsigned long)ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003573 unsigned long min_addr = PAGE_OFFSET;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003574 unsigned long align_mask = BYTES_PER_WORD - 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003575 unsigned long size = cachep->buffer_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003576 struct page *page;
3577
3578 if (unlikely(addr < min_addr))
3579 goto out;
3580 if (unlikely(addr > (unsigned long)high_memory - size))
3581 goto out;
3582 if (unlikely(addr & align_mask))
3583 goto out;
3584 if (unlikely(!kern_addr_valid(addr)))
3585 goto out;
3586 if (unlikely(!kern_addr_valid(addr + size - 1)))
3587 goto out;
3588 page = virt_to_page(ptr);
3589 if (unlikely(!PageSlab(page)))
3590 goto out;
Pekka Enberg065d41c2005-11-13 16:06:46 -08003591 if (unlikely(page_get_cache(page) != cachep))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003592 goto out;
3593 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003594out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003595 return 0;
3596}
3597
3598#ifdef CONFIG_NUMA
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003599void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3600{
3601 return __cache_alloc_node(cachep, flags, nodeid,
3602 __builtin_return_address(0));
3603}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003604EXPORT_SYMBOL(kmem_cache_alloc_node);
3605
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003606static __always_inline void *
3607__do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003608{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003609 struct kmem_cache *cachep;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003610
3611 cachep = kmem_find_general_cachep(size, flags);
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003612 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3613 return cachep;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003614 return kmem_cache_alloc_node(cachep, flags, node);
3615}
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003616
3617#ifdef CONFIG_DEBUG_SLAB
3618void *__kmalloc_node(size_t size, gfp_t flags, int node)
3619{
3620 return __do_kmalloc_node(size, flags, node,
3621 __builtin_return_address(0));
3622}
Christoph Hellwigdbe5e692006-09-25 23:31:36 -07003623EXPORT_SYMBOL(__kmalloc_node);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003624
3625void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003626 int node, unsigned long caller)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003627{
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003628 return __do_kmalloc_node(size, flags, node, (void *)caller);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003629}
3630EXPORT_SYMBOL(__kmalloc_node_track_caller);
3631#else
3632void *__kmalloc_node(size_t size, gfp_t flags, int node)
3633{
3634 return __do_kmalloc_node(size, flags, node, NULL);
3635}
3636EXPORT_SYMBOL(__kmalloc_node);
3637#endif /* CONFIG_DEBUG_SLAB */
3638#endif /* CONFIG_NUMA */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003639
3640/**
Paul Drynoff800590f2006-06-23 02:03:48 -07003641 * __do_kmalloc - allocate memory
Linus Torvalds1da177e2005-04-16 15:20:36 -07003642 * @size: how many bytes of memory are required.
Paul Drynoff800590f2006-06-23 02:03:48 -07003643 * @flags: the type of memory to allocate (see kmalloc).
Randy Dunlap911851e2006-03-22 00:08:14 -08003644 * @caller: function caller for debug tracking of the caller
Linus Torvalds1da177e2005-04-16 15:20:36 -07003645 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003646static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3647 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003648{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003649 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003650
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003651 /* If you want to save a few bytes .text space: replace
3652 * __ with kmem_.
3653 * Then kmalloc uses the uninlined functions instead of the inline
3654 * functions.
3655 */
3656 cachep = __find_general_cachep(size, flags);
Linus Torvaldsa5c96d82007-07-19 13:17:15 -07003657 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3658 return cachep;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003659 return __cache_alloc(cachep, flags, caller);
3660}
3661
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003662
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003663#ifdef CONFIG_DEBUG_SLAB
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003664void *__kmalloc(size_t size, gfp_t flags)
3665{
Al Viro871751e2006-03-25 03:06:39 -08003666 return __do_kmalloc(size, flags, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003667}
3668EXPORT_SYMBOL(__kmalloc);
3669
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003670void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003671{
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003672 return __do_kmalloc(size, flags, (void *)caller);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003673}
3674EXPORT_SYMBOL(__kmalloc_track_caller);
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003675
3676#else
3677void *__kmalloc(size_t size, gfp_t flags)
3678{
3679 return __do_kmalloc(size, flags, NULL);
3680}
3681EXPORT_SYMBOL(__kmalloc);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003682#endif
3683
Linus Torvalds1da177e2005-04-16 15:20:36 -07003684/**
3685 * kmem_cache_free - Deallocate an object
3686 * @cachep: The cache the allocation was from.
