blob: 4cbf8bb135571f620ee76b9e76cdf1186ba24889 [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
89#include <linux/config.h>
90#include <linux/slab.h>
91#include <linux/mm.h>
92#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>
Christoph Lametere498be72005-09-09 13:03:32 -0700106#include <linux/nodemask.h>
Christoph Lameterdc85da12006-01-18 17:42:36 -0800107#include <linux/mempolicy.h>
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800108#include <linux/mutex.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700109
110#include <asm/uaccess.h>
111#include <asm/cacheflush.h>
112#include <asm/tlbflush.h>
113#include <asm/page.h>
114
115/*
116 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL,
117 * SLAB_RED_ZONE & SLAB_POISON.
118 * 0 for faster, smaller code (especially in the critical paths).
119 *
120 * STATS - 1 to collect stats for /proc/slabinfo.
121 * 0 for faster, smaller code (especially in the critical paths).
122 *
123 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
124 */
125
126#ifdef CONFIG_DEBUG_SLAB
127#define DEBUG 1
128#define STATS 1
129#define FORCED_DEBUG 1
130#else
131#define DEBUG 0
132#define STATS 0
133#define FORCED_DEBUG 0
134#endif
135
Linus Torvalds1da177e2005-04-16 15:20:36 -0700136/* Shouldn't this be in a header file somewhere? */
137#define BYTES_PER_WORD sizeof(void *)
138
139#ifndef cache_line_size
140#define cache_line_size() L1_CACHE_BYTES
141#endif
142
143#ifndef ARCH_KMALLOC_MINALIGN
144/*
145 * Enforce a minimum alignment for the kmalloc caches.
146 * Usually, the kmalloc caches are cache_line_size() aligned, except when
147 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
148 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
149 * alignment larger than BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that.
150 * Note that this flag disables some debug features.
151 */
152#define ARCH_KMALLOC_MINALIGN 0
153#endif
154
155#ifndef ARCH_SLAB_MINALIGN
156/*
157 * Enforce a minimum alignment for all caches.
158 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
159 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
160 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
161 * some debug features.
162 */
163#define ARCH_SLAB_MINALIGN 0
164#endif
165
166#ifndef ARCH_KMALLOC_FLAGS
167#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
168#endif
169
170/* Legal flag mask for kmem_cache_create(). */
171#if DEBUG
172# define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \
173 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
Christoph Lameterac2b8982006-03-22 00:08:15 -0800174 SLAB_CACHE_DMA | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700175 SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \
176 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
Paul Jackson101a5002006-03-24 03:16:07 -0800177 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700178#else
Christoph Lameterac2b8982006-03-22 00:08:15 -0800179# define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700180 SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \
181 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
Paul Jackson101a5002006-03-24 03:16:07 -0800182 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700183#endif
184
185/*
186 * kmem_bufctl_t:
187 *
188 * Bufctl's are used for linking objs within a slab
189 * linked offsets.
190 *
191 * This implementation relies on "struct page" for locating the cache &
192 * slab an object belongs to.
193 * This allows the bufctl structure to be small (one int), but limits
194 * the number of objects a slab (not a cache) can contain when off-slab
195 * bufctls are used. The limit is the size of the largest general cache
196 * that does not use off-slab slabs.
197 * For 32bit archs with 4 kB pages, is this 56.
198 * This is not serious, as it is only for large objects, when it is unwise
199 * to have too many per slab.
200 * Note: This limit can be raised by introducing a general cache whose size
201 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
202 */
203
Kyle Moffettfa5b08d2005-09-03 15:55:03 -0700204typedef unsigned int kmem_bufctl_t;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700205#define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
206#define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
Al Viro871751e2006-03-25 03:06:39 -0800207#define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
208#define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700209
210/* Max number of objs-per-slab for caches which use off-slab slabs.
211 * Needed to avoid a possible looping condition in cache_grow().
212 */
213static unsigned long offslab_limit;
214
215/*
216 * struct slab
217 *
218 * Manages the objs in a slab. Placed either at the beginning of mem allocated
219 * for a slab, or allocated from an general cache.
220 * Slabs are chained into three list: fully used, partial, fully free slabs.
221 */
222struct slab {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800223 struct list_head list;
224 unsigned long colouroff;
225 void *s_mem; /* including colour offset */
226 unsigned int inuse; /* num of objs active in slab */
227 kmem_bufctl_t free;
228 unsigned short nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700229};
230
231/*
232 * struct slab_rcu
233 *
234 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
235 * arrange for kmem_freepages to be called via RCU. This is useful if
236 * we need to approach a kernel structure obliquely, from its address
237 * obtained without the usual locking. We can lock the structure to
238 * stabilize it and check it's still at the given address, only if we
239 * can be sure that the memory has not been meanwhile reused for some
240 * other kind of object (which our subsystem's lock might corrupt).
241 *
242 * rcu_read_lock before reading the address, then rcu_read_unlock after
243 * taking the spinlock within the structure expected at that address.
244 *
245 * We assume struct slab_rcu can overlay struct slab when destroying.
246 */
247struct slab_rcu {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800248 struct rcu_head head;
Pekka Enberg343e0d72006-02-01 03:05:50 -0800249 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800250 void *addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700251};
252
253/*
254 * struct array_cache
255 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700256 * Purpose:
257 * - LIFO ordering, to hand out cache-warm objects from _alloc
258 * - reduce the number of linked list operations
259 * - reduce spinlock operations
260 *
261 * The limit is stored in the per-cpu structure to reduce the data cache
262 * footprint.
263 *
264 */
265struct array_cache {
266 unsigned int avail;
267 unsigned int limit;
268 unsigned int batchcount;
269 unsigned int touched;
Christoph Lametere498be72005-09-09 13:03:32 -0700270 spinlock_t lock;
Andrew Mortona737b3e2006-03-22 00:08:11 -0800271 void *entry[0]; /*
272 * Must have this definition in here for the proper
273 * alignment of array_cache. Also simplifies accessing
274 * the entries.
275 * [0] is for gcc 2.95. It should really be [].
276 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700277};
278
Andrew Mortona737b3e2006-03-22 00:08:11 -0800279/*
280 * bootstrap: The caches do not work without cpuarrays anymore, but the
281 * cpuarrays are allocated from the generic caches...
Linus Torvalds1da177e2005-04-16 15:20:36 -0700282 */
283#define BOOT_CPUCACHE_ENTRIES 1
284struct arraycache_init {
285 struct array_cache cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800286 void *entries[BOOT_CPUCACHE_ENTRIES];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700287};
288
289/*
Christoph Lametere498be72005-09-09 13:03:32 -0700290 * The slab lists for all objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700291 */
292struct kmem_list3 {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800293 struct list_head slabs_partial; /* partial list first, better asm code */
294 struct list_head slabs_full;
295 struct list_head slabs_free;
296 unsigned long free_objects;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800297 unsigned int free_limit;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800298 unsigned int colour_next; /* Per-node cache coloring */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800299 spinlock_t list_lock;
300 struct array_cache *shared; /* shared per node */
301 struct array_cache **alien; /* on other nodes */
Christoph Lameter35386e32006-03-22 00:09:05 -0800302 unsigned long next_reap; /* updated without locking */
303 int free_touched; /* updated without locking */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700304};
305
Christoph Lametere498be72005-09-09 13:03:32 -0700306/*
307 * Need this for bootstrapping a per node allocator.
308 */
309#define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
310struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
311#define CACHE_CACHE 0
312#define SIZE_AC 1
313#define SIZE_L3 (1 + MAX_NUMNODES)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700314
Christoph Lametere498be72005-09-09 13:03:32 -0700315/*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800316 * This function must be completely optimized away if a constant is passed to
317 * it. Mostly the same as what is in linux/slab.h except it returns an index.
Christoph Lametere498be72005-09-09 13:03:32 -0700318 */
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700319static __always_inline int index_of(const size_t size)
Christoph Lametere498be72005-09-09 13:03:32 -0700320{
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800321 extern void __bad_size(void);
322
Christoph Lametere498be72005-09-09 13:03:32 -0700323 if (__builtin_constant_p(size)) {
324 int i = 0;
325
326#define CACHE(x) \
327 if (size <=x) \
328 return i; \
329 else \
330 i++;
331#include "linux/kmalloc_sizes.h"
332#undef CACHE
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800333 __bad_size();
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700334 } else
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800335 __bad_size();
Christoph Lametere498be72005-09-09 13:03:32 -0700336 return 0;
337}
338
339#define INDEX_AC index_of(sizeof(struct arraycache_init))
340#define INDEX_L3 index_of(sizeof(struct kmem_list3))
341
Pekka Enberg5295a742006-02-01 03:05:48 -0800342static void kmem_list3_init(struct kmem_list3 *parent)
Christoph Lametere498be72005-09-09 13:03:32 -0700343{
344 INIT_LIST_HEAD(&parent->slabs_full);
345 INIT_LIST_HEAD(&parent->slabs_partial);
346 INIT_LIST_HEAD(&parent->slabs_free);
347 parent->shared = NULL;
348 parent->alien = NULL;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800349 parent->colour_next = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700350 spin_lock_init(&parent->list_lock);
351 parent->free_objects = 0;
352 parent->free_touched = 0;
353}
354
Andrew Mortona737b3e2006-03-22 00:08:11 -0800355#define MAKE_LIST(cachep, listp, slab, nodeid) \
356 do { \
357 INIT_LIST_HEAD(listp); \
358 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
Christoph Lametere498be72005-09-09 13:03:32 -0700359 } while (0)
360
Andrew Mortona737b3e2006-03-22 00:08:11 -0800361#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
362 do { \
Christoph Lametere498be72005-09-09 13:03:32 -0700363 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
364 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
365 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
366 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700367
368/*
Pekka Enberg343e0d72006-02-01 03:05:50 -0800369 * struct kmem_cache
Linus Torvalds1da177e2005-04-16 15:20:36 -0700370 *
371 * manages a cache.
372 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800373
Pekka J Enberg2109a2d2005-11-07 00:58:01 -0800374struct kmem_cache {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700375/* 1) per-cpu data, touched during every alloc/free */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800376 struct array_cache *array[NR_CPUS];
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800377/* 2) Cache tunables. Protected by cache_chain_mutex */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800378 unsigned int batchcount;
379 unsigned int limit;
380 unsigned int shared;
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800381
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800382 unsigned int buffer_size;
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800383/* 3) touched by every alloc & free from the backend */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800384 struct kmem_list3 *nodelists[MAX_NUMNODES];
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800385
Andrew Mortona737b3e2006-03-22 00:08:11 -0800386 unsigned int flags; /* constant flags */
387 unsigned int num; /* # of objs per slab */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700388
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800389/* 4) cache_grow/shrink */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700390 /* order of pgs per slab (2^n) */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800391 unsigned int gfporder;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700392
393 /* force GFP flags, e.g. GFP_DMA */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800394 gfp_t gfpflags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700395
Andrew Mortona737b3e2006-03-22 00:08:11 -0800396 size_t colour; /* cache colouring range */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800397 unsigned int colour_off; /* colour offset */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800398 struct kmem_cache *slabp_cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800399 unsigned int slab_size;
Andrew Mortona737b3e2006-03-22 00:08:11 -0800400 unsigned int dflags; /* dynamic flags */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700401
402 /* constructor func */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800403 void (*ctor) (void *, struct kmem_cache *, unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700404
405 /* de-constructor func */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800406 void (*dtor) (void *, struct kmem_cache *, unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700407
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800408/* 5) cache creation/removal */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800409 const char *name;
410 struct list_head next;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700411
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800412/* 6) statistics */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700413#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800414 unsigned long num_active;
415 unsigned long num_allocations;
416 unsigned long high_mark;
417 unsigned long grown;
418 unsigned long reaped;
419 unsigned long errors;
420 unsigned long max_freeable;
421 unsigned long node_allocs;
422 unsigned long node_frees;
423 atomic_t allochit;
424 atomic_t allocmiss;
425 atomic_t freehit;
426 atomic_t freemiss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700427#endif
428#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800429 /*
430 * If debugging is enabled, then the allocator can add additional
431 * fields and/or padding to every object. buffer_size contains the total
432 * object size including these internal fields, the following two
433 * variables contain the offset to the user object and its size.
434 */
435 int obj_offset;
436 int obj_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700437#endif
438};
439
440#define CFLGS_OFF_SLAB (0x80000000UL)
441#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
442
443#define BATCHREFILL_LIMIT 16
Andrew Mortona737b3e2006-03-22 00:08:11 -0800444/*
445 * Optimization question: fewer reaps means less probability for unnessary
446 * cpucache drain/refill cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700447 *
Adrian Bunkdc6f3f22005-11-08 16:44:08 +0100448 * OTOH the cpuarrays can contain lots of objects,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700449 * which could lock up otherwise freeable slabs.
450 */
451#define REAPTIMEOUT_CPUC (2*HZ)
452#define REAPTIMEOUT_LIST3 (4*HZ)
453
454#if STATS
455#define STATS_INC_ACTIVE(x) ((x)->num_active++)
456#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
457#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
458#define STATS_INC_GROWN(x) ((x)->grown++)
459#define STATS_INC_REAPED(x) ((x)->reaped++)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800460#define STATS_SET_HIGH(x) \
461 do { \
462 if ((x)->num_active > (x)->high_mark) \
463 (x)->high_mark = (x)->num_active; \
464 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700465#define STATS_INC_ERR(x) ((x)->errors++)
466#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
Christoph Lametere498be72005-09-09 13:03:32 -0700467#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800468#define STATS_SET_FREEABLE(x, i) \
469 do { \
470 if ((x)->max_freeable < i) \
471 (x)->max_freeable = i; \
472 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700473#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
474#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
475#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
476#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
477#else
478#define STATS_INC_ACTIVE(x) do { } while (0)
479#define STATS_DEC_ACTIVE(x) do { } while (0)
480#define STATS_INC_ALLOCED(x) do { } while (0)
481#define STATS_INC_GROWN(x) do { } while (0)
482#define STATS_INC_REAPED(x) do { } while (0)
483#define STATS_SET_HIGH(x) do { } while (0)
484#define STATS_INC_ERR(x) do { } while (0)
485#define STATS_INC_NODEALLOCS(x) do { } while (0)
Christoph Lametere498be72005-09-09 13:03:32 -0700486#define STATS_INC_NODEFREES(x) do { } while (0)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800487#define STATS_SET_FREEABLE(x, i) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700488#define STATS_INC_ALLOCHIT(x) do { } while (0)
489#define STATS_INC_ALLOCMISS(x) do { } while (0)
490#define STATS_INC_FREEHIT(x) do { } while (0)
491#define STATS_INC_FREEMISS(x) do { } while (0)
492#endif
493
494#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -0800495/*
496 * Magic nums for obj red zoning.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700497 * Placed in the first word before and the first word after an obj.
498 */
499#define RED_INACTIVE 0x5A2CF071UL /* when obj is inactive */
500#define RED_ACTIVE 0x170FC2A5UL /* when obj is active */
501
502/* ...and for poisoning */
503#define POISON_INUSE 0x5a /* for use-uninitialised poisoning */
504#define POISON_FREE 0x6b /* for use-after-free poisoning */
505#define POISON_END 0xa5 /* end-byte of poisoning */
506
Andrew Mortona737b3e2006-03-22 00:08:11 -0800507/*
508 * memory layout of objects:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700509 * 0 : objp
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800510 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
Linus Torvalds1da177e2005-04-16 15:20:36 -0700511 * the end of an object is aligned with the end of the real
512 * allocation. Catches writes behind the end of the allocation.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800513 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700514 * redzone word.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800515 * cachep->obj_offset: The real object.
