blob: a684778b2b41af38e06aa47e5fe66a7a988ad4c7 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * linux/mm/slab.c
3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
5 *
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
7 *
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
10 *
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
13 *
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
21 *
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
27 *
28 * This means, that your constructor is used only for newly allocated
29 * slabs and you must pass objects with the same intializations to
30 * kmem_cache_free.
31 *
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
35 *
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
38 * partial slabs
39 * empty slabs with no allocated objects
40 *
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
43 *
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
46 *
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
52 *
Andrew Mortona737b3e2006-03-22 00:08:11 -080053 * The c_cpuarray may not be read with enabled local interrupts -
Linus Torvalds1da177e2005-04-16 15:20:36 -070054 * it's changed with a smp_call_function().
55 *
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
Pekka Enberg343e0d72006-02-01 03:05:50 -080058 * Several members in struct kmem_cache and struct slab never change, they
Linus Torvalds1da177e2005-04-16 15:20:36 -070059 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
63 *
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
66 * his patch.
67 *
68 * Further notes from the original documentation:
69 *
70 * 11 April '97. Started multi-threading - markhe
Ingo Molnarfc0abb12006-01-18 17:42:33 -080071 * The global cache-chain is protected by the mutex 'cache_chain_mutex'.
Linus Torvalds1da177e2005-04-16 15:20:36 -070072 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
75 *
76 * At present, each engine can be growing a cache. This should be blocked.
77 *
Christoph Lametere498be72005-09-09 13:03:32 -070078 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
83 *
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
Linus Torvalds1da177e2005-04-16 15:20:36 -070087 */
88
Linus Torvalds1da177e2005-04-16 15:20:36 -070089#include <linux/slab.h>
90#include <linux/mm.h>
Randy Dunlapc9cf5522006-06-27 02:53:52 -070091#include <linux/poison.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070092#include <linux/swap.h>
93#include <linux/cache.h>
94#include <linux/interrupt.h>
95#include <linux/init.h>
96#include <linux/compiler.h>
Paul Jackson101a5002006-03-24 03:16:07 -080097#include <linux/cpuset.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070098#include <linux/seq_file.h>
99#include <linux/notifier.h>
100#include <linux/kallsyms.h>
101#include <linux/cpu.h>
102#include <linux/sysctl.h>
103#include <linux/module.h>
104#include <linux/rcupdate.h>
Paulo Marques543537b2005-06-23 00:09:02 -0700105#include <linux/string.h>
Andrew Morton138ae662006-12-06 20:36:41 -0800106#include <linux/uaccess.h>
Christoph Lametere498be72005-09-09 13:03:32 -0700107#include <linux/nodemask.h>
Christoph Lameterdc85da12006-01-18 17:42:36 -0800108#include <linux/mempolicy.h>
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800109#include <linux/mutex.h>
Akinobu Mita8a8b6502006-12-08 02:39:44 -0800110#include <linux/fault-inject.h>
Ingo Molnare7eebaf2006-06-27 02:54:55 -0700111#include <linux/rtmutex.h>
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800112#include <linux/reciprocal_div.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700113
Linus Torvalds1da177e2005-04-16 15:20:36 -0700114#include <asm/cacheflush.h>
115#include <asm/tlbflush.h>
116#include <asm/page.h>
117
118/*
Christoph Lameter50953fe2007-05-06 14:50:16 -0700119 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700120 * 0 for faster, smaller code (especially in the critical paths).
121 *
122 * STATS - 1 to collect stats for /proc/slabinfo.
123 * 0 for faster, smaller code (especially in the critical paths).
124 *
125 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
126 */
127
128#ifdef CONFIG_DEBUG_SLAB
129#define DEBUG 1
130#define STATS 1
131#define FORCED_DEBUG 1
132#else
133#define DEBUG 0
134#define STATS 0
135#define FORCED_DEBUG 0
136#endif
137
Linus Torvalds1da177e2005-04-16 15:20:36 -0700138/* Shouldn't this be in a header file somewhere? */
139#define BYTES_PER_WORD sizeof(void *)
David Woodhouse87a927c2007-07-04 21:26:44 -0400140#define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700141
142#ifndef cache_line_size
143#define cache_line_size() L1_CACHE_BYTES
144#endif
145
146#ifndef ARCH_KMALLOC_MINALIGN
147/*
148 * Enforce a minimum alignment for the kmalloc caches.
149 * Usually, the kmalloc caches are cache_line_size() aligned, except when
150 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
151 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
David Woodhouseb46b8f12007-05-08 00:22:59 -0700152 * alignment larger than the alignment of a 64-bit integer.
153 * ARCH_KMALLOC_MINALIGN allows that.
154 * Note that increasing this value may disable some debug features.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700155 */
David Woodhouseb46b8f12007-05-08 00:22:59 -0700156#define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700157#endif
158
159#ifndef ARCH_SLAB_MINALIGN
160/*
161 * Enforce a minimum alignment for all caches.
162 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
163 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
164 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
165 * some debug features.
166 */
167#define ARCH_SLAB_MINALIGN 0
168#endif
169
170#ifndef ARCH_KMALLOC_FLAGS
171#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
172#endif
173
174/* Legal flag mask for kmem_cache_create(). */
175#if DEBUG
Christoph Lameter50953fe2007-05-06 14:50:16 -0700176# define CREATE_MASK (SLAB_RED_ZONE | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700177 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
Christoph Lameterac2b8982006-03-22 00:08:15 -0800178 SLAB_CACHE_DMA | \
Christoph Lameter5af60832007-05-06 14:49:56 -0700179 SLAB_STORE_USER | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700180 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
Paul Jackson101a5002006-03-24 03:16:07 -0800181 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700182#else
Christoph Lameterac2b8982006-03-22 00:08:15 -0800183# define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
Christoph Lameter5af60832007-05-06 14:49:56 -0700184 SLAB_CACHE_DMA | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700185 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
Paul Jackson101a5002006-03-24 03:16:07 -0800186 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700187#endif
188
189/*
190 * kmem_bufctl_t:
191 *
192 * Bufctl's are used for linking objs within a slab
193 * linked offsets.
194 *
195 * This implementation relies on "struct page" for locating the cache &
196 * slab an object belongs to.
197 * This allows the bufctl structure to be small (one int), but limits
198 * the number of objects a slab (not a cache) can contain when off-slab
199 * bufctls are used. The limit is the size of the largest general cache
200 * that does not use off-slab slabs.
201 * For 32bit archs with 4 kB pages, is this 56.
202 * This is not serious, as it is only for large objects, when it is unwise
203 * to have too many per slab.
204 * Note: This limit can be raised by introducing a general cache whose size
205 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
206 */
207
Kyle Moffettfa5b08d2005-09-03 15:55:03 -0700208typedef unsigned int kmem_bufctl_t;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700209#define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
210#define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
Al Viro871751e2006-03-25 03:06:39 -0800211#define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
212#define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700213
Linus Torvalds1da177e2005-04-16 15:20:36 -0700214/*
215 * struct slab
216 *
217 * Manages the objs in a slab. Placed either at the beginning of mem allocated
218 * for a slab, or allocated from an general cache.
219 * Slabs are chained into three list: fully used, partial, fully free slabs.
220 */
221struct slab {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800222 struct list_head list;
223 unsigned long colouroff;
224 void *s_mem; /* including colour offset */
225 unsigned int inuse; /* num of objs active in slab */
226 kmem_bufctl_t free;
227 unsigned short nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700228};
229
230/*
231 * struct slab_rcu
232 *
233 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
234 * arrange for kmem_freepages to be called via RCU. This is useful if
235 * we need to approach a kernel structure obliquely, from its address
236 * obtained without the usual locking. We can lock the structure to
237 * stabilize it and check it's still at the given address, only if we
238 * can be sure that the memory has not been meanwhile reused for some
239 * other kind of object (which our subsystem's lock might corrupt).
240 *
241 * rcu_read_lock before reading the address, then rcu_read_unlock after
242 * taking the spinlock within the structure expected at that address.
243 *
244 * We assume struct slab_rcu can overlay struct slab when destroying.
245 */
246struct slab_rcu {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800247 struct rcu_head head;
Pekka Enberg343e0d72006-02-01 03:05:50 -0800248 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800249 void *addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700250};
251
252/*
253 * struct array_cache
254 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700255 * Purpose:
256 * - LIFO ordering, to hand out cache-warm objects from _alloc
257 * - reduce the number of linked list operations
258 * - reduce spinlock operations
259 *
260 * The limit is stored in the per-cpu structure to reduce the data cache
261 * footprint.
262 *
263 */
264struct array_cache {
265 unsigned int avail;
266 unsigned int limit;
267 unsigned int batchcount;
268 unsigned int touched;
Christoph Lametere498be72005-09-09 13:03:32 -0700269 spinlock_t lock;
Andrew Mortona737b3e2006-03-22 00:08:11 -0800270 void *entry[0]; /*
271 * Must have this definition in here for the proper
272 * alignment of array_cache. Also simplifies accessing
273 * the entries.
274 * [0] is for gcc 2.95. It should really be [].
275 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700276};
277
Andrew Mortona737b3e2006-03-22 00:08:11 -0800278/*
279 * bootstrap: The caches do not work without cpuarrays anymore, but the
280 * cpuarrays are allocated from the generic caches...
Linus Torvalds1da177e2005-04-16 15:20:36 -0700281 */
282#define BOOT_CPUCACHE_ENTRIES 1
283struct arraycache_init {
284 struct array_cache cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800285 void *entries[BOOT_CPUCACHE_ENTRIES];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700286};
287
288/*
Christoph Lametere498be72005-09-09 13:03:32 -0700289 * The slab lists for all objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700290 */
291struct kmem_list3 {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800292 struct list_head slabs_partial; /* partial list first, better asm code */
293 struct list_head slabs_full;
294 struct list_head slabs_free;
295 unsigned long free_objects;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800296 unsigned int free_limit;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800297 unsigned int colour_next; /* Per-node cache coloring */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800298 spinlock_t list_lock;
299 struct array_cache *shared; /* shared per node */
300 struct array_cache **alien; /* on other nodes */
Christoph Lameter35386e32006-03-22 00:09:05 -0800301 unsigned long next_reap; /* updated without locking */
302 int free_touched; /* updated without locking */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700303};
304
Christoph Lametere498be72005-09-09 13:03:32 -0700305/*
306 * Need this for bootstrapping a per node allocator.
307 */
308#define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
309struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
310#define CACHE_CACHE 0
311#define SIZE_AC 1
312#define SIZE_L3 (1 + MAX_NUMNODES)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700313
Christoph Lametered11d9e2006-06-30 01:55:45 -0700314static int drain_freelist(struct kmem_cache *cache,
315 struct kmem_list3 *l3, int tofree);
316static void free_block(struct kmem_cache *cachep, void **objpp, int len,
317 int node);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -0700318static int enable_cpucache(struct kmem_cache *cachep);
David Howells65f27f32006-11-22 14:55:48 +0000319static void cache_reap(struct work_struct *unused);
Christoph Lametered11d9e2006-06-30 01:55:45 -0700320
Christoph Lametere498be72005-09-09 13:03:32 -0700321/*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800322 * This function must be completely optimized away if a constant is passed to
323 * it. Mostly the same as what is in linux/slab.h except it returns an index.
Christoph Lametere498be72005-09-09 13:03:32 -0700324 */
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700325static __always_inline int index_of(const size_t size)
Christoph Lametere498be72005-09-09 13:03:32 -0700326{
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800327 extern void __bad_size(void);
328
Christoph Lametere498be72005-09-09 13:03:32 -0700329 if (__builtin_constant_p(size)) {
330 int i = 0;
331
332#define CACHE(x) \
333 if (size <=x) \
334 return i; \
335 else \
336 i++;
337#include "linux/kmalloc_sizes.h"
338#undef CACHE
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800339 __bad_size();
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700340 } else
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800341 __bad_size();
Christoph Lametere498be72005-09-09 13:03:32 -0700342 return 0;
343}
344
Ingo Molnare0a42722006-06-23 02:03:46 -0700345static int slab_early_init = 1;
346
Christoph Lametere498be72005-09-09 13:03:32 -0700347#define INDEX_AC index_of(sizeof(struct arraycache_init))
348#define INDEX_L3 index_of(sizeof(struct kmem_list3))
349
Pekka Enberg5295a742006-02-01 03:05:48 -0800350static void kmem_list3_init(struct kmem_list3 *parent)
Christoph Lametere498be72005-09-09 13:03:32 -0700351{
352 INIT_LIST_HEAD(&parent->slabs_full);
353 INIT_LIST_HEAD(&parent->slabs_partial);
354 INIT_LIST_HEAD(&parent->slabs_free);
355 parent->shared = NULL;
356 parent->alien = NULL;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800357 parent->colour_next = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700358 spin_lock_init(&parent->list_lock);
359 parent->free_objects = 0;
360 parent->free_touched = 0;
361}
362
Andrew Mortona737b3e2006-03-22 00:08:11 -0800363#define MAKE_LIST(cachep, listp, slab, nodeid) \
364 do { \
365 INIT_LIST_HEAD(listp); \
366 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
Christoph Lametere498be72005-09-09 13:03:32 -0700367 } while (0)
368
Andrew Mortona737b3e2006-03-22 00:08:11 -0800369#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
370 do { \
Christoph Lametere498be72005-09-09 13:03:32 -0700371 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
372 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
373 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
374 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700375
376/*
Pekka Enberg343e0d72006-02-01 03:05:50 -0800377 * struct kmem_cache
Linus Torvalds1da177e2005-04-16 15:20:36 -0700378 *
379 * manages a cache.
380 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800381
Pekka J Enberg2109a2d2005-11-07 00:58:01 -0800382struct kmem_cache {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700383/* 1) per-cpu data, touched during every alloc/free */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800384 struct array_cache *array[NR_CPUS];
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800385/* 2) Cache tunables. Protected by cache_chain_mutex */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800386 unsigned int batchcount;
387 unsigned int limit;
388 unsigned int shared;
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800389
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800390 unsigned int buffer_size;
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800391 u32 reciprocal_buffer_size;
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800392/* 3) touched by every alloc & free from the backend */
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800393
Andrew Mortona737b3e2006-03-22 00:08:11 -0800394 unsigned int flags; /* constant flags */
395 unsigned int num; /* # of objs per slab */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700396
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800397/* 4) cache_grow/shrink */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700398 /* order of pgs per slab (2^n) */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800399 unsigned int gfporder;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700400
401 /* force GFP flags, e.g. GFP_DMA */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800402 gfp_t gfpflags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700403
Andrew Mortona737b3e2006-03-22 00:08:11 -0800404 size_t colour; /* cache colouring range */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800405 unsigned int colour_off; /* colour offset */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800406 struct kmem_cache *slabp_cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800407 unsigned int slab_size;
Andrew Mortona737b3e2006-03-22 00:08:11 -0800408 unsigned int dflags; /* dynamic flags */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700409
410 /* constructor func */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800411 void (*ctor) (void *, struct kmem_cache *, unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700412
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800413/* 5) cache creation/removal */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800414 const char *name;
415 struct list_head next;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700416
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800417/* 6) statistics */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700418#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800419 unsigned long num_active;
420 unsigned long num_allocations;
421 unsigned long high_mark;
422 unsigned long grown;
423 unsigned long reaped;
424 unsigned long errors;
425 unsigned long max_freeable;
426 unsigned long node_allocs;
427 unsigned long node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700428 unsigned long node_overflow;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800429 atomic_t allochit;
430 atomic_t allocmiss;
431 atomic_t freehit;
432 atomic_t freemiss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700433#endif
434#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800435 /*
436 * If debugging is enabled, then the allocator can add additional
437 * fields and/or padding to every object. buffer_size contains the total
438 * object size including these internal fields, the following two
439 * variables contain the offset to the user object and its size.
440 */
441 int obj_offset;
442 int obj_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700443#endif
Eric Dumazet8da34302007-05-06 14:49:29 -0700444 /*
445 * We put nodelists[] at the end of kmem_cache, because we want to size
446 * this array to nr_node_ids slots instead of MAX_NUMNODES
447 * (see kmem_cache_init())
448 * We still use [MAX_NUMNODES] and not [1] or [0] because cache_cache
449 * is statically defined, so we reserve the max number of nodes.
450 */
451 struct kmem_list3 *nodelists[MAX_NUMNODES];
452 /*
453 * Do not add fields after nodelists[]
454 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700455};
456
457#define CFLGS_OFF_SLAB (0x80000000UL)
458#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
459
460#define BATCHREFILL_LIMIT 16
Andrew Mortona737b3e2006-03-22 00:08:11 -0800461/*
462 * Optimization question: fewer reaps means less probability for unnessary
463 * cpucache drain/refill cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700464 *
Adrian Bunkdc6f3f22005-11-08 16:44:08 +0100465 * OTOH the cpuarrays can contain lots of objects,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700466 * which could lock up otherwise freeable slabs.
467 */
468#define REAPTIMEOUT_CPUC (2*HZ)
469#define REAPTIMEOUT_LIST3 (4*HZ)
470
471#if STATS
472#define STATS_INC_ACTIVE(x) ((x)->num_active++)
473#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
474#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
475#define STATS_INC_GROWN(x) ((x)->grown++)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700476#define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
Andrew Mortona737b3e2006-03-22 00:08:11 -0800477#define STATS_SET_HIGH(x) \
478 do { \
479 if ((x)->num_active > (x)->high_mark) \
480 (x)->high_mark = (x)->num_active; \
481 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700482#define STATS_INC_ERR(x) ((x)->errors++)
483#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
Christoph Lametere498be72005-09-09 13:03:32 -0700484#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700485#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800486#define STATS_SET_FREEABLE(x, i) \
487 do { \
488 if ((x)->max_freeable < i) \
489 (x)->max_freeable = i; \
490 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700491#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
492#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
493#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
494#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
495#else
496#define STATS_INC_ACTIVE(x) do { } while (0)
497#define STATS_DEC_ACTIVE(x) do { } while (0)
498#define STATS_INC_ALLOCED(x) do { } while (0)
499#define STATS_INC_GROWN(x) do { } while (0)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700500#define STATS_ADD_REAPED(x,y) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700501#define STATS_SET_HIGH(x) do { } while (0)
502#define STATS_INC_ERR(x) do { } while (0)
503#define STATS_INC_NODEALLOCS(x) do { } while (0)
Christoph Lametere498be72005-09-09 13:03:32 -0700504#define STATS_INC_NODEFREES(x) do { } while (0)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700505#define STATS_INC_ACOVERFLOW(x) do { } while (0)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800506#define STATS_SET_FREEABLE(x, i) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700507#define STATS_INC_ALLOCHIT(x) do { } while (0)
508#define STATS_INC_ALLOCMISS(x) do { } while (0)
509#define STATS_INC_FREEHIT(x) do { } while (0)
510#define STATS_INC_FREEMISS(x) do { } while (0)
511#endif
512
513#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -0700514
Andrew Mortona737b3e2006-03-22 00:08:11 -0800515/*
516 * memory layout of objects:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700517 * 0 : objp
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800518 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
Linus Torvalds1da177e2005-04-16 15:20:36 -0700519 * the end of an object is aligned with the end of the real
520 * allocation. Catches writes behind the end of the allocation.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800521 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700522 * redzone word.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800523 * cachep->obj_offset: The real object.
524 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
Andrew Mortona737b3e2006-03-22 00:08:11 -0800525 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
526 * [BYTES_PER_WORD long]
Linus Torvalds1da177e2005-04-16 15:20:36 -0700527 */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800528static int obj_offset(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700529{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800530 return cachep->obj_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700531}
532
Pekka Enberg343e0d72006-02-01 03:05:50 -0800533static int obj_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700534{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800535 return cachep->obj_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700536}
537
David Woodhouseb46b8f12007-05-08 00:22:59 -0700538static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700539{
540 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
David Woodhouseb46b8f12007-05-08 00:22:59 -0700541 return (unsigned long long*) (objp + obj_offset(cachep) -
542 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700543}
544
David Woodhouseb46b8f12007-05-08 00:22:59 -0700545static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700546{
547 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
548 if (cachep->flags & SLAB_STORE_USER)
David Woodhouseb46b8f12007-05-08 00:22:59 -0700549 return (unsigned long long *)(objp + cachep->buffer_size -
550 sizeof(unsigned long long) -
David Woodhouse87a927c2007-07-04 21:26:44 -0400551 REDZONE_ALIGN);
David Woodhouseb46b8f12007-05-08 00:22:59 -0700552 return (unsigned long long *) (objp + cachep->buffer_size -
553 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700554}
555
Pekka Enberg343e0d72006-02-01 03:05:50 -0800556static void **dbg_userword(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700557{
558 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800559 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700560}
561
562#else
563
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800564#define obj_offset(x) 0
565#define obj_size(cachep) (cachep->buffer_size)
David Woodhouseb46b8f12007-05-08 00:22:59 -0700566#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
567#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
Linus Torvalds1da177e2005-04-16 15:20:36 -0700568#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
569
570#endif
571
572/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700573 * Do not go above this order unless 0 objects fit into the slab.