3687 * @objp: The previously allocated object.
3688 *
3689 * Free an object which was previously allocated from this
3690 * cache.
3691 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003692void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003693{
3694 unsigned long flags;
3695
3696 local_irq_save(flags);
Ingo Molnar898552c2007-02-10 01:44:57 -08003697 debug_check_no_locks_freed(objp, obj_size(cachep));
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07003698 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3699 debug_check_no_obj_freed(objp, obj_size(cachep));
Ingo Molnar873623d2006-07-13 14:44:38 +02003700 __cache_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003701 local_irq_restore(flags);
3702}
3703EXPORT_SYMBOL(kmem_cache_free);
3704
3705/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003706 * kfree - free previously allocated memory
3707 * @objp: pointer returned by kmalloc.
3708 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003709 * If @objp is NULL, no operation is performed.
3710 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003711 * Don't free memory not originally allocated by kmalloc()
3712 * or you will run into trouble.
3713 */
3714void kfree(const void *objp)
3715{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003716 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003717 unsigned long flags;
3718
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003719 if (unlikely(ZERO_OR_NULL_PTR(objp)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003720 return;
3721 local_irq_save(flags);
3722 kfree_debugcheck(objp);
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003723 c = virt_to_cache(objp);
Ingo Molnarf9b84042006-06-27 02:54:49 -07003724 debug_check_no_locks_freed(objp, obj_size(c));
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07003725 debug_check_no_obj_freed(objp, obj_size(c));
Ingo Molnar873623d2006-07-13 14:44:38 +02003726 __cache_free(c, (void *)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003727 local_irq_restore(flags);
3728}
3729EXPORT_SYMBOL(kfree);
3730
Pekka Enberg343e0d72006-02-01 03:05:50 -08003731unsigned int kmem_cache_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003732{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003733 return obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003734}
3735EXPORT_SYMBOL(kmem_cache_size);
3736
Pekka Enberg343e0d72006-02-01 03:05:50 -08003737const char *kmem_cache_name(struct kmem_cache *cachep)
Arnaldo Carvalho de Melo19449722005-06-18 22:46:19 -07003738{
3739 return cachep->name;
3740}
3741EXPORT_SYMBOL_GPL(kmem_cache_name);
3742
Christoph Lametere498be72005-09-09 13:03:32 -07003743/*
Simon Arlott183ff222007-10-20 01:27:18 +02003744 * This initializes kmem_list3 or resizes various caches for all nodes.
Christoph Lametere498be72005-09-09 13:03:32 -07003745 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003746static int alloc_kmemlist(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07003747{
3748 int node;
3749 struct kmem_list3 *l3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003750 struct array_cache *new_shared;
Paul Menage3395ee02006-12-06 20:32:16 -08003751 struct array_cache **new_alien = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003752
Mel Gorman9c09a952008-01-24 05:49:54 -08003753 for_each_online_node(node) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003754
Paul Menage3395ee02006-12-06 20:32:16 -08003755 if (use_alien_caches) {
3756 new_alien = alloc_alien_cache(node, cachep->limit);
3757 if (!new_alien)
3758 goto fail;
3759 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003760
Eric Dumazet63109842007-05-06 14:49:28 -07003761 new_shared = NULL;
3762 if (cachep->shared) {
3763 new_shared = alloc_arraycache(node,
Christoph Lameter0718dc22006-03-25 03:06:47 -08003764 cachep->shared*cachep->batchcount,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003765 0xbaadf00d);
Eric Dumazet63109842007-05-06 14:49:28 -07003766 if (!new_shared) {
3767 free_alien_cache(new_alien);
3768 goto fail;
3769 }
Christoph Lameter0718dc22006-03-25 03:06:47 -08003770 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003771
Andrew Mortona737b3e2006-03-22 00:08:11 -08003772 l3 = cachep->nodelists[node];
3773 if (l3) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003774 struct array_cache *shared = l3->shared;
3775
Christoph Lametere498be72005-09-09 13:03:32 -07003776 spin_lock_irq(&l3->list_lock);
3777
Christoph Lametercafeb022006-03-25 03:06:46 -08003778 if (shared)
Christoph Lameter0718dc22006-03-25 03:06:47 -08003779 free_block(cachep, shared->entry,
3780 shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07003781
Christoph Lametercafeb022006-03-25 03:06:46 -08003782 l3->shared = new_shared;
3783 if (!l3->alien) {
Christoph Lametere498be72005-09-09 13:03:32 -07003784 l3->alien = new_alien;
3785 new_alien = NULL;
3786 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003787 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003788 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003789 spin_unlock_irq(&l3->list_lock);
Christoph Lametercafeb022006-03-25 03:06:46 -08003790 kfree(shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003791 free_alien_cache(new_alien);
3792 continue;
3793 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08003794 l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node);
Christoph Lameter0718dc22006-03-25 03:06:47 -08003795 if (!l3) {