516 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
Andrew Mortona737b3e2006-03-22 00:08:11 -0800517 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
518 * [BYTES_PER_WORD long]
Linus Torvalds1da177e2005-04-16 15:20:36 -0700519 */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800520static int obj_offset(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700521{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800522 return cachep->obj_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700523}
524
Pekka Enberg343e0d72006-02-01 03:05:50 -0800525static int obj_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700526{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800527 return cachep->obj_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700528}
529
Pekka Enberg343e0d72006-02-01 03:05:50 -0800530static unsigned long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700531{
532 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800533 return (unsigned long*) (objp+obj_offset(cachep)-BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700534}
535
Pekka Enberg343e0d72006-02-01 03:05:50 -0800536static unsigned long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700537{
538 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
539 if (cachep->flags & SLAB_STORE_USER)
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800540 return (unsigned long *)(objp + cachep->buffer_size -
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800541 2 * BYTES_PER_WORD);
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800542 return (unsigned long *)(objp + cachep->buffer_size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700543}
544
Pekka Enberg343e0d72006-02-01 03:05:50 -0800545static void **dbg_userword(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700546{
547 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800548 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700549}
550
551#else
552
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800553#define obj_offset(x) 0
554#define obj_size(cachep) (cachep->buffer_size)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700555#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;})
556#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;})
557#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
558
559#endif
560
561/*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800562 * Maximum size of an obj (in 2^order pages) and absolute limit for the gfp
563 * order.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700564 */
565#if defined(CONFIG_LARGE_ALLOCS)
566#define MAX_OBJ_ORDER 13 /* up to 32Mb */
567#define MAX_GFP_ORDER 13 /* up to 32Mb */
568#elif defined(CONFIG_MMU)
569#define MAX_OBJ_ORDER 5 /* 32 pages */
570#define MAX_GFP_ORDER 5 /* 32 pages */
571#else
572#define MAX_OBJ_ORDER 8 /* up to 1Mb */
573#define MAX_GFP_ORDER 8 /* up to 1Mb */
574#endif
575
576/*
577 * Do not go above this order unless 0 objects fit into the slab.
578 */
579#define BREAK_GFP_ORDER_HI 1
580#define BREAK_GFP_ORDER_LO 0
581static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
582
Andrew Mortona737b3e2006-03-22 00:08:11 -0800583/*
584 * Functions for storing/retrieving the cachep and or slab from the page
585 * allocator. These are used to find the slab an obj belongs to. With kfree(),
586 * these are used to find the cache which an obj belongs to.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700587 */
Pekka Enberg065d41c2005-11-13 16:06:46 -0800588static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
589{
590 page->lru.next = (struct list_head *)cache;
591}
592
593static inline struct kmem_cache *page_get_cache(struct page *page)
594{
Nick Piggin84097512006-03-22 00:08:34 -0800595 if (unlikely(PageCompound(page)))
596 page = (struct page *)page_private(page);
Pekka Enberg065d41c2005-11-13 16:06:46 -0800597 return (struct kmem_cache *)page->lru.next;
598}
599
600static inline void page_set_slab(struct page *page, struct slab *slab)
601{
602 page->lru.prev = (struct list_head *)slab;
603}
604
605static inline struct slab *page_get_slab(struct page *page)
606{
Nick Piggin84097512006-03-22 00:08:34 -0800607 if (unlikely(PageCompound(page)))
608 page = (struct page *)page_private(page);
Pekka Enberg065d41c2005-11-13 16:06:46 -0800609 return (struct slab *)page->lru.prev;
610}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700611
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800612static inline struct kmem_cache *virt_to_cache(const void *obj)
613{
614 struct page *page = virt_to_page(obj);
615 return page_get_cache(page);
616}
617
618static inline struct slab *virt_to_slab(const void *obj)
619{
620 struct page *page = virt_to_page(obj);
621 return page_get_slab(page);
622}
623
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800624static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
625 unsigned int idx)
626{
627 return slab->s_mem + cache->buffer_size * idx;
628}
629
630static inline unsigned int obj_to_index(struct kmem_cache *cache,
631 struct slab *slab, void *obj)
632{
633 return (unsigned)(obj - slab->s_mem) / cache->buffer_size;
634}
635
Andrew Mortona737b3e2006-03-22 00:08:11 -0800636/*
637 * These are the default caches for kmalloc. Custom caches can have other sizes.
638 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700639struct cache_sizes malloc_sizes[] = {
640#define CACHE(x) { .cs_size = (x) },
641#include <linux/kmalloc_sizes.h>
642 CACHE(ULONG_MAX)
643#undef CACHE
644};
645EXPORT_SYMBOL(malloc_sizes);
646
647/* Must match cache_sizes above. Out of line to keep cache footprint low. */
648struct cache_names {
649 char *name;
650 char *name_dma;
651};
652
653static struct cache_names __initdata cache_names[] = {
654#define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
655#include <linux/kmalloc_sizes.h>
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800656 {NULL,}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700657#undef CACHE
658};
659
660static struct arraycache_init initarray_cache __initdata =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800661 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700662static struct arraycache_init initarray_generic =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800663 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700664
665/* internal cache of cache description objs */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800666static struct kmem_cache cache_cache = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800667 .batchcount = 1,
668 .limit = BOOT_CPUCACHE_ENTRIES,
669 .shared = 1,
Pekka Enberg343e0d72006-02-01 03:05:50 -0800670 .buffer_size = sizeof(struct kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800671 .name = "kmem_cache",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700672#if DEBUG
Pekka Enberg343e0d72006-02-01 03:05:50 -0800673 .obj_size = sizeof(struct kmem_cache),
Linus Torvalds1da177e2005-04-16 15:20:36 -0700674#endif
675};
676
677/* Guard access to the cache-chain. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800678static DEFINE_MUTEX(cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700679static struct list_head cache_chain;
680
681/*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800682 * vm_enough_memory() looks at this to determine how many slab-allocated pages
683 * are possibly freeable under pressure
Linus Torvalds1da177e2005-04-16 15:20:36 -0700684 *
685 * SLAB_RECLAIM_ACCOUNT turns this on per-slab
686 */
687atomic_t slab_reclaim_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700688
689/*
690 * chicken and egg problem: delay the per-cpu array allocation
691 * until the general caches are up.
692 */
693static enum {
694 NONE,
Christoph Lametere498be72005-09-09 13:03:32 -0700695 PARTIAL_AC,
696 PARTIAL_L3,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700697 FULL
698} g_cpucache_up;
699
700static DEFINE_PER_CPU(struct work_struct, reap_work);
701
Andrew Mortona737b3e2006-03-22 00:08:11 -0800702static void free_block(struct kmem_cache *cachep, void **objpp, int len,
703 int node);
Pekka Enberg343e0d72006-02-01 03:05:50 -0800704static void enable_cpucache(struct kmem_cache *cachep);
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800705static void cache_reap(void *unused);
Pekka Enberg343e0d72006-02-01 03:05:50 -0800706static int __node_shrink(struct kmem_cache *cachep, int node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700707
Pekka Enberg343e0d72006-02-01 03:05:50 -0800708static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700709{
710 return cachep->array[smp_processor_id()];
711}
712
Andrew Mortona737b3e2006-03-22 00:08:11 -0800713static inline struct kmem_cache *__find_general_cachep(size_t size,
714 gfp_t gfpflags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700715{
716 struct cache_sizes *csizep = malloc_sizes;
717
718#if DEBUG
719 /* This happens if someone tries to call
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800720 * kmem_cache_create(), or __kmalloc(), before
721 * the generic caches are initialized.
722 */
Alok Katariac7e43c72005-09-14 12:17:53 -0700723 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700724#endif
725 while (size > csizep->cs_size)
726 csizep++;
727
728 /*
Martin Hicks0abf40c2005-09-03 15:54:54 -0700729 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
Linus Torvalds1da177e2005-04-16 15:20:36 -0700730 * has cs_{dma,}cachep==NULL. Thus no special case
731 * for large kmalloc calls required.
732 */
733 if (unlikely(gfpflags & GFP_DMA))
734 return csizep->cs_dmacachep;
735 return csizep->cs_cachep;
736}
737
Pekka Enberg343e0d72006-02-01 03:05:50 -0800738struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700739{
740 return __find_general_cachep(size, gfpflags);
741}
742EXPORT_SYMBOL(kmem_find_general_cachep);
743
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800744static size_t slab_mgmt_size(size_t nr_objs, size_t align)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700745{
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800746 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
747}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700748
Andrew Mortona737b3e2006-03-22 00:08:11 -0800749/*
750 * Calculate the number of objects and left-over bytes for a given buffer size.
751 */
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800752static void cache_estimate(unsigned long gfporder, size_t buffer_size,
753 size_t align, int flags, size_t *left_over,
754 unsigned int *num)
755{
756 int nr_objs;
757 size_t mgmt_size;
758 size_t slab_size = PAGE_SIZE << gfporder;
759
760 /*
761 * The slab management structure can be either off the slab or
762 * on it. For the latter case, the memory allocated for a
763 * slab is used for:
764 *
765 * - The struct slab
766 * - One kmem_bufctl_t for each object
767 * - Padding to respect alignment of @align
768 * - @buffer_size bytes for each object
769 *
770 * If the slab management structure is off the slab, then the
771 * alignment will already be calculated into the size. Because
772 * the slabs are all pages aligned, the objects will be at the
773 * correct alignment when allocated.
774 */
775 if (flags & CFLGS_OFF_SLAB) {
776 mgmt_size = 0;
777 nr_objs = slab_size / buffer_size;
778
779 if (nr_objs > SLAB_LIMIT)
780 nr_objs = SLAB_LIMIT;
781 } else {
782 /*
783 * Ignore padding for the initial guess. The padding
784 * is at most @align-1 bytes, and @buffer_size is at
785 * least @align. In the worst case, this result will
786 * be one greater than the number of objects that fit
787 * into the memory allocation when taking the padding
788 * into account.
789 */
790 nr_objs = (slab_size - sizeof(struct slab)) /
791 (buffer_size + sizeof(kmem_bufctl_t));
792
793 /*
794 * This calculated number will be either the right
795 * amount, or one greater than what we want.
796 */
797 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
798 > slab_size)
799 nr_objs--;
800
801 if (nr_objs > SLAB_LIMIT)
802 nr_objs = SLAB_LIMIT;
803
804 mgmt_size = slab_mgmt_size(nr_objs, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700805 }
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800806 *num = nr_objs;
807 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700808}
809
810#define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
811
Andrew Mortona737b3e2006-03-22 00:08:11 -0800812static void __slab_error(const char *function, struct kmem_cache *cachep,
813 char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700814{
815 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800816 function, cachep->name, msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700817 dump_stack();
818}
819
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800820#ifdef CONFIG_NUMA
821/*
822 * Special reaping functions for NUMA systems called from cache_reap().
823 * These take care of doing round robin flushing of alien caches (containing
824 * objects freed on different nodes from which they were allocated) and the
825 * flushing of remote pcps by calling drain_node_pages.
826 */
827static DEFINE_PER_CPU(unsigned long, reap_node);
828
829static void init_reap_node(int cpu)
830{
831 int node;
832
833 node = next_node(cpu_to_node(cpu), node_online_map);
834 if (node == MAX_NUMNODES)
Paul Jackson442295c2006-03-22 00:09:11 -0800835 node = first_node(node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800836
837 __get_cpu_var(reap_node) = node;
838}
839
840static void next_reap_node(void)
841{
842 int node = __get_cpu_var(reap_node);
843
844 /*
845 * Also drain per cpu pages on remote zones
846 */
847 if (node != numa_node_id())
848 drain_node_pages(node);
849
850 node = next_node(node, node_online_map);
851 if (unlikely(node >= MAX_NUMNODES))
852 node = first_node(node_online_map);
853 __get_cpu_var(reap_node) = node;
854}
855
856#else
857#define init_reap_node(cpu) do { } while (0)
858#define next_reap_node(void) do { } while (0)
859#endif
860
Linus Torvalds1da177e2005-04-16 15:20:36 -0700861/*
862 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
863 * via the workqueue/eventd.
864 * Add the CPU number into the expiration time to minimize the possibility of
865 * the CPUs getting into lockstep and contending for the global cache chain
866 * lock.
867 */
868static void __devinit start_cpu_timer(int cpu)
869{
870 struct work_struct *reap_work = &per_cpu(reap_work, cpu);
871
872 /*
873 * When this gets called from do_initcalls via cpucache_init(),
874 * init_workqueues() has already run, so keventd will be setup
875 * at that time.
876 */
877 if (keventd_up() && reap_work->func == NULL) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800878 init_reap_node(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700879 INIT_WORK(reap_work, cache_reap, NULL);
880 schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu);
881 }
882}
883
Christoph Lametere498be72005-09-09 13:03:32 -0700884static struct array_cache *alloc_arraycache(int node, int entries,
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800885 int batchcount)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700886{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800887 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700888 struct array_cache *nc = NULL;
889
Christoph Lametere498be72005-09-09 13:03:32 -0700890 nc = kmalloc_node(memsize, GFP_KERNEL, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700891 if (nc) {
892 nc->avail = 0;
893 nc->limit = entries;
894 nc->batchcount = batchcount;
895 nc->touched = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700896 spin_lock_init(&nc->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700897 }
898 return nc;
899}
900
Christoph Lameter3ded1752006-03-25 03:06:44 -0800901/*
902 * Transfer objects in one arraycache to another.
903 * Locking must be handled by the caller.
904 *
905 * Return the number of entries transferred.
906 */
907static int transfer_objects(struct array_cache *to,
908 struct array_cache *from, unsigned int max)
909{
910 /* Figure out how many entries to transfer */
911 int nr = min(min(from->avail, max), to->limit - to->avail);
912
913 if (!nr)
914 return 0;
915
916 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
917 sizeof(void *) *nr);
918
919 from->avail -= nr;
920 to->avail += nr;
921 to->touched = 1;
922 return nr;
923}
924
Christoph Lametere498be72005-09-09 13:03:32 -0700925#ifdef CONFIG_NUMA
Pekka Enberg343e0d72006-02-01 03:05:50 -0800926static void *__cache_alloc_node(struct kmem_cache *, gfp_t, int);
Paul Jacksonc61afb12006-03-24 03:16:08 -0800927static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
Christoph Lameterdc85da12006-01-18 17:42:36 -0800928
Pekka Enberg5295a742006-02-01 03:05:48 -0800929static struct array_cache **alloc_alien_cache(int node, int limit)
Christoph Lametere498be72005-09-09 13:03:32 -0700930{
931 struct array_cache **ac_ptr;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800932 int memsize = sizeof(void *) * MAX_NUMNODES;
Christoph Lametere498be72005-09-09 13:03:32 -0700933 int i;
934
935 if (limit > 1)
936 limit = 12;
937 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
938 if (ac_ptr) {
939 for_each_node(i) {
940 if (i == node || !node_online(i)) {
941 ac_ptr[i] = NULL;
942 continue;
943 }
944 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
945 if (!ac_ptr[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800946 for (i--; i <= 0; i--)
Christoph Lametere498be72005-09-09 13:03:32 -0700947 kfree(ac_ptr[i]);
948 kfree(ac_ptr);
949 return NULL;
950 }
951 }
952 }
953 return ac_ptr;
954}
955
Pekka Enberg5295a742006-02-01 03:05:48 -0800956static void free_alien_cache(struct array_cache **ac_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -0700957{
958 int i;
959
960 if (!ac_ptr)
961 return;
Christoph Lametere498be72005-09-09 13:03:32 -0700962 for_each_node(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800963 kfree(ac_ptr[i]);
Christoph Lametere498be72005-09-09 13:03:32 -0700964 kfree(ac_ptr);
965}
966
Pekka Enberg343e0d72006-02-01 03:05:50 -0800967static void __drain_alien_cache(struct kmem_cache *cachep,
Pekka Enberg5295a742006-02-01 03:05:48 -0800968 struct array_cache *ac, int node)
Christoph Lametere498be72005-09-09 13:03:32 -0700969{
970 struct kmem_list3 *rl3 = cachep->nodelists[node];
971
972 if (ac->avail) {
973 spin_lock(&rl3->list_lock);
Christoph Lametere00946f2006-03-25 03:06:45 -0800974 /*
975 * Stuff objects into the remote nodes shared array first.