574 */
575#define BREAK_GFP_ORDER_HI 1
576#define BREAK_GFP_ORDER_LO 0
577static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
578
Andrew Mortona737b3e2006-03-22 00:08:11 -0800579/*
580 * Functions for storing/retrieving the cachep and or slab from the page
581 * allocator. These are used to find the slab an obj belongs to. With kfree(),
582 * these are used to find the cache which an obj belongs to.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700583 */
Pekka Enberg065d41c2005-11-13 16:06:46 -0800584static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
585{
586 page->lru.next = (struct list_head *)cache;
587}
588
589static inline struct kmem_cache *page_get_cache(struct page *page)
590{
Christoph Lameterd85f3382007-05-06 14:49:39 -0700591 page = compound_head(page);
Pekka Enbergddc2e812006-06-23 02:03:40 -0700592 BUG_ON(!PageSlab(page));
Pekka Enberg065d41c2005-11-13 16:06:46 -0800593 return (struct kmem_cache *)page->lru.next;
594}
595
596static inline void page_set_slab(struct page *page, struct slab *slab)
597{
598 page->lru.prev = (struct list_head *)slab;
599}
600
601static inline struct slab *page_get_slab(struct page *page)
602{
Pekka Enbergddc2e812006-06-23 02:03:40 -0700603 BUG_ON(!PageSlab(page));
Pekka Enberg065d41c2005-11-13 16:06:46 -0800604 return (struct slab *)page->lru.prev;
605}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700606
Pekka Enberg6ed5eb22006-02-01 03:05:49 -0800607static inline struct kmem_cache *virt_to_cache(const void *obj)
608{
Christoph Lameterb49af682007-05-06 14:49:41 -0700609 struct page *page = virt_to_head_page(obj);
Pekka Enberg6ed5eb22006-02-01 03:05:49 -0800610 return page_get_cache(page);
611}
612
613static inline struct slab *virt_to_slab(const void *obj)
614{
Christoph Lameterb49af682007-05-06 14:49:41 -0700615 struct page *page = virt_to_head_page(obj);
Pekka Enberg6ed5eb22006-02-01 03:05:49 -0800616 return page_get_slab(page);
617}
618
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800619static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
620 unsigned int idx)
621{
622 return slab->s_mem + cache->buffer_size * idx;
623}
624
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800625/*
626 * We want to avoid an expensive divide : (offset / cache->buffer_size)
627 * Using the fact that buffer_size is a constant for a particular cache,
628 * we can replace (offset / cache->buffer_size) by
629 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
630 */
631static inline unsigned int obj_to_index(const struct kmem_cache *cache,
632 const struct slab *slab, void *obj)
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800633{
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800634 u32 offset = (obj - slab->s_mem);
635 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800636}
637
Andrew Mortona737b3e2006-03-22 00:08:11 -0800638/*
639 * These are the default caches for kmalloc. Custom caches can have other sizes.
640 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700641struct cache_sizes malloc_sizes[] = {
642#define CACHE(x) { .cs_size = (x) },
643#include <linux/kmalloc_sizes.h>
644 CACHE(ULONG_MAX)
645#undef CACHE
646};
647EXPORT_SYMBOL(malloc_sizes);
648
649/* Must match cache_sizes above. Out of line to keep cache footprint low. */
650struct cache_names {
651 char *name;
652 char *name_dma;
653};
654
655static struct cache_names __initdata cache_names[] = {
656#define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
657#include <linux/kmalloc_sizes.h>
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800658 {NULL,}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700659#undef CACHE
660};
661
662static struct arraycache_init initarray_cache __initdata =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800663 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700664static struct arraycache_init initarray_generic =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800665 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700666
667/* internal cache of cache description objs */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800668static struct kmem_cache cache_cache = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800669 .batchcount = 1,
670 .limit = BOOT_CPUCACHE_ENTRIES,
671 .shared = 1,
Pekka Enberg343e0d72006-02-01 03:05:50 -0800672 .buffer_size = sizeof(struct kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800673 .name = "kmem_cache",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700674};
675
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700676#define BAD_ALIEN_MAGIC 0x01020304ul
677
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200678#ifdef CONFIG_LOCKDEP
679
680/*
681 * Slab sometimes uses the kmalloc slabs to store the slab headers
682 * for other slabs "off slab".
683 * The locking for this is tricky in that it nests within the locks
684 * of all other slabs in a few places; to deal with this special
685 * locking we put on-slab caches into a separate lock-class.
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700686 *
687 * We set lock class for alien array caches which are up during init.
688 * The lock annotation will be lost if all cpus of a node goes down and
689 * then comes back up during hotplug
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200690 */
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700691static struct lock_class_key on_slab_l3_key;
692static struct lock_class_key on_slab_alc_key;
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200693
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700694static inline void init_lock_keys(void)
695
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200696{
697 int q;
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700698 struct cache_sizes *s = malloc_sizes;
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200699
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700700 while (s->cs_size != ULONG_MAX) {
701 for_each_node(q) {
702 struct array_cache **alc;
703 int r;
704 struct kmem_list3 *l3 = s->cs_cachep->nodelists[q];
705 if (!l3 || OFF_SLAB(s->cs_cachep))
706 continue;
707 lockdep_set_class(&l3->list_lock, &on_slab_l3_key);
708 alc = l3->alien;
709 /*
710 * FIXME: This check for BAD_ALIEN_MAGIC
711 * should go away when common slab code is taught to
712 * work even without alien caches.
713 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
714 * for alloc_alien_cache,
715 */
716 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
717 continue;
718 for_each_node(r) {
719 if (alc[r])
720 lockdep_set_class(&alc[r]->lock,
721 &on_slab_alc_key);
722 }
723 }
724 s++;
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200725 }
726}
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200727#else
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700728static inline void init_lock_keys(void)
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200729{
730}
731#endif
732
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -0800733/*
734 * 1. Guard access to the cache-chain.
735 * 2. Protect sanity of cpu_online_map against cpu hotplug events
736 */
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800737static DEFINE_MUTEX(cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700738static struct list_head cache_chain;
739
740/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700741 * chicken and egg problem: delay the per-cpu array allocation
742 * until the general caches are up.
743 */
744static enum {
745 NONE,
Christoph Lametere498be72005-09-09 13:03:32 -0700746 PARTIAL_AC,
747 PARTIAL_L3,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700748 FULL
749} g_cpucache_up;
750
Mike Kravetz39d24e62006-05-15 09:44:13 -0700751/*
752 * used by boot code to determine if it can use slab based allocator
753 */
754int slab_is_available(void)
755{
756 return g_cpucache_up == FULL;
757}
758
David Howells52bad642006-11-22 14:54:01 +0000759static DEFINE_PER_CPU(struct delayed_work, reap_work);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700760
Pekka Enberg343e0d72006-02-01 03:05:50 -0800761static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700762{
763 return cachep->array[smp_processor_id()];
764}
765
Andrew Mortona737b3e2006-03-22 00:08:11 -0800766static inline struct kmem_cache *__find_general_cachep(size_t size,
767 gfp_t gfpflags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700768{
769 struct cache_sizes *csizep = malloc_sizes;
770
771#if DEBUG
772 /* This happens if someone tries to call
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800773 * kmem_cache_create(), or __kmalloc(), before
774 * the generic caches are initialized.
775 */
Alok Katariac7e43c72005-09-14 12:17:53 -0700776 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700777#endif
Christoph Lameter6cb8f912007-07-17 04:03:22 -0700778 if (!size)
779 return ZERO_SIZE_PTR;
780
Linus Torvalds1da177e2005-04-16 15:20:36 -0700781 while (size > csizep->cs_size)
782 csizep++;
783
784 /*
Martin Hicks0abf40c2005-09-03 15:54:54 -0700785 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
Linus Torvalds1da177e2005-04-16 15:20:36 -0700786 * has cs_{dma,}cachep==NULL. Thus no special case
787 * for large kmalloc calls required.
788 */
Christoph Lameter4b51d662007-02-10 01:43:10 -0800789#ifdef CONFIG_ZONE_DMA
Linus Torvalds1da177e2005-04-16 15:20:36 -0700790 if (unlikely(gfpflags & GFP_DMA))
791 return csizep->cs_dmacachep;
Christoph Lameter4b51d662007-02-10 01:43:10 -0800792#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700793 return csizep->cs_cachep;
794}
795
Adrian Bunkb2213852006-09-25 23:31:02 -0700796static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700797{
798 return __find_general_cachep(size, gfpflags);
799}
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700800
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800801static size_t slab_mgmt_size(size_t nr_objs, size_t align)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700802{
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800803 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
804}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700805
Andrew Mortona737b3e2006-03-22 00:08:11 -0800806/*
807 * Calculate the number of objects and left-over bytes for a given buffer size.
808 */
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800809static void cache_estimate(unsigned long gfporder, size_t buffer_size,
810 size_t align, int flags, size_t *left_over,
811 unsigned int *num)
812{
813 int nr_objs;
814 size_t mgmt_size;
815 size_t slab_size = PAGE_SIZE << gfporder;
816
817 /*
818 * The slab management structure can be either off the slab or
819 * on it. For the latter case, the memory allocated for a
820 * slab is used for:
821 *
822 * - The struct slab
823 * - One kmem_bufctl_t for each object
824 * - Padding to respect alignment of @align
825 * - @buffer_size bytes for each object
826 *
827 * If the slab management structure is off the slab, then the
828 * alignment will already be calculated into the size. Because
829 * the slabs are all pages aligned, the objects will be at the
830 * correct alignment when allocated.
831 */
832 if (flags & CFLGS_OFF_SLAB) {
833 mgmt_size = 0;
834 nr_objs = slab_size / buffer_size;
835
836 if (nr_objs > SLAB_LIMIT)
837 nr_objs = SLAB_LIMIT;
838 } else {
839 /*
840 * Ignore padding for the initial guess. The padding
841 * is at most @align-1 bytes, and @buffer_size is at
842 * least @align. In the worst case, this result will
843 * be one greater than the number of objects that fit
844 * into the memory allocation when taking the padding
845 * into account.
846 */
847 nr_objs = (slab_size - sizeof(struct slab)) /
848 (buffer_size + sizeof(kmem_bufctl_t));
849
850 /*
851 * This calculated number will be either the right
852 * amount, or one greater than what we want.
853 */
854 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
855 > slab_size)
856 nr_objs--;
857
858 if (nr_objs > SLAB_LIMIT)
859 nr_objs = SLAB_LIMIT;
860
861 mgmt_size = slab_mgmt_size(nr_objs, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700862 }
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800863 *num = nr_objs;
864 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700865}
866
867#define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
868
Andrew Mortona737b3e2006-03-22 00:08:11 -0800869static void __slab_error(const char *function, struct kmem_cache *cachep,
870 char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700871{
872 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800873 function, cachep->name, msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700874 dump_stack();
875}
876
Paul Menage3395ee02006-12-06 20:32:16 -0800877/*
878 * By default on NUMA we use alien caches to stage the freeing of
879 * objects allocated from other nodes. This causes massive memory
880 * inefficiencies when using fake NUMA setup to split memory into a
881 * large number of small nodes, so it can be disabled on the command
882 * line
883 */
884
885static int use_alien_caches __read_mostly = 1;
886static int __init noaliencache_setup(char *s)
887{
888 use_alien_caches = 0;
889 return 1;
890}
891__setup("noaliencache", noaliencache_setup);
892
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800893#ifdef CONFIG_NUMA
894/*
895 * Special reaping functions for NUMA systems called from cache_reap().
896 * These take care of doing round robin flushing of alien caches (containing
897 * objects freed on different nodes from which they were allocated) and the
898 * flushing of remote pcps by calling drain_node_pages.
899 */
900static DEFINE_PER_CPU(unsigned long, reap_node);
901
902static void init_reap_node(int cpu)
903{
904 int node;
905
906 node = next_node(cpu_to_node(cpu), node_online_map);
907 if (node == MAX_NUMNODES)
Paul Jackson442295c2006-03-22 00:09:11 -0800908 node = first_node(node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800909
Daniel Yeisley7f6b8872006-11-02 22:07:14 -0800910 per_cpu(reap_node, cpu) = node;
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800911}
912
913static void next_reap_node(void)
914{
915 int node = __get_cpu_var(reap_node);
916
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800917 node = next_node(node, node_online_map);
918 if (unlikely(node >= MAX_NUMNODES))
919 node = first_node(node_online_map);
920 __get_cpu_var(reap_node) = node;
921}
922
923#else
924#define init_reap_node(cpu) do { } while (0)
925#define next_reap_node(void) do { } while (0)
926#endif
927
Linus Torvalds1da177e2005-04-16 15:20:36 -0700928/*
929 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
930 * via the workqueue/eventd.
931 * Add the CPU number into the expiration time to minimize the possibility of
932 * the CPUs getting into lockstep and contending for the global cache chain
933 * lock.
934 */
Adrian Bunk897e6792007-07-15 23:38:20 -0700935static void __cpuinit start_cpu_timer(int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700936{
David Howells52bad642006-11-22 14:54:01 +0000937 struct delayed_work *reap_work = &per_cpu(reap_work, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700938
939 /*
940 * When this gets called from do_initcalls via cpucache_init(),
941 * init_workqueues() has already run, so keventd will be setup
942 * at that time.
943 */
David Howells52bad642006-11-22 14:54:01 +0000944 if (keventd_up() && reap_work->work.func == NULL) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800945 init_reap_node(cpu);
David Howells65f27f32006-11-22 14:55:48 +0000946 INIT_DELAYED_WORK(reap_work, cache_reap);
Arjan van de Ven2b284212006-12-10 02:21:28 -0800947 schedule_delayed_work_on(cpu, reap_work,
948 __round_jiffies_relative(HZ, cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700949 }
950}
951
Christoph Lametere498be72005-09-09 13:03:32 -0700952static struct array_cache *alloc_arraycache(int node, int entries,
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800953 int batchcount)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700954{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800955 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700956 struct array_cache *nc = NULL;
957
Christoph Lametere498be72005-09-09 13:03:32 -0700958 nc = kmalloc_node(memsize, GFP_KERNEL, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700959 if (nc) {
960 nc->avail = 0;
961 nc->limit = entries;
962 nc->batchcount = batchcount;
963 nc->touched = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700964 spin_lock_init(&nc->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700965 }
966 return nc;
967}
968
Christoph Lameter3ded1752006-03-25 03:06:44 -0800969/*
970 * Transfer objects in one arraycache to another.
971 * Locking must be handled by the caller.
972 *
973 * Return the number of entries transferred.
974 */
975static int transfer_objects(struct array_cache *to,
976 struct array_cache *from, unsigned int max)
977{
978 /* Figure out how many entries to transfer */
979 int nr = min(min(from->avail, max), to->limit - to->avail);
980
981 if (!nr)
982 return 0;
983
984 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
985 sizeof(void *) *nr);
986
987 from->avail -= nr;
988 to->avail += nr;
989 to->touched = 1;
990 return nr;
991}
992
Christoph Lameter765c4502006-09-27 01:50:08 -0700993#ifndef CONFIG_NUMA
994
995#define drain_alien_cache(cachep, alien) do { } while (0)
996#define reap_alien(cachep, l3) do { } while (0)
997
998static inline struct array_cache **alloc_alien_cache(int node, int limit)
999{
1000 return (struct array_cache **)BAD_ALIEN_MAGIC;
1001}
1002
1003static inline void free_alien_cache(struct array_cache **ac_ptr)
1004{
1005}
1006
1007static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1008{
1009 return 0;
1010}
1011
1012static inline void *alternate_node_alloc(struct kmem_cache *cachep,
1013 gfp_t flags)
1014{
1015 return NULL;
1016}
1017
Christoph Hellwig8b98c162006-12-06 20:32:30 -08001018static inline void *____cache_alloc_node(struct kmem_cache *cachep,
Christoph Lameter765c4502006-09-27 01:50:08 -07001019 gfp_t flags, int nodeid)
1020{
1021 return NULL;
1022}
1023
1024#else /* CONFIG_NUMA */
1025
Christoph Hellwig8b98c162006-12-06 20:32:30 -08001026static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
Paul Jacksonc61afb12006-03-24 03:16:08 -08001027static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
Christoph Lameterdc85da12006-01-18 17:42:36 -08001028
Pekka Enberg5295a742006-02-01 03:05:48 -08001029static struct array_cache **alloc_alien_cache(int node, int limit)
Christoph Lametere498be72005-09-09 13:03:32 -07001030{
1031 struct array_cache **ac_ptr;
Christoph Lameter8ef82862007-02-20 13:57:52 -08001032 int memsize = sizeof(void *) * nr_node_ids;
Christoph Lametere498be72005-09-09 13:03:32 -07001033 int i;
1034
1035 if (limit > 1)
1036 limit = 12;
1037 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
1038 if (ac_ptr) {
1039 for_each_node(i) {
1040 if (i == node || !node_online(i)) {
1041 ac_ptr[i] = NULL;
1042 continue;
1043 }
1044 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
1045 if (!ac_ptr[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001046 for (i--; i <= 0; i--)
Christoph Lametere498be72005-09-09 13:03:32 -07001047 kfree(ac_ptr[i]);
1048 kfree(ac_ptr);
1049 return NULL;
1050 }
1051 }
1052 }
1053 return ac_ptr;
1054}
1055
Pekka Enberg5295a742006-02-01 03:05:48 -08001056static void free_alien_cache(struct array_cache **ac_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -07001057{
1058 int i;
1059
1060 if (!ac_ptr)
1061 return;
Christoph Lametere498be72005-09-09 13:03:32 -07001062 for_each_node(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001063 kfree(ac_ptr[i]);
Christoph Lametere498be72005-09-09 13:03:32 -07001064 kfree(ac_ptr);
1065}
1066
Pekka Enberg343e0d72006-02-01 03:05:50 -08001067static void __drain_alien_cache(struct kmem_cache *cachep,
Pekka Enberg5295a742006-02-01 03:05:48 -08001068 struct array_cache *ac, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07001069{
1070 struct kmem_list3 *rl3 = cachep->nodelists[node];
1071
1072 if (ac->avail) {
1073 spin_lock(&rl3->list_lock);
Christoph Lametere00946f2006-03-25 03:06:45 -08001074 /*
1075 * Stuff objects into the remote nodes shared array first.
1076 * That way we could avoid the overhead of putting the objects
1077 * into the free lists and getting them back later.
1078 */
shin, jacob693f7d32006-04-28 10:54:37 -05001079 if (rl3->shared)
1080 transfer_objects(rl3->shared, ac, ac->limit);
Christoph Lametere00946f2006-03-25 03:06:45 -08001081
Christoph Lameterff694162005-09-22 21:44:02 -07001082 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001083 ac->avail = 0;
1084 spin_unlock(&rl3->list_lock);
1085 }
1086}
1087
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001088/*
1089 * Called from cache_reap() to regularly drain alien caches round robin.
1090 */
1091static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1092{
1093 int node = __get_cpu_var(reap_node);
1094
1095 if (l3->alien) {
1096 struct array_cache *ac = l3->alien[node];
Christoph Lametere00946f2006-03-25 03:06:45 -08001097
1098 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001099 __drain_alien_cache(cachep, ac, node);
1100 spin_unlock_irq(&ac->lock);
1101 }
1102 }
1103}
1104
Andrew Mortona737b3e2006-03-22 00:08:11 -08001105static void drain_alien_cache(struct kmem_cache *cachep,
1106 struct array_cache **alien)
Christoph Lametere498be72005-09-09 13:03:32 -07001107{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001108 int i = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07001109 struct array_cache *ac;
1110 unsigned long flags;
1111
1112 for_each_online_node(i) {
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001113 ac = alien[i];
Christoph Lametere498be72005-09-09 13:03:32 -07001114 if (ac) {
1115 spin_lock_irqsave(&ac->lock, flags);
1116 __drain_alien_cache(cachep, ac, i);
1117 spin_unlock_irqrestore(&ac->lock, flags);
1118 }
1119 }
1120}
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001121
Ingo Molnar873623d2006-07-13 14:44:38 +02001122static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001123{
1124 struct slab *slabp = virt_to_slab(objp);
1125 int nodeid = slabp->nodeid;
1126 struct kmem_list3 *l3;
1127 struct array_cache *alien = NULL;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001128 int node;
1129
1130 node = numa_node_id();
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001131
1132 /*
1133 * Make sure we are not freeing a object from another node to the array
1134 * cache on this cpu.