3796 free_alien_cache(new_alien);
3797 kfree(new_shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003798 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003799 }
Christoph Lametere498be72005-09-09 13:03:32 -07003800
3801 kmem_list3_init(l3);
3802 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Andrew Mortona737b3e2006-03-22 00:08:11 -08003803 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003804 l3->shared = new_shared;
Christoph Lametere498be72005-09-09 13:03:32 -07003805 l3->alien = new_alien;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003806 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003807 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003808 cachep->nodelists[node] = l3;
3809 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003810 return 0;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003811
Andrew Mortona737b3e2006-03-22 00:08:11 -08003812fail:
Christoph Lameter0718dc22006-03-25 03:06:47 -08003813 if (!cachep->next.next) {
3814 /* Cache is not active yet. Roll back what we did */
3815 node--;
3816 while (node >= 0) {
3817 if (cachep->nodelists[node]) {
3818 l3 = cachep->nodelists[node];
3819
3820 kfree(l3->shared);
3821 free_alien_cache(l3->alien);
3822 kfree(l3);
3823 cachep->nodelists[node] = NULL;
3824 }
3825 node--;
3826 }
3827 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003828 return -ENOMEM;
Christoph Lametere498be72005-09-09 13:03:32 -07003829}
3830
Linus Torvalds1da177e2005-04-16 15:20:36 -07003831struct ccupdate_struct {
Pekka Enberg343e0d72006-02-01 03:05:50 -08003832 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003833 struct array_cache *new[NR_CPUS];
3834};
3835
3836static void do_ccupdate_local(void *info)
3837{
Andrew Mortona737b3e2006-03-22 00:08:11 -08003838 struct ccupdate_struct *new = info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003839 struct array_cache *old;
3840
3841 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003842 old = cpu_cache_get(new->cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003843
Linus Torvalds1da177e2005-04-16 15:20:36 -07003844 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3845 new->new[smp_processor_id()] = old;
3846}
3847
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003848/* Always called with the cache_chain_mutex held */
Andrew Mortona737b3e2006-03-22 00:08:11 -08003849static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3850 int batchcount, int shared)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003851{
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003852 struct ccupdate_struct *new;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003853 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003854
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003855 new = kzalloc(sizeof(*new), GFP_KERNEL);
3856 if (!new)
3857 return -ENOMEM;
3858
Christoph Lametere498be72005-09-09 13:03:32 -07003859 for_each_online_cpu(i) {
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003860 new->new[i] = alloc_arraycache(cpu_to_node(i), limit,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003861 batchcount);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003862 if (!new->new[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003863 for (i--; i >= 0; i--)
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003864 kfree(new->new[i]);
3865 kfree(new);
Christoph Lametere498be72005-09-09 13:03:32 -07003866 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003867 }
3868 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003869 new->cachep = cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003870
Jens Axboe15c8b6c2008-05-09 09:39:44 +02003871 on_each_cpu(do_ccupdate_local, (void *)new, 1);
Christoph Lametere498be72005-09-09 13:03:32 -07003872
Linus Torvalds1da177e2005-04-16 15:20:36 -07003873 check_irq_on();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003874 cachep->batchcount = batchcount;
3875 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07003876 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003877
Christoph Lametere498be72005-09-09 13:03:32 -07003878 for_each_online_cpu(i) {
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003879 struct array_cache *ccold = new->new[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003880 if (!ccold)
3881 continue;
Christoph Lametere498be72005-09-09 13:03:32 -07003882 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -07003883 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
Christoph Lametere498be72005-09-09 13:03:32 -07003884 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003885 kfree(ccold);
3886 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003887 kfree(new);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003888 return alloc_kmemlist(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003889}
3890
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003891/* Called with cache_chain_mutex held always */
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003892static int enable_cpucache(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003893{
3894 int err;
3895 int limit, shared;
3896
Andrew Mortona737b3e2006-03-22 00:08:11 -08003897 /*
3898 * The head array serves three purposes:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003899 * - create a LIFO ordering, i.e. return objects that are cache-warm
3900 * - reduce the number of spinlock operations.