976 * That way we could avoid the overhead of putting the objects
977 * into the free lists and getting them back later.
978 */
979 transfer_objects(rl3->shared, ac, ac->limit);
980
Christoph Lameterff694162005-09-22 21:44:02 -0700981 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -0700982 ac->avail = 0;
983 spin_unlock(&rl3->list_lock);
984 }
985}
986
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800987/*
988 * Called from cache_reap() to regularly drain alien caches round robin.
989 */
990static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
991{
992 int node = __get_cpu_var(reap_node);
993
994 if (l3->alien) {
995 struct array_cache *ac = l3->alien[node];
Christoph Lametere00946f2006-03-25 03:06:45 -0800996
997 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800998 __drain_alien_cache(cachep, ac, node);
999 spin_unlock_irq(&ac->lock);
1000 }
1001 }
1002}
1003
Andrew Mortona737b3e2006-03-22 00:08:11 -08001004static void drain_alien_cache(struct kmem_cache *cachep,
1005 struct array_cache **alien)
Christoph Lametere498be72005-09-09 13:03:32 -07001006{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001007 int i = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07001008 struct array_cache *ac;
1009 unsigned long flags;
1010
1011 for_each_online_node(i) {
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001012 ac = alien[i];
Christoph Lametere498be72005-09-09 13:03:32 -07001013 if (ac) {
1014 spin_lock_irqsave(&ac->lock, flags);
1015 __drain_alien_cache(cachep, ac, i);
1016 spin_unlock_irqrestore(&ac->lock, flags);
1017 }
1018 }
1019}
1020#else
Linus Torvalds7a21ef62006-02-05 11:26:38 -08001021
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001022#define drain_alien_cache(cachep, alien) do { } while (0)
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001023#define reap_alien(cachep, l3) do { } while (0)
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001024
Linus Torvalds7a21ef62006-02-05 11:26:38 -08001025static inline struct array_cache **alloc_alien_cache(int node, int limit)
1026{
1027 return (struct array_cache **) 0x01020304ul;
1028}
1029
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001030static inline void free_alien_cache(struct array_cache **ac_ptr)
1031{
1032}
Linus Torvalds7a21ef62006-02-05 11:26:38 -08001033
Christoph Lametere498be72005-09-09 13:03:32 -07001034#endif
1035
Linus Torvalds1da177e2005-04-16 15:20:36 -07001036static int __devinit cpuup_callback(struct notifier_block *nfb,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001037 unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001038{
1039 long cpu = (long)hcpu;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001040 struct kmem_cache *cachep;
Christoph Lametere498be72005-09-09 13:03:32 -07001041 struct kmem_list3 *l3 = NULL;
1042 int node = cpu_to_node(cpu);
1043 int memsize = sizeof(struct kmem_list3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001044
1045 switch (action) {
1046 case CPU_UP_PREPARE:
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001047 mutex_lock(&cache_chain_mutex);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001048 /*
1049 * We need to do this right in the beginning since
Christoph Lametere498be72005-09-09 13:03:32 -07001050 * alloc_arraycache's are going to use this list.
1051 * kmalloc_node allows us to add the slab to the right
1052 * kmem_list3 and not this cpu's kmem_list3
1053 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001054
Christoph Lametere498be72005-09-09 13:03:32 -07001055 list_for_each_entry(cachep, &cache_chain, next) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08001056 /*
1057 * Set up the size64 kmemlist for cpu before we can
Christoph Lametere498be72005-09-09 13:03:32 -07001058 * begin anything. Make sure some other cpu on this
1059 * node has not already allocated this
1060 */
1061 if (!cachep->nodelists[node]) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08001062 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1063 if (!l3)
Christoph Lametere498be72005-09-09 13:03:32 -07001064 goto bad;
1065 kmem_list3_init(l3);
1066 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001067 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -07001068
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001069 /*
1070 * The l3s don't come and go as CPUs come and
1071 * go. cache_chain_mutex is sufficient
1072 * protection here.
1073 */
Christoph Lametere498be72005-09-09 13:03:32 -07001074 cachep->nodelists[node] = l3;
1075 }
1076
1077 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1078 cachep->nodelists[node]->free_limit =
Andrew Mortona737b3e2006-03-22 00:08:11 -08001079 (1 + nr_cpus_node(node)) *
1080 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07001081 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1082 }
1083
Andrew Mortona737b3e2006-03-22 00:08:11 -08001084 /*
1085 * Now we can go ahead with allocating the shared arrays and
1086 * array caches
1087 */
Christoph Lametere498be72005-09-09 13:03:32 -07001088 list_for_each_entry(cachep, &cache_chain, next) {
Tobias Klausercd105df2006-01-08 01:00:59 -08001089 struct array_cache *nc;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001090 struct array_cache *shared;
1091 struct array_cache **alien;
Tobias Klausercd105df2006-01-08 01:00:59 -08001092
Christoph Lametere498be72005-09-09 13:03:32 -07001093 nc = alloc_arraycache(node, cachep->limit,
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001094 cachep->batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001095 if (!nc)
1096 goto bad;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001097 shared = alloc_arraycache(node,
1098 cachep->shared * cachep->batchcount,
1099 0xbaadf00d);
1100 if (!shared)
1101 goto bad;
Linus Torvalds7a21ef62006-02-05 11:26:38 -08001102
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001103 alien = alloc_alien_cache(node, cachep->limit);
1104 if (!alien)
1105 goto bad;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001106 cachep->array[cpu] = nc;
Christoph Lametere498be72005-09-09 13:03:32 -07001107 l3 = cachep->nodelists[node];
1108 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07001109
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001110 spin_lock_irq(&l3->list_lock);
1111 if (!l3->shared) {
1112 /*
1113 * We are serialised from CPU_DEAD or
1114 * CPU_UP_CANCELLED by the cpucontrol lock
1115 */
1116 l3->shared = shared;
1117 shared = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07001118 }
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001119#ifdef CONFIG_NUMA
1120 if (!l3->alien) {
1121 l3->alien = alien;
1122 alien = NULL;
1123 }
1124#endif
1125 spin_unlock_irq(&l3->list_lock);
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001126 kfree(shared);
1127 free_alien_cache(alien);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001128 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001129 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001130 break;
1131 case CPU_ONLINE:
1132 start_cpu_timer(cpu);
1133 break;
1134#ifdef CONFIG_HOTPLUG_CPU
1135 case CPU_DEAD:
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001136 /*
1137 * Even if all the cpus of a node are down, we don't free the
1138 * kmem_list3 of any cache. This to avoid a race between
1139 * cpu_down, and a kmalloc allocation from another cpu for
1140 * memory from the node of the cpu going down. The list3
1141 * structure is usually allocated from kmem_cache_create() and
1142 * gets destroyed at kmem_cache_destroy().
1143 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001144 /* fall thru */
1145 case CPU_UP_CANCELED:
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001146 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001147 list_for_each_entry(cachep, &cache_chain, next) {
1148 struct array_cache *nc;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001149 struct array_cache *shared;
1150 struct array_cache **alien;
Christoph Lametere498be72005-09-09 13:03:32 -07001151 cpumask_t mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001152
Christoph Lametere498be72005-09-09 13:03:32 -07001153 mask = node_to_cpumask(node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001154 /* cpu is dead; no one can alloc from it. */
1155 nc = cachep->array[cpu];
1156 cachep->array[cpu] = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07001157 l3 = cachep->nodelists[node];
1158
1159 if (!l3)
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001160 goto free_array_cache;
Christoph Lametere498be72005-09-09 13:03:32 -07001161
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08001162 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07001163
1164 /* Free limit for this kmem_list3 */
1165 l3->free_limit -= cachep->batchcount;
1166 if (nc)
Christoph Lameterff694162005-09-22 21:44:02 -07001167 free_block(cachep, nc->entry, nc->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001168
1169 if (!cpus_empty(mask)) {
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08001170 spin_unlock_irq(&l3->list_lock);
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001171 goto free_array_cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001172 }
Christoph Lametere498be72005-09-09 13:03:32 -07001173
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001174 shared = l3->shared;
1175 if (shared) {
Christoph Lametere498be72005-09-09 13:03:32 -07001176 free_block(cachep, l3->shared->entry,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001177 l3->shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001178 l3->shared = NULL;
1179 }
Christoph Lametere498be72005-09-09 13:03:32 -07001180
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001181 alien = l3->alien;
1182 l3->alien = NULL;
1183
1184 spin_unlock_irq(&l3->list_lock);
1185
1186 kfree(shared);
1187 if (alien) {
1188 drain_alien_cache(cachep, alien);
1189 free_alien_cache(alien);
Christoph Lametere498be72005-09-09 13:03:32 -07001190 }
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001191free_array_cache:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001192 kfree(nc);
1193 }
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001194 /*
1195 * In the previous loop, all the objects were freed to
1196 * the respective cache's slabs, now we can go ahead and
1197 * shrink each nodelist to its limit.
1198 */
1199 list_for_each_entry(cachep, &cache_chain, next) {
1200 l3 = cachep->nodelists[node];
1201 if (!l3)
1202 continue;
1203 spin_lock_irq(&l3->list_lock);
1204 /* free slabs belonging to this node */
1205 __node_shrink(cachep, node);
1206 spin_unlock_irq(&l3->list_lock);
1207 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001208 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001209 break;
1210#endif
1211 }
1212 return NOTIFY_OK;
Andrew Mortona737b3e2006-03-22 00:08:11 -08001213bad:
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001214 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001215 return NOTIFY_BAD;
1216}
1217
1218static struct notifier_block cpucache_notifier = { &cpuup_callback, NULL, 0 };
1219
Christoph Lametere498be72005-09-09 13:03:32 -07001220/*
1221 * swap the static kmem_list3 with kmalloced memory
1222 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001223static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1224 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001225{
1226 struct kmem_list3 *ptr;
1227
1228 BUG_ON(cachep->nodelists[nodeid] != list);
1229 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1230 BUG_ON(!ptr);
1231
1232 local_irq_disable();
1233 memcpy(ptr, list, sizeof(struct kmem_list3));
1234 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1235 cachep->nodelists[nodeid] = ptr;
1236 local_irq_enable();
1237}
1238
Andrew Mortona737b3e2006-03-22 00:08:11 -08001239/*
1240 * Initialisation. Called after the page allocator have been initialised and
1241 * before smp_init().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001242 */
1243void __init kmem_cache_init(void)
1244{
1245 size_t left_over;
1246 struct cache_sizes *sizes;
1247 struct cache_names *names;
Christoph Lametere498be72005-09-09 13:03:32 -07001248 int i;
Jack Steiner07ed76b2006-03-07 21:55:46 -08001249 int order;
Christoph Lametere498be72005-09-09 13:03:32 -07001250
1251 for (i = 0; i < NUM_INIT_LISTS; i++) {
1252 kmem_list3_init(&initkmem_list3[i]);
1253 if (i < MAX_NUMNODES)
1254 cache_cache.nodelists[i] = NULL;
1255 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001256
1257 /*
1258 * Fragmentation resistance on low memory - only use bigger
1259 * page orders on machines with more than 32MB of memory.
1260 */
1261 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1262 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1263
Linus Torvalds1da177e2005-04-16 15:20:36 -07001264 /* Bootstrap is tricky, because several objects are allocated
1265 * from caches that do not exist yet:
Andrew Mortona737b3e2006-03-22 00:08:11 -08001266 * 1) initialize the cache_cache cache: it contains the struct
1267 * kmem_cache structures of all caches, except cache_cache itself:
1268 * cache_cache is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001269 * Initially an __init data area is used for the head array and the
1270 * kmem_list3 structures, it's replaced with a kmalloc allocated
1271 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001272 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001273 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001274 * An __init data area is used for the head array.
1275 * 3) Create the remaining kmalloc caches, with minimally sized
1276 * head arrays.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001277 * 4) Replace the __init data head arrays for cache_cache and the first
1278 * kmalloc cache with kmalloc allocated arrays.
Christoph Lametere498be72005-09-09 13:03:32 -07001279 * 5) Replace the __init data for kmem_list3 for cache_cache and
1280 * the other cache's with kmalloc allocated memory.
1281 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001282 */
1283
1284 /* 1) create the cache_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001285 INIT_LIST_HEAD(&cache_chain);
1286 list_add(&cache_cache.next, &cache_chain);
1287 cache_cache.colour_off = cache_line_size();
1288 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
Christoph Lametere498be72005-09-09 13:03:32 -07001289 cache_cache.nodelists[numa_node_id()] = &initkmem_list3[CACHE_CACHE];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001290
Andrew Mortona737b3e2006-03-22 00:08:11 -08001291 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1292 cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001293
Jack Steiner07ed76b2006-03-07 21:55:46 -08001294 for (order = 0; order < MAX_ORDER; order++) {
1295 cache_estimate(order, cache_cache.buffer_size,
1296 cache_line_size(), 0, &left_over, &cache_cache.num);
1297 if (cache_cache.num)
1298 break;
1299 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001300 if (!cache_cache.num)
1301 BUG();
Jack Steiner07ed76b2006-03-07 21:55:46 -08001302 cache_cache.gfporder = order;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001303 cache_cache.colour = left_over / cache_cache.colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001304 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1305 sizeof(struct slab), cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001306
1307 /* 2+3) create the kmalloc caches */
1308 sizes = malloc_sizes;
1309 names = cache_names;
1310
Andrew Mortona737b3e2006-03-22 00:08:11 -08001311 /*
1312 * Initialize the caches that provide memory for the array cache and the
1313 * kmem_list3 structures first. Without this, further allocations will
1314 * bug.
Christoph Lametere498be72005-09-09 13:03:32 -07001315 */
1316
1317 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001318 sizes[INDEX_AC].cs_size,
1319 ARCH_KMALLOC_MINALIGN,
1320 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1321 NULL, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07001322
Andrew Mortona737b3e2006-03-22 00:08:11 -08001323 if (INDEX_AC != INDEX_L3) {
Christoph Lametere498be72005-09-09 13:03:32 -07001324 sizes[INDEX_L3].cs_cachep =
Andrew Mortona737b3e2006-03-22 00:08:11 -08001325 kmem_cache_create(names[INDEX_L3].name,
1326 sizes[INDEX_L3].cs_size,
1327 ARCH_KMALLOC_MINALIGN,
1328 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1329 NULL, NULL);
1330 }
Christoph Lametere498be72005-09-09 13:03:32 -07001331
Linus Torvalds1da177e2005-04-16 15:20:36 -07001332 while (sizes->cs_size != ULONG_MAX) {
Christoph Lametere498be72005-09-09 13:03:32 -07001333 /*
1334 * For performance, all the general caches are L1 aligned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001335 * This should be particularly beneficial on SMP boxes, as it
1336 * eliminates "false sharing".
1337 * Note for systems short on memory removing the alignment will
Christoph Lametere498be72005-09-09 13:03:32 -07001338 * allow tighter packing of the smaller caches.