1135 */
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001136 if (likely(slabp->nodeid == node))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001137 return 0;
1138
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001139 l3 = cachep->nodelists[node];
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001140 STATS_INC_NODEFREES(cachep);
1141 if (l3->alien && l3->alien[nodeid]) {
1142 alien = l3->alien[nodeid];
Ingo Molnar873623d2006-07-13 14:44:38 +02001143 spin_lock(&alien->lock);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001144 if (unlikely(alien->avail == alien->limit)) {
1145 STATS_INC_ACOVERFLOW(cachep);
1146 __drain_alien_cache(cachep, alien, nodeid);
1147 }
1148 alien->entry[alien->avail++] = objp;
1149 spin_unlock(&alien->lock);
1150 } else {
1151 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1152 free_block(cachep, &objp, 1, nodeid);
1153 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1154 }
1155 return 1;
1156}
Christoph Lametere498be72005-09-09 13:03:32 -07001157#endif
1158
Chandra Seetharaman8c78f302006-07-30 03:03:35 -07001159static int __cpuinit cpuup_callback(struct notifier_block *nfb,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001160 unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001161{
1162 long cpu = (long)hcpu;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001163 struct kmem_cache *cachep;
Christoph Lametere498be72005-09-09 13:03:32 -07001164 struct kmem_list3 *l3 = NULL;
1165 int node = cpu_to_node(cpu);
David Howellsea02e3d2007-07-19 01:49:09 -07001166 const int memsize = sizeof(struct kmem_list3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001167
1168 switch (action) {
Heiko Carstens38c3bd92007-05-09 02:34:05 -07001169 case CPU_LOCK_ACQUIRE:
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001170 mutex_lock(&cache_chain_mutex);
Heiko Carstens38c3bd92007-05-09 02:34:05 -07001171 break;
1172 case CPU_UP_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001173 case CPU_UP_PREPARE_FROZEN:
Andrew Mortona737b3e2006-03-22 00:08:11 -08001174 /*
1175 * We need to do this right in the beginning since
Christoph Lametere498be72005-09-09 13:03:32 -07001176 * alloc_arraycache's are going to use this list.
1177 * kmalloc_node allows us to add the slab to the right
1178 * kmem_list3 and not this cpu's kmem_list3
1179 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001180
Christoph Lametere498be72005-09-09 13:03:32 -07001181 list_for_each_entry(cachep, &cache_chain, next) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08001182 /*
1183 * Set up the size64 kmemlist for cpu before we can
Christoph Lametere498be72005-09-09 13:03:32 -07001184 * begin anything. Make sure some other cpu on this
1185 * node has not already allocated this
1186 */
1187 if (!cachep->nodelists[node]) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08001188 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1189 if (!l3)
Christoph Lametere498be72005-09-09 13:03:32 -07001190 goto bad;
1191 kmem_list3_init(l3);
1192 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001193 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -07001194
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001195 /*
1196 * The l3s don't come and go as CPUs come and
1197 * go. cache_chain_mutex is sufficient
1198 * protection here.
1199 */
Christoph Lametere498be72005-09-09 13:03:32 -07001200 cachep->nodelists[node] = l3;
1201 }
1202
1203 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1204 cachep->nodelists[node]->free_limit =
Andrew Mortona737b3e2006-03-22 00:08:11 -08001205 (1 + nr_cpus_node(node)) *
1206 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07001207 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1208 }
1209
Andrew Mortona737b3e2006-03-22 00:08:11 -08001210 /*
1211 * Now we can go ahead with allocating the shared arrays and
1212 * array caches
1213 */
Christoph Lametere498be72005-09-09 13:03:32 -07001214 list_for_each_entry(cachep, &cache_chain, next) {
Tobias Klausercd105df2006-01-08 01:00:59 -08001215 struct array_cache *nc;
Eric Dumazet63109842007-05-06 14:49:28 -07001216 struct array_cache *shared = NULL;
Paul Menage3395ee02006-12-06 20:32:16 -08001217 struct array_cache **alien = NULL;
Tobias Klausercd105df2006-01-08 01:00:59 -08001218
Christoph Lametere498be72005-09-09 13:03:32 -07001219 nc = alloc_arraycache(node, cachep->limit,
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001220 cachep->batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001221 if (!nc)
1222 goto bad;
Eric Dumazet63109842007-05-06 14:49:28 -07001223 if (cachep->shared) {
1224 shared = alloc_arraycache(node,
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001225 cachep->shared * cachep->batchcount,
1226 0xbaadf00d);
Eric Dumazet63109842007-05-06 14:49:28 -07001227 if (!shared)
1228 goto bad;
1229 }
Paul Menage3395ee02006-12-06 20:32:16 -08001230 if (use_alien_caches) {
1231 alien = alloc_alien_cache(node, cachep->limit);
1232 if (!alien)
1233 goto bad;
1234 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001235 cachep->array[cpu] = nc;
Christoph Lametere498be72005-09-09 13:03:32 -07001236 l3 = cachep->nodelists[node];
1237 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07001238
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001239 spin_lock_irq(&l3->list_lock);
1240 if (!l3->shared) {
1241 /*
1242 * We are serialised from CPU_DEAD or
1243 * CPU_UP_CANCELLED by the cpucontrol lock
1244 */
1245 l3->shared = shared;
1246 shared = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07001247 }
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001248#ifdef CONFIG_NUMA
1249 if (!l3->alien) {
1250 l3->alien = alien;
1251 alien = NULL;
1252 }
1253#endif
1254 spin_unlock_irq(&l3->list_lock);
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001255 kfree(shared);
1256 free_alien_cache(alien);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001257 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001258 break;
1259 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001260 case CPU_ONLINE_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001261 start_cpu_timer(cpu);
1262 break;
1263#ifdef CONFIG_HOTPLUG_CPU
Christoph Lameter5830c592007-05-09 02:34:22 -07001264 case CPU_DOWN_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001265 case CPU_DOWN_PREPARE_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001266 /*
1267 * Shutdown cache reaper. Note that the cache_chain_mutex is
1268 * held so that if cache_reap() is invoked it cannot do
1269 * anything expensive but will only modify reap_work
1270 * and reschedule the timer.
1271 */
1272 cancel_rearming_delayed_work(&per_cpu(reap_work, cpu));
1273 /* Now the cache_reaper is guaranteed to be not running. */
1274 per_cpu(reap_work, cpu).work.func = NULL;
1275 break;
1276 case CPU_DOWN_FAILED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001277 case CPU_DOWN_FAILED_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001278 start_cpu_timer(cpu);
1279 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001280 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001281 case CPU_DEAD_FROZEN:
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001282 /*
1283 * Even if all the cpus of a node are down, we don't free the
1284 * kmem_list3 of any cache. This to avoid a race between
1285 * cpu_down, and a kmalloc allocation from another cpu for
1286 * memory from the node of the cpu going down. The list3
1287 * structure is usually allocated from kmem_cache_create() and
1288 * gets destroyed at kmem_cache_destroy().
1289 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001290 /* fall thru */
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08001291#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001292 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001293 case CPU_UP_CANCELED_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001294 list_for_each_entry(cachep, &cache_chain, next) {
1295 struct array_cache *nc;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001296 struct array_cache *shared;
1297 struct array_cache **alien;
Christoph Lametere498be72005-09-09 13:03:32 -07001298 cpumask_t mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001299
Christoph Lametere498be72005-09-09 13:03:32 -07001300 mask = node_to_cpumask(node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001301 /* cpu is dead; no one can alloc from it. */
1302 nc = cachep->array[cpu];
1303 cachep->array[cpu] = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07001304 l3 = cachep->nodelists[node];
1305
1306 if (!l3)
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001307 goto free_array_cache;
Christoph Lametere498be72005-09-09 13:03:32 -07001308
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08001309 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07001310
1311 /* Free limit for this kmem_list3 */
1312 l3->free_limit -= cachep->batchcount;
1313 if (nc)
Christoph Lameterff694162005-09-22 21:44:02 -07001314 free_block(cachep, nc->entry, nc->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001315
1316 if (!cpus_empty(mask)) {
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08001317 spin_unlock_irq(&l3->list_lock);
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001318 goto free_array_cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001319 }
Christoph Lametere498be72005-09-09 13:03:32 -07001320
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001321 shared = l3->shared;
1322 if (shared) {
Eric Dumazet63109842007-05-06 14:49:28 -07001323 free_block(cachep, shared->entry,
1324 shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001325 l3->shared = NULL;
1326 }
Christoph Lametere498be72005-09-09 13:03:32 -07001327
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001328 alien = l3->alien;
1329 l3->alien = NULL;
1330
1331 spin_unlock_irq(&l3->list_lock);
1332
1333 kfree(shared);
1334 if (alien) {
1335 drain_alien_cache(cachep, alien);
1336 free_alien_cache(alien);
Christoph Lametere498be72005-09-09 13:03:32 -07001337 }
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001338free_array_cache:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001339 kfree(nc);
1340 }
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001341 /*
1342 * In the previous loop, all the objects were freed to
1343 * the respective cache's slabs, now we can go ahead and
1344 * shrink each nodelist to its limit.
1345 */
1346 list_for_each_entry(cachep, &cache_chain, next) {
1347 l3 = cachep->nodelists[node];
1348 if (!l3)
1349 continue;
Christoph Lametered11d9e2006-06-30 01:55:45 -07001350 drain_freelist(cachep, l3, l3->free_objects);
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001351 }
Heiko Carstens38c3bd92007-05-09 02:34:05 -07001352 break;
1353 case CPU_LOCK_RELEASE:
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001354 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001355 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001356 }
1357 return NOTIFY_OK;
Andrew Mortona737b3e2006-03-22 00:08:11 -08001358bad:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001359 return NOTIFY_BAD;
1360}
1361
Chandra Seetharaman74b85f32006-06-27 02:54:09 -07001362static struct notifier_block __cpuinitdata cpucache_notifier = {
1363 &cpuup_callback, NULL, 0
1364};
Linus Torvalds1da177e2005-04-16 15:20:36 -07001365
Christoph Lametere498be72005-09-09 13:03:32 -07001366/*
1367 * swap the static kmem_list3 with kmalloced memory
1368 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001369static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1370 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001371{
1372 struct kmem_list3 *ptr;
1373
Christoph Lametere498be72005-09-09 13:03:32 -07001374 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1375 BUG_ON(!ptr);
1376
1377 local_irq_disable();
1378 memcpy(ptr, list, sizeof(struct kmem_list3));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001379 /*
1380 * Do not assume that spinlocks can be initialized via memcpy:
1381 */
1382 spin_lock_init(&ptr->list_lock);
1383
Christoph Lametere498be72005-09-09 13:03:32 -07001384 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1385 cachep->nodelists[nodeid] = ptr;
1386 local_irq_enable();
1387}
1388
Andrew Mortona737b3e2006-03-22 00:08:11 -08001389/*
1390 * Initialisation. Called after the page allocator have been initialised and
1391 * before smp_init().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001392 */
1393void __init kmem_cache_init(void)
1394{
1395 size_t left_over;
1396 struct cache_sizes *sizes;
1397 struct cache_names *names;
Christoph Lametere498be72005-09-09 13:03:32 -07001398 int i;
Jack Steiner07ed76b2006-03-07 21:55:46 -08001399 int order;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001400 int node;
Christoph Lametere498be72005-09-09 13:03:32 -07001401
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001402 if (num_possible_nodes() == 1)
1403 use_alien_caches = 0;
1404
Christoph Lametere498be72005-09-09 13:03:32 -07001405 for (i = 0; i < NUM_INIT_LISTS; i++) {
1406 kmem_list3_init(&initkmem_list3[i]);
1407 if (i < MAX_NUMNODES)
1408 cache_cache.nodelists[i] = NULL;
1409 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001410
1411 /*
1412 * Fragmentation resistance on low memory - only use bigger
1413 * page orders on machines with more than 32MB of memory.
1414 */
1415 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1416 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1417
Linus Torvalds1da177e2005-04-16 15:20:36 -07001418 /* Bootstrap is tricky, because several objects are allocated
1419 * from caches that do not exist yet:
Andrew Mortona737b3e2006-03-22 00:08:11 -08001420 * 1) initialize the cache_cache cache: it contains the struct
1421 * kmem_cache structures of all caches, except cache_cache itself:
1422 * cache_cache is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001423 * Initially an __init data area is used for the head array and the
1424 * kmem_list3 structures, it's replaced with a kmalloc allocated
1425 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001426 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001427 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001428 * An __init data area is used for the head array.
1429 * 3) Create the remaining kmalloc caches, with minimally sized
1430 * head arrays.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001431 * 4) Replace the __init data head arrays for cache_cache and the first
1432 * kmalloc cache with kmalloc allocated arrays.
Christoph Lametere498be72005-09-09 13:03:32 -07001433 * 5) Replace the __init data for kmem_list3 for cache_cache and
1434 * the other cache's with kmalloc allocated memory.
1435 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001436 */
1437
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001438 node = numa_node_id();
1439
Linus Torvalds1da177e2005-04-16 15:20:36 -07001440 /* 1) create the cache_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001441 INIT_LIST_HEAD(&cache_chain);
1442 list_add(&cache_cache.next, &cache_chain);
1443 cache_cache.colour_off = cache_line_size();
1444 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001445 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001446
Eric Dumazet8da34302007-05-06 14:49:29 -07001447 /*
1448 * struct kmem_cache size depends on nr_node_ids, which
1449 * can be less than MAX_NUMNODES.
1450 */
1451 cache_cache.buffer_size = offsetof(struct kmem_cache, nodelists) +
1452 nr_node_ids * sizeof(struct kmem_list3 *);
1453#if DEBUG
1454 cache_cache.obj_size = cache_cache.buffer_size;
1455#endif
Andrew Mortona737b3e2006-03-22 00:08:11 -08001456 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1457 cache_line_size());
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08001458 cache_cache.reciprocal_buffer_size =
1459 reciprocal_value(cache_cache.buffer_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001460
Jack Steiner07ed76b2006-03-07 21:55:46 -08001461 for (order = 0; order < MAX_ORDER; order++) {
1462 cache_estimate(order, cache_cache.buffer_size,
1463 cache_line_size(), 0, &left_over, &cache_cache.num);
1464 if (cache_cache.num)
1465 break;
1466 }
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02001467 BUG_ON(!cache_cache.num);
Jack Steiner07ed76b2006-03-07 21:55:46 -08001468 cache_cache.gfporder = order;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001469 cache_cache.colour = left_over / cache_cache.colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001470 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1471 sizeof(struct slab), cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001472
1473 /* 2+3) create the kmalloc caches */
1474 sizes = malloc_sizes;
1475 names = cache_names;
1476
Andrew Mortona737b3e2006-03-22 00:08:11 -08001477 /*
1478 * Initialize the caches that provide memory for the array cache and the
1479 * kmem_list3 structures first. Without this, further allocations will
1480 * bug.
Christoph Lametere498be72005-09-09 13:03:32 -07001481 */
1482
1483 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001484 sizes[INDEX_AC].cs_size,
1485 ARCH_KMALLOC_MINALIGN,
1486 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001487 NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07001488
Andrew Mortona737b3e2006-03-22 00:08:11 -08001489 if (INDEX_AC != INDEX_L3) {
Christoph Lametere498be72005-09-09 13:03:32 -07001490 sizes[INDEX_L3].cs_cachep =
Andrew Mortona737b3e2006-03-22 00:08:11 -08001491 kmem_cache_create(names[INDEX_L3].name,
1492 sizes[INDEX_L3].cs_size,
1493 ARCH_KMALLOC_MINALIGN,
1494 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001495 NULL);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001496 }
Christoph Lametere498be72005-09-09 13:03:32 -07001497
Ingo Molnare0a42722006-06-23 02:03:46 -07001498 slab_early_init = 0;
1499
Linus Torvalds1da177e2005-04-16 15:20:36 -07001500 while (sizes->cs_size != ULONG_MAX) {
Christoph Lametere498be72005-09-09 13:03:32 -07001501 /*
1502 * For performance, all the general caches are L1 aligned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001503 * This should be particularly beneficial on SMP boxes, as it
1504 * eliminates "false sharing".
1505 * Note for systems short on memory removing the alignment will
Christoph Lametere498be72005-09-09 13:03:32 -07001506 * allow tighter packing of the smaller caches.
1507 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001508 if (!sizes->cs_cachep) {
Christoph Lametere498be72005-09-09 13:03:32 -07001509 sizes->cs_cachep = kmem_cache_create(names->name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001510 sizes->cs_size,
1511 ARCH_KMALLOC_MINALIGN,
1512 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001513 NULL);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001514 }
Christoph Lameter4b51d662007-02-10 01:43:10 -08001515#ifdef CONFIG_ZONE_DMA
1516 sizes->cs_dmacachep = kmem_cache_create(
1517 names->name_dma,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001518 sizes->cs_size,
1519 ARCH_KMALLOC_MINALIGN,
1520 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1521 SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001522 NULL);
Christoph Lameter4b51d662007-02-10 01:43:10 -08001523#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001524 sizes++;
1525 names++;
1526 }
1527 /* 4) Replace the bootstrap head arrays */
1528 {
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001529 struct array_cache *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001530
Linus Torvalds1da177e2005-04-16 15:20:36 -07001531 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001532
Linus Torvalds1da177e2005-04-16 15:20:36 -07001533 local_irq_disable();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001534 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1535 memcpy(ptr, cpu_cache_get(&cache_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001536 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001537 /*
1538 * Do not assume that spinlocks can be initialized via memcpy:
1539 */
1540 spin_lock_init(&ptr->lock);
1541
Linus Torvalds1da177e2005-04-16 15:20:36 -07001542 cache_cache.array[smp_processor_id()] = ptr;
1543 local_irq_enable();
Christoph Lametere498be72005-09-09 13:03:32 -07001544
Linus Torvalds1da177e2005-04-16 15:20:36 -07001545 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001546
Linus Torvalds1da177e2005-04-16 15:20:36 -07001547 local_irq_disable();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001548 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001549 != &initarray_generic.cache);
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001550 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001551 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001552 /*
1553 * Do not assume that spinlocks can be initialized via memcpy:
1554 */
1555 spin_lock_init(&ptr->lock);
1556
Christoph Lametere498be72005-09-09 13:03:32 -07001557 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001558 ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001559 local_irq_enable();
1560 }
Christoph Lametere498be72005-09-09 13:03:32 -07001561 /* 5) Replace the bootstrap kmem_list3's */
1562 {
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001563 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001564
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001565 /* Replace the static kmem_list3 structures for the boot cpu */
1566 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE], node);
1567
1568 for_each_online_node(nid) {
Christoph Lametere498be72005-09-09 13:03:32 -07001569 init_list(malloc_sizes[INDEX_AC].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001570 &initkmem_list3[SIZE_AC + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001571
1572 if (INDEX_AC != INDEX_L3) {
1573 init_list(malloc_sizes[INDEX_L3].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001574 &initkmem_list3[SIZE_L3 + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001575 }
1576 }
1577 }
1578
1579 /* 6) resize the head arrays to their final sizes */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001580 {
Pekka Enberg343e0d72006-02-01 03:05:50 -08001581 struct kmem_cache *cachep;
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001582 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001583 list_for_each_entry(cachep, &cache_chain, next)
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07001584 if (enable_cpucache(cachep))
1585 BUG();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001586 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001587 }
1588
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001589 /* Annotate slab for lockdep -- annotate the malloc caches */
1590 init_lock_keys();
1591
1592
Linus Torvalds1da177e2005-04-16 15:20:36 -07001593 /* Done! */
1594 g_cpucache_up = FULL;
1595
Andrew Mortona737b3e2006-03-22 00:08:11 -08001596 /*
1597 * Register a cpu startup notifier callback that initializes
1598 * cpu_cache_get for all new cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07001599 */
1600 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001601
Andrew Mortona737b3e2006-03-22 00:08:11 -08001602 /*
1603 * The reap timers are started later, with a module init call: That part
1604 * of the kernel is not yet operational.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001605 */
1606}
1607
1608static int __init cpucache_init(void)
1609{
1610 int cpu;
1611
Andrew Mortona737b3e2006-03-22 00:08:11 -08001612 /*
1613 * Register the timers that return unneeded pages to the page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -07001614 */
Christoph Lametere498be72005-09-09 13:03:32 -07001615 for_each_online_cpu(cpu)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001616 start_cpu_timer(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001617 return 0;
1618}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001619__initcall(cpucache_init);
1620
1621/*
1622 * Interface to system's page allocator. No need to hold the cache-lock.