Andrew Mortona737b3e2006-03-22 00:08:11 -08003901 * - reduce the number of linked list operations on the slab and
Linus Torvalds1da177e2005-04-16 15:20:36 -07003902 * bufctl chains: array operations are cheaper.
3903 * The numbers are guessed, we should auto-tune as described by
3904 * Bonwick.
3905 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003906 if (cachep->buffer_size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003907 limit = 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003908 else if (cachep->buffer_size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003909 limit = 8;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003910 else if (cachep->buffer_size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003911 limit = 24;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003912 else if (cachep->buffer_size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003913 limit = 54;
3914 else
3915 limit = 120;
3916
Andrew Mortona737b3e2006-03-22 00:08:11 -08003917 /*
3918 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
Linus Torvalds1da177e2005-04-16 15:20:36 -07003919 * allocation behaviour: Most allocs on one cpu, most free operations
3920 * on another cpu. For these cases, an efficient object passing between
3921 * cpus is necessary. This is provided by a shared array. The array
3922 * replaces Bonwick's magazine layer.
3923 * On uniprocessor, it's functionally equivalent (but less efficient)
3924 * to a larger limit. Thus disabled by default.
3925 */
3926 shared = 0;
Eric Dumazet364fbb22007-05-06 14:49:27 -07003927 if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003928 shared = 8;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003929
3930#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003931 /*
3932 * With debugging enabled, large batchcount lead to excessively long
3933 * periods with disabled local interrupts. Limit the batchcount
Linus Torvalds1da177e2005-04-16 15:20:36 -07003934 */
3935 if (limit > 32)
3936 limit = 32;
3937#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003938 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003939 if (err)
3940 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003941 cachep->name, -err);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003942 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003943}
3944
Christoph Lameter1b552532006-03-22 00:09:07 -08003945/*
3946 * Drain an array if it contains any elements taking the l3 lock only if
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003947 * necessary. Note that the l3 listlock also protects the array_cache
3948 * if drain_array() is used on the shared array.
Christoph Lameter1b552532006-03-22 00:09:07 -08003949 */
3950void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
3951 struct array_cache *ac, int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003952{
3953 int tofree;
3954
Christoph Lameter1b552532006-03-22 00:09:07 -08003955 if (!ac || !ac->avail)
3956 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003957 if (ac->touched && !force) {
3958 ac->touched = 0;
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003959 } else {
Christoph Lameter1b552532006-03-22 00:09:07 -08003960 spin_lock_irq(&l3->list_lock);
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003961 if (ac->avail) {
3962 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3963 if (tofree > ac->avail)
3964 tofree = (ac->avail + 1) / 2;
3965 free_block(cachep, ac->entry, tofree, node);
3966 ac->avail -= tofree;
3967 memmove(ac->entry, &(ac->entry[tofree]),
3968 sizeof(void *) * ac->avail);
3969 }
Christoph Lameter1b552532006-03-22 00:09:07 -08003970 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003971 }
3972}
3973
3974/**
3975 * cache_reap - Reclaim memory from caches.
Randy Dunlap05fb6bf2007-02-28 20:12:13 -08003976 * @w: work descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07003977 *
3978 * Called from workqueue/eventd every few seconds.
3979 * Purpose:
3980 * - clear the per-cpu caches for this CPU.
3981 * - return freeable pages to the main free memory pool.