1339 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001340 if (!sizes->cs_cachep) {
Christoph Lametere498be72005-09-09 13:03:32 -07001341 sizes->cs_cachep = kmem_cache_create(names->name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001342 sizes->cs_size,
1343 ARCH_KMALLOC_MINALIGN,
1344 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1345 NULL, NULL);
1346 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001347
1348 /* Inc off-slab bufctl limit until the ceiling is hit. */
1349 if (!(OFF_SLAB(sizes->cs_cachep))) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001350 offslab_limit = sizes->cs_size - sizeof(struct slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001351 offslab_limit /= sizeof(kmem_bufctl_t);
1352 }
1353
1354 sizes->cs_dmacachep = kmem_cache_create(names->name_dma,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001355 sizes->cs_size,
1356 ARCH_KMALLOC_MINALIGN,
1357 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1358 SLAB_PANIC,
1359 NULL, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001360 sizes++;
1361 names++;
1362 }
1363 /* 4) Replace the bootstrap head arrays */
1364 {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001365 void *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001366
Linus Torvalds1da177e2005-04-16 15:20:36 -07001367 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001368
Linus Torvalds1da177e2005-04-16 15:20:36 -07001369 local_irq_disable();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001370 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1371 memcpy(ptr, cpu_cache_get(&cache_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001372 sizeof(struct arraycache_init));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001373 cache_cache.array[smp_processor_id()] = ptr;
1374 local_irq_enable();
Christoph Lametere498be72005-09-09 13:03:32 -07001375
Linus Torvalds1da177e2005-04-16 15:20:36 -07001376 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001377
Linus Torvalds1da177e2005-04-16 15:20:36 -07001378 local_irq_disable();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001379 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001380 != &initarray_generic.cache);
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001381 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001382 sizeof(struct arraycache_init));
Christoph Lametere498be72005-09-09 13:03:32 -07001383 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001384 ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001385 local_irq_enable();
1386 }
Christoph Lametere498be72005-09-09 13:03:32 -07001387 /* 5) Replace the bootstrap kmem_list3's */
1388 {
1389 int node;
1390 /* Replace the static kmem_list3 structures for the boot cpu */
1391 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE],
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001392 numa_node_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001393
Christoph Lametere498be72005-09-09 13:03:32 -07001394 for_each_online_node(node) {
1395 init_list(malloc_sizes[INDEX_AC].cs_cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001396 &initkmem_list3[SIZE_AC + node], node);
Christoph Lametere498be72005-09-09 13:03:32 -07001397
1398 if (INDEX_AC != INDEX_L3) {
1399 init_list(malloc_sizes[INDEX_L3].cs_cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001400 &initkmem_list3[SIZE_L3 + node],
1401 node);
Christoph Lametere498be72005-09-09 13:03:32 -07001402 }
1403 }
1404 }
1405
1406 /* 6) resize the head arrays to their final sizes */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001407 {
Pekka Enberg343e0d72006-02-01 03:05:50 -08001408 struct kmem_cache *cachep;
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001409 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001410 list_for_each_entry(cachep, &cache_chain, next)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001411 enable_cpucache(cachep);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001412 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001413 }
1414
1415 /* Done! */
1416 g_cpucache_up = FULL;
1417
Andrew Mortona737b3e2006-03-22 00:08:11 -08001418 /*
1419 * Register a cpu startup notifier callback that initializes
1420 * cpu_cache_get for all new cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07001421 */
1422 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001423
Andrew Mortona737b3e2006-03-22 00:08:11 -08001424 /*
1425 * The reap timers are started later, with a module init call: That part
1426 * of the kernel is not yet operational.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001427 */
1428}
1429
1430static int __init cpucache_init(void)
1431{
1432 int cpu;
1433
Andrew Mortona737b3e2006-03-22 00:08:11 -08001434 /*
1435 * Register the timers that return unneeded pages to the page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -07001436 */
Christoph Lametere498be72005-09-09 13:03:32 -07001437 for_each_online_cpu(cpu)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001438 start_cpu_timer(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001439 return 0;
1440}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001441__initcall(cpucache_init);
1442
1443/*
1444 * Interface to system's page allocator. No need to hold the cache-lock.
1445 *
1446 * If we requested dmaable memory, we will get it. Even if we
1447 * did not request dmaable memory, we might get it, but that
1448 * would be relatively rare and ignorable.
1449 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001450static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001451{
1452 struct page *page;
1453 void *addr;
1454 int i;
1455
1456 flags |= cachep->gfpflags;
Christoph Lameter50c85a12005-11-13 16:06:47 -08001457 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001458 if (!page)
1459 return NULL;
1460 addr = page_address(page);
1461
1462 i = (1 << cachep->gfporder);
1463 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1464 atomic_add(i, &slab_reclaim_pages);
1465 add_page_state(nr_slab, i);
1466 while (i--) {
Nick Pigginf205b2f2006-03-22 00:08:02 -08001467 __SetPageSlab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001468 page++;
1469 }
1470 return addr;
1471}
1472
1473/*
1474 * Interface to system's page release.
1475 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001476static void kmem_freepages(struct kmem_cache *cachep, void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001477{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001478 unsigned long i = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001479 struct page *page = virt_to_page(addr);
1480 const unsigned long nr_freed = i;
1481
1482 while (i--) {
Nick Pigginf205b2f2006-03-22 00:08:02 -08001483 BUG_ON(!PageSlab(page));
1484 __ClearPageSlab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001485 page++;
1486 }
1487 sub_page_state(nr_slab, nr_freed);
1488 if (current->reclaim_state)
1489 current->reclaim_state->reclaimed_slab += nr_freed;
1490 free_pages((unsigned long)addr, cachep->gfporder);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001491 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1492 atomic_sub(1 << cachep->gfporder, &slab_reclaim_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001493}
1494
1495static void kmem_rcu_free(struct rcu_head *head)
1496{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001497 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001498 struct kmem_cache *cachep = slab_rcu->cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001499
1500 kmem_freepages(cachep, slab_rcu->addr);
1501 if (OFF_SLAB(cachep))
1502 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1503}
1504
1505#if DEBUG
1506
1507#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001508static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001509 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001510{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001511 int size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001512
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001513 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001514
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001515 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001516 return;
1517
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001518 *addr++ = 0x12345678;
1519 *addr++ = caller;
1520 *addr++ = smp_processor_id();
1521 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001522 {
1523 unsigned long *sptr = &caller;
1524 unsigned long svalue;
1525
1526 while (!kstack_end(sptr)) {
1527 svalue = *sptr++;
1528 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001529 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001530 size -= sizeof(unsigned long);
1531 if (size <= sizeof(unsigned long))
1532 break;
1533 }
1534 }
1535
1536 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001537 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001538}
1539#endif
1540
Pekka Enberg343e0d72006-02-01 03:05:50 -08001541static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001542{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001543 int size = obj_size(cachep);
1544 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001545
1546 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001547 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001548}
1549
1550static void dump_line(char *data, int offset, int limit)
1551{
1552 int i;
1553 printk(KERN_ERR "%03x:", offset);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001554 for (i = 0; i < limit; i++)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001555 printk(" %02x", (unsigned char)data[offset + i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001556 printk("\n");
1557}
1558#endif
1559
1560#if DEBUG
1561
Pekka Enberg343e0d72006-02-01 03:05:50 -08001562static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001563{
1564 int i, size;
1565 char *realobj;
1566
1567 if (cachep->flags & SLAB_RED_ZONE) {
1568 printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001569 *dbg_redzone1(cachep, objp),
1570 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001571 }
1572
1573 if (cachep->flags & SLAB_STORE_USER) {
1574 printk(KERN_ERR "Last user: [<%p>]",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001575 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001576 print_symbol("(%s)",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001577 (unsigned long)*dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001578 printk("\n");
1579 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001580 realobj = (char *)objp + obj_offset(cachep);
1581 size = obj_size(cachep);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001582 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001583 int limit;
1584 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001585 if (i + limit > size)
1586 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001587 dump_line(realobj, i, limit);
1588 }
1589}
1590
Pekka Enberg343e0d72006-02-01 03:05:50 -08001591static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001592{
1593 char *realobj;
1594 int size, i;
1595 int lines = 0;
1596
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001597 realobj = (char *)objp + obj_offset(cachep);
1598 size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001599
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001600 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001601 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001602 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001603 exp = POISON_END;
1604 if (realobj[i] != exp) {
1605 int limit;
1606 /* Mismatch ! */
1607 /* Print header */
1608 if (lines == 0) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001609 printk(KERN_ERR
Andrew Mortona737b3e2006-03-22 00:08:11 -08001610 "Slab corruption: start=%p, len=%d\n",
1611 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001612 print_objinfo(cachep, objp, 0);
1613 }
1614 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001615 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001616 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001617 if (i + limit > size)
1618 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001619 dump_line(realobj, i, limit);
1620 i += 16;
1621 lines++;
1622 /* Limit to 5 lines */
1623 if (lines > 5)
1624 break;
1625 }
1626 }
1627 if (lines != 0) {
1628 /* Print some data about the neighboring objects, if they
1629 * exist:
1630 */
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08001631 struct slab *slabp = virt_to_slab(objp);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001632 unsigned int objnr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001633
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001634 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001635 if (objnr) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001636 objp = index_to_obj(cachep, slabp, objnr - 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001637 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001638 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001639 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001640 print_objinfo(cachep, objp, 2);
1641 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001642 if (objnr + 1 < cachep->num) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001643 objp = index_to_obj(cachep, slabp, objnr + 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001644 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001645 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001646 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001647 print_objinfo(cachep, objp, 2);
1648 }
1649 }
1650}
1651#endif
1652
Linus Torvalds1da177e2005-04-16 15:20:36 -07001653#if DEBUG
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001654/**
Randy Dunlap911851e2006-03-22 00:08:14 -08001655 * slab_destroy_objs - destroy a slab and its objects
1656 * @cachep: cache pointer being destroyed
1657 * @slabp: slab pointer being destroyed
1658 *
1659 * Call the registered destructor for each object in a slab that is being
1660 * destroyed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001661 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001662static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001663{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001664 int i;
1665 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001666 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001667
1668 if (cachep->flags & SLAB_POISON) {
1669#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08001670 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1671 OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001672 kernel_map_pages(virt_to_page(objp),
Andrew Mortona737b3e2006-03-22 00:08:11 -08001673 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001674 else
1675 check_poison_obj(cachep, objp);
1676#else
1677 check_poison_obj(cachep, objp);
1678#endif
1679 }
1680 if (cachep->flags & SLAB_RED_ZONE) {
1681 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1682 slab_error(cachep, "start of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001683 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001684 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1685 slab_error(cachep, "end of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001686 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001687 }
1688 if (cachep->dtor && !(cachep->flags & SLAB_POISON))
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001689 (cachep->dtor) (objp + obj_offset(cachep), cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001690 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001691}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001692#else
Pekka Enberg343e0d72006-02-01 03:05:50 -08001693static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001694{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001695 if (cachep->dtor) {
1696 int i;
1697 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001698 void *objp = index_to_obj(cachep, slabp, i);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001699 (cachep->dtor) (objp, cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001700 }
1701 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001702}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001703#endif
1704
Randy Dunlap911851e2006-03-22 00:08:14 -08001705/**
1706 * slab_destroy - destroy and release all objects in a slab
1707 * @cachep: cache pointer being destroyed
1708 * @slabp: slab pointer being destroyed
1709 *
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001710 * Destroy all the objs in a slab, and release the mem back to the system.
Andrew Mortona737b3e2006-03-22 00:08:11 -08001711 * Before calling the slab must have been unlinked from the cache. The
1712 * cache-lock is not held/needed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001713 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001714static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001715{
1716 void *addr = slabp->s_mem - slabp->colouroff;
1717
1718 slab_destroy_objs(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001719 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1720 struct slab_rcu *slab_rcu;
1721
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001722 slab_rcu = (struct slab_rcu *)slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001723 slab_rcu->cachep = cachep;
1724 slab_rcu->addr = addr;
1725 call_rcu(&slab_rcu->head, kmem_rcu_free);
1726 } else {
1727 kmem_freepages(cachep, addr);
1728 if (OFF_SLAB(cachep))
1729 kmem_cache_free(cachep->slabp_cache, slabp);
1730 }
1731}
1732
Andrew Mortona737b3e2006-03-22 00:08:11 -08001733/*
1734 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1735 * size of kmem_list3.
1736 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001737static void set_up_list3s(struct kmem_cache *cachep, int index)
Christoph Lametere498be72005-09-09 13:03:32 -07001738{
1739 int node;
1740
1741 for_each_online_node(node) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001742 cachep->nodelists[node] = &initkmem_list3[index + node];
Christoph Lametere498be72005-09-09 13:03:32 -07001743 cachep->nodelists[node]->next_reap = jiffies +
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001744 REAPTIMEOUT_LIST3 +
1745 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -07001746 }
1747}
1748
Linus Torvalds1da177e2005-04-16 15:20:36 -07001749/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08001750 * calculate_slab_order - calculate size (page order) of slabs
1751 * @cachep: pointer to the cache that is being created
1752 * @size: size of objects to be created in this cache.
1753 * @align: required alignment for the objects.
1754 * @flags: slab allocation flags
1755 *
1756 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001757 *
1758 * This could be made much more intelligent. For now, try to avoid using
1759 * high order pages for slabs. When the gfp() functions are more friendly
1760 * towards high-order requests, this should be changed.
1761 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001762static size_t calculate_slab_order(struct kmem_cache *cachep,
Randy Dunlapee13d782006-02-01 03:05:53 -08001763 size_t size, size_t align, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001764{
1765 size_t left_over = 0;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001766 int gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001767
Andrew Mortona737b3e2006-03-22 00:08:11 -08001768 for (gfporder = 0; gfporder <= MAX_GFP_ORDER; gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001769 unsigned int num;
1770 size_t remainder;
1771
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001772 cache_estimate(gfporder, size, align, flags, &remainder, &num);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001773 if (!num)
1774 continue;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001775
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001776 /* More than offslab_limit objects will cause problems */
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001777 if ((flags & CFLGS_OFF_SLAB) && num > offslab_limit)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001778 break;
1779
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001780 /* Found something acceptable - save it away */
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001781 cachep->num = num;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001782 cachep->gfporder = gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001783 left_over = remainder;
1784
1785 /*
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08001786 * A VFS-reclaimable slab tends to have most allocations
1787 * as GFP_NOFS and we really don't want to have to be allocating
1788 * higher-order pages when we are unable to shrink dcache.
1789 */
1790 if (flags & SLAB_RECLAIM_ACCOUNT)
1791 break;
1792
1793 /*
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001794 * Large number of objects is good, but very large slabs are
1795 * currently bad for the gfp()s.
1796 */
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001797 if (gfporder >= slab_break_gfp_order)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001798 break;
1799
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001800 /*
1801 * Acceptable internal fragmentation?
1802 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001803 if (left_over * 8 <= (PAGE_SIZE << gfporder))
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001804 break;
1805 }
1806 return left_over;
1807}
1808
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08001809static void setup_cpu_cache(struct kmem_cache *cachep)
1810{
1811 if (g_cpucache_up == FULL) {
1812 enable_cpucache(cachep);
1813 return;
1814 }
1815 if (g_cpucache_up == NONE) {
1816 /*
1817 * Note: the first kmem_cache_create must create the cache
1818 * that's used by kmalloc(24), otherwise the creation of
1819 * further caches will BUG().
1820 */
1821 cachep->array[smp_processor_id()] = &initarray_generic.cache;
1822
1823 /*
1824 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
1825 * the first cache, then we need to set up all its list3s,
1826 * otherwise the creation of further caches will BUG().