1623 *
1624 * If we requested dmaable memory, we will get it. Even if we
1625 * did not request dmaable memory, we might get it, but that
1626 * would be relatively rare and ignorable.
1627 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001628static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001629{
1630 struct page *page;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001631 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001632 int i;
1633
Luke Yangd6fef9d2006-04-10 22:52:56 -07001634#ifndef CONFIG_MMU
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001635 /*
1636 * Nommu uses slab's for process anonymous memory allocations, and thus
1637 * requires __GFP_COMP to properly refcount higher order allocations
Luke Yangd6fef9d2006-04-10 22:52:56 -07001638 */
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001639 flags |= __GFP_COMP;
Luke Yangd6fef9d2006-04-10 22:52:56 -07001640#endif
Christoph Lameter765c4502006-09-27 01:50:08 -07001641
Christoph Lameter3c517a62006-12-06 20:33:29 -08001642 flags |= cachep->gfpflags;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001643
1644 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001645 if (!page)
1646 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001647
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001648 nr_pages = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001649 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
Christoph Lameter972d1a72006-09-25 23:31:51 -07001650 add_zone_page_state(page_zone(page),
1651 NR_SLAB_RECLAIMABLE, nr_pages);
1652 else
1653 add_zone_page_state(page_zone(page),
1654 NR_SLAB_UNRECLAIMABLE, nr_pages);
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001655 for (i = 0; i < nr_pages; i++)
1656 __SetPageSlab(page + i);
1657 return page_address(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001658}
1659
1660/*
1661 * Interface to system's page release.
1662 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001663static void kmem_freepages(struct kmem_cache *cachep, void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001664{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001665 unsigned long i = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001666 struct page *page = virt_to_page(addr);
1667 const unsigned long nr_freed = i;
1668
Christoph Lameter972d1a72006-09-25 23:31:51 -07001669 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1670 sub_zone_page_state(page_zone(page),
1671 NR_SLAB_RECLAIMABLE, nr_freed);
1672 else
1673 sub_zone_page_state(page_zone(page),
1674 NR_SLAB_UNRECLAIMABLE, nr_freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001675 while (i--) {
Nick Pigginf205b2f2006-03-22 00:08:02 -08001676 BUG_ON(!PageSlab(page));
1677 __ClearPageSlab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001678 page++;
1679 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001680 if (current->reclaim_state)
1681 current->reclaim_state->reclaimed_slab += nr_freed;
1682 free_pages((unsigned long)addr, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001683}
1684
1685static void kmem_rcu_free(struct rcu_head *head)
1686{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001687 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001688 struct kmem_cache *cachep = slab_rcu->cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001689
1690 kmem_freepages(cachep, slab_rcu->addr);
1691 if (OFF_SLAB(cachep))
1692 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1693}
1694
1695#if DEBUG
1696
1697#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001698static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001699 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001700{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001701 int size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001702
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001703 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001704
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001705 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001706 return;
1707
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001708 *addr++ = 0x12345678;
1709 *addr++ = caller;
1710 *addr++ = smp_processor_id();
1711 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001712 {
1713 unsigned long *sptr = &caller;
1714 unsigned long svalue;
1715
1716 while (!kstack_end(sptr)) {
1717 svalue = *sptr++;
1718 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001719 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001720 size -= sizeof(unsigned long);
1721 if (size <= sizeof(unsigned long))
1722 break;
1723 }
1724 }
1725
1726 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001727 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001728}
1729#endif
1730
Pekka Enberg343e0d72006-02-01 03:05:50 -08001731static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001732{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001733 int size = obj_size(cachep);
1734 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001735
1736 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001737 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001738}
1739
1740static void dump_line(char *data, int offset, int limit)
1741{
1742 int i;
Dave Jonesaa83aa42006-09-29 01:59:51 -07001743 unsigned char error = 0;
1744 int bad_count = 0;
1745
Linus Torvalds1da177e2005-04-16 15:20:36 -07001746 printk(KERN_ERR "%03x:", offset);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001747 for (i = 0; i < limit; i++) {
1748 if (data[offset + i] != POISON_FREE) {
1749 error = data[offset + i];
1750 bad_count++;
1751 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001752 printk(" %02x", (unsigned char)data[offset + i]);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001753 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001754 printk("\n");
Dave Jonesaa83aa42006-09-29 01:59:51 -07001755
1756 if (bad_count == 1) {
1757 error ^= POISON_FREE;
1758 if (!(error & (error - 1))) {
1759 printk(KERN_ERR "Single bit error detected. Probably "
1760 "bad RAM.\n");
1761#ifdef CONFIG_X86
1762 printk(KERN_ERR "Run memtest86+ or a similar memory "
1763 "test tool.\n");
1764#else
1765 printk(KERN_ERR "Run a memory test tool.\n");
1766#endif
1767 }
1768 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001769}
1770#endif
1771
1772#if DEBUG
1773
Pekka Enberg343e0d72006-02-01 03:05:50 -08001774static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001775{
1776 int i, size;
1777 char *realobj;
1778
1779 if (cachep->flags & SLAB_RED_ZONE) {
David Woodhouseb46b8f12007-05-08 00:22:59 -07001780 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001781 *dbg_redzone1(cachep, objp),
1782 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001783 }
1784
1785 if (cachep->flags & SLAB_STORE_USER) {
1786 printk(KERN_ERR "Last user: [<%p>]",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001787 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001788 print_symbol("(%s)",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001789 (unsigned long)*dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001790 printk("\n");
1791 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001792 realobj = (char *)objp + obj_offset(cachep);
1793 size = obj_size(cachep);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001794 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001795 int limit;
1796 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001797 if (i + limit > size)
1798 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001799 dump_line(realobj, i, limit);
1800 }
1801}
1802
Pekka Enberg343e0d72006-02-01 03:05:50 -08001803static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001804{
1805 char *realobj;
1806 int size, i;
1807 int lines = 0;
1808
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001809 realobj = (char *)objp + obj_offset(cachep);
1810 size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001811
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001812 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001813 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001814 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001815 exp = POISON_END;
1816 if (realobj[i] != exp) {
1817 int limit;
1818 /* Mismatch ! */
1819 /* Print header */
1820 if (lines == 0) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001821 printk(KERN_ERR
David Howellse94a40c2007-04-02 23:46:28 +01001822 "Slab corruption: %s start=%p, len=%d\n",
1823 cachep->name, realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001824 print_objinfo(cachep, objp, 0);
1825 }
1826 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001827 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001828 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001829 if (i + limit > size)
1830 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001831 dump_line(realobj, i, limit);
1832 i += 16;
1833 lines++;
1834 /* Limit to 5 lines */
1835 if (lines > 5)
1836 break;
1837 }
1838 }
1839 if (lines != 0) {
1840 /* Print some data about the neighboring objects, if they
1841 * exist:
1842 */
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08001843 struct slab *slabp = virt_to_slab(objp);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001844 unsigned int objnr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001845
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001846 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001847 if (objnr) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001848 objp = index_to_obj(cachep, slabp, objnr - 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001849 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001850 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001851 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001852 print_objinfo(cachep, objp, 2);
1853 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001854 if (objnr + 1 < cachep->num) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001855 objp = index_to_obj(cachep, slabp, objnr + 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001856 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001857 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001858 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001859 print_objinfo(cachep, objp, 2);
1860 }
1861 }
1862}
1863#endif
1864
Linus Torvalds1da177e2005-04-16 15:20:36 -07001865#if DEBUG
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001866/**
Randy Dunlap911851e2006-03-22 00:08:14 -08001867 * slab_destroy_objs - destroy a slab and its objects
1868 * @cachep: cache pointer being destroyed
1869 * @slabp: slab pointer being destroyed
1870 *
1871 * Call the registered destructor for each object in a slab that is being
1872 * destroyed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001873 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001874static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001875{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001876 int i;
1877 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001878 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001879
1880 if (cachep->flags & SLAB_POISON) {
1881#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08001882 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1883 OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001884 kernel_map_pages(virt_to_page(objp),
Andrew Mortona737b3e2006-03-22 00:08:11 -08001885 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001886 else
1887 check_poison_obj(cachep, objp);
1888#else
1889 check_poison_obj(cachep, objp);
1890#endif
1891 }
1892 if (cachep->flags & SLAB_RED_ZONE) {
1893 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1894 slab_error(cachep, "start of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001895 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001896 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1897 slab_error(cachep, "end of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001898 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001899 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001900 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001901}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001902#else
Pekka Enberg343e0d72006-02-01 03:05:50 -08001903static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001904{
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001905}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001906#endif
1907
Randy Dunlap911851e2006-03-22 00:08:14 -08001908/**
1909 * slab_destroy - destroy and release all objects in a slab
1910 * @cachep: cache pointer being destroyed
1911 * @slabp: slab pointer being destroyed
1912 *
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001913 * Destroy all the objs in a slab, and release the mem back to the system.
Andrew Mortona737b3e2006-03-22 00:08:11 -08001914 * Before calling the slab must have been unlinked from the cache. The
1915 * cache-lock is not held/needed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001916 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001917static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001918{
1919 void *addr = slabp->s_mem - slabp->colouroff;
1920
1921 slab_destroy_objs(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001922 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1923 struct slab_rcu *slab_rcu;
1924
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001925 slab_rcu = (struct slab_rcu *)slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001926 slab_rcu->cachep = cachep;
1927 slab_rcu->addr = addr;
1928 call_rcu(&slab_rcu->head, kmem_rcu_free);
1929 } else {
1930 kmem_freepages(cachep, addr);
Ingo Molnar873623d2006-07-13 14:44:38 +02001931 if (OFF_SLAB(cachep))
1932 kmem_cache_free(cachep->slabp_cache, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001933 }
1934}
1935
Andrew Mortona737b3e2006-03-22 00:08:11 -08001936/*
1937 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1938 * size of kmem_list3.
1939 */
Andrew Mortona3a02be2007-05-06 14:49:31 -07001940static void __init set_up_list3s(struct kmem_cache *cachep, int index)
Christoph Lametere498be72005-09-09 13:03:32 -07001941{
1942 int node;
1943
1944 for_each_online_node(node) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001945 cachep->nodelists[node] = &initkmem_list3[index + node];
Christoph Lametere498be72005-09-09 13:03:32 -07001946 cachep->nodelists[node]->next_reap = jiffies +
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001947 REAPTIMEOUT_LIST3 +
1948 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -07001949 }
1950}
1951
Christoph Lameter117f6eb2006-09-25 23:31:37 -07001952static void __kmem_cache_destroy(struct kmem_cache *cachep)
1953{
1954 int i;
1955 struct kmem_list3 *l3;
1956
1957 for_each_online_cpu(i)
1958 kfree(cachep->array[i]);
1959
1960 /* NUMA: free the list3 structures */
1961 for_each_online_node(i) {
1962 l3 = cachep->nodelists[i];
1963 if (l3) {
1964 kfree(l3->shared);
1965 free_alien_cache(l3->alien);
1966 kfree(l3);
1967 }
1968 }
1969 kmem_cache_free(&cache_cache, cachep);
1970}
1971
1972
Linus Torvalds1da177e2005-04-16 15:20:36 -07001973/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08001974 * calculate_slab_order - calculate size (page order) of slabs
1975 * @cachep: pointer to the cache that is being created
1976 * @size: size of objects to be created in this cache.
1977 * @align: required alignment for the objects.
1978 * @flags: slab allocation flags
1979 *
1980 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001981 *
1982 * This could be made much more intelligent. For now, try to avoid using
1983 * high order pages for slabs. When the gfp() functions are more friendly
1984 * towards high-order requests, this should be changed.
1985 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001986static size_t calculate_slab_order(struct kmem_cache *cachep,
Randy Dunlapee13d782006-02-01 03:05:53 -08001987 size_t size, size_t align, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001988{
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001989 unsigned long offslab_limit;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001990 size_t left_over = 0;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001991 int gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001992
Christoph Lameter0aa817f2007-05-16 22:11:01 -07001993 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001994 unsigned int num;
1995 size_t remainder;
1996
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001997 cache_estimate(gfporder, size, align, flags, &remainder, &num);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001998 if (!num)
1999 continue;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002000
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002001 if (flags & CFLGS_OFF_SLAB) {
2002 /*
2003 * Max number of objs-per-slab for caches which
2004 * use off-slab slabs. Needed to avoid a possible
2005 * looping condition in cache_grow().
2006 */
2007 offslab_limit = size - sizeof(struct slab);
2008 offslab_limit /= sizeof(kmem_bufctl_t);
2009
2010 if (num > offslab_limit)
2011 break;
2012 }
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002013
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002014 /* Found something acceptable - save it away */
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002015 cachep->num = num;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002016 cachep->gfporder = gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002017 left_over = remainder;
2018
2019 /*
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002020 * A VFS-reclaimable slab tends to have most allocations
2021 * as GFP_NOFS and we really don't want to have to be allocating
2022 * higher-order pages when we are unable to shrink dcache.
2023 */
2024 if (flags & SLAB_RECLAIM_ACCOUNT)
2025 break;
2026
2027 /*
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002028 * Large number of objects is good, but very large slabs are
2029 * currently bad for the gfp()s.
2030 */
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002031 if (gfporder >= slab_break_gfp_order)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002032 break;
2033
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002034 /*
2035 * Acceptable internal fragmentation?
2036 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002037 if (left_over * 8 <= (PAGE_SIZE << gfporder))
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002038 break;
2039 }
2040 return left_over;
2041}
2042
Sam Ravnborg38bdc322007-05-17 23:48:19 +02002043static int __init_refok setup_cpu_cache(struct kmem_cache *cachep)
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002044{
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002045 if (g_cpucache_up == FULL)
2046 return enable_cpucache(cachep);
2047
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002048 if (g_cpucache_up == NONE) {
2049 /*
2050 * Note: the first kmem_cache_create must create the cache
2051 * that's used by kmalloc(24), otherwise the creation of
2052 * further caches will BUG().
2053 */
2054 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2055
2056 /*
2057 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2058 * the first cache, then we need to set up all its list3s,
2059 * otherwise the creation of further caches will BUG().
2060 */
2061 set_up_list3s(cachep, SIZE_AC);
2062 if (INDEX_AC == INDEX_L3)
2063 g_cpucache_up = PARTIAL_L3;
2064 else
2065 g_cpucache_up = PARTIAL_AC;
2066 } else {
2067 cachep->array[smp_processor_id()] =
2068 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
2069
2070 if (g_cpucache_up == PARTIAL_AC) {
2071 set_up_list3s(cachep, SIZE_L3);
2072 g_cpucache_up = PARTIAL_L3;
2073 } else {
2074 int node;
2075 for_each_online_node(node) {
2076 cachep->nodelists[node] =
2077 kmalloc_node(sizeof(struct kmem_list3),
2078 GFP_KERNEL, node);
2079 BUG_ON(!cachep->nodelists[node]);
2080 kmem_list3_init(cachep->nodelists[node]);
2081 }
2082 }
2083 }
2084 cachep->nodelists[numa_node_id()]->next_reap =
2085 jiffies + REAPTIMEOUT_LIST3 +
2086 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2087
2088 cpu_cache_get(cachep)->avail = 0;
2089 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2090 cpu_cache_get(cachep)->batchcount = 1;
2091 cpu_cache_get(cachep)->touched = 0;
2092 cachep->batchcount = 1;
2093 cachep->limit = BOOT_CPUCACHE_ENTRIES;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002094 return 0;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002095}
2096
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002097/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07002098 * kmem_cache_create - Create a cache.
2099 * @name: A string which is used in /proc/slabinfo to identify this cache.
2100 * @size: The size of objects to be created in this cache.
2101 * @align: The required alignment for the objects.
2102 * @flags: SLAB flags
2103 * @ctor: A constructor for the objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002104 *
2105 * Returns a ptr to the cache on success, NULL on failure.
2106 * Cannot be called within a int, but can be interrupted.
Paul Mundt20c2df82007-07-20 10:11:58 +09002107 * The @ctor is run when new pages are allocated by the cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002108 *
2109 * @name must be valid until the cache is destroyed. This implies that
Andrew Mortona737b3e2006-03-22 00:08:11 -08002110 * the module calling this has to destroy the cache before getting unloaded.
2111 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002112 * The flags are
2113 *
2114 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2115 * to catch references to uninitialised memory.
2116 *
2117 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2118 * for buffer overruns.
2119 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002120 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2121 * cacheline. This can be beneficial if you're counting cycles as closely
2122 * as davem.
2123 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002124struct kmem_cache *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002125kmem_cache_create (const char *name, size_t size, size_t align,
Andrew Mortona737b3e2006-03-22 00:08:11 -08002126 unsigned long flags,
Paul Mundt20c2df82007-07-20 10:11:58 +09002127 void (*ctor)(void*, struct kmem_cache *, unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002128{
2129 size_t left_over, slab_size, ralign;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07002130 struct kmem_cache *cachep = NULL, *pc;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002131
2132 /*
2133 * Sanity checks... these are all serious usage bugs.
2134 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002135 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
Paul Mundt20c2df82007-07-20 10:11:58 +09002136 size > KMALLOC_MAX_SIZE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002137 printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__,
2138 name);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002139 BUG();
2140 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002141
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002142 /*
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002143 * We use cache_chain_mutex to ensure a consistent view of
2144 * cpu_online_map as well. Please see cpuup_callback
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002145 */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002146 mutex_lock(&cache_chain_mutex);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002147
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07002148 list_for_each_entry(pc, &cache_chain, next) {
Andrew Morton4f12bb42005-11-07 00:58:00 -08002149 char tmp;
2150 int res;
2151
2152 /*
2153 * This happens when the module gets unloaded and doesn't
2154 * destroy its slab cache and no-one else reuses the vmalloc
2155 * area of the module. Print a warning.
2156 */
Andrew Morton138ae662006-12-06 20:36:41 -08002157 res = probe_kernel_address(pc->name, tmp);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002158 if (res) {
matzeb4169522007-05-06 14:49:52 -07002159 printk(KERN_ERR
2160 "SLAB: cache with size %d has lost its name\n",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002161 pc->buffer_size);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002162 continue;
2163 }
2164
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002165 if (!strcmp(pc->name, name)) {
matzeb4169522007-05-06 14:49:52 -07002166 printk(KERN_ERR
2167 "kmem_cache_create: duplicate cache %s\n", name);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002168 dump_stack();
2169 goto oops;
2170 }
2171 }
2172
Linus Torvalds1da177e2005-04-16 15:20:36 -07002173#if DEBUG
2174 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002175#if FORCED_DEBUG
2176 /*
2177 * Enable redzoning and last user accounting, except for caches with
2178 * large objects, if the increased size would increase the object size
2179 * above the next power of two: caches with object sizes just above a
2180 * power of two have a significant amount of internal fragmentation.
2181 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002182 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2183 2 * sizeof(unsigned long long)))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002184 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002185 if (!(flags & SLAB_DESTROY_BY_RCU))
2186 flags |= SLAB_POISON;
2187#endif
2188 if (flags & SLAB_DESTROY_BY_RCU)
2189 BUG_ON(flags & SLAB_POISON);
2190#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002191 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002192 * Always checks flags, a caller might be expecting debug support which
2193 * isn't available.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002194 */
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002195 BUG_ON(flags & ~CREATE_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002196
Andrew Mortona737b3e2006-03-22 00:08:11 -08002197 /*
2198 * Check that size is in terms of words. This is needed to avoid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002199 * unaligned accesses for some archs when redzoning is used, and makes
2200 * sure any on-slab bufctl's are also correctly aligned.