3982 *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003983 * If we cannot acquire the cache chain mutex then just give up - we'll try
3984 * again on the next iteration.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003985 */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08003986static void cache_reap(struct work_struct *w)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003987{
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07003988 struct kmem_cache *searchp;
Christoph Lametere498be72005-09-09 13:03:32 -07003989 struct kmem_list3 *l3;
Christoph Lameteraab22072006-03-22 00:09:06 -08003990 int node = numa_node_id();
Christoph Lameter7c5cae32007-02-10 01:42:55 -08003991 struct delayed_work *work =
3992 container_of(w, struct delayed_work, work);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003993
Christoph Lameter7c5cae32007-02-10 01:42:55 -08003994 if (!mutex_trylock(&cache_chain_mutex))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003995 /* Give up. Setup the next iteration. */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08003996 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003997
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07003998 list_for_each_entry(searchp, &cache_chain, next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003999 check_irq_on();
4000
Christoph Lameter35386e32006-03-22 00:09:05 -08004001 /*
4002 * We only take the l3 lock if absolutely necessary and we
4003 * have established with reasonable certainty that
4004 * we can do some work if the lock was obtained.
4005 */
Christoph Lameteraab22072006-03-22 00:09:06 -08004006 l3 = searchp->nodelists[node];
Christoph Lameter35386e32006-03-22 00:09:05 -08004007
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004008 reap_alien(searchp, l3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004009
Christoph Lameteraab22072006-03-22 00:09:06 -08004010 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004011
Christoph Lameter35386e32006-03-22 00:09:05 -08004012 /*
4013 * These are racy checks but it does not matter
4014 * if we skip one check or scan twice.
4015 */
Christoph Lametere498be72005-09-09 13:03:32 -07004016 if (time_after(l3->next_reap, jiffies))
Christoph Lameter35386e32006-03-22 00:09:05 -08004017 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004018
Christoph Lametere498be72005-09-09 13:03:32 -07004019 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004020
Christoph Lameteraab22072006-03-22 00:09:06 -08004021 drain_array(searchp, l3, l3->shared, 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004022
Christoph Lametered11d9e2006-06-30 01:55:45 -07004023 if (l3->free_touched)
Christoph Lametere498be72005-09-09 13:03:32 -07004024 l3->free_touched = 0;
Christoph Lametered11d9e2006-06-30 01:55:45 -07004025 else {
4026 int freed;
4027
4028 freed = drain_freelist(searchp, l3, (l3->free_limit +
4029 5 * searchp->num - 1) / (5 * searchp->num));
4030 STATS_ADD_REAPED(searchp, freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004031 }
Christoph Lameter35386e32006-03-22 00:09:05 -08004032next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004033 cond_resched();
4034 }
4035 check_irq_on();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004036 mutex_unlock(&cache_chain_mutex);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004037 next_reap_node();
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004038out:
Andrew Mortona737b3e2006-03-22 00:08:11 -08004039 /* Set up the next iteration */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004040 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004041}
4042
Linus Torvalds158a9622008-01-02 13:04:48 -08004043#ifdef CONFIG_SLABINFO
Linus Torvalds1da177e2005-04-16 15:20:36 -07004044
Pekka Enberg85289f92006-01-08 01:00:36 -08004045static void print_slabinfo_header(struct seq_file *m)
4046{
4047 /*
4048 * Output format version, so at least we can change it
4049 * without _too_ many complaints.