1827 */
1828 set_up_list3s(cachep, SIZE_AC);
1829 if (INDEX_AC == INDEX_L3)
1830 g_cpucache_up = PARTIAL_L3;
1831 else
1832 g_cpucache_up = PARTIAL_AC;
1833 } else {
1834 cachep->array[smp_processor_id()] =
1835 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1836
1837 if (g_cpucache_up == PARTIAL_AC) {
1838 set_up_list3s(cachep, SIZE_L3);
1839 g_cpucache_up = PARTIAL_L3;
1840 } else {
1841 int node;
1842 for_each_online_node(node) {
1843 cachep->nodelists[node] =
1844 kmalloc_node(sizeof(struct kmem_list3),
1845 GFP_KERNEL, node);
1846 BUG_ON(!cachep->nodelists[node]);
1847 kmem_list3_init(cachep->nodelists[node]);
1848 }
1849 }
1850 }
1851 cachep->nodelists[numa_node_id()]->next_reap =
1852 jiffies + REAPTIMEOUT_LIST3 +
1853 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1854
1855 cpu_cache_get(cachep)->avail = 0;
1856 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1857 cpu_cache_get(cachep)->batchcount = 1;
1858 cpu_cache_get(cachep)->touched = 0;
1859 cachep->batchcount = 1;
1860 cachep->limit = BOOT_CPUCACHE_ENTRIES;
1861}
1862
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001863/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07001864 * kmem_cache_create - Create a cache.
1865 * @name: A string which is used in /proc/slabinfo to identify this cache.
1866 * @size: The size of objects to be created in this cache.
1867 * @align: The required alignment for the objects.
1868 * @flags: SLAB flags
1869 * @ctor: A constructor for the objects.
1870 * @dtor: A destructor for the objects.
1871 *
1872 * Returns a ptr to the cache on success, NULL on failure.
1873 * Cannot be called within a int, but can be interrupted.
1874 * The @ctor is run when new pages are allocated by the cache
1875 * and the @dtor is run before the pages are handed back.
1876 *
1877 * @name must be valid until the cache is destroyed. This implies that
Andrew Mortona737b3e2006-03-22 00:08:11 -08001878 * the module calling this has to destroy the cache before getting unloaded.
1879 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001880 * The flags are
1881 *
1882 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1883 * to catch references to uninitialised memory.
1884 *
1885 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1886 * for buffer overruns.
1887 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001888 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1889 * cacheline. This can be beneficial if you're counting cycles as closely
1890 * as davem.
1891 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001892struct kmem_cache *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001893kmem_cache_create (const char *name, size_t size, size_t align,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001894 unsigned long flags,
1895 void (*ctor)(void*, struct kmem_cache *, unsigned long),
Pekka Enberg343e0d72006-02-01 03:05:50 -08001896 void (*dtor)(void*, struct kmem_cache *, unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001897{
1898 size_t left_over, slab_size, ralign;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001899 struct kmem_cache *cachep = NULL;
Andrew Morton4f12bb42005-11-07 00:58:00 -08001900 struct list_head *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001901
1902 /*
1903 * Sanity checks... these are all serious usage bugs.
1904 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001905 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001906 (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08001907 printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__,
1908 name);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001909 BUG();
1910 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001911
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08001912 /*
1913 * Prevent CPUs from coming and going.
1914 * lock_cpu_hotplug() nests outside cache_chain_mutex
1915 */
1916 lock_cpu_hotplug();
1917
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001918 mutex_lock(&cache_chain_mutex);
Andrew Morton4f12bb42005-11-07 00:58:00 -08001919
1920 list_for_each(p, &cache_chain) {
Pekka Enberg343e0d72006-02-01 03:05:50 -08001921 struct kmem_cache *pc = list_entry(p, struct kmem_cache, next);
Andrew Morton4f12bb42005-11-07 00:58:00 -08001922 mm_segment_t old_fs = get_fs();
1923 char tmp;
1924 int res;
1925
1926 /*
1927 * This happens when the module gets unloaded and doesn't
1928 * destroy its slab cache and no-one else reuses the vmalloc
1929 * area of the module. Print a warning.
1930 */
1931 set_fs(KERNEL_DS);
1932 res = __get_user(tmp, pc->name);
1933 set_fs(old_fs);
1934 if (res) {
1935 printk("SLAB: cache with size %d has lost its name\n",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001936 pc->buffer_size);
Andrew Morton4f12bb42005-11-07 00:58:00 -08001937 continue;
1938 }
1939
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001940 if (!strcmp(pc->name, name)) {
Andrew Morton4f12bb42005-11-07 00:58:00 -08001941 printk("kmem_cache_create: duplicate cache %s\n", name);
1942 dump_stack();
1943 goto oops;
1944 }
1945 }
1946
Linus Torvalds1da177e2005-04-16 15:20:36 -07001947#if DEBUG
1948 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
1949 if ((flags & SLAB_DEBUG_INITIAL) && !ctor) {
1950 /* No constructor, but inital state check requested */
1951 printk(KERN_ERR "%s: No con, but init state check "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001952 "requested - %s\n", __FUNCTION__, name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001953 flags &= ~SLAB_DEBUG_INITIAL;
1954 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001955#if FORCED_DEBUG
1956 /*
1957 * Enable redzoning and last user accounting, except for caches with
1958 * large objects, if the increased size would increase the object size
1959 * above the next power of two: caches with object sizes just above a
1960 * power of two have a significant amount of internal fragmentation.
1961 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001962 if (size < 4096 || fls(size - 1) == fls(size-1 + 3 * BYTES_PER_WORD))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001963 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001964 if (!(flags & SLAB_DESTROY_BY_RCU))
1965 flags |= SLAB_POISON;
1966#endif
1967 if (flags & SLAB_DESTROY_BY_RCU)
1968 BUG_ON(flags & SLAB_POISON);
1969#endif
1970 if (flags & SLAB_DESTROY_BY_RCU)
1971 BUG_ON(dtor);
1972
1973 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08001974 * Always checks flags, a caller might be expecting debug support which
1975 * isn't available.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001976 */
1977 if (flags & ~CREATE_MASK)
1978 BUG();
1979
Andrew Mortona737b3e2006-03-22 00:08:11 -08001980 /*
1981 * Check that size is in terms of words. This is needed to avoid
Linus Torvalds1da177e2005-04-16 15:20:36 -07001982 * unaligned accesses for some archs when redzoning is used, and makes
1983 * sure any on-slab bufctl's are also correctly aligned.
1984 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001985 if (size & (BYTES_PER_WORD - 1)) {
1986 size += (BYTES_PER_WORD - 1);
1987 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001988 }
1989
Andrew Mortona737b3e2006-03-22 00:08:11 -08001990 /* calculate the final buffer alignment: */
1991
Linus Torvalds1da177e2005-04-16 15:20:36 -07001992 /* 1) arch recommendation: can be overridden for debug */
1993 if (flags & SLAB_HWCACHE_ALIGN) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08001994 /*
1995 * Default alignment: as specified by the arch code. Except if
1996 * an object is really small, then squeeze multiple objects into
1997 * one cacheline.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001998 */
1999 ralign = cache_line_size();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002000 while (size <= ralign / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002001 ralign /= 2;
2002 } else {
2003 ralign = BYTES_PER_WORD;
2004 }
2005 /* 2) arch mandated alignment: disables debug if necessary */
2006 if (ralign < ARCH_SLAB_MINALIGN) {
2007 ralign = ARCH_SLAB_MINALIGN;
2008 if (ralign > BYTES_PER_WORD)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002009 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002010 }
2011 /* 3) caller mandated alignment: disables debug if necessary */
2012 if (ralign < align) {
2013 ralign = align;
2014 if (ralign > BYTES_PER_WORD)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002015 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002016 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08002017 /*
2018 * 4) Store it. Note that the debug code below can reduce
Linus Torvalds1da177e2005-04-16 15:20:36 -07002019 * the alignment to BYTES_PER_WORD.
2020 */
2021 align = ralign;
2022
2023 /* Get cache's description obj. */
Pekka Enbergc5e3b832006-03-25 03:06:43 -08002024 cachep = kmem_cache_zalloc(&cache_cache, SLAB_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002025 if (!cachep)
Andrew Morton4f12bb42005-11-07 00:58:00 -08002026 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002027
2028#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002029 cachep->obj_size = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002030
2031 if (flags & SLAB_RED_ZONE) {
2032 /* redzoning only works with word aligned caches */
2033 align = BYTES_PER_WORD;
2034
2035 /* add space for red zone words */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002036 cachep->obj_offset += BYTES_PER_WORD;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002037 size += 2 * BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002038 }
2039 if (flags & SLAB_STORE_USER) {
2040 /* user store requires word alignment and
2041 * one word storage behind the end of the real
2042 * object.
2043 */
2044 align = BYTES_PER_WORD;
2045 size += BYTES_PER_WORD;
2046 }
2047#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002048 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002049 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2050 cachep->obj_offset += PAGE_SIZE - size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002051 size = PAGE_SIZE;
2052 }
2053#endif
2054#endif
2055
2056 /* Determine if the slab management is 'on' or 'off' slab. */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002057 if (size >= (PAGE_SIZE >> 3))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002058 /*
2059 * Size is large, assume best to place the slab management obj
2060 * off-slab (should allow better packing of objs).
2061 */
2062 flags |= CFLGS_OFF_SLAB;
2063
2064 size = ALIGN(size, align);
2065
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002066 left_over = calculate_slab_order(cachep, size, align, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002067
2068 if (!cachep->num) {
2069 printk("kmem_cache_create: couldn't create cache %s.\n", name);
2070 kmem_cache_free(&cache_cache, cachep);
2071 cachep = NULL;
Andrew Morton4f12bb42005-11-07 00:58:00 -08002072 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002073 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002074 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2075 + sizeof(struct slab), align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002076
2077 /*
2078 * If the slab has been placed off-slab, and we have enough space then
2079 * move it on-slab. This is at the expense of any extra colouring.
2080 */
2081 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2082 flags &= ~CFLGS_OFF_SLAB;
2083 left_over -= slab_size;
2084 }
2085
2086 if (flags & CFLGS_OFF_SLAB) {
2087 /* really off slab. No need for manual alignment */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002088 slab_size =
2089 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002090 }
2091
2092 cachep->colour_off = cache_line_size();
2093 /* Offset must be a multiple of the alignment. */
2094 if (cachep->colour_off < align)
2095 cachep->colour_off = align;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002096 cachep->colour = left_over / cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002097 cachep->slab_size = slab_size;
2098 cachep->flags = flags;
2099 cachep->gfpflags = 0;
2100 if (flags & SLAB_CACHE_DMA)
2101 cachep->gfpflags |= GFP_DMA;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002102 cachep->buffer_size = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002103
2104 if (flags & CFLGS_OFF_SLAB)
Victor Fuscob2d55072005-09-10 00:26:36 -07002105 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002106 cachep->ctor = ctor;
2107 cachep->dtor = dtor;
2108 cachep->name = name;
2109
Linus Torvalds1da177e2005-04-16 15:20:36 -07002110
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002111 setup_cpu_cache(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002112
Linus Torvalds1da177e2005-04-16 15:20:36 -07002113 /* cache setup completed, link it into the list */
2114 list_add(&cachep->next, &cache_chain);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002115oops:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002116 if (!cachep && (flags & SLAB_PANIC))
2117 panic("kmem_cache_create(): failed to create slab `%s'\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002118 name);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002119 mutex_unlock(&cache_chain_mutex);
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002120 unlock_cpu_hotplug();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002121 return cachep;
2122}
2123EXPORT_SYMBOL(kmem_cache_create);
2124
2125#if DEBUG
2126static void check_irq_off(void)
2127{
2128 BUG_ON(!irqs_disabled());
2129}
2130
2131static void check_irq_on(void)
2132{
2133 BUG_ON(irqs_disabled());
2134}
2135
Pekka Enberg343e0d72006-02-01 03:05:50 -08002136static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002137{
2138#ifdef CONFIG_SMP
2139 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002140 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002141#endif
2142}
Christoph Lametere498be72005-09-09 13:03:32 -07002143
Pekka Enberg343e0d72006-02-01 03:05:50 -08002144static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002145{
2146#ifdef CONFIG_SMP
2147 check_irq_off();
2148 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2149#endif
2150}
2151
Linus Torvalds1da177e2005-04-16 15:20:36 -07002152#else
2153#define check_irq_off() do { } while(0)
2154#define check_irq_on() do { } while(0)
2155#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002156#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002157#endif
2158
Christoph Lameteraab22072006-03-22 00:09:06 -08002159static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2160 struct array_cache *ac,
2161 int force, int node);
2162
Linus Torvalds1da177e2005-04-16 15:20:36 -07002163static void do_drain(void *arg)
2164{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002165 struct kmem_cache *cachep = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002166 struct array_cache *ac;
Christoph Lameterff694162005-09-22 21:44:02 -07002167 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002168
2169 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002170 ac = cpu_cache_get(cachep);
Christoph Lameterff694162005-09-22 21:44:02 -07002171 spin_lock(&cachep->nodelists[node]->list_lock);
2172 free_block(cachep, ac->entry, ac->avail, node);
2173 spin_unlock(&cachep->nodelists[node]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002174 ac->avail = 0;
2175}
2176
Pekka Enberg343e0d72006-02-01 03:05:50 -08002177static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002178{
Christoph Lametere498be72005-09-09 13:03:32 -07002179 struct kmem_list3 *l3;
2180 int node;
2181
Andrew Mortona07fa392006-03-22 00:08:17 -08002182 on_each_cpu(do_drain, cachep, 1, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002183 check_irq_on();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002184 for_each_online_node(node) {
Christoph Lametere498be72005-09-09 13:03:32 -07002185 l3 = cachep->nodelists[node];
2186 if (l3) {
Christoph Lameteraab22072006-03-22 00:09:06 -08002187 drain_array(cachep, l3, l3->shared, 1, node);
Christoph Lametere498be72005-09-09 13:03:32 -07002188 if (l3->alien)
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08002189 drain_alien_cache(cachep, l3->alien);
Christoph Lametere498be72005-09-09 13:03:32 -07002190 }
2191 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002192}
2193
Pekka Enberg343e0d72006-02-01 03:05:50 -08002194static int __node_shrink(struct kmem_cache *cachep, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002195{
2196 struct slab *slabp;
Christoph Lametere498be72005-09-09 13:03:32 -07002197 struct kmem_list3 *l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002198 int ret;
2199
Christoph Lametere498be72005-09-09 13:03:32 -07002200 for (;;) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002201 struct list_head *p;
2202
Christoph Lametere498be72005-09-09 13:03:32 -07002203 p = l3->slabs_free.prev;
2204 if (p == &l3->slabs_free)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002205 break;
2206
Christoph Lametere498be72005-09-09 13:03:32 -07002207 slabp = list_entry(l3->slabs_free.prev, struct slab, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002208#if DEBUG
2209 if (slabp->inuse)
2210 BUG();
2211#endif
2212 list_del(&slabp->list);
2213
Christoph Lametere498be72005-09-09 13:03:32 -07002214 l3->free_objects -= cachep->num;
2215 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002216 slab_destroy(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07002217 spin_lock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002218 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002219 ret = !list_empty(&l3->slabs_full) || !list_empty(&l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002220 return ret;
2221}
2222
Pekka Enberg343e0d72006-02-01 03:05:50 -08002223static int __cache_shrink(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07002224{
2225 int ret = 0, i = 0;
2226 struct kmem_list3 *l3;
2227
2228 drain_cpu_caches(cachep);
2229
2230 check_irq_on();
2231 for_each_online_node(i) {
2232 l3 = cachep->nodelists[i];
2233 if (l3) {
2234 spin_lock_irq(&l3->list_lock);
2235 ret += __node_shrink(cachep, i);
2236 spin_unlock_irq(&l3->list_lock);
2237 }
2238 }
2239 return (ret ? 1 : 0);
2240}
2241
Linus Torvalds1da177e2005-04-16 15:20:36 -07002242/**
2243 * kmem_cache_shrink - Shrink a cache.