2201 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002202 if (size & (BYTES_PER_WORD - 1)) {
2203 size += (BYTES_PER_WORD - 1);
2204 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002205 }
2206
Andrew Mortona737b3e2006-03-22 00:08:11 -08002207 /* calculate the final buffer alignment: */
2208
Linus Torvalds1da177e2005-04-16 15:20:36 -07002209 /* 1) arch recommendation: can be overridden for debug */
2210 if (flags & SLAB_HWCACHE_ALIGN) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002211 /*
2212 * Default alignment: as specified by the arch code. Except if
2213 * an object is really small, then squeeze multiple objects into
2214 * one cacheline.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002215 */
2216 ralign = cache_line_size();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002217 while (size <= ralign / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002218 ralign /= 2;
2219 } else {
2220 ralign = BYTES_PER_WORD;
2221 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002222
2223 /*
David Woodhouse87a927c2007-07-04 21:26:44 -04002224 * Redzoning and user store require word alignment or possibly larger.
2225 * Note this will be overridden by architecture or caller mandated
2226 * alignment if either is greater than BYTES_PER_WORD.
Pekka Enbergca5f9702006-09-25 23:31:25 -07002227 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002228 if (flags & SLAB_STORE_USER)
2229 ralign = BYTES_PER_WORD;
2230
2231 if (flags & SLAB_RED_ZONE) {
2232 ralign = REDZONE_ALIGN;
2233 /* If redzoning, ensure that the second redzone is suitably
2234 * aligned, by adjusting the object size accordingly. */
2235 size += REDZONE_ALIGN - 1;
2236 size &= ~(REDZONE_ALIGN - 1);
2237 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002238
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002239 /* 2) arch mandated alignment */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002240 if (ralign < ARCH_SLAB_MINALIGN) {
2241 ralign = ARCH_SLAB_MINALIGN;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002242 }
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002243 /* 3) caller mandated alignment */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002244 if (ralign < align) {
2245 ralign = align;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002246 }
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002247 /* disable debug if necessary */
David Woodhouseb46b8f12007-05-08 00:22:59 -07002248 if (ralign > __alignof__(unsigned long long))
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002249 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002250 /*
Pekka Enbergca5f9702006-09-25 23:31:25 -07002251 * 4) Store it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002252 */
2253 align = ralign;
2254
2255 /* Get cache's description obj. */
Christoph Lametere94b1762006-12-06 20:33:17 -08002256 cachep = kmem_cache_zalloc(&cache_cache, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002257 if (!cachep)
Andrew Morton4f12bb42005-11-07 00:58:00 -08002258 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002259
2260#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002261 cachep->obj_size = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002262
Pekka Enbergca5f9702006-09-25 23:31:25 -07002263 /*
2264 * Both debugging options require word-alignment which is calculated
2265 * into align above.
2266 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002267 if (flags & SLAB_RED_ZONE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002268 /* add space for red zone words */
David Woodhouseb46b8f12007-05-08 00:22:59 -07002269 cachep->obj_offset += sizeof(unsigned long long);
2270 size += 2 * sizeof(unsigned long long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002271 }
2272 if (flags & SLAB_STORE_USER) {
Pekka Enbergca5f9702006-09-25 23:31:25 -07002273 /* user store requires one word storage behind the end of
David Woodhouse87a927c2007-07-04 21:26:44 -04002274 * the real object. But if the second red zone needs to be
2275 * aligned to 64 bits, we must allow that much space.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002276 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002277 if (flags & SLAB_RED_ZONE)
2278 size += REDZONE_ALIGN;
2279 else
2280 size += BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002281 }
2282#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002283 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002284 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2285 cachep->obj_offset += PAGE_SIZE - size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002286 size = PAGE_SIZE;
2287 }
2288#endif
2289#endif
2290
Ingo Molnare0a42722006-06-23 02:03:46 -07002291 /*
2292 * Determine if the slab management is 'on' or 'off' slab.
2293 * (bootstrapping cannot cope with offslab caches so don't do
2294 * it too early on.)
2295 */
2296 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002297 /*
2298 * Size is large, assume best to place the slab management obj
2299 * off-slab (should allow better packing of objs).
2300 */
2301 flags |= CFLGS_OFF_SLAB;
2302
2303 size = ALIGN(size, align);
2304
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002305 left_over = calculate_slab_order(cachep, size, align, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002306
2307 if (!cachep->num) {
matzeb4169522007-05-06 14:49:52 -07002308 printk(KERN_ERR
2309 "kmem_cache_create: couldn't create cache %s.\n", name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002310 kmem_cache_free(&cache_cache, cachep);
2311 cachep = NULL;
Andrew Morton4f12bb42005-11-07 00:58:00 -08002312 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002313 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002314 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2315 + sizeof(struct slab), align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002316
2317 /*
2318 * If the slab has been placed off-slab, and we have enough space then
2319 * move it on-slab. This is at the expense of any extra colouring.
2320 */
2321 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2322 flags &= ~CFLGS_OFF_SLAB;
2323 left_over -= slab_size;
2324 }
2325
2326 if (flags & CFLGS_OFF_SLAB) {
2327 /* really off slab. No need for manual alignment */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002328 slab_size =
2329 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002330 }
2331
2332 cachep->colour_off = cache_line_size();
2333 /* Offset must be a multiple of the alignment. */
2334 if (cachep->colour_off < align)
2335 cachep->colour_off = align;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002336 cachep->colour = left_over / cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002337 cachep->slab_size = slab_size;
2338 cachep->flags = flags;
2339 cachep->gfpflags = 0;
Christoph Lameter4b51d662007-02-10 01:43:10 -08002340 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002341 cachep->gfpflags |= GFP_DMA;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002342 cachep->buffer_size = size;
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08002343 cachep->reciprocal_buffer_size = reciprocal_value(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002344
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002345 if (flags & CFLGS_OFF_SLAB) {
Victor Fuscob2d55072005-09-10 00:26:36 -07002346 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002347 /*
2348 * This is a possibility for one of the malloc_sizes caches.
2349 * But since we go off slab only for object size greater than
2350 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2351 * this should not happen at all.
2352 * But leave a BUG_ON for some lucky dude.
2353 */
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002354 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002355 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002356 cachep->ctor = ctor;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002357 cachep->name = name;
2358
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002359 if (setup_cpu_cache(cachep)) {
2360 __kmem_cache_destroy(cachep);
2361 cachep = NULL;
2362 goto oops;
2363 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002364
Linus Torvalds1da177e2005-04-16 15:20:36 -07002365 /* cache setup completed, link it into the list */
2366 list_add(&cachep->next, &cache_chain);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002367oops:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002368 if (!cachep && (flags & SLAB_PANIC))
2369 panic("kmem_cache_create(): failed to create slab `%s'\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002370 name);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002371 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002372 return cachep;
2373}
2374EXPORT_SYMBOL(kmem_cache_create);
2375
2376#if DEBUG
2377static void check_irq_off(void)
2378{
2379 BUG_ON(!irqs_disabled());
2380}
2381
2382static void check_irq_on(void)
2383{
2384 BUG_ON(irqs_disabled());
2385}
2386
Pekka Enberg343e0d72006-02-01 03:05:50 -08002387static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002388{
2389#ifdef CONFIG_SMP
2390 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002391 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002392#endif
2393}
Christoph Lametere498be72005-09-09 13:03:32 -07002394
Pekka Enberg343e0d72006-02-01 03:05:50 -08002395static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002396{
2397#ifdef CONFIG_SMP
2398 check_irq_off();
2399 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2400#endif
2401}
2402
Linus Torvalds1da177e2005-04-16 15:20:36 -07002403#else
2404#define check_irq_off() do { } while(0)
2405#define check_irq_on() do { } while(0)
2406#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002407#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002408#endif
2409
Christoph Lameteraab22072006-03-22 00:09:06 -08002410static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2411 struct array_cache *ac,
2412 int force, int node);
2413
Linus Torvalds1da177e2005-04-16 15:20:36 -07002414static void do_drain(void *arg)
2415{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002416 struct kmem_cache *cachep = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002417 struct array_cache *ac;
Christoph Lameterff694162005-09-22 21:44:02 -07002418 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002419
2420 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002421 ac = cpu_cache_get(cachep);
Christoph Lameterff694162005-09-22 21:44:02 -07002422 spin_lock(&cachep->nodelists[node]->list_lock);
2423 free_block(cachep, ac->entry, ac->avail, node);
2424 spin_unlock(&cachep->nodelists[node]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002425 ac->avail = 0;
2426}
2427
Pekka Enberg343e0d72006-02-01 03:05:50 -08002428static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002429{
Christoph Lametere498be72005-09-09 13:03:32 -07002430 struct kmem_list3 *l3;
2431 int node;
2432
Andrew Mortona07fa392006-03-22 00:08:17 -08002433 on_each_cpu(do_drain, cachep, 1, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002434 check_irq_on();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002435 for_each_online_node(node) {
Christoph Lametere498be72005-09-09 13:03:32 -07002436 l3 = cachep->nodelists[node];
Roland Dreiera4523a82006-05-15 11:41:00 -07002437 if (l3 && l3->alien)
2438 drain_alien_cache(cachep, l3->alien);
2439 }
2440
2441 for_each_online_node(node) {
2442 l3 = cachep->nodelists[node];
2443 if (l3)
Christoph Lameteraab22072006-03-22 00:09:06 -08002444 drain_array(cachep, l3, l3->shared, 1, node);
Christoph Lametere498be72005-09-09 13:03:32 -07002445 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002446}
2447
Christoph Lametered11d9e2006-06-30 01:55:45 -07002448/*
2449 * Remove slabs from the list of free slabs.
2450 * Specify the number of slabs to drain in tofree.
2451 *
2452 * Returns the actual number of slabs released.
2453 */
2454static int drain_freelist(struct kmem_cache *cache,
2455 struct kmem_list3 *l3, int tofree)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002456{
Christoph Lametered11d9e2006-06-30 01:55:45 -07002457 struct list_head *p;
2458 int nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002459 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002460
Christoph Lametered11d9e2006-06-30 01:55:45 -07002461 nr_freed = 0;
2462 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002463
Christoph Lametered11d9e2006-06-30 01:55:45 -07002464 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002465 p = l3->slabs_free.prev;
Christoph Lametered11d9e2006-06-30 01:55:45 -07002466 if (p == &l3->slabs_free) {
2467 spin_unlock_irq(&l3->list_lock);
2468 goto out;
2469 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002470
Christoph Lametered11d9e2006-06-30 01:55:45 -07002471 slabp = list_entry(p, struct slab, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002472#if DEBUG
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002473 BUG_ON(slabp->inuse);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002474#endif
2475 list_del(&slabp->list);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002476 /*
2477 * Safe to drop the lock. The slab is no longer linked
2478 * to the cache.
2479 */
2480 l3->free_objects -= cache->num;
Christoph Lametere498be72005-09-09 13:03:32 -07002481 spin_unlock_irq(&l3->list_lock);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002482 slab_destroy(cache, slabp);
2483 nr_freed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002484 }
Christoph Lametered11d9e2006-06-30 01:55:45 -07002485out:
2486 return nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002487}
2488
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002489/* Called with cache_chain_mutex held to protect against cpu hotplug */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002490static int __cache_shrink(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07002491{
2492 int ret = 0, i = 0;
2493 struct kmem_list3 *l3;
2494
2495 drain_cpu_caches(cachep);
2496
2497 check_irq_on();
2498 for_each_online_node(i) {
2499 l3 = cachep->nodelists[i];
Christoph Lametered11d9e2006-06-30 01:55:45 -07002500 if (!l3)
2501 continue;
2502
2503 drain_freelist(cachep, l3, l3->free_objects);
2504
2505 ret += !list_empty(&l3->slabs_full) ||
2506 !list_empty(&l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07002507 }
2508 return (ret ? 1 : 0);
2509}
2510
Linus Torvalds1da177e2005-04-16 15:20:36 -07002511/**
2512 * kmem_cache_shrink - Shrink a cache.
2513 * @cachep: The cache to shrink.
2514 *
2515 * Releases as many slabs as possible for a cache.
2516 * To help debugging, a zero exit status indicates all slabs were released.
2517 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002518int kmem_cache_shrink(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002519{
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002520 int ret;
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002521 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002522
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002523 mutex_lock(&cache_chain_mutex);
2524 ret = __cache_shrink(cachep);
2525 mutex_unlock(&cache_chain_mutex);
2526 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002527}
2528EXPORT_SYMBOL(kmem_cache_shrink);
2529
2530/**
2531 * kmem_cache_destroy - delete a cache
2532 * @cachep: the cache to destroy
2533 *
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08002534 * Remove a &struct kmem_cache object from the slab cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002535 *
2536 * It is expected this function will be called by a module when it is
2537 * unloaded. This will remove the cache completely, and avoid a duplicate
2538 * cache being allocated each time a module is loaded and unloaded, if the
2539 * module doesn't have persistent in-kernel storage across loads and unloads.
2540 *
2541 * The cache must be empty before calling this function.
2542 *
2543 * The caller must guarantee that noone will allocate memory from the cache
2544 * during the kmem_cache_destroy().
2545 */
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002546void kmem_cache_destroy(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002547{
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002548 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002549
Linus Torvalds1da177e2005-04-16 15:20:36 -07002550 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002551 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002552 /*
2553 * the chain is never empty, cache_cache is never destroyed
2554 */
2555 list_del(&cachep->next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002556 if (__cache_shrink(cachep)) {
2557 slab_error(cachep, "Can't free all objects");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002558 list_add(&cachep->next, &cache_chain);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002559 mutex_unlock(&cache_chain_mutex);
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002560 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002561 }
2562
2563 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
Paul E. McKenneyfbd568a3e2005-05-01 08:59:04 -07002564 synchronize_rcu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002565
Christoph Lameter117f6eb2006-09-25 23:31:37 -07002566 __kmem_cache_destroy(cachep);
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002567 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002568}
2569EXPORT_SYMBOL(kmem_cache_destroy);
2570
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002571/*
2572 * Get the memory for a slab management obj.
2573 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2574 * always come from malloc_sizes caches. The slab descriptor cannot
2575 * come from the same cache which is getting created because,
2576 * when we are searching for an appropriate cache for these
2577 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2578 * If we are creating a malloc_sizes cache here it would not be visible to
2579 * kmem_find_general_cachep till the initialization is complete.
2580 * Hence we cannot have slabp_cache same as the original cache.
2581 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002582static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002583 int colour_off, gfp_t local_flags,
2584 int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002585{
2586 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002587
Linus Torvalds1da177e2005-04-16 15:20:36 -07002588 if (OFF_SLAB(cachep)) {
2589 /* Slab management obj is off-slab. */
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002590 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
Christoph Lameter3c517a62006-12-06 20:33:29 -08002591 local_flags & ~GFP_THISNODE, nodeid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002592 if (!slabp)
2593 return NULL;
2594 } else {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002595 slabp = objp + colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002596 colour_off += cachep->slab_size;
2597 }
2598 slabp->inuse = 0;
2599 slabp->colouroff = colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002600 slabp->s_mem = objp + colour_off;
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002601 slabp->nodeid = nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002602 return slabp;
2603}
2604
2605static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2606{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002607 return (kmem_bufctl_t *) (slabp + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002608}
2609
Pekka Enberg343e0d72006-02-01 03:05:50 -08002610static void cache_init_objs(struct kmem_cache *cachep,
Christoph Lametera35afb82007-05-16 22:10:57 -07002611 struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002612{
2613 int i;
2614
2615 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002616 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002617#if DEBUG
2618 /* need to poison the objs? */
2619 if (cachep->flags & SLAB_POISON)
2620 poison_obj(cachep, objp, POISON_FREE);
2621 if (cachep->flags & SLAB_STORE_USER)
2622 *dbg_userword(cachep, objp) = NULL;
2623
2624 if (cachep->flags & SLAB_RED_ZONE) {
2625 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2626 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2627 }
2628 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002629 * Constructors are not allowed to allocate memory from the same
2630 * cache which they are a constructor for. Otherwise, deadlock.
2631 * They must also be threaded.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002632 */
2633 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002634 cachep->ctor(objp + obj_offset(cachep), cachep,
Christoph Lametera35afb82007-05-16 22:10:57 -07002635 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002636
2637 if (cachep->flags & SLAB_RED_ZONE) {
2638 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2639 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002640 " end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002641 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2642 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002643 " start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002644 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08002645 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2646 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002647 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002648 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002649#else
2650 if (cachep->ctor)
Christoph Lametera35afb82007-05-16 22:10:57 -07002651 cachep->ctor(objp, cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002652#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002653 slab_bufctl(slabp)[i] = i + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002654 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002655 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002656 slabp->free = 0;
2657}
2658
Pekka Enberg343e0d72006-02-01 03:05:50 -08002659static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002660{
Christoph Lameter4b51d662007-02-10 01:43:10 -08002661 if (CONFIG_ZONE_DMA_FLAG) {
2662 if (flags & GFP_DMA)
2663 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2664 else
2665 BUG_ON(cachep->gfpflags & GFP_DMA);
2666 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002667}
2668
Andrew Mortona737b3e2006-03-22 00:08:11 -08002669static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2670 int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002671{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002672 void *objp = index_to_obj(cachep, slabp, slabp->free);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002673 kmem_bufctl_t next;
2674
2675 slabp->inuse++;
2676 next = slab_bufctl(slabp)[slabp->free];
2677#if DEBUG
2678 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2679 WARN_ON(slabp->nodeid != nodeid);
2680#endif
2681 slabp->free = next;
2682
2683 return objp;
2684}
2685
Andrew Mortona737b3e2006-03-22 00:08:11 -08002686static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2687 void *objp, int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002688{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002689 unsigned int objnr = obj_to_index(cachep, slabp, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002690
2691#if DEBUG
2692 /* Verify that the slab belongs to the intended node */
2693 WARN_ON(slabp->nodeid != nodeid);
2694
Al Viro871751e2006-03-25 03:06:39 -08002695 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
Matthew Dobson78d382d2006-02-01 03:05:47 -08002696 printk(KERN_ERR "slab: double free detected in cache "
Andrew Mortona737b3e2006-03-22 00:08:11 -08002697 "'%s', objp %p\n", cachep->name, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002698 BUG();
2699 }
2700#endif
2701 slab_bufctl(slabp)[objnr] = slabp->free;
2702 slabp->free = objnr;
2703 slabp->inuse--;
2704}
2705
Pekka Enberg47768742006-06-23 02:03:07 -07002706/*
2707 * Map pages beginning at addr to the given cache and slab. This is required
2708 * for the slab allocator to be able to lookup the cache and slab of a
2709 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2710 */
2711static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2712 void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002713{
Pekka Enberg47768742006-06-23 02:03:07 -07002714 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002715 struct page *page;
2716
Pekka Enberg47768742006-06-23 02:03:07 -07002717 page = virt_to_page(addr);
Nick Piggin84097512006-03-22 00:08:34 -08002718
Pekka Enberg47768742006-06-23 02:03:07 -07002719 nr_pages = 1;
Nick Piggin84097512006-03-22 00:08:34 -08002720 if (likely(!PageCompound(page)))
Pekka Enberg47768742006-06-23 02:03:07 -07002721 nr_pages <<= cache->gfporder;
2722
Linus Torvalds1da177e2005-04-16 15:20:36 -07002723 do {
Pekka Enberg47768742006-06-23 02:03:07 -07002724 page_set_cache(page, cache);
2725 page_set_slab(page, slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002726 page++;
Pekka Enberg47768742006-06-23 02:03:07 -07002727 } while (--nr_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002728}
2729
2730/*
2731 * Grow (by 1) the number of slabs within a cache. This is called by
2732 * kmem_cache_alloc() when there are no active objs left in a cache.