4050 */
4051#if STATS
4052 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4053#else
4054 seq_puts(m, "slabinfo - version: 2.1\n");
4055#endif
4056 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4057 "<objperslab> <pagesperslab>");
4058 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4059 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4060#if STATS
4061 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004062 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
Pekka Enberg85289f92006-01-08 01:00:36 -08004063 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4064#endif
4065 seq_putc(m, '\n');
4066}
4067
Linus Torvalds1da177e2005-04-16 15:20:36 -07004068static void *s_start(struct seq_file *m, loff_t *pos)
4069{
4070 loff_t n = *pos;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004071
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004072 mutex_lock(&cache_chain_mutex);
Pekka Enberg85289f92006-01-08 01:00:36 -08004073 if (!n)
4074 print_slabinfo_header(m);
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004075
4076 return seq_list_start(&cache_chain, *pos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004077}
4078
4079static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4080{
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004081 return seq_list_next(p, &cache_chain, pos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004082}
4083
4084static void s_stop(struct seq_file *m, void *p)
4085{
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004086 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004087}
4088
4089static int s_show(struct seq_file *m, void *p)
4090{
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004091 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004092 struct slab *slabp;
4093 unsigned long active_objs;
4094 unsigned long num_objs;
4095 unsigned long active_slabs = 0;
4096 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004097 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004098 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07004099 int node;
4100 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004101
Linus Torvalds1da177e2005-04-16 15:20:36 -07004102 active_objs = 0;
4103 num_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004104 for_each_online_node(node) {
4105 l3 = cachep->nodelists[node];
4106 if (!l3)
4107 continue;
4108
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004109 check_irq_on();
4110 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07004111
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004112 list_for_each_entry(slabp, &l3->slabs_full, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004113 if (slabp->inuse != cachep->num && !error)
4114 error = "slabs_full accounting error";
4115 active_objs += cachep->num;
4116 active_slabs++;
4117 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004118 list_for_each_entry(slabp, &l3->slabs_partial, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004119 if (slabp->inuse == cachep->num && !error)
4120 error = "slabs_partial inuse accounting error";
4121 if (!slabp->inuse && !error)
4122 error = "slabs_partial/inuse accounting error";
4123 active_objs += slabp->inuse;
4124 active_slabs++;
4125 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004126 list_for_each_entry(slabp, &l3->slabs_free, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004127 if (slabp->inuse && !error)
4128 error = "slabs_free/inuse accounting error";
4129 num_slabs++;
4130 }
4131 free_objects += l3->free_objects;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08004132 if (l3->shared)
4133 shared_avail += l3->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07004134
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004135 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004136 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004137 num_slabs += active_slabs;
4138 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07004139 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004140 error = "free_objects accounting error";
4141
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004142 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004143 if (error)
4144 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4145
4146 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004147 name, active_objs, num_objs, cachep->buffer_size,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004148 cachep->num, (1 << cachep->gfporder));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004149 seq_printf(m, " : tunables %4u %4u %4u",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004150 cachep->limit, cachep->batchcount, cachep->shared);
Christoph Lametere498be72005-09-09 13:03:32 -07004151 seq_printf(m, " : slabdata %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004152 active_slabs, num_slabs, shared_avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004153#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004154 { /* list3 stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004155 unsigned long high = cachep->high_mark;
4156 unsigned long allocs = cachep->num_allocations;
4157 unsigned long grown = cachep->grown;
4158 unsigned long reaped = cachep->reaped;
4159 unsigned long errors = cachep->errors;