2244 * @cachep: The cache to shrink.
2245 *
2246 * Releases as many slabs as possible for a cache.
2247 * To help debugging, a zero exit status indicates all slabs were released.
2248 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002249int kmem_cache_shrink(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002250{
2251 if (!cachep || in_interrupt())
2252 BUG();
2253
2254 return __cache_shrink(cachep);
2255}
2256EXPORT_SYMBOL(kmem_cache_shrink);
2257
2258/**
2259 * kmem_cache_destroy - delete a cache
2260 * @cachep: the cache to destroy
2261 *
Pekka Enberg343e0d72006-02-01 03:05:50 -08002262 * Remove a struct kmem_cache object from the slab cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002263 * Returns 0 on success.
2264 *
2265 * It is expected this function will be called by a module when it is
2266 * unloaded. This will remove the cache completely, and avoid a duplicate
2267 * cache being allocated each time a module is loaded and unloaded, if the
2268 * module doesn't have persistent in-kernel storage across loads and unloads.
2269 *
2270 * The cache must be empty before calling this function.
2271 *
2272 * The caller must guarantee that noone will allocate memory from the cache
2273 * during the kmem_cache_destroy().
2274 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002275int kmem_cache_destroy(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002276{
2277 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07002278 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002279
2280 if (!cachep || in_interrupt())
2281 BUG();
2282
2283 /* Don't let CPUs to come and go */
2284 lock_cpu_hotplug();
2285
2286 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002287 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002288 /*
2289 * the chain is never empty, cache_cache is never destroyed
2290 */
2291 list_del(&cachep->next);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002292 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002293
2294 if (__cache_shrink(cachep)) {
2295 slab_error(cachep, "Can't free all objects");
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002296 mutex_lock(&cache_chain_mutex);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002297 list_add(&cachep->next, &cache_chain);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002298 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002299 unlock_cpu_hotplug();
2300 return 1;
2301 }
2302
2303 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
Paul E. McKenneyfbd568a3e2005-05-01 08:59:04 -07002304 synchronize_rcu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002305
Christoph Lametere498be72005-09-09 13:03:32 -07002306 for_each_online_cpu(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002307 kfree(cachep->array[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002308
2309 /* NUMA: free the list3 structures */
Christoph Lametere498be72005-09-09 13:03:32 -07002310 for_each_online_node(i) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002311 l3 = cachep->nodelists[i];
2312 if (l3) {
Christoph Lametere498be72005-09-09 13:03:32 -07002313 kfree(l3->shared);
2314 free_alien_cache(l3->alien);
2315 kfree(l3);
2316 }
2317 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002318 kmem_cache_free(&cache_cache, cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002319 unlock_cpu_hotplug();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002320 return 0;
2321}
2322EXPORT_SYMBOL(kmem_cache_destroy);
2323
2324/* Get the memory for a slab management obj. */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002325static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002326 int colour_off, gfp_t local_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002327{
2328 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002329
Linus Torvalds1da177e2005-04-16 15:20:36 -07002330 if (OFF_SLAB(cachep)) {
2331 /* Slab management obj is off-slab. */
2332 slabp = kmem_cache_alloc(cachep->slabp_cache, local_flags);
2333 if (!slabp)
2334 return NULL;
2335 } else {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002336 slabp = objp + colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002337 colour_off += cachep->slab_size;
2338 }
2339 slabp->inuse = 0;
2340 slabp->colouroff = colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002341 slabp->s_mem = objp + colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002342 return slabp;
2343}
2344
2345static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2346{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002347 return (kmem_bufctl_t *) (slabp + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002348}
2349
Pekka Enberg343e0d72006-02-01 03:05:50 -08002350static void cache_init_objs(struct kmem_cache *cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002351 struct slab *slabp, unsigned long ctor_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002352{
2353 int i;
2354
2355 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002356 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002357#if DEBUG
2358 /* need to poison the objs? */
2359 if (cachep->flags & SLAB_POISON)
2360 poison_obj(cachep, objp, POISON_FREE);
2361 if (cachep->flags & SLAB_STORE_USER)
2362 *dbg_userword(cachep, objp) = NULL;
2363
2364 if (cachep->flags & SLAB_RED_ZONE) {
2365 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2366 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2367 }
2368 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002369 * Constructors are not allowed to allocate memory from the same
2370 * cache which they are a constructor for. Otherwise, deadlock.
2371 * They must also be threaded.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002372 */
2373 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002374 cachep->ctor(objp + obj_offset(cachep), cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002375 ctor_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002376
2377 if (cachep->flags & SLAB_RED_ZONE) {
2378 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2379 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002380 " end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002381 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2382 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002383 " start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002384 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08002385 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2386 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002387 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002388 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002389#else
2390 if (cachep->ctor)
2391 cachep->ctor(objp, cachep, ctor_flags);
2392#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002393 slab_bufctl(slabp)[i] = i + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002394 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002395 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002396 slabp->free = 0;
2397}
2398
Pekka Enberg343e0d72006-02-01 03:05:50 -08002399static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002400{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002401 if (flags & SLAB_DMA)
2402 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2403 else
2404 BUG_ON(cachep->gfpflags & GFP_DMA);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002405}
2406
Andrew Mortona737b3e2006-03-22 00:08:11 -08002407static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2408 int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002409{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002410 void *objp = index_to_obj(cachep, slabp, slabp->free);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002411 kmem_bufctl_t next;
2412
2413 slabp->inuse++;
2414 next = slab_bufctl(slabp)[slabp->free];
2415#if DEBUG
2416 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2417 WARN_ON(slabp->nodeid != nodeid);
2418#endif
2419 slabp->free = next;
2420
2421 return objp;
2422}
2423
Andrew Mortona737b3e2006-03-22 00:08:11 -08002424static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2425 void *objp, int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002426{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002427 unsigned int objnr = obj_to_index(cachep, slabp, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002428
2429#if DEBUG
2430 /* Verify that the slab belongs to the intended node */
2431 WARN_ON(slabp->nodeid != nodeid);
2432
Al Viro871751e2006-03-25 03:06:39 -08002433 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
Matthew Dobson78d382d2006-02-01 03:05:47 -08002434 printk(KERN_ERR "slab: double free detected in cache "
Andrew Mortona737b3e2006-03-22 00:08:11 -08002435 "'%s', objp %p\n", cachep->name, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002436 BUG();
2437 }
2438#endif
2439 slab_bufctl(slabp)[objnr] = slabp->free;
2440 slabp->free = objnr;
2441 slabp->inuse--;
2442}
2443
Andrew Mortona737b3e2006-03-22 00:08:11 -08002444static void set_slab_attr(struct kmem_cache *cachep, struct slab *slabp,
2445 void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002446{
2447 int i;
2448 struct page *page;
2449
2450 /* Nasty!!!!!! I hope this is OK. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002451 page = virt_to_page(objp);
Nick Piggin84097512006-03-22 00:08:34 -08002452
2453 i = 1;
2454 if (likely(!PageCompound(page)))
2455 i <<= cachep->gfporder;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002456 do {
Pekka Enberg065d41c2005-11-13 16:06:46 -08002457 page_set_cache(page, cachep);
2458 page_set_slab(page, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002459 page++;
2460 } while (--i);
2461}
2462
2463/*
2464 * Grow (by 1) the number of slabs within a cache. This is called by
2465 * kmem_cache_alloc() when there are no active objs left in a cache.
2466 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002467static int cache_grow(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002468{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002469 struct slab *slabp;
2470 void *objp;
2471 size_t offset;
2472 gfp_t local_flags;
2473 unsigned long ctor_flags;
Christoph Lametere498be72005-09-09 13:03:32 -07002474 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002475
Andrew Mortona737b3e2006-03-22 00:08:11 -08002476 /*
2477 * Be lazy and only check for valid flags here, keeping it out of the
2478 * critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002479 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002480 if (flags & ~(SLAB_DMA | SLAB_LEVEL_MASK | SLAB_NO_GROW))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002481 BUG();
2482 if (flags & SLAB_NO_GROW)
2483 return 0;
2484
2485 ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2486 local_flags = (flags & SLAB_LEVEL_MASK);
2487 if (!(local_flags & __GFP_WAIT))
2488 /*
2489 * Not allowed to sleep. Need to tell a constructor about
2490 * this - it might need to know...
2491 */
2492 ctor_flags |= SLAB_CTOR_ATOMIC;
2493
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002494 /* Take the l3 list lock to change the colour_next on this node */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002495 check_irq_off();
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002496 l3 = cachep->nodelists[nodeid];
2497 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002498
2499 /* Get colour for the slab, and cal the next value. */
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002500 offset = l3->colour_next;
2501 l3->colour_next++;
2502 if (l3->colour_next >= cachep->colour)
2503 l3->colour_next = 0;
2504 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002505
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002506 offset *= cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002507
2508 if (local_flags & __GFP_WAIT)
2509 local_irq_enable();
2510
2511 /*
2512 * The test for missing atomic flag is performed here, rather than
2513 * the more obvious place, simply to reduce the critical path length
2514 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2515 * will eventually be caught here (where it matters).
2516 */
2517 kmem_flagcheck(cachep, flags);
2518
Andrew Mortona737b3e2006-03-22 00:08:11 -08002519 /*
2520 * Get mem for the objs. Attempt to allocate a physical page from
2521 * 'nodeid'.
Christoph Lametere498be72005-09-09 13:03:32 -07002522 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002523 objp = kmem_getpages(cachep, flags, nodeid);
2524 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002525 goto failed;
2526
2527 /* Get slab management. */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002528 slabp = alloc_slabmgmt(cachep, objp, offset, local_flags);
2529 if (!slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002530 goto opps1;
2531
Christoph Lametere498be72005-09-09 13:03:32 -07002532 slabp->nodeid = nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002533 set_slab_attr(cachep, slabp, objp);
2534
2535 cache_init_objs(cachep, slabp, ctor_flags);
2536
2537 if (local_flags & __GFP_WAIT)
2538 local_irq_disable();
2539 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002540 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002541
2542 /* Make slab active. */
Christoph Lametere498be72005-09-09 13:03:32 -07002543 list_add_tail(&slabp->list, &(l3->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002544 STATS_INC_GROWN(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002545 l3->free_objects += cachep->num;
2546 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002547 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002548opps1:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002549 kmem_freepages(cachep, objp);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002550failed:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002551 if (local_flags & __GFP_WAIT)
2552 local_irq_disable();
2553 return 0;
2554}
2555
2556#if DEBUG
2557
2558/*
2559 * Perform extra freeing checks:
2560 * - detect bad pointers.
2561 * - POISON/RED_ZONE checking
2562 * - destructor calls, for caches with POISON+dtor
2563 */
2564static void kfree_debugcheck(const void *objp)
2565{
2566 struct page *page;
2567
2568 if (!virt_addr_valid(objp)) {
2569 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002570 (unsigned long)objp);
2571 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002572 }
2573 page = virt_to_page(objp);
2574 if (!PageSlab(page)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002575 printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n",
2576 (unsigned long)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002577 BUG();
2578 }
2579}
2580
Pekka Enberg343e0d72006-02-01 03:05:50 -08002581static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002582 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002583{
2584 struct page *page;
2585 unsigned int objnr;
2586 struct slab *slabp;
2587
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002588 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002589 kfree_debugcheck(objp);
2590 page = virt_to_page(objp);
2591
Pekka Enberg065d41c2005-11-13 16:06:46 -08002592 if (page_get_cache(page) != cachep) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002593 printk(KERN_ERR "mismatch in kmem_cache_free: expected "
2594 "cache %p, got %p\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002595 page_get_cache(page), cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002596 printk(KERN_ERR "%p is %s.\n", cachep, cachep->name);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002597 printk(KERN_ERR "%p is %s.\n", page_get_cache(page),
2598 page_get_cache(page)->name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002599 WARN_ON(1);
2600 }
Pekka Enberg065d41c2005-11-13 16:06:46 -08002601 slabp = page_get_slab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002602
2603 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002604 if (*dbg_redzone1(cachep, objp) != RED_ACTIVE ||
2605 *dbg_redzone2(cachep, objp) != RED_ACTIVE) {
2606 slab_error(cachep, "double free, or memory outside"
2607 " object was overwritten");
2608 printk(KERN_ERR "%p: redzone 1:0x%lx, "
2609 "redzone 2:0x%lx.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002610 objp, *dbg_redzone1(cachep, objp),
2611 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002612 }
2613 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2614 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2615 }
2616 if (cachep->flags & SLAB_STORE_USER)
2617 *dbg_userword(cachep, objp) = caller;
2618
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002619 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002620
2621 BUG_ON(objnr >= cachep->num);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002622 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002623
2624 if (cachep->flags & SLAB_DEBUG_INITIAL) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002625 /*
2626 * Need to call the slab's constructor so the caller can
2627 * perform a verify of its state (debugging). Called without
2628 * the cache-lock held.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002629 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002630 cachep->ctor(objp + obj_offset(cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002631 cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002632 }
2633 if (cachep->flags & SLAB_POISON && cachep->dtor) {
2634 /* we want to cache poison the object,
2635 * call the destruction callback
2636 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002637 cachep->dtor(objp + obj_offset(cachep), cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002638 }
Al Viro871751e2006-03-25 03:06:39 -08002639#ifdef CONFIG_DEBUG_SLAB_LEAK
2640 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2641#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002642 if (cachep->flags & SLAB_POISON) {
2643#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08002644 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002645 store_stackinfo(cachep, objp, (unsigned long)caller);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002646 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002647 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002648 } else {
2649 poison_obj(cachep, objp, POISON_FREE);
2650 }
2651#else
2652 poison_obj(cachep, objp, POISON_FREE);
2653#endif
2654 }
2655 return objp;
2656}
2657
Pekka Enberg343e0d72006-02-01 03:05:50 -08002658static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002659{
2660 kmem_bufctl_t i;
2661 int entries = 0;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002662
Linus Torvalds1da177e2005-04-16 15:20:36 -07002663 /* Check slab's freelist to see if this obj is there. */
2664 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2665 entries++;
2666 if (entries > cachep->num || i >= cachep->num)
2667 goto bad;
2668 }
2669 if (entries != cachep->num - slabp->inuse) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002670bad:
2671 printk(KERN_ERR "slab: Internal list corruption detected in "
2672 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2673 cachep->name, cachep->num, slabp, slabp->inuse);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002674 for (i = 0;
Linus Torvalds264132b2006-03-06 12:10:07 -08002675 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002676 i++) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002677 if (i % 16 == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002678 printk("\n%03x:", i);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002679 printk(" %02x", ((unsigned char *)slabp)[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002680 }
2681 printk("\n");
2682 BUG();
2683 }
2684}
2685#else
2686#define kfree_debugcheck(x) do { } while(0)