2733 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002734static int cache_grow(struct kmem_cache *cachep,
2735 gfp_t flags, int nodeid, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002736{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002737 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002738 size_t offset;
2739 gfp_t local_flags;
Christoph Lametere498be72005-09-09 13:03:32 -07002740 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002741
Andrew Mortona737b3e2006-03-22 00:08:11 -08002742 /*
2743 * Be lazy and only check for valid flags here, keeping it out of the
2744 * critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002745 */
Christoph Lameterd07dbea2007-07-17 04:03:23 -07002746 BUG_ON(flags & ~(GFP_DMA | __GFP_ZERO | GFP_LEVEL_MASK));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002747
Christoph Lametera06d72c2006-12-06 20:33:12 -08002748 local_flags = (flags & GFP_LEVEL_MASK);
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002749 /* Take the l3 list lock to change the colour_next on this node */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002750 check_irq_off();
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002751 l3 = cachep->nodelists[nodeid];
2752 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002753
2754 /* Get colour for the slab, and cal the next value. */
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002755 offset = l3->colour_next;
2756 l3->colour_next++;
2757 if (l3->colour_next >= cachep->colour)
2758 l3->colour_next = 0;
2759 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002760
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002761 offset *= cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002762
2763 if (local_flags & __GFP_WAIT)
2764 local_irq_enable();
2765
2766 /*
2767 * The test for missing atomic flag is performed here, rather than
2768 * the more obvious place, simply to reduce the critical path length
2769 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2770 * will eventually be caught here (where it matters).
2771 */
2772 kmem_flagcheck(cachep, flags);
2773
Andrew Mortona737b3e2006-03-22 00:08:11 -08002774 /*
2775 * Get mem for the objs. Attempt to allocate a physical page from
2776 * 'nodeid'.
Christoph Lametere498be72005-09-09 13:03:32 -07002777 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002778 if (!objp)
Andrew Mortonb8c1c5d2007-07-24 12:02:40 -07002779 objp = kmem_getpages(cachep, local_flags, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002780 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002781 goto failed;
2782
2783 /* Get slab management. */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002784 slabp = alloc_slabmgmt(cachep, objp, offset,
2785 local_flags & ~GFP_THISNODE, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002786 if (!slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002787 goto opps1;
2788
Christoph Lametere498be72005-09-09 13:03:32 -07002789 slabp->nodeid = nodeid;
Pekka Enberg47768742006-06-23 02:03:07 -07002790 slab_map_pages(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002791
Christoph Lametera35afb82007-05-16 22:10:57 -07002792 cache_init_objs(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002793
2794 if (local_flags & __GFP_WAIT)
2795 local_irq_disable();
2796 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002797 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002798
2799 /* Make slab active. */
Christoph Lametere498be72005-09-09 13:03:32 -07002800 list_add_tail(&slabp->list, &(l3->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002801 STATS_INC_GROWN(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002802 l3->free_objects += cachep->num;
2803 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002804 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002805opps1:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002806 kmem_freepages(cachep, objp);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002807failed:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002808 if (local_flags & __GFP_WAIT)
2809 local_irq_disable();
2810 return 0;
2811}
2812
2813#if DEBUG
2814
2815/*
2816 * Perform extra freeing checks:
2817 * - detect bad pointers.
2818 * - POISON/RED_ZONE checking
Linus Torvalds1da177e2005-04-16 15:20:36 -07002819 */
2820static void kfree_debugcheck(const void *objp)
2821{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002822 if (!virt_addr_valid(objp)) {
2823 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002824 (unsigned long)objp);
2825 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002826 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002827}
2828
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002829static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2830{
David Woodhouseb46b8f12007-05-08 00:22:59 -07002831 unsigned long long redzone1, redzone2;
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002832
2833 redzone1 = *dbg_redzone1(cache, obj);
2834 redzone2 = *dbg_redzone2(cache, obj);
2835
2836 /*
2837 * Redzone is ok.
2838 */
2839 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2840 return;
2841
2842 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2843 slab_error(cache, "double free detected");
2844 else
2845 slab_error(cache, "memory outside object was overwritten");
2846
David Woodhouseb46b8f12007-05-08 00:22:59 -07002847 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002848 obj, redzone1, redzone2);
2849}
2850
Pekka Enberg343e0d72006-02-01 03:05:50 -08002851static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002852 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002853{
2854 struct page *page;
2855 unsigned int objnr;
2856 struct slab *slabp;
2857
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002858 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002859 kfree_debugcheck(objp);
Christoph Lameterb49af682007-05-06 14:49:41 -07002860 page = virt_to_head_page(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002861
Pekka Enberg065d41c2005-11-13 16:06:46 -08002862 slabp = page_get_slab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002863
2864 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002865 verify_redzone_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002866 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2867 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2868 }
2869 if (cachep->flags & SLAB_STORE_USER)
2870 *dbg_userword(cachep, objp) = caller;
2871
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002872 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002873
2874 BUG_ON(objnr >= cachep->num);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002875 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002876
Al Viro871751e2006-03-25 03:06:39 -08002877#ifdef CONFIG_DEBUG_SLAB_LEAK
2878 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2879#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002880 if (cachep->flags & SLAB_POISON) {
2881#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08002882 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002883 store_stackinfo(cachep, objp, (unsigned long)caller);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002884 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002885 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002886 } else {
2887 poison_obj(cachep, objp, POISON_FREE);
2888 }
2889#else
2890 poison_obj(cachep, objp, POISON_FREE);
2891#endif
2892 }
2893 return objp;
2894}
2895
Pekka Enberg343e0d72006-02-01 03:05:50 -08002896static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002897{
2898 kmem_bufctl_t i;
2899 int entries = 0;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002900
Linus Torvalds1da177e2005-04-16 15:20:36 -07002901 /* Check slab's freelist to see if this obj is there. */
2902 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2903 entries++;
2904 if (entries > cachep->num || i >= cachep->num)
2905 goto bad;
2906 }
2907 if (entries != cachep->num - slabp->inuse) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002908bad:
2909 printk(KERN_ERR "slab: Internal list corruption detected in "
2910 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2911 cachep->name, cachep->num, slabp, slabp->inuse);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002912 for (i = 0;
Linus Torvalds264132b2006-03-06 12:10:07 -08002913 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002914 i++) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002915 if (i % 16 == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002916 printk("\n%03x:", i);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002917 printk(" %02x", ((unsigned char *)slabp)[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002918 }
2919 printk("\n");
2920 BUG();
2921 }
2922}
2923#else
2924#define kfree_debugcheck(x) do { } while(0)
2925#define cache_free_debugcheck(x,objp,z) (objp)
2926#define check_slabp(x,y) do { } while(0)
2927#endif
2928
Pekka Enberg343e0d72006-02-01 03:05:50 -08002929static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002930{
2931 int batchcount;
2932 struct kmem_list3 *l3;
2933 struct array_cache *ac;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002934 int node;
2935
2936 node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002937
2938 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002939 ac = cpu_cache_get(cachep);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002940retry:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002941 batchcount = ac->batchcount;
2942 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002943 /*
2944 * If there was little recent activity on this cache, then
2945 * perform only a partial refill. Otherwise we could generate
2946 * refill bouncing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002947 */
2948 batchcount = BATCHREFILL_LIMIT;
2949 }
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002950 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002951
Christoph Lametere498be72005-09-09 13:03:32 -07002952 BUG_ON(ac->avail > 0 || !l3);
2953 spin_lock(&l3->list_lock);
2954
Christoph Lameter3ded1752006-03-25 03:06:44 -08002955 /* See if we can refill from the shared array */
2956 if (l3->shared && transfer_objects(ac, l3->shared, batchcount))
2957 goto alloc_done;
2958
Linus Torvalds1da177e2005-04-16 15:20:36 -07002959 while (batchcount > 0) {
2960 struct list_head *entry;
2961 struct slab *slabp;
2962 /* Get slab alloc is to come from. */
2963 entry = l3->slabs_partial.next;
2964 if (entry == &l3->slabs_partial) {
2965 l3->free_touched = 1;
2966 entry = l3->slabs_free.next;
2967 if (entry == &l3->slabs_free)
2968 goto must_grow;
2969 }
2970
2971 slabp = list_entry(entry, struct slab, list);
2972 check_slabp(cachep, slabp);
2973 check_spinlock_acquired(cachep);
Pekka Enberg714b8172007-05-06 14:49:03 -07002974
2975 /*
2976 * The slab was either on partial or free list so
2977 * there must be at least one object available for
2978 * allocation.
2979 */
2980 BUG_ON(slabp->inuse < 0 || slabp->inuse >= cachep->num);
2981
Linus Torvalds1da177e2005-04-16 15:20:36 -07002982 while (slabp->inuse < cachep->num && batchcount--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002983 STATS_INC_ALLOCED(cachep);
2984 STATS_INC_ACTIVE(cachep);
2985 STATS_SET_HIGH(cachep);
2986
Matthew Dobson78d382d2006-02-01 03:05:47 -08002987 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002988 node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002989 }
2990 check_slabp(cachep, slabp);
2991
2992 /* move slabp to correct slabp list: */
2993 list_del(&slabp->list);
2994 if (slabp->free == BUFCTL_END)
2995 list_add(&slabp->list, &l3->slabs_full);
2996 else
2997 list_add(&slabp->list, &l3->slabs_partial);
2998 }
2999
Andrew Mortona737b3e2006-03-22 00:08:11 -08003000must_grow:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003001 l3->free_objects -= ac->avail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003002alloc_done:
Christoph Lametere498be72005-09-09 13:03:32 -07003003 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003004
3005 if (unlikely(!ac->avail)) {
3006 int x;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003007 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07003008
Andrew Mortona737b3e2006-03-22 00:08:11 -08003009 /* cache_grow can reenable interrupts, then ac could change. */
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003010 ac = cpu_cache_get(cachep);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003011 if (!x && ac->avail == 0) /* no objects in sight? abort */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003012 return NULL;
3013
Andrew Mortona737b3e2006-03-22 00:08:11 -08003014 if (!ac->avail) /* objects refilled by interrupt? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003015 goto retry;
3016 }
3017 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07003018 return ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003019}
3020
Andrew Mortona737b3e2006-03-22 00:08:11 -08003021static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3022 gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003023{
3024 might_sleep_if(flags & __GFP_WAIT);
3025#if DEBUG
3026 kmem_flagcheck(cachep, flags);
3027#endif
3028}
3029
3030#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003031static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3032 gfp_t flags, void *objp, void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003033{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003034 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003035 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003036 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003037#ifdef CONFIG_DEBUG_PAGEALLOC
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003038 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003039 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003040 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003041 else
3042 check_poison_obj(cachep, objp);
3043#else
3044 check_poison_obj(cachep, objp);
3045#endif
3046 poison_obj(cachep, objp, POISON_INUSE);
3047 }
3048 if (cachep->flags & SLAB_STORE_USER)
3049 *dbg_userword(cachep, objp) = caller;
3050
3051 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003052 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3053 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3054 slab_error(cachep, "double free, or memory outside"
3055 " object was overwritten");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003056 printk(KERN_ERR
David Woodhouseb46b8f12007-05-08 00:22:59 -07003057 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08003058 objp, *dbg_redzone1(cachep, objp),
3059 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003060 }
3061 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3062 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3063 }
Al Viro871751e2006-03-25 03:06:39 -08003064#ifdef CONFIG_DEBUG_SLAB_LEAK
3065 {
3066 struct slab *slabp;
3067 unsigned objnr;
3068
Christoph Lameterb49af682007-05-06 14:49:41 -07003069 slabp = page_get_slab(virt_to_head_page(objp));
Al Viro871751e2006-03-25 03:06:39 -08003070 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3071 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3072 }
3073#endif
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003074 objp += obj_offset(cachep);
Christoph Lameter4f104932007-05-06 14:50:17 -07003075 if (cachep->ctor && cachep->flags & SLAB_POISON)
Christoph Lametera35afb82007-05-16 22:10:57 -07003076 cachep->ctor(objp, cachep, 0);
Kevin Hilmana44b56d2006-12-06 20:32:11 -08003077#if ARCH_SLAB_MINALIGN
3078 if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) {
3079 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3080 objp, ARCH_SLAB_MINALIGN);
3081 }
3082#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003083 return objp;
3084}
3085#else
3086#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3087#endif
3088
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003089#ifdef CONFIG_FAILSLAB
3090
3091static struct failslab_attr {
3092
3093 struct fault_attr attr;
3094
3095 u32 ignore_gfp_wait;
3096#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3097 struct dentry *ignore_gfp_wait_file;
3098#endif
3099
3100} failslab = {
3101 .attr = FAULT_ATTR_INITIALIZER,
Don Mullis6b1b60f2006-12-08 02:39:53 -08003102 .ignore_gfp_wait = 1,
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003103};
3104
3105static int __init setup_failslab(char *str)
3106{
3107 return setup_fault_attr(&failslab.attr, str);
3108}
3109__setup("failslab=", setup_failslab);
3110
3111static int should_failslab(struct kmem_cache *cachep, gfp_t flags)
3112{
3113 if (cachep == &cache_cache)
3114 return 0;
3115 if (flags & __GFP_NOFAIL)
3116 return 0;
3117 if (failslab.ignore_gfp_wait && (flags & __GFP_WAIT))
3118 return 0;
3119
3120 return should_fail(&failslab.attr, obj_size(cachep));
3121}
3122
3123#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3124
3125static int __init failslab_debugfs(void)
3126{
3127 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
3128 struct dentry *dir;
3129 int err;
3130
Akinobu Mita824ebef2007-05-06 14:49:58 -07003131 err = init_fault_attr_dentries(&failslab.attr, "failslab");
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003132 if (err)
3133 return err;
3134 dir = failslab.attr.dentries.dir;
3135
3136 failslab.ignore_gfp_wait_file =
3137 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3138 &failslab.ignore_gfp_wait);
3139
3140 if (!failslab.ignore_gfp_wait_file) {
3141 err = -ENOMEM;
3142 debugfs_remove(failslab.ignore_gfp_wait_file);
3143 cleanup_fault_attr_dentries(&failslab.attr);
3144 }
3145
3146 return err;
3147}
3148
3149late_initcall(failslab_debugfs);
3150
3151#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3152
3153#else /* CONFIG_FAILSLAB */
3154
3155static inline int should_failslab(struct kmem_cache *cachep, gfp_t flags)
3156{
3157 return 0;
3158}
3159
3160#endif /* CONFIG_FAILSLAB */
3161
Pekka Enberg343e0d72006-02-01 03:05:50 -08003162static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003163{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003164 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003165 struct array_cache *ac;
3166
Alok N Kataria5c382302005-09-27 21:45:46 -07003167 check_irq_off();
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003168
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003169 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003170 if (likely(ac->avail)) {
3171 STATS_INC_ALLOCHIT(cachep);
3172 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07003173 objp = ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003174 } else {
3175 STATS_INC_ALLOCMISS(cachep);
3176 objp = cache_alloc_refill(cachep, flags);
3177 }
Alok N Kataria5c382302005-09-27 21:45:46 -07003178 return objp;
3179}
3180
Christoph Lametere498be72005-09-09 13:03:32 -07003181#ifdef CONFIG_NUMA
3182/*
Paul Jacksonb2455392006-03-24 03:16:12 -08003183 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
Paul Jacksonc61afb12006-03-24 03:16:08 -08003184 *
3185 * If we are in_interrupt, then process context, including cpusets and
3186 * mempolicy, may not apply and should not be used for allocation policy.
3187 */
3188static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3189{
3190 int nid_alloc, nid_here;
3191
Christoph Lameter765c4502006-09-27 01:50:08 -07003192 if (in_interrupt() || (flags & __GFP_THISNODE))
Paul Jacksonc61afb12006-03-24 03:16:08 -08003193 return NULL;
3194 nid_alloc = nid_here = numa_node_id();
3195 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3196 nid_alloc = cpuset_mem_spread_node();
3197 else if (current->mempolicy)
3198 nid_alloc = slab_node(current->mempolicy);
3199 if (nid_alloc != nid_here)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003200 return ____cache_alloc_node(cachep, flags, nid_alloc);
Paul Jacksonc61afb12006-03-24 03:16:08 -08003201 return NULL;
3202}
3203
3204/*
Christoph Lameter765c4502006-09-27 01:50:08 -07003205 * Fallback function if there was no memory available and no objects on a
Christoph Lameter3c517a62006-12-06 20:33:29 -08003206 * certain node and fall back is permitted. First we scan all the
3207 * available nodelists for available objects. If that fails then we
3208 * perform an allocation without specifying a node. This allows the page
3209 * allocator to do its reclaim / fallback magic. We then insert the
3210 * slab into the proper nodelist and then allocate from it.
Christoph Lameter765c4502006-09-27 01:50:08 -07003211 */
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003212static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
Christoph Lameter765c4502006-09-27 01:50:08 -07003213{
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003214 struct zonelist *zonelist;
3215 gfp_t local_flags;
Christoph Lameter765c4502006-09-27 01:50:08 -07003216 struct zone **z;
3217 void *obj = NULL;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003218 int nid;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003219
3220 if (flags & __GFP_THISNODE)
3221 return NULL;
3222
3223 zonelist = &NODE_DATA(slab_node(current->mempolicy))
3224 ->node_zonelists[gfp_zone(flags)];
3225 local_flags = (flags & GFP_LEVEL_MASK);
Christoph Lameter765c4502006-09-27 01:50:08 -07003226
Christoph Lameter3c517a62006-12-06 20:33:29 -08003227retry:
3228 /*
3229 * Look through allowed nodes for objects available
3230 * from existing per node queues.
3231 */
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003232 for (z = zonelist->zones; *z && !obj; z++) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003233 nid = zone_to_nid(*z);
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003234
Paul Jackson02a0e532006-12-13 00:34:25 -08003235 if (cpuset_zone_allowed_hardwall(*z, flags) &&
Christoph Lameter3c517a62006-12-06 20:33:29 -08003236 cache->nodelists[nid] &&
3237 cache->nodelists[nid]->free_objects)
3238 obj = ____cache_alloc_node(cache,
3239 flags | GFP_THISNODE, nid);
3240 }
3241
Christoph Lametercfce6602007-05-06 14:50:17 -07003242 if (!obj) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003243 /*
3244 * This allocation will be performed within the constraints
3245 * of the current cpuset / memory policy requirements.
3246 * We may trigger various forms of reclaim on the allowed
3247 * set and go into memory reserves if necessary.
3248 */
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003249 if (local_flags & __GFP_WAIT)
3250 local_irq_enable();
3251 kmem_flagcheck(cache, flags);
Christoph Lameter3c517a62006-12-06 20:33:29 -08003252 obj = kmem_getpages(cache, flags, -1);
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003253 if (local_flags & __GFP_WAIT)
3254 local_irq_disable();
Christoph Lameter3c517a62006-12-06 20:33:29 -08003255 if (obj) {
3256 /*
3257 * Insert into the appropriate per node queues
3258 */
3259 nid = page_to_nid(virt_to_page(obj));
3260 if (cache_grow(cache, flags, nid, obj)) {
3261 obj = ____cache_alloc_node(cache,
3262 flags | GFP_THISNODE, nid);
3263 if (!obj)
3264 /*
3265 * Another processor may allocate the
3266 * objects in the slab since we are
3267 * not holding any locks.
3268 */
3269 goto retry;
3270 } else {
Hugh Dickinsb6a60452007-01-05 16:36:36 -08003271 /* cache_grow already freed obj */
Christoph Lameter3c517a62006-12-06 20:33:29 -08003272 obj = NULL;
3273 }
3274 }
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003275 }
Christoph Lameter765c4502006-09-27 01:50:08 -07003276 return obj;
3277}
3278
3279/*
Christoph Lametere498be72005-09-09 13:03:32 -07003280 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07003281 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003282static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003283 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07003284{
3285 struct list_head *entry;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003286 struct slab *slabp;
3287 struct kmem_list3 *l3;
3288 void *obj;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003289 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003290
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003291 l3 = cachep->nodelists[nodeid];
3292 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07003293
Andrew Mortona737b3e2006-03-22 00:08:11 -08003294retry:
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003295 check_irq_off();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003296 spin_lock(&l3->list_lock);
3297 entry = l3->slabs_partial.next;
3298 if (entry == &l3->slabs_partial) {
3299 l3->free_touched = 1;
3300 entry = l3->slabs_free.next;
3301 if (entry == &l3->slabs_free)
3302 goto must_grow;
3303 }
Christoph Lametere498be72005-09-09 13:03:32 -07003304
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003305 slabp = list_entry(entry, struct slab, list);
3306 check_spinlock_acquired_node(cachep, nodeid);
3307 check_slabp(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07003308
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003309 STATS_INC_NODEALLOCS(cachep);
3310 STATS_INC_ACTIVE(cachep);
3311 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003312
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003313 BUG_ON(slabp->inuse == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07003314
Matthew Dobson78d382d2006-02-01 03:05:47 -08003315 obj = slab_get_obj(cachep, slabp, nodeid);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003316 check_slabp(cachep, slabp);
3317 l3->free_objects--;
3318 /* move slabp to correct slabp list: */
3319 list_del(&slabp->list);
Christoph Lametere498be72005-09-09 13:03:32 -07003320
Andrew Mortona737b3e2006-03-22 00:08:11 -08003321 if (slabp->free == BUFCTL_END)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003322 list_add(&slabp->list, &l3->slabs_full);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003323 else
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003324 list_add(&slabp->list, &l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07003325
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003326 spin_unlock(&l3->list_lock);
3327 goto done;
Christoph Lametere498be72005-09-09 13:03:32 -07003328
Andrew Mortona737b3e2006-03-22 00:08:11 -08003329must_grow:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003330 spin_unlock(&l3->list_lock);
Christoph Lameter3c517a62006-12-06 20:33:29 -08003331 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
Christoph Lameter765c4502006-09-27 01:50:08 -07003332 if (x)
3333 goto retry;
Christoph Lametere498be72005-09-09 13:03:32 -07003334
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003335 return fallback_alloc(cachep, flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003336
Andrew Mortona737b3e2006-03-22 00:08:11 -08003337done:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003338 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07003339}
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003340
3341/**
3342 * kmem_cache_alloc_node - Allocate an object on the specified node
3343 * @cachep: The cache to allocate from.