4160 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004161 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07004162 unsigned long node_frees = cachep->node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004163 unsigned long overflows = cachep->node_overflow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004164
Christoph Lametere498be72005-09-09 13:03:32 -07004165 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004166 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown,
Andrew Mortona737b3e2006-03-22 00:08:11 -08004167 reaped, errors, max_freeable, node_allocs,
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004168 node_frees, overflows);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004169 }
4170 /* cpu stats */
4171 {
4172 unsigned long allochit = atomic_read(&cachep->allochit);
4173 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4174 unsigned long freehit = atomic_read(&cachep->freehit);
4175 unsigned long freemiss = atomic_read(&cachep->freemiss);
4176
4177 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004178 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004179 }
4180#endif
4181 seq_putc(m, '\n');
Linus Torvalds1da177e2005-04-16 15:20:36 -07004182 return 0;
4183}
4184
4185/*
4186 * slabinfo_op - iterator that generates /proc/slabinfo
4187 *
4188 * Output layout:
4189 * cache-name
4190 * num-active-objs
4191 * total-objs
4192 * object size
4193 * num-active-slabs
4194 * total-slabs
4195 * num-pages-per-slab
4196 * + further values on SMP and with statistics enabled
4197 */
4198
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04004199static const struct seq_operations slabinfo_op = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004200 .start = s_start,
4201 .next = s_next,
4202 .stop = s_stop,
4203 .show = s_show,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004204};
4205
4206#define MAX_SLABINFO_WRITE 128
4207/**
4208 * slabinfo_write - Tuning for the slab allocator
4209 * @file: unused
4210 * @buffer: user buffer
4211 * @count: data length
4212 * @ppos: unused
4213 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004214ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4215 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004216{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004217 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004218 int limit, batchcount, shared, res;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004219 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004220
Linus Torvalds1da177e2005-04-16 15:20:36 -07004221 if (count > MAX_SLABINFO_WRITE)
4222 return -EINVAL;
4223 if (copy_from_user(&kbuf, buffer, count))
4224 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004225 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07004226
4227 tmp = strchr(kbuf, ' ');
4228 if (!tmp)
4229 return -EINVAL;
4230 *tmp = '\0';
4231 tmp++;
4232 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4233 return -EINVAL;
4234
4235 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004236 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004237 res = -EINVAL;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004238 list_for_each_entry(cachep, &cache_chain, next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004239 if (!strcmp(cachep->name, kbuf)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08004240 if (limit < 1 || batchcount < 1 ||
4241 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07004242 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004243 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07004244 res = do_tune_cpucache(cachep, limit,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004245 batchcount, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004246 }
4247 break;
4248 }
4249 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004250 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004251 if (res >= 0)
4252 res = count;
4253 return res;
4254}
Al Viro871751e2006-03-25 03:06:39 -08004255
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04004256static int slabinfo_open(struct inode *inode, struct file *file)
4257{
4258 return seq_open(file, &slabinfo_op);
4259}
4260
4261static const struct file_operations proc_slabinfo_operations = {
4262 .open = slabinfo_open,
4263 .read = seq_read,
4264 .write = slabinfo_write,
4265 .llseek = seq_lseek,
4266 .release = seq_release,
4267};
4268
Al Viro871751e2006-03-25 03:06:39 -08004269#ifdef CONFIG_DEBUG_SLAB_LEAK
4270
4271static void *leaks_start(struct seq_file *m, loff_t *pos)
4272{
Al Viro871751e2006-03-25 03:06:39 -08004273 mutex_lock(&cache_chain_mutex);
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004274 return seq_list_start(&cache_chain, *pos);
Al Viro871751e2006-03-25 03:06:39 -08004275}
4276
4277static inline int add_caller(unsigned long *n, unsigned long v)
4278{
4279 unsigned long *p;
4280 int l;
4281 if (!v)
4282 return 1;
4283 l = n[1];
4284 p = n + 2;
4285 while (l) {
4286 int i = l/2;
4287 unsigned long *q = p + 2 * i;
4288 if (*q == v) {
4289 q[1]++;
4290 return 1;
4291 }
4292 if (*q > v) {
4293 l = i;
4294 } else {
4295 p = q + 2;
4296 l -= i + 1;
4297 }
4298 }
4299 if (++n[1] == n[0])
4300 return 0;
4301 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4302 p[0] = v;
4303 p[1] = 1;
4304 return 1;
4305}
4306
4307static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4308{
4309 void *p;
4310 int i;
4311 if (n[0] == n[1])
4312 return;
4313 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4314 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4315 continue;