2687#define cache_free_debugcheck(x,objp,z) (objp)
2688#define check_slabp(x,y) do { } while(0)
2689#endif
2690
Pekka Enberg343e0d72006-02-01 03:05:50 -08002691static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002692{
2693 int batchcount;
2694 struct kmem_list3 *l3;
2695 struct array_cache *ac;
2696
2697 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002698 ac = cpu_cache_get(cachep);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002699retry:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002700 batchcount = ac->batchcount;
2701 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002702 /*
2703 * If there was little recent activity on this cache, then
2704 * perform only a partial refill. Otherwise we could generate
2705 * refill bouncing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002706 */
2707 batchcount = BATCHREFILL_LIMIT;
2708 }
Christoph Lametere498be72005-09-09 13:03:32 -07002709 l3 = cachep->nodelists[numa_node_id()];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002710
Christoph Lametere498be72005-09-09 13:03:32 -07002711 BUG_ON(ac->avail > 0 || !l3);
2712 spin_lock(&l3->list_lock);
2713
Christoph Lameter3ded1752006-03-25 03:06:44 -08002714 /* See if we can refill from the shared array */
2715 if (l3->shared && transfer_objects(ac, l3->shared, batchcount))
2716 goto alloc_done;
2717
Linus Torvalds1da177e2005-04-16 15:20:36 -07002718 while (batchcount > 0) {
2719 struct list_head *entry;
2720 struct slab *slabp;
2721 /* Get slab alloc is to come from. */
2722 entry = l3->slabs_partial.next;
2723 if (entry == &l3->slabs_partial) {
2724 l3->free_touched = 1;
2725 entry = l3->slabs_free.next;
2726 if (entry == &l3->slabs_free)
2727 goto must_grow;
2728 }
2729
2730 slabp = list_entry(entry, struct slab, list);
2731 check_slabp(cachep, slabp);
2732 check_spinlock_acquired(cachep);
2733 while (slabp->inuse < cachep->num && batchcount--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002734 STATS_INC_ALLOCED(cachep);
2735 STATS_INC_ACTIVE(cachep);
2736 STATS_SET_HIGH(cachep);
2737
Matthew Dobson78d382d2006-02-01 03:05:47 -08002738 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
2739 numa_node_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002740 }
2741 check_slabp(cachep, slabp);
2742
2743 /* move slabp to correct slabp list: */
2744 list_del(&slabp->list);
2745 if (slabp->free == BUFCTL_END)
2746 list_add(&slabp->list, &l3->slabs_full);
2747 else
2748 list_add(&slabp->list, &l3->slabs_partial);
2749 }
2750
Andrew Mortona737b3e2006-03-22 00:08:11 -08002751must_grow:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002752 l3->free_objects -= ac->avail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002753alloc_done:
Christoph Lametere498be72005-09-09 13:03:32 -07002754 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002755
2756 if (unlikely(!ac->avail)) {
2757 int x;
Christoph Lametere498be72005-09-09 13:03:32 -07002758 x = cache_grow(cachep, flags, numa_node_id());
2759
Andrew Mortona737b3e2006-03-22 00:08:11 -08002760 /* cache_grow can reenable interrupts, then ac could change. */
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002761 ac = cpu_cache_get(cachep);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002762 if (!x && ac->avail == 0) /* no objects in sight? abort */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002763 return NULL;
2764
Andrew Mortona737b3e2006-03-22 00:08:11 -08002765 if (!ac->avail) /* objects refilled by interrupt? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002766 goto retry;
2767 }
2768 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07002769 return ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002770}
2771
Andrew Mortona737b3e2006-03-22 00:08:11 -08002772static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
2773 gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002774{
2775 might_sleep_if(flags & __GFP_WAIT);
2776#if DEBUG
2777 kmem_flagcheck(cachep, flags);
2778#endif
2779}
2780
2781#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08002782static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
2783 gfp_t flags, void *objp, void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002784{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002785 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002786 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002787 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002788#ifdef CONFIG_DEBUG_PAGEALLOC
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002789 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002790 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002791 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002792 else
2793 check_poison_obj(cachep, objp);
2794#else
2795 check_poison_obj(cachep, objp);
2796#endif
2797 poison_obj(cachep, objp, POISON_INUSE);
2798 }
2799 if (cachep->flags & SLAB_STORE_USER)
2800 *dbg_userword(cachep, objp) = caller;
2801
2802 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002803 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
2804 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2805 slab_error(cachep, "double free, or memory outside"
2806 " object was overwritten");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002807 printk(KERN_ERR
Andrew Mortona737b3e2006-03-22 00:08:11 -08002808 "%p: redzone 1:0x%lx, redzone 2:0x%lx\n",
2809 objp, *dbg_redzone1(cachep, objp),
2810 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002811 }
2812 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2813 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2814 }
Al Viro871751e2006-03-25 03:06:39 -08002815#ifdef CONFIG_DEBUG_SLAB_LEAK
2816 {
2817 struct slab *slabp;
2818 unsigned objnr;
2819
2820 slabp = page_get_slab(virt_to_page(objp));
2821 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
2822 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
2823 }
2824#endif
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002825 objp += obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002826 if (cachep->ctor && cachep->flags & SLAB_POISON) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002827 unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002828
2829 if (!(flags & __GFP_WAIT))
2830 ctor_flags |= SLAB_CTOR_ATOMIC;
2831
2832 cachep->ctor(objp, cachep, ctor_flags);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002833 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002834 return objp;
2835}
2836#else
2837#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2838#endif
2839
Pekka Enberg343e0d72006-02-01 03:05:50 -08002840static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002841{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002842 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002843 struct array_cache *ac;
2844
Christoph Lameterdc85da12006-01-18 17:42:36 -08002845#ifdef CONFIG_NUMA
Paul Jacksonb2455392006-03-24 03:16:12 -08002846 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
Paul Jacksonc61afb12006-03-24 03:16:08 -08002847 objp = alternate_node_alloc(cachep, flags);
2848 if (objp != NULL)
2849 return objp;
Paul Jackson101a5002006-03-24 03:16:07 -08002850 }
Christoph Lameterdc85da12006-01-18 17:42:36 -08002851#endif
2852
Alok N Kataria5c382302005-09-27 21:45:46 -07002853 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002854 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002855 if (likely(ac->avail)) {
2856 STATS_INC_ALLOCHIT(cachep);
2857 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07002858 objp = ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002859 } else {
2860 STATS_INC_ALLOCMISS(cachep);
2861 objp = cache_alloc_refill(cachep, flags);
2862 }
Alok N Kataria5c382302005-09-27 21:45:46 -07002863 return objp;
2864}
2865
Andrew Mortona737b3e2006-03-22 00:08:11 -08002866static __always_inline void *__cache_alloc(struct kmem_cache *cachep,
2867 gfp_t flags, void *caller)
Alok N Kataria5c382302005-09-27 21:45:46 -07002868{
2869 unsigned long save_flags;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002870 void *objp;
Alok N Kataria5c382302005-09-27 21:45:46 -07002871
2872 cache_alloc_debugcheck_before(cachep, flags);
2873
2874 local_irq_save(save_flags);
2875 objp = ____cache_alloc(cachep, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002876 local_irq_restore(save_flags);
Eric Dumazet34342e82005-09-03 15:55:06 -07002877 objp = cache_alloc_debugcheck_after(cachep, flags, objp,
Pekka Enberg7fd6b142006-02-01 03:05:52 -08002878 caller);
Eric Dumazet34342e82005-09-03 15:55:06 -07002879 prefetchw(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002880 return objp;
2881}
2882
Christoph Lametere498be72005-09-09 13:03:32 -07002883#ifdef CONFIG_NUMA
2884/*
Paul Jacksonb2455392006-03-24 03:16:12 -08002885 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
Paul Jacksonc61afb12006-03-24 03:16:08 -08002886 *
2887 * If we are in_interrupt, then process context, including cpusets and
2888 * mempolicy, may not apply and should not be used for allocation policy.
2889 */
2890static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
2891{
2892 int nid_alloc, nid_here;
2893
2894 if (in_interrupt())
2895 return NULL;
2896 nid_alloc = nid_here = numa_node_id();
2897 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
2898 nid_alloc = cpuset_mem_spread_node();
2899 else if (current->mempolicy)
2900 nid_alloc = slab_node(current->mempolicy);
2901 if (nid_alloc != nid_here)
2902 return __cache_alloc_node(cachep, flags, nid_alloc);
2903 return NULL;
2904}
2905
2906/*
Christoph Lametere498be72005-09-09 13:03:32 -07002907 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002908 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002909static void *__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
2910 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07002911{
2912 struct list_head *entry;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002913 struct slab *slabp;
2914 struct kmem_list3 *l3;
2915 void *obj;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002916 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002917
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002918 l3 = cachep->nodelists[nodeid];
2919 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07002920
Andrew Mortona737b3e2006-03-22 00:08:11 -08002921retry:
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08002922 check_irq_off();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002923 spin_lock(&l3->list_lock);
2924 entry = l3->slabs_partial.next;
2925 if (entry == &l3->slabs_partial) {
2926 l3->free_touched = 1;
2927 entry = l3->slabs_free.next;
2928 if (entry == &l3->slabs_free)
2929 goto must_grow;
2930 }
Christoph Lametere498be72005-09-09 13:03:32 -07002931
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002932 slabp = list_entry(entry, struct slab, list);
2933 check_spinlock_acquired_node(cachep, nodeid);
2934 check_slabp(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07002935
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002936 STATS_INC_NODEALLOCS(cachep);
2937 STATS_INC_ACTIVE(cachep);
2938 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002939
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002940 BUG_ON(slabp->inuse == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07002941
Matthew Dobson78d382d2006-02-01 03:05:47 -08002942 obj = slab_get_obj(cachep, slabp, nodeid);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002943 check_slabp(cachep, slabp);
2944 l3->free_objects--;
2945 /* move slabp to correct slabp list: */
2946 list_del(&slabp->list);
Christoph Lametere498be72005-09-09 13:03:32 -07002947
Andrew Mortona737b3e2006-03-22 00:08:11 -08002948 if (slabp->free == BUFCTL_END)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002949 list_add(&slabp->list, &l3->slabs_full);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002950 else
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002951 list_add(&slabp->list, &l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07002952
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002953 spin_unlock(&l3->list_lock);
2954 goto done;
Christoph Lametere498be72005-09-09 13:03:32 -07002955
Andrew Mortona737b3e2006-03-22 00:08:11 -08002956must_grow:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002957 spin_unlock(&l3->list_lock);
2958 x = cache_grow(cachep, flags, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07002959
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002960 if (!x)
2961 return NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07002962
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002963 goto retry;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002964done:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002965 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07002966}
2967#endif
2968
2969/*
2970 * Caller needs to acquire correct kmem_list's list_lock
2971 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002972static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002973 int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002974{
2975 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07002976 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002977
2978 for (i = 0; i < nr_objects; i++) {
2979 void *objp = objpp[i];
2980 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002981
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08002982 slabp = virt_to_slab(objp);
Christoph Lameterff694162005-09-22 21:44:02 -07002983 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002984 list_del(&slabp->list);
Christoph Lameterff694162005-09-22 21:44:02 -07002985 check_spinlock_acquired_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002986 check_slabp(cachep, slabp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002987 slab_put_obj(cachep, slabp, objp, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002988 STATS_DEC_ACTIVE(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002989 l3->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002990 check_slabp(cachep, slabp);
2991
2992 /* fixup slab chains */
2993 if (slabp->inuse == 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07002994 if (l3->free_objects > l3->free_limit) {
2995 l3->free_objects -= cachep->num;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002996 slab_destroy(cachep, slabp);
2997 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07002998 list_add(&slabp->list, &l3->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002999 }
3000 } else {
3001 /* Unconditionally move a slab to the end of the
3002 * partial list on free - maximum time for the
3003 * other objects to be freed, too.
3004 */
Christoph Lametere498be72005-09-09 13:03:32 -07003005 list_add_tail(&slabp->list, &l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003006 }
3007 }
3008}
3009
Pekka Enberg343e0d72006-02-01 03:05:50 -08003010static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003011{
3012 int batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07003013 struct kmem_list3 *l3;
Christoph Lameterff694162005-09-22 21:44:02 -07003014 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003015
3016 batchcount = ac->batchcount;
3017#if DEBUG
3018 BUG_ON(!batchcount || batchcount > ac->avail);
3019#endif
3020 check_irq_off();
Christoph Lameterff694162005-09-22 21:44:02 -07003021 l3 = cachep->nodelists[node];
Christoph Lametere498be72005-09-09 13:03:32 -07003022 spin_lock(&l3->list_lock);
3023 if (l3->shared) {
3024 struct array_cache *shared_array = l3->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003025 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003026 if (max) {
3027 if (batchcount > max)
3028 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07003029 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003030 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003031 shared_array->avail += batchcount;
3032 goto free_done;
3033 }
3034 }
3035
Christoph Lameterff694162005-09-22 21:44:02 -07003036 free_block(cachep, ac->entry, batchcount, node);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003037free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003038#if STATS
3039 {
3040 int i = 0;
3041 struct list_head *p;
3042
Christoph Lametere498be72005-09-09 13:03:32 -07003043 p = l3->slabs_free.next;
3044 while (p != &(l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003045 struct slab *slabp;
3046
3047 slabp = list_entry(p, struct slab, list);
3048 BUG_ON(slabp->inuse);
3049
3050 i++;
3051 p = p->next;
3052 }
3053 STATS_SET_FREEABLE(cachep, i);
3054 }
3055#endif
Christoph Lametere498be72005-09-09 13:03:32 -07003056 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003057 ac->avail -= batchcount;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003058 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003059}
3060
3061/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08003062 * Release an obj back to its cache. If the obj has a constructed state, it must
3063 * be in this state _before_ it is released. Called with disabled ints.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003064 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003065static inline void __cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003066{
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003067 struct array_cache *ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003068
3069 check_irq_off();
3070 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3071
Christoph Lametere498be72005-09-09 13:03:32 -07003072 /* Make sure we are not freeing a object from another
3073 * node to the array cache on this cpu.
3074 */
3075#ifdef CONFIG_NUMA
3076 {
3077 struct slab *slabp;
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003078 slabp = virt_to_slab(objp);
Christoph Lametere498be72005-09-09 13:03:32 -07003079 if (unlikely(slabp->nodeid != numa_node_id())) {
3080 struct array_cache *alien = NULL;
3081 int nodeid = slabp->nodeid;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003082 struct kmem_list3 *l3;
Christoph Lametere498be72005-09-09 13:03:32 -07003083
Andrew Mortona737b3e2006-03-22 00:08:11 -08003084 l3 = cachep->nodelists[numa_node_id()];
Christoph Lametere498be72005-09-09 13:03:32 -07003085 STATS_INC_NODEFREES(cachep);
3086 if (l3->alien && l3->alien[nodeid]) {
3087 alien = l3->alien[nodeid];
3088 spin_lock(&alien->lock);
3089 if (unlikely(alien->avail == alien->limit))
3090 __drain_alien_cache(cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003091 alien, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07003092 alien->entry[alien->avail++] = objp;
3093 spin_unlock(&alien->lock);
3094 } else {
3095 spin_lock(&(cachep->nodelists[nodeid])->
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003096 list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -07003097 free_block(cachep, &objp, 1, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07003098 spin_unlock(&(cachep->nodelists[nodeid])->
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003099 list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003100 }
3101 return;
3102 }
3103 }
3104#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003105 if (likely(ac->avail < ac->limit)) {
3106 STATS_INC_FREEHIT(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003107 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003108 return;
3109 } else {
3110 STATS_INC_FREEMISS(cachep);
3111 cache_flusharray(cachep, ac);
Christoph Lametere498be72005-09-09 13:03:32 -07003112 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003113 }
3114}
3115
3116/**
3117 * kmem_cache_alloc - Allocate an object
3118 * @cachep: The cache to allocate from.
3119 * @flags: See kmalloc().
3120 *
3121 * Allocate an object from this cache. The flags are only relevant
3122 * if the cache has no available objects.