3344 * @flags: See kmalloc().
3345 * @nodeid: node number of the target node.
3346 * @caller: return address of caller, used for debug information
3347 *
3348 * Identical to kmem_cache_alloc but it will allocate memory on the given
3349 * node, which can improve the performance for cpu bound structures.
3350 *
3351 * Fallback to other node is possible if __GFP_THISNODE is not set.
3352 */
3353static __always_inline void *
3354__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3355 void *caller)
3356{
3357 unsigned long save_flags;
3358 void *ptr;
3359
Akinobu Mita824ebef2007-05-06 14:49:58 -07003360 if (should_failslab(cachep, flags))
3361 return NULL;
3362
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003363 cache_alloc_debugcheck_before(cachep, flags);
3364 local_irq_save(save_flags);
3365
3366 if (unlikely(nodeid == -1))
3367 nodeid = numa_node_id();
3368
3369 if (unlikely(!cachep->nodelists[nodeid])) {
3370 /* Node not bootstrapped yet */
3371 ptr = fallback_alloc(cachep, flags);
3372 goto out;
3373 }
3374
3375 if (nodeid == numa_node_id()) {
3376 /*
3377 * Use the locally cached objects if possible.
3378 * However ____cache_alloc does not allow fallback
3379 * to other nodes. It may fail while we still have
3380 * objects on other nodes available.
3381 */
3382 ptr = ____cache_alloc(cachep, flags);
3383 if (ptr)
3384 goto out;
3385 }
3386 /* ___cache_alloc_node can fall back to other nodes */
3387 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3388 out:
3389 local_irq_restore(save_flags);
3390 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3391
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003392 if (unlikely((flags & __GFP_ZERO) && ptr))
3393 memset(ptr, 0, obj_size(cachep));
3394
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003395 return ptr;
3396}
3397
3398static __always_inline void *
3399__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3400{
3401 void *objp;
3402
3403 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3404 objp = alternate_node_alloc(cache, flags);
3405 if (objp)
3406 goto out;
3407 }
3408 objp = ____cache_alloc(cache, flags);
3409
3410 /*
3411 * We may just have run out of memory on the local node.
3412 * ____cache_alloc_node() knows how to locate memory on other nodes
3413 */
3414 if (!objp)
3415 objp = ____cache_alloc_node(cache, flags, numa_node_id());
3416
3417 out:
3418 return objp;
3419}
3420#else
3421
3422static __always_inline void *
3423__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3424{
3425 return ____cache_alloc(cachep, flags);
3426}
3427
3428#endif /* CONFIG_NUMA */
3429
3430static __always_inline void *
3431__cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3432{
3433 unsigned long save_flags;
3434 void *objp;
3435
Akinobu Mita824ebef2007-05-06 14:49:58 -07003436 if (should_failslab(cachep, flags))
3437 return NULL;
3438
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003439 cache_alloc_debugcheck_before(cachep, flags);
3440 local_irq_save(save_flags);
3441 objp = __do_cache_alloc(cachep, flags);
3442 local_irq_restore(save_flags);
3443 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3444 prefetchw(objp);
3445
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003446 if (unlikely((flags & __GFP_ZERO) && objp))
3447 memset(objp, 0, obj_size(cachep));
3448
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003449 return objp;
3450}
Christoph Lametere498be72005-09-09 13:03:32 -07003451
3452/*
3453 * Caller needs to acquire correct kmem_list's list_lock
3454 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003455static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003456 int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003457{
3458 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07003459 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003460
3461 for (i = 0; i < nr_objects; i++) {
3462 void *objp = objpp[i];
3463 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003464
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08003465 slabp = virt_to_slab(objp);
Christoph Lameterff694162005-09-22 21:44:02 -07003466 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003467 list_del(&slabp->list);
Christoph Lameterff694162005-09-22 21:44:02 -07003468 check_spinlock_acquired_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003469 check_slabp(cachep, slabp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08003470 slab_put_obj(cachep, slabp, objp, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003471 STATS_DEC_ACTIVE(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003472 l3->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003473 check_slabp(cachep, slabp);
3474
3475 /* fixup slab chains */
3476 if (slabp->inuse == 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07003477 if (l3->free_objects > l3->free_limit) {
3478 l3->free_objects -= cachep->num;
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07003479 /* No need to drop any previously held
3480 * lock here, even if we have a off-slab slab
3481 * descriptor it is guaranteed to come from
3482 * a different cache, refer to comments before
3483 * alloc_slabmgmt.
3484 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003485 slab_destroy(cachep, slabp);
3486 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07003487 list_add(&slabp->list, &l3->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003488 }
3489 } else {
3490 /* Unconditionally move a slab to the end of the
3491 * partial list on free - maximum time for the
3492 * other objects to be freed, too.
3493 */
Christoph Lametere498be72005-09-09 13:03:32 -07003494 list_add_tail(&slabp->list, &l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003495 }
3496 }
3497}
3498
Pekka Enberg343e0d72006-02-01 03:05:50 -08003499static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003500{
3501 int batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07003502 struct kmem_list3 *l3;
Christoph Lameterff694162005-09-22 21:44:02 -07003503 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003504
3505 batchcount = ac->batchcount;
3506#if DEBUG
3507 BUG_ON(!batchcount || batchcount > ac->avail);
3508#endif
3509 check_irq_off();
Christoph Lameterff694162005-09-22 21:44:02 -07003510 l3 = cachep->nodelists[node];
Ingo Molnar873623d2006-07-13 14:44:38 +02003511 spin_lock(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003512 if (l3->shared) {
3513 struct array_cache *shared_array = l3->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003514 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003515 if (max) {
3516 if (batchcount > max)
3517 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07003518 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003519 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003520 shared_array->avail += batchcount;
3521 goto free_done;
3522 }
3523 }
3524
Christoph Lameterff694162005-09-22 21:44:02 -07003525 free_block(cachep, ac->entry, batchcount, node);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003526free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003527#if STATS
3528 {
3529 int i = 0;
3530 struct list_head *p;
3531
Christoph Lametere498be72005-09-09 13:03:32 -07003532 p = l3->slabs_free.next;
3533 while (p != &(l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003534 struct slab *slabp;
3535
3536 slabp = list_entry(p, struct slab, list);
3537 BUG_ON(slabp->inuse);
3538
3539 i++;
3540 p = p->next;
3541 }
3542 STATS_SET_FREEABLE(cachep, i);
3543 }
3544#endif
Christoph Lametere498be72005-09-09 13:03:32 -07003545 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003546 ac->avail -= batchcount;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003547 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003548}
3549
3550/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08003551 * Release an obj back to its cache. If the obj has a constructed state, it must
3552 * be in this state _before_ it is released. Called with disabled ints.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003553 */
Ingo Molnar873623d2006-07-13 14:44:38 +02003554static inline void __cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003555{
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003556 struct array_cache *ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003557
3558 check_irq_off();
3559 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3560
Christoph Lameter3cdc0ed2007-06-08 13:46:46 -07003561 if (cache_free_alien(cachep, objp))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07003562 return;
Christoph Lametere498be72005-09-09 13:03:32 -07003563
Linus Torvalds1da177e2005-04-16 15:20:36 -07003564 if (likely(ac->avail < ac->limit)) {
3565 STATS_INC_FREEHIT(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003566 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003567 return;
3568 } else {
3569 STATS_INC_FREEMISS(cachep);
3570 cache_flusharray(cachep, ac);
Christoph Lametere498be72005-09-09 13:03:32 -07003571 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003572 }
3573}
3574
3575/**
3576 * kmem_cache_alloc - Allocate an object
3577 * @cachep: The cache to allocate from.
3578 * @flags: See kmalloc().
3579 *
3580 * Allocate an object from this cache. The flags are only relevant
3581 * if the cache has no available objects.
3582 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003583void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003584{
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003585 return __cache_alloc(cachep, flags, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003586}
3587EXPORT_SYMBOL(kmem_cache_alloc);
3588
3589/**
3590 * kmem_ptr_validate - check if an untrusted pointer might
3591 * be a slab entry.
3592 * @cachep: the cache we're checking against
3593 * @ptr: pointer to validate
3594 *
3595 * This verifies that the untrusted pointer looks sane:
3596 * it is _not_ a guarantee that the pointer is actually
3597 * part of the slab cache in question, but it at least
3598 * validates that the pointer can be dereferenced and
3599 * looks half-way sane.
3600 *
3601 * Currently only used for dentry validation.
3602 */
Christoph Lameterb7f869a2006-12-22 01:06:44 -08003603int kmem_ptr_validate(struct kmem_cache *cachep, const void *ptr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003604{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003605 unsigned long addr = (unsigned long)ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003606 unsigned long min_addr = PAGE_OFFSET;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003607 unsigned long align_mask = BYTES_PER_WORD - 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003608 unsigned long size = cachep->buffer_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003609 struct page *page;
3610
3611 if (unlikely(addr < min_addr))
3612 goto out;
3613 if (unlikely(addr > (unsigned long)high_memory - size))
3614 goto out;
3615 if (unlikely(addr & align_mask))
3616 goto out;
3617 if (unlikely(!kern_addr_valid(addr)))
3618 goto out;
3619 if (unlikely(!kern_addr_valid(addr + size - 1)))
3620 goto out;
3621 page = virt_to_page(ptr);
3622 if (unlikely(!PageSlab(page)))
3623 goto out;
Pekka Enberg065d41c2005-11-13 16:06:46 -08003624 if (unlikely(page_get_cache(page) != cachep))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003625 goto out;
3626 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003627out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003628 return 0;
3629}
3630
3631#ifdef CONFIG_NUMA
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003632void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3633{
3634 return __cache_alloc_node(cachep, flags, nodeid,
3635 __builtin_return_address(0));
3636}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003637EXPORT_SYMBOL(kmem_cache_alloc_node);
3638
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003639static __always_inline void *
3640__do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003641{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003642 struct kmem_cache *cachep;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003643
3644 cachep = kmem_find_general_cachep(size, flags);
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003645 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3646 return cachep;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003647 return kmem_cache_alloc_node(cachep, flags, node);
3648}
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003649
3650#ifdef CONFIG_DEBUG_SLAB
3651void *__kmalloc_node(size_t size, gfp_t flags, int node)
3652{
3653 return __do_kmalloc_node(size, flags, node,
3654 __builtin_return_address(0));
3655}
Christoph Hellwigdbe5e692006-09-25 23:31:36 -07003656EXPORT_SYMBOL(__kmalloc_node);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003657
3658void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3659 int node, void *caller)
3660{
3661 return __do_kmalloc_node(size, flags, node, caller);
3662}
3663EXPORT_SYMBOL(__kmalloc_node_track_caller);
3664#else
3665void *__kmalloc_node(size_t size, gfp_t flags, int node)
3666{
3667 return __do_kmalloc_node(size, flags, node, NULL);
3668}
3669EXPORT_SYMBOL(__kmalloc_node);
3670#endif /* CONFIG_DEBUG_SLAB */
3671#endif /* CONFIG_NUMA */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003672
3673/**
Paul Drynoff800590f2006-06-23 02:03:48 -07003674 * __do_kmalloc - allocate memory
Linus Torvalds1da177e2005-04-16 15:20:36 -07003675 * @size: how many bytes of memory are required.
Paul Drynoff800590f2006-06-23 02:03:48 -07003676 * @flags: the type of memory to allocate (see kmalloc).
Randy Dunlap911851e2006-03-22 00:08:14 -08003677 * @caller: function caller for debug tracking of the caller
Linus Torvalds1da177e2005-04-16 15:20:36 -07003678 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003679static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3680 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003681{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003682 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003683
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003684 /* If you want to save a few bytes .text space: replace
3685 * __ with kmem_.
3686 * Then kmalloc uses the uninlined functions instead of the inline
3687 * functions.
3688 */
3689 cachep = __find_general_cachep(size, flags);
Linus Torvaldsa5c96d82007-07-19 13:17:15 -07003690 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3691 return cachep;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003692 return __cache_alloc(cachep, flags, caller);
3693}
3694
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003695
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003696#ifdef CONFIG_DEBUG_SLAB
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003697void *__kmalloc(size_t size, gfp_t flags)
3698{
Al Viro871751e2006-03-25 03:06:39 -08003699 return __do_kmalloc(size, flags, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003700}
3701EXPORT_SYMBOL(__kmalloc);
3702
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003703void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller)
3704{
3705 return __do_kmalloc(size, flags, caller);
3706}
3707EXPORT_SYMBOL(__kmalloc_track_caller);
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003708
3709#else
3710void *__kmalloc(size_t size, gfp_t flags)
3711{
3712 return __do_kmalloc(size, flags, NULL);
3713}
3714EXPORT_SYMBOL(__kmalloc);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003715#endif
3716
Linus Torvalds1da177e2005-04-16 15:20:36 -07003717/**
3718 * kmem_cache_free - Deallocate an object
3719 * @cachep: The cache the allocation was from.
3720 * @objp: The previously allocated object.
3721 *
3722 * Free an object which was previously allocated from this
3723 * cache.
3724 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003725void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003726{
3727 unsigned long flags;
3728
Pekka Enbergddc2e812006-06-23 02:03:40 -07003729 BUG_ON(virt_to_cache(objp) != cachep);
3730
Linus Torvalds1da177e2005-04-16 15:20:36 -07003731 local_irq_save(flags);
Ingo Molnar898552c2007-02-10 01:44:57 -08003732 debug_check_no_locks_freed(objp, obj_size(cachep));
Ingo Molnar873623d2006-07-13 14:44:38 +02003733 __cache_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003734 local_irq_restore(flags);
3735}
3736EXPORT_SYMBOL(kmem_cache_free);
3737
3738/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003739 * kfree - free previously allocated memory
3740 * @objp: pointer returned by kmalloc.
3741 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003742 * If @objp is NULL, no operation is performed.
3743 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003744 * Don't free memory not originally allocated by kmalloc()
3745 * or you will run into trouble.
3746 */
3747void kfree(const void *objp)
3748{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003749 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003750 unsigned long flags;
3751
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003752 if (unlikely(ZERO_OR_NULL_PTR(objp)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003753 return;
3754 local_irq_save(flags);
3755 kfree_debugcheck(objp);
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08003756 c = virt_to_cache(objp);
Ingo Molnarf9b84042006-06-27 02:54:49 -07003757 debug_check_no_locks_freed(objp, obj_size(c));
Ingo Molnar873623d2006-07-13 14:44:38 +02003758 __cache_free(c, (void *)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003759 local_irq_restore(flags);
3760}
3761EXPORT_SYMBOL(kfree);
3762
Pekka Enberg343e0d72006-02-01 03:05:50 -08003763unsigned int kmem_cache_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003764{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003765 return obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003766}
3767EXPORT_SYMBOL(kmem_cache_size);
3768
Pekka Enberg343e0d72006-02-01 03:05:50 -08003769const char *kmem_cache_name(struct kmem_cache *cachep)
Arnaldo Carvalho de Melo19449722005-06-18 22:46:19 -07003770{
3771 return cachep->name;
3772}
3773EXPORT_SYMBOL_GPL(kmem_cache_name);
3774
Christoph Lametere498be72005-09-09 13:03:32 -07003775/*
Christoph Lameter0718dc22006-03-25 03:06:47 -08003776 * This initializes kmem_list3 or resizes varioius caches for all nodes.
Christoph Lametere498be72005-09-09 13:03:32 -07003777 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003778static int alloc_kmemlist(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07003779{
3780 int node;
3781 struct kmem_list3 *l3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003782 struct array_cache *new_shared;
Paul Menage3395ee02006-12-06 20:32:16 -08003783 struct array_cache **new_alien = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003784
3785 for_each_online_node(node) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003786
Paul Menage3395ee02006-12-06 20:32:16 -08003787 if (use_alien_caches) {
3788 new_alien = alloc_alien_cache(node, cachep->limit);
3789 if (!new_alien)
3790 goto fail;
3791 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003792
Eric Dumazet63109842007-05-06 14:49:28 -07003793 new_shared = NULL;
3794 if (cachep->shared) {
3795 new_shared = alloc_arraycache(node,
Christoph Lameter0718dc22006-03-25 03:06:47 -08003796 cachep->shared*cachep->batchcount,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003797 0xbaadf00d);
Eric Dumazet63109842007-05-06 14:49:28 -07003798 if (!new_shared) {
3799 free_alien_cache(new_alien);
3800 goto fail;
3801 }
Christoph Lameter0718dc22006-03-25 03:06:47 -08003802 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003803
Andrew Mortona737b3e2006-03-22 00:08:11 -08003804 l3 = cachep->nodelists[node];
3805 if (l3) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003806 struct array_cache *shared = l3->shared;
3807
Christoph Lametere498be72005-09-09 13:03:32 -07003808 spin_lock_irq(&l3->list_lock);
3809
Christoph Lametercafeb022006-03-25 03:06:46 -08003810 if (shared)
Christoph Lameter0718dc22006-03-25 03:06:47 -08003811 free_block(cachep, shared->entry,
3812 shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07003813
Christoph Lametercafeb022006-03-25 03:06:46 -08003814 l3->shared = new_shared;
3815 if (!l3->alien) {
Christoph Lametere498be72005-09-09 13:03:32 -07003816 l3->alien = new_alien;
3817 new_alien = NULL;
3818 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003819 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003820 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003821 spin_unlock_irq(&l3->list_lock);
Christoph Lametercafeb022006-03-25 03:06:46 -08003822 kfree(shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003823 free_alien_cache(new_alien);
3824 continue;
3825 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08003826 l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node);
Christoph Lameter0718dc22006-03-25 03:06:47 -08003827 if (!l3) {
3828 free_alien_cache(new_alien);
3829 kfree(new_shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003830 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003831 }
Christoph Lametere498be72005-09-09 13:03:32 -07003832
3833 kmem_list3_init(l3);
3834 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Andrew Mortona737b3e2006-03-22 00:08:11 -08003835 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003836 l3->shared = new_shared;
Christoph Lametere498be72005-09-09 13:03:32 -07003837 l3->alien = new_alien;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003838 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003839 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003840 cachep->nodelists[node] = l3;
3841 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003842 return 0;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003843
Andrew Mortona737b3e2006-03-22 00:08:11 -08003844fail:
Christoph Lameter0718dc22006-03-25 03:06:47 -08003845 if (!cachep->next.next) {
3846 /* Cache is not active yet. Roll back what we did */
3847 node--;
3848 while (node >= 0) {
3849 if (cachep->nodelists[node]) {
3850 l3 = cachep->nodelists[node];
3851
3852 kfree(l3->shared);
3853 free_alien_cache(l3->alien);
3854 kfree(l3);
3855 cachep->nodelists[node] = NULL;
3856 }
3857 node--;
3858 }
3859 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003860 return -ENOMEM;
Christoph Lametere498be72005-09-09 13:03:32 -07003861}
3862
Linus Torvalds1da177e2005-04-16 15:20:36 -07003863struct ccupdate_struct {
Pekka Enberg343e0d72006-02-01 03:05:50 -08003864 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003865 struct array_cache *new[NR_CPUS];
3866};
3867
3868static void do_ccupdate_local(void *info)
3869{
Andrew Mortona737b3e2006-03-22 00:08:11 -08003870 struct ccupdate_struct *new = info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003871 struct array_cache *old;
3872
3873 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003874 old = cpu_cache_get(new->cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003875
Linus Torvalds1da177e2005-04-16 15:20:36 -07003876 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3877 new->new[smp_processor_id()] = old;
3878}
3879
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003880/* Always called with the cache_chain_mutex held */
Andrew Mortona737b3e2006-03-22 00:08:11 -08003881static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3882 int batchcount, int shared)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003883{
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003884 struct ccupdate_struct *new;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003885 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003886
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003887 new = kzalloc(sizeof(*new), GFP_KERNEL);
3888 if (!new)
3889 return -ENOMEM;
3890
Christoph Lametere498be72005-09-09 13:03:32 -07003891 for_each_online_cpu(i) {
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003892 new->new[i] = alloc_arraycache(cpu_to_node(i), limit,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003893 batchcount);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003894 if (!new->new[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003895 for (i--; i >= 0; i--)
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003896 kfree(new->new[i]);
3897 kfree(new);
Christoph Lametere498be72005-09-09 13:03:32 -07003898 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003899 }
3900 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003901 new->cachep = cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003902
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003903 on_each_cpu(do_ccupdate_local, (void *)new, 1, 1);
Christoph Lametere498be72005-09-09 13:03:32 -07003904
Linus Torvalds1da177e2005-04-16 15:20:36 -07003905 check_irq_on();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003906 cachep->batchcount = batchcount;
3907 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07003908 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003909
Christoph Lametere498be72005-09-09 13:03:32 -07003910 for_each_online_cpu(i) {
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003911 struct array_cache *ccold = new->new[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003912 if (!ccold)
3913 continue;
Christoph Lametere498be72005-09-09 13:03:32 -07003914 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -07003915 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
Christoph Lametere498be72005-09-09 13:03:32 -07003916 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003917 kfree(ccold);
3918 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003919 kfree(new);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003920 return alloc_kmemlist(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003921}
3922
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003923/* Called with cache_chain_mutex held always */
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003924static int enable_cpucache(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003925{
3926 int err;
3927 int limit, shared;
3928
Andrew Mortona737b3e2006-03-22 00:08:11 -08003929 /*
3930 * The head array serves three purposes:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003931 * - create a LIFO ordering, i.e. return objects that are cache-warm
3932 * - reduce the number of spinlock operations.