4316 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4317 return;
4318 }
4319}
4320
4321static void show_symbol(struct seq_file *m, unsigned long address)
4322{
4323#ifdef CONFIG_KALLSYMS
Al Viro871751e2006-03-25 03:06:39 -08004324 unsigned long offset, size;
Tejun Heo9281ace2007-07-17 04:03:51 -07004325 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
Al Viro871751e2006-03-25 03:06:39 -08004326
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004327 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
Al Viro871751e2006-03-25 03:06:39 -08004328 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004329 if (modname[0])
Al Viro871751e2006-03-25 03:06:39 -08004330 seq_printf(m, " [%s]", modname);
4331 return;
4332 }
4333#endif
4334 seq_printf(m, "%p", (void *)address);
4335}
4336
4337static int leaks_show(struct seq_file *m, void *p)
4338{
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004339 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
Al Viro871751e2006-03-25 03:06:39 -08004340 struct slab *slabp;
4341 struct kmem_list3 *l3;
4342 const char *name;
4343 unsigned long *n = m->private;
4344 int node;
4345 int i;
4346
4347 if (!(cachep->flags & SLAB_STORE_USER))
4348 return 0;
4349 if (!(cachep->flags & SLAB_RED_ZONE))
4350 return 0;
4351
4352 /* OK, we can do it */
4353
4354 n[1] = 0;
4355
4356 for_each_online_node(node) {
4357 l3 = cachep->nodelists[node];
4358 if (!l3)
4359 continue;
4360
4361 check_irq_on();
4362 spin_lock_irq(&l3->list_lock);
4363
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004364 list_for_each_entry(slabp, &l3->slabs_full, list)
Al Viro871751e2006-03-25 03:06:39 -08004365 handle_slab(n, cachep, slabp);
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004366 list_for_each_entry(slabp, &l3->slabs_partial, list)
Al Viro871751e2006-03-25 03:06:39 -08004367 handle_slab(n, cachep, slabp);
Al Viro871751e2006-03-25 03:06:39 -08004368 spin_unlock_irq(&l3->list_lock);
4369 }
4370 name = cachep->name;
4371 if (n[0] == n[1]) {
4372 /* Increase the buffer size */
4373 mutex_unlock(&cache_chain_mutex);
4374 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4375 if (!m->private) {
4376 /* Too bad, we are really out */
4377 m->private = n;
4378 mutex_lock(&cache_chain_mutex);
4379 return -ENOMEM;
4380 }
4381 *(unsigned long *)m->private = n[0] * 2;
4382 kfree(n);
4383 mutex_lock(&cache_chain_mutex);
4384 /* Now make sure this entry will be retried */
4385 m->count = m->size;
4386 return 0;
4387 }
4388 for (i = 0; i < n[1]; i++) {
4389 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4390 show_symbol(m, n[2*i+2]);
4391 seq_putc(m, '\n');
4392 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004393
Al Viro871751e2006-03-25 03:06:39 -08004394 return 0;
4395}
4396
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004397static const struct seq_operations slabstats_op = {
Al Viro871751e2006-03-25 03:06:39 -08004398 .start = leaks_start,
4399 .next = s_next,
4400 .stop = s_stop,
4401 .show = leaks_show,
4402};
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004403
4404static int slabstats_open(struct inode *inode, struct file *file)
4405{
4406 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4407 int ret = -ENOMEM;
4408 if (n) {
4409 ret = seq_open(file, &slabstats_op);
4410 if (!ret) {
4411 struct seq_file *m = file->private_data;
4412 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4413 m->private = n;
4414 n = NULL;
4415 }
4416 kfree(n);
4417 }
4418 return ret;
4419}
4420
4421static const struct file_operations proc_slabstats_operations = {
4422 .open = slabstats_open,
4423 .read = seq_read,
4424 .llseek = seq_lseek,
4425 .release = seq_release_private,
4426};
Al Viro871751e2006-03-25 03:06:39 -08004427#endif
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004428
4429static int __init slab_proc_init(void)
4430{
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04004431 proc_create("slabinfo",S_IWUSR|S_IRUGO,NULL,&proc_slabinfo_operations);
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004432#ifdef CONFIG_DEBUG_SLAB_LEAK
4433 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4434#endif
4435 return 0;
4436}
4437module_init(slab_proc_init);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004438#endif
4439
Manfred Spraul00e145b2005-09-03 15:55:07 -07004440/**
4441 * ksize - get the actual amount of memory allocated for a given object
4442 * @objp: Pointer to the object
4443 *
4444 * kmalloc may internally round up allocations and return more memory
4445 * than requested. ksize() can be used to determine the actual amount of
4446 * memory allocated. The caller may use this additional memory, even though
4447 * a smaller amount of memory was initially specified with the kmalloc call.
4448 * The caller must guarantee that objp points to a valid object previously
4449 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4450 * must not be freed during the duration of the call.
4451 */
Pekka Enbergfd76bab2007-05-06 14:48:40 -07004452size_t ksize(const void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004453{
Christoph Lameteref8b4522007-10-16 01:24:46 -07004454 BUG_ON(!objp);
4455 if (unlikely(objp == ZERO_SIZE_PTR))
Manfred Spraul00e145b2005-09-03 15:55:07 -07004456 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004457
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08004458 return obj_size(virt_to_cache(objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004459}