Andrew Mortona737b3e2006-03-22 00:08:11 -08003933 * - reduce the number of linked list operations on the slab and
Linus Torvalds1da177e2005-04-16 15:20:36 -07003934 * bufctl chains: array operations are cheaper.
3935 * The numbers are guessed, we should auto-tune as described by
3936 * Bonwick.
3937 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003938 if (cachep->buffer_size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003939 limit = 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003940 else if (cachep->buffer_size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003941 limit = 8;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003942 else if (cachep->buffer_size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003943 limit = 24;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003944 else if (cachep->buffer_size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003945 limit = 54;
3946 else
3947 limit = 120;
3948
Andrew Mortona737b3e2006-03-22 00:08:11 -08003949 /*
3950 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
Linus Torvalds1da177e2005-04-16 15:20:36 -07003951 * allocation behaviour: Most allocs on one cpu, most free operations
3952 * on another cpu. For these cases, an efficient object passing between
3953 * cpus is necessary. This is provided by a shared array. The array
3954 * replaces Bonwick's magazine layer.
3955 * On uniprocessor, it's functionally equivalent (but less efficient)
3956 * to a larger limit. Thus disabled by default.
3957 */
3958 shared = 0;
Eric Dumazet364fbb22007-05-06 14:49:27 -07003959 if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003960 shared = 8;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003961
3962#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003963 /*
3964 * With debugging enabled, large batchcount lead to excessively long
3965 * periods with disabled local interrupts. Limit the batchcount
Linus Torvalds1da177e2005-04-16 15:20:36 -07003966 */
3967 if (limit > 32)
3968 limit = 32;
3969#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003970 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003971 if (err)
3972 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003973 cachep->name, -err);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003974 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003975}
3976
Christoph Lameter1b552532006-03-22 00:09:07 -08003977/*
3978 * Drain an array if it contains any elements taking the l3 lock only if
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003979 * necessary. Note that the l3 listlock also protects the array_cache
3980 * if drain_array() is used on the shared array.
Christoph Lameter1b552532006-03-22 00:09:07 -08003981 */
3982void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
3983 struct array_cache *ac, int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003984{
3985 int tofree;
3986
Christoph Lameter1b552532006-03-22 00:09:07 -08003987 if (!ac || !ac->avail)
3988 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003989 if (ac->touched && !force) {
3990 ac->touched = 0;
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003991 } else {
Christoph Lameter1b552532006-03-22 00:09:07 -08003992 spin_lock_irq(&l3->list_lock);
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003993 if (ac->avail) {
3994 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3995 if (tofree > ac->avail)
3996 tofree = (ac->avail + 1) / 2;
3997 free_block(cachep, ac->entry, tofree, node);
3998 ac->avail -= tofree;
3999 memmove(ac->entry, &(ac->entry[tofree]),
4000 sizeof(void *) * ac->avail);
4001 }
Christoph Lameter1b552532006-03-22 00:09:07 -08004002 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004003 }
4004}
4005
4006/**
4007 * cache_reap - Reclaim memory from caches.
Randy Dunlap05fb6bf2007-02-28 20:12:13 -08004008 * @w: work descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07004009 *
4010 * Called from workqueue/eventd every few seconds.
4011 * Purpose:
4012 * - clear the per-cpu caches for this CPU.
4013 * - return freeable pages to the main free memory pool.
4014 *
Andrew Mortona737b3e2006-03-22 00:08:11 -08004015 * If we cannot acquire the cache chain mutex then just give up - we'll try
4016 * again on the next iteration.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004017 */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004018static void cache_reap(struct work_struct *w)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004019{
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004020 struct kmem_cache *searchp;
Christoph Lametere498be72005-09-09 13:03:32 -07004021 struct kmem_list3 *l3;
Christoph Lameteraab22072006-03-22 00:09:06 -08004022 int node = numa_node_id();
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004023 struct delayed_work *work =
4024 container_of(w, struct delayed_work, work);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004025
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004026 if (!mutex_trylock(&cache_chain_mutex))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004027 /* Give up. Setup the next iteration. */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004028 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004029
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004030 list_for_each_entry(searchp, &cache_chain, next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004031 check_irq_on();
4032
Christoph Lameter35386e32006-03-22 00:09:05 -08004033 /*
4034 * We only take the l3 lock if absolutely necessary and we
4035 * have established with reasonable certainty that
4036 * we can do some work if the lock was obtained.
4037 */
Christoph Lameteraab22072006-03-22 00:09:06 -08004038 l3 = searchp->nodelists[node];
Christoph Lameter35386e32006-03-22 00:09:05 -08004039
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004040 reap_alien(searchp, l3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004041
Christoph Lameteraab22072006-03-22 00:09:06 -08004042 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004043
Christoph Lameter35386e32006-03-22 00:09:05 -08004044 /*
4045 * These are racy checks but it does not matter
4046 * if we skip one check or scan twice.
4047 */
Christoph Lametere498be72005-09-09 13:03:32 -07004048 if (time_after(l3->next_reap, jiffies))
Christoph Lameter35386e32006-03-22 00:09:05 -08004049 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004050
Christoph Lametere498be72005-09-09 13:03:32 -07004051 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004052
Christoph Lameteraab22072006-03-22 00:09:06 -08004053 drain_array(searchp, l3, l3->shared, 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004054
Christoph Lametered11d9e2006-06-30 01:55:45 -07004055 if (l3->free_touched)
Christoph Lametere498be72005-09-09 13:03:32 -07004056 l3->free_touched = 0;
Christoph Lametered11d9e2006-06-30 01:55:45 -07004057 else {
4058 int freed;
4059
4060 freed = drain_freelist(searchp, l3, (l3->free_limit +
4061 5 * searchp->num - 1) / (5 * searchp->num));
4062 STATS_ADD_REAPED(searchp, freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004063 }
Christoph Lameter35386e32006-03-22 00:09:05 -08004064next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004065 cond_resched();
4066 }
4067 check_irq_on();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004068 mutex_unlock(&cache_chain_mutex);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004069 next_reap_node();
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004070out:
Andrew Mortona737b3e2006-03-22 00:08:11 -08004071 /* Set up the next iteration */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004072 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004073}
4074
4075#ifdef CONFIG_PROC_FS
4076
Pekka Enberg85289f92006-01-08 01:00:36 -08004077static void print_slabinfo_header(struct seq_file *m)
4078{
4079 /*
4080 * Output format version, so at least we can change it
4081 * without _too_ many complaints.
4082 */
4083#if STATS
4084 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4085#else
4086 seq_puts(m, "slabinfo - version: 2.1\n");
4087#endif
4088 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4089 "<objperslab> <pagesperslab>");
4090 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4091 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4092#if STATS
4093 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004094 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
Pekka Enberg85289f92006-01-08 01:00:36 -08004095 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4096#endif
4097 seq_putc(m, '\n');
4098}
4099
Linus Torvalds1da177e2005-04-16 15:20:36 -07004100static void *s_start(struct seq_file *m, loff_t *pos)
4101{
4102 loff_t n = *pos;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004103
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004104 mutex_lock(&cache_chain_mutex);
Pekka Enberg85289f92006-01-08 01:00:36 -08004105 if (!n)
4106 print_slabinfo_header(m);
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004107
4108 return seq_list_start(&cache_chain, *pos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004109}
4110
4111static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4112{
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004113 return seq_list_next(p, &cache_chain, pos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004114}
4115
4116static void s_stop(struct seq_file *m, void *p)
4117{
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004118 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004119}
4120
4121static int s_show(struct seq_file *m, void *p)
4122{
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004123 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004124 struct slab *slabp;
4125 unsigned long active_objs;
4126 unsigned long num_objs;
4127 unsigned long active_slabs = 0;
4128 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004129 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004130 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07004131 int node;
4132 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004133
Linus Torvalds1da177e2005-04-16 15:20:36 -07004134 active_objs = 0;
4135 num_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004136 for_each_online_node(node) {
4137 l3 = cachep->nodelists[node];
4138 if (!l3)
4139 continue;
4140
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004141 check_irq_on();
4142 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07004143
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004144 list_for_each_entry(slabp, &l3->slabs_full, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004145 if (slabp->inuse != cachep->num && !error)
4146 error = "slabs_full accounting error";
4147 active_objs += cachep->num;
4148 active_slabs++;
4149 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004150 list_for_each_entry(slabp, &l3->slabs_partial, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004151 if (slabp->inuse == cachep->num && !error)
4152 error = "slabs_partial inuse accounting error";
4153 if (!slabp->inuse && !error)
4154 error = "slabs_partial/inuse accounting error";
4155 active_objs += slabp->inuse;
4156 active_slabs++;
4157 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004158 list_for_each_entry(slabp, &l3->slabs_free, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004159 if (slabp->inuse && !error)
4160 error = "slabs_free/inuse accounting error";
4161 num_slabs++;
4162 }
4163 free_objects += l3->free_objects;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08004164 if (l3->shared)
4165 shared_avail += l3->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07004166
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004167 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004168 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004169 num_slabs += active_slabs;
4170 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07004171 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004172 error = "free_objects accounting error";
4173
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004174 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004175 if (error)
4176 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4177
4178 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004179 name, active_objs, num_objs, cachep->buffer_size,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004180 cachep->num, (1 << cachep->gfporder));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004181 seq_printf(m, " : tunables %4u %4u %4u",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004182 cachep->limit, cachep->batchcount, cachep->shared);
Christoph Lametere498be72005-09-09 13:03:32 -07004183 seq_printf(m, " : slabdata %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004184 active_slabs, num_slabs, shared_avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004185#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004186 { /* list3 stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004187 unsigned long high = cachep->high_mark;
4188 unsigned long allocs = cachep->num_allocations;
4189 unsigned long grown = cachep->grown;
4190 unsigned long reaped = cachep->reaped;
4191 unsigned long errors = cachep->errors;
4192 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004193 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07004194 unsigned long node_frees = cachep->node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004195 unsigned long overflows = cachep->node_overflow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004196
Christoph Lametere498be72005-09-09 13:03:32 -07004197 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004198 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown,
Andrew Mortona737b3e2006-03-22 00:08:11 -08004199 reaped, errors, max_freeable, node_allocs,
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004200 node_frees, overflows);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004201 }
4202 /* cpu stats */
4203 {
4204 unsigned long allochit = atomic_read(&cachep->allochit);
4205 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4206 unsigned long freehit = atomic_read(&cachep->freehit);
4207 unsigned long freemiss = atomic_read(&cachep->freemiss);
4208
4209 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004210 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004211 }
4212#endif
4213 seq_putc(m, '\n');
Linus Torvalds1da177e2005-04-16 15:20:36 -07004214 return 0;
4215}
4216
4217/*
4218 * slabinfo_op - iterator that generates /proc/slabinfo
4219 *
4220 * Output layout:
4221 * cache-name
4222 * num-active-objs
4223 * total-objs
4224 * object size
4225 * num-active-slabs
4226 * total-slabs
4227 * num-pages-per-slab
4228 * + further values on SMP and with statistics enabled
4229 */
4230
Helge Deller15ad7cd2006-12-06 20:40:36 -08004231const struct seq_operations slabinfo_op = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004232 .start = s_start,
4233 .next = s_next,
4234 .stop = s_stop,
4235 .show = s_show,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004236};
4237
4238#define MAX_SLABINFO_WRITE 128
4239/**
4240 * slabinfo_write - Tuning for the slab allocator
4241 * @file: unused
4242 * @buffer: user buffer
4243 * @count: data length
4244 * @ppos: unused
4245 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004246ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4247 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004248{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004249 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004250 int limit, batchcount, shared, res;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004251 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004252
Linus Torvalds1da177e2005-04-16 15:20:36 -07004253 if (count > MAX_SLABINFO_WRITE)
4254 return -EINVAL;
4255 if (copy_from_user(&kbuf, buffer, count))
4256 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004257 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07004258
4259 tmp = strchr(kbuf, ' ');
4260 if (!tmp)
4261 return -EINVAL;
4262 *tmp = '\0';
4263 tmp++;
4264 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4265 return -EINVAL;
4266
4267 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004268 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004269 res = -EINVAL;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004270 list_for_each_entry(cachep, &cache_chain, next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004271 if (!strcmp(cachep->name, kbuf)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08004272 if (limit < 1 || batchcount < 1 ||
4273 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07004274 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004275 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07004276 res = do_tune_cpucache(cachep, limit,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004277 batchcount, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004278 }
4279 break;
4280 }
4281 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004282 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004283 if (res >= 0)
4284 res = count;
4285 return res;
4286}
Al Viro871751e2006-03-25 03:06:39 -08004287
4288#ifdef CONFIG_DEBUG_SLAB_LEAK
4289
4290static void *leaks_start(struct seq_file *m, loff_t *pos)
4291{
Al Viro871751e2006-03-25 03:06:39 -08004292 mutex_lock(&cache_chain_mutex);
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004293 return seq_list_start(&cache_chain, *pos);
Al Viro871751e2006-03-25 03:06:39 -08004294}
4295
4296static inline int add_caller(unsigned long *n, unsigned long v)
4297{
4298 unsigned long *p;
4299 int l;
4300 if (!v)
4301 return 1;
4302 l = n[1];
4303 p = n + 2;
4304 while (l) {
4305 int i = l/2;
4306 unsigned long *q = p + 2 * i;
4307 if (*q == v) {
4308 q[1]++;
4309 return 1;
4310 }
4311 if (*q > v) {
4312 l = i;
4313 } else {
4314 p = q + 2;
4315 l -= i + 1;
4316 }
4317 }
4318 if (++n[1] == n[0])
4319 return 0;
4320 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4321 p[0] = v;
4322 p[1] = 1;
4323 return 1;
4324}
4325
4326static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4327{
4328 void *p;
4329 int i;
4330 if (n[0] == n[1])
4331 return;
4332 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4333 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4334 continue;
4335 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4336 return;
4337 }
4338}
4339
4340static void show_symbol(struct seq_file *m, unsigned long address)
4341{
4342#ifdef CONFIG_KALLSYMS
Al Viro871751e2006-03-25 03:06:39 -08004343 unsigned long offset, size;
Tejun Heo9281ace2007-07-17 04:03:51 -07004344 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
Al Viro871751e2006-03-25 03:06:39 -08004345
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004346 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
Al Viro871751e2006-03-25 03:06:39 -08004347 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004348 if (modname[0])
Al Viro871751e2006-03-25 03:06:39 -08004349 seq_printf(m, " [%s]", modname);
4350 return;
4351 }
4352#endif
4353 seq_printf(m, "%p", (void *)address);
4354}
4355
4356static int leaks_show(struct seq_file *m, void *p)
4357{
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004358 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
Al Viro871751e2006-03-25 03:06:39 -08004359 struct slab *slabp;
4360 struct kmem_list3 *l3;
4361 const char *name;
4362 unsigned long *n = m->private;
4363 int node;
4364 int i;
4365
4366 if (!(cachep->flags & SLAB_STORE_USER))
4367 return 0;
4368 if (!(cachep->flags & SLAB_RED_ZONE))
4369 return 0;
4370
4371 /* OK, we can do it */
4372
4373 n[1] = 0;
4374
4375 for_each_online_node(node) {
4376 l3 = cachep->nodelists[node];
4377 if (!l3)
4378 continue;
4379
4380 check_irq_on();
4381 spin_lock_irq(&l3->list_lock);
4382
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004383 list_for_each_entry(slabp, &l3->slabs_full, list)
Al Viro871751e2006-03-25 03:06:39 -08004384 handle_slab(n, cachep, slabp);
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004385 list_for_each_entry(slabp, &l3->slabs_partial, list)
Al Viro871751e2006-03-25 03:06:39 -08004386 handle_slab(n, cachep, slabp);
Al Viro871751e2006-03-25 03:06:39 -08004387 spin_unlock_irq(&l3->list_lock);
4388 }
4389 name = cachep->name;
4390 if (n[0] == n[1]) {
4391 /* Increase the buffer size */
4392 mutex_unlock(&cache_chain_mutex);
4393 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4394 if (!m->private) {
4395 /* Too bad, we are really out */
4396 m->private = n;
4397 mutex_lock(&cache_chain_mutex);
4398 return -ENOMEM;
4399 }
4400 *(unsigned long *)m->private = n[0] * 2;
4401 kfree(n);
4402 mutex_lock(&cache_chain_mutex);
4403 /* Now make sure this entry will be retried */
4404 m->count = m->size;
4405 return 0;
4406 }
4407 for (i = 0; i < n[1]; i++) {
4408 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4409 show_symbol(m, n[2*i+2]);
4410 seq_putc(m, '\n');
4411 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004412
Al Viro871751e2006-03-25 03:06:39 -08004413 return 0;
4414}
4415
Helge Deller15ad7cd2006-12-06 20:40:36 -08004416const struct seq_operations slabstats_op = {
Al Viro871751e2006-03-25 03:06:39 -08004417 .start = leaks_start,
4418 .next = s_next,
4419 .stop = s_stop,
4420 .show = leaks_show,
4421};
4422#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004423#endif
4424
Manfred Spraul00e145b2005-09-03 15:55:07 -07004425/**
4426 * ksize - get the actual amount of memory allocated for a given object
4427 * @objp: Pointer to the object
4428 *
4429 * kmalloc may internally round up allocations and return more memory
4430 * than requested. ksize() can be used to determine the actual amount of
4431 * memory allocated. The caller may use this additional memory, even though
4432 * a smaller amount of memory was initially specified with the kmalloc call.
4433 * The caller must guarantee that objp points to a valid object previously
4434 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4435 * must not be freed during the duration of the call.
4436 */
Pekka Enbergfd76bab2007-05-06 14:48:40 -07004437size_t ksize(const void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004438{
Christoph Lameter6cb8f912007-07-17 04:03:22 -07004439 if (unlikely(ZERO_OR_NULL_PTR(objp)))
Manfred Spraul00e145b2005-09-03 15:55:07 -07004440 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004441
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08004442 return obj_size(virt_to_cache(objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004443}