3123 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003124void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003125{
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003126 return __cache_alloc(cachep, flags, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003127}
3128EXPORT_SYMBOL(kmem_cache_alloc);
3129
3130/**
Pekka Enberga8c0f9a2006-03-25 03:06:42 -08003131 * kmem_cache_alloc - Allocate an object. The memory is set to zero.
3132 * @cache: The cache to allocate from.
3133 * @flags: See kmalloc().
3134 *
3135 * Allocate an object from this cache and set the allocated memory to zero.
3136 * The flags are only relevant if the cache has no available objects.
3137 */
3138void *kmem_cache_zalloc(struct kmem_cache *cache, gfp_t flags)
3139{
3140 void *ret = __cache_alloc(cache, flags, __builtin_return_address(0));
3141 if (ret)
3142 memset(ret, 0, obj_size(cache));
3143 return ret;
3144}
3145EXPORT_SYMBOL(kmem_cache_zalloc);
3146
3147/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003148 * kmem_ptr_validate - check if an untrusted pointer might
3149 * be a slab entry.
3150 * @cachep: the cache we're checking against
3151 * @ptr: pointer to validate
3152 *
3153 * This verifies that the untrusted pointer looks sane:
3154 * it is _not_ a guarantee that the pointer is actually
3155 * part of the slab cache in question, but it at least
3156 * validates that the pointer can be dereferenced and
3157 * looks half-way sane.
3158 *
3159 * Currently only used for dentry validation.
3160 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003161int fastcall kmem_ptr_validate(struct kmem_cache *cachep, void *ptr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003162{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003163 unsigned long addr = (unsigned long)ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003164 unsigned long min_addr = PAGE_OFFSET;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003165 unsigned long align_mask = BYTES_PER_WORD - 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003166 unsigned long size = cachep->buffer_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003167 struct page *page;
3168
3169 if (unlikely(addr < min_addr))
3170 goto out;
3171 if (unlikely(addr > (unsigned long)high_memory - size))
3172 goto out;
3173 if (unlikely(addr & align_mask))
3174 goto out;
3175 if (unlikely(!kern_addr_valid(addr)))
3176 goto out;
3177 if (unlikely(!kern_addr_valid(addr + size - 1)))
3178 goto out;
3179 page = virt_to_page(ptr);
3180 if (unlikely(!PageSlab(page)))
3181 goto out;
Pekka Enberg065d41c2005-11-13 16:06:46 -08003182 if (unlikely(page_get_cache(page) != cachep))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003183 goto out;
3184 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003185out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003186 return 0;
3187}
3188
3189#ifdef CONFIG_NUMA
3190/**
3191 * kmem_cache_alloc_node - Allocate an object on the specified node
3192 * @cachep: The cache to allocate from.
3193 * @flags: See kmalloc().
3194 * @nodeid: node number of the target node.
3195 *
3196 * Identical to kmem_cache_alloc, except that this function is slow
3197 * and can sleep. And it will allocate memory on the given node, which
3198 * can improve the performance for cpu bound structures.
Christoph Lametere498be72005-09-09 13:03:32 -07003199 * New and improved: it will now make sure that the object gets
3200 * put on the correct node list so that there is no false sharing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003201 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003202void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003203{
Christoph Lametere498be72005-09-09 13:03:32 -07003204 unsigned long save_flags;
3205 void *ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003206
Christoph Lametere498be72005-09-09 13:03:32 -07003207 cache_alloc_debugcheck_before(cachep, flags);
3208 local_irq_save(save_flags);
Christoph Lameter18f820f2006-02-01 03:05:43 -08003209
3210 if (nodeid == -1 || nodeid == numa_node_id() ||
Andrew Mortona737b3e2006-03-22 00:08:11 -08003211 !cachep->nodelists[nodeid])
Alok N Kataria5c382302005-09-27 21:45:46 -07003212 ptr = ____cache_alloc(cachep, flags);
3213 else
3214 ptr = __cache_alloc_node(cachep, flags, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07003215 local_irq_restore(save_flags);
Christoph Lameter18f820f2006-02-01 03:05:43 -08003216
3217 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr,
3218 __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003219
Christoph Lametere498be72005-09-09 13:03:32 -07003220 return ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003221}
3222EXPORT_SYMBOL(kmem_cache_alloc_node);
3223
Al Virodd0fc662005-10-07 07:46:04 +01003224void *kmalloc_node(size_t size, gfp_t flags, int node)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003225{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003226 struct kmem_cache *cachep;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003227
3228 cachep = kmem_find_general_cachep(size, flags);
3229 if (unlikely(cachep == NULL))
3230 return NULL;
3231 return kmem_cache_alloc_node(cachep, flags, node);
3232}
3233EXPORT_SYMBOL(kmalloc_node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003234#endif
3235
3236/**
3237 * kmalloc - allocate memory
3238 * @size: how many bytes of memory are required.
3239 * @flags: the type of memory to allocate.
Randy Dunlap911851e2006-03-22 00:08:14 -08003240 * @caller: function caller for debug tracking of the caller
Linus Torvalds1da177e2005-04-16 15:20:36 -07003241 *
3242 * kmalloc is the normal method of allocating memory
3243 * in the kernel.
3244 *
3245 * The @flags argument may be one of:
3246 *
3247 * %GFP_USER - Allocate memory on behalf of user. May sleep.
3248 *
3249 * %GFP_KERNEL - Allocate normal kernel ram. May sleep.
3250 *
3251 * %GFP_ATOMIC - Allocation will not sleep. Use inside interrupt handlers.
3252 *
3253 * Additionally, the %GFP_DMA flag may be set to indicate the memory
3254 * must be suitable for DMA. This can mean different things on different
3255 * platforms. For example, on i386, it means that the memory must come
3256 * from the first 16MB.
3257 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003258static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3259 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003260{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003261 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003262
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003263 /* If you want to save a few bytes .text space: replace
3264 * __ with kmem_.
3265 * Then kmalloc uses the uninlined functions instead of the inline
3266 * functions.
3267 */
3268 cachep = __find_general_cachep(size, flags);
Andrew Mortondbdb9042005-09-23 13:24:10 -07003269 if (unlikely(cachep == NULL))
3270 return NULL;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003271 return __cache_alloc(cachep, flags, caller);
3272}
3273
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003274
3275void *__kmalloc(size_t size, gfp_t flags)
3276{
Al Viro871751e2006-03-25 03:06:39 -08003277#ifndef CONFIG_DEBUG_SLAB
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003278 return __do_kmalloc(size, flags, NULL);
Al Viro871751e2006-03-25 03:06:39 -08003279#else
3280 return __do_kmalloc(size, flags, __builtin_return_address(0));
3281#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003282}
3283EXPORT_SYMBOL(__kmalloc);
3284
Al Viro871751e2006-03-25 03:06:39 -08003285#ifdef CONFIG_DEBUG_SLAB
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003286void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller)
3287{
3288 return __do_kmalloc(size, flags, caller);
3289}
3290EXPORT_SYMBOL(__kmalloc_track_caller);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003291#endif
3292
Linus Torvalds1da177e2005-04-16 15:20:36 -07003293#ifdef CONFIG_SMP
3294/**
3295 * __alloc_percpu - allocate one copy of the object for every present
3296 * cpu in the system, zeroing them.
3297 * Objects should be dereferenced using the per_cpu_ptr macro only.
3298 *
3299 * @size: how many bytes of memory are required.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003300 */
Pekka Enbergf9f75002006-01-08 01:00:33 -08003301void *__alloc_percpu(size_t size)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003302{
3303 int i;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003304 struct percpu_data *pdata = kmalloc(sizeof(*pdata), GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003305
3306 if (!pdata)
3307 return NULL;
3308
Christoph Lametere498be72005-09-09 13:03:32 -07003309 /*
3310 * Cannot use for_each_online_cpu since a cpu may come online
3311 * and we have no way of figuring out how to fix the array
3312 * that we have allocated then....
3313 */
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08003314 for_each_possible_cpu(i) {
Christoph Lametere498be72005-09-09 13:03:32 -07003315 int node = cpu_to_node(i);
3316
3317 if (node_online(node))
3318 pdata->ptrs[i] = kmalloc_node(size, GFP_KERNEL, node);
3319 else
3320 pdata->ptrs[i] = kmalloc(size, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003321
3322 if (!pdata->ptrs[i])
3323 goto unwind_oom;
3324 memset(pdata->ptrs[i], 0, size);
3325 }
3326
3327 /* Catch derefs w/o wrappers */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003328 return (void *)(~(unsigned long)pdata);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003329
Andrew Mortona737b3e2006-03-22 00:08:11 -08003330unwind_oom:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003331 while (--i >= 0) {
3332 if (!cpu_possible(i))
3333 continue;
3334 kfree(pdata->ptrs[i]);
3335 }
3336 kfree(pdata);
3337 return NULL;
3338}
3339EXPORT_SYMBOL(__alloc_percpu);
3340#endif
3341
3342/**
3343 * kmem_cache_free - Deallocate an object
3344 * @cachep: The cache the allocation was from.
3345 * @objp: The previously allocated object.
3346 *
3347 * Free an object which was previously allocated from this
3348 * cache.
3349 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003350void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003351{
3352 unsigned long flags;
3353
3354 local_irq_save(flags);
3355 __cache_free(cachep, objp);
3356 local_irq_restore(flags);
3357}
3358EXPORT_SYMBOL(kmem_cache_free);
3359
3360/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003361 * kfree - free previously allocated memory
3362 * @objp: pointer returned by kmalloc.
3363 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003364 * If @objp is NULL, no operation is performed.
3365 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003366 * Don't free memory not originally allocated by kmalloc()
3367 * or you will run into trouble.
3368 */
3369void kfree(const void *objp)
3370{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003371 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003372 unsigned long flags;
3373
3374 if (unlikely(!objp))
3375 return;
3376 local_irq_save(flags);
3377 kfree_debugcheck(objp);
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003378 c = virt_to_cache(objp);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003379 mutex_debug_check_no_locks_freed(objp, obj_size(c));
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003380 __cache_free(c, (void *)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003381 local_irq_restore(flags);
3382}
3383EXPORT_SYMBOL(kfree);
3384
3385#ifdef CONFIG_SMP
3386/**
3387 * free_percpu - free previously allocated percpu memory
3388 * @objp: pointer returned by alloc_percpu.
3389 *
3390 * Don't free memory not originally allocated by alloc_percpu()
3391 * The complemented objp is to check for that.
3392 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003393void free_percpu(const void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003394{
3395 int i;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003396 struct percpu_data *p = (struct percpu_data *)(~(unsigned long)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003397
Christoph Lametere498be72005-09-09 13:03:32 -07003398 /*
3399 * We allocate for all cpus so we cannot use for online cpu here.
3400 */
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08003401 for_each_possible_cpu(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003402 kfree(p->ptrs[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003403 kfree(p);
3404}
3405EXPORT_SYMBOL(free_percpu);
3406#endif
3407
Pekka Enberg343e0d72006-02-01 03:05:50 -08003408unsigned int kmem_cache_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003409{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003410 return obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003411}
3412EXPORT_SYMBOL(kmem_cache_size);
3413
Pekka Enberg343e0d72006-02-01 03:05:50 -08003414const char *kmem_cache_name(struct kmem_cache *cachep)
Arnaldo Carvalho de Melo19449722005-06-18 22:46:19 -07003415{
3416 return cachep->name;
3417}
3418EXPORT_SYMBOL_GPL(kmem_cache_name);
3419
Christoph Lametere498be72005-09-09 13:03:32 -07003420/*
Christoph Lameter0718dc22006-03-25 03:06:47 -08003421 * This initializes kmem_list3 or resizes varioius caches for all nodes.
Christoph Lametere498be72005-09-09 13:03:32 -07003422 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003423static int alloc_kmemlist(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07003424{
3425 int node;
3426 struct kmem_list3 *l3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003427 struct array_cache *new_shared;
3428 struct array_cache **new_alien;
Christoph Lametere498be72005-09-09 13:03:32 -07003429
3430 for_each_online_node(node) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003431
Andrew Mortona737b3e2006-03-22 00:08:11 -08003432 new_alien = alloc_alien_cache(node, cachep->limit);
3433 if (!new_alien)
Christoph Lametere498be72005-09-09 13:03:32 -07003434 goto fail;
Christoph Lametercafeb022006-03-25 03:06:46 -08003435
Christoph Lameter0718dc22006-03-25 03:06:47 -08003436 new_shared = alloc_arraycache(node,
3437 cachep->shared*cachep->batchcount,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003438 0xbaadf00d);
Christoph Lameter0718dc22006-03-25 03:06:47 -08003439 if (!new_shared) {
3440 free_alien_cache(new_alien);
Christoph Lametere498be72005-09-09 13:03:32 -07003441 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003442 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003443
Andrew Mortona737b3e2006-03-22 00:08:11 -08003444 l3 = cachep->nodelists[node];
3445 if (l3) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003446 struct array_cache *shared = l3->shared;
3447
Christoph Lametere498be72005-09-09 13:03:32 -07003448 spin_lock_irq(&l3->list_lock);
3449
Christoph Lametercafeb022006-03-25 03:06:46 -08003450 if (shared)
Christoph Lameter0718dc22006-03-25 03:06:47 -08003451 free_block(cachep, shared->entry,
3452 shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07003453
Christoph Lametercafeb022006-03-25 03:06:46 -08003454 l3->shared = new_shared;
3455 if (!l3->alien) {
Christoph Lametere498be72005-09-09 13:03:32 -07003456 l3->alien = new_alien;
3457 new_alien = NULL;
3458 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003459 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003460 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003461 spin_unlock_irq(&l3->list_lock);
Christoph Lametercafeb022006-03-25 03:06:46 -08003462 kfree(shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003463 free_alien_cache(new_alien);
3464 continue;
3465 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08003466 l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node);
Christoph Lameter0718dc22006-03-25 03:06:47 -08003467 if (!l3) {
3468 free_alien_cache(new_alien);
3469 kfree(new_shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003470 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003471 }
Christoph Lametere498be72005-09-09 13:03:32 -07003472
3473 kmem_list3_init(l3);
3474 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Andrew Mortona737b3e2006-03-22 00:08:11 -08003475 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003476 l3->shared = new_shared;
Christoph Lametere498be72005-09-09 13:03:32 -07003477 l3->alien = new_alien;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003478 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003479 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003480 cachep->nodelists[node] = l3;
3481 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003482 return 0;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003483
Andrew Mortona737b3e2006-03-22 00:08:11 -08003484fail:
Christoph Lameter0718dc22006-03-25 03:06:47 -08003485 if (!cachep->next.next) {
3486 /* Cache is not active yet. Roll back what we did */
3487 node--;
3488 while (node >= 0) {
3489 if (cachep->nodelists[node]) {
3490 l3 = cachep->nodelists[node];
3491
3492 kfree(l3->shared);
3493 free_alien_cache(l3->alien);
3494 kfree(l3);
3495 cachep->nodelists[node] = NULL;
3496 }
3497 node--;
3498 }
3499 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003500 return -ENOMEM;
Christoph Lametere498be72005-09-09 13:03:32 -07003501}
3502
Linus Torvalds1da177e2005-04-16 15:20:36 -07003503struct ccupdate_struct {
Pekka Enberg343e0d72006-02-01 03:05:50 -08003504 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003505 struct array_cache *new[NR_CPUS];
3506};
3507
3508static void do_ccupdate_local(void *info)
3509{
Andrew Mortona737b3e2006-03-22 00:08:11 -08003510 struct ccupdate_struct *new = info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003511 struct array_cache *old;
3512
3513 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003514 old = cpu_cache_get(new->cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003515
Linus Torvalds1da177e2005-04-16 15:20:36 -07003516 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3517 new->new[smp_processor_id()] = old;
3518}
3519
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003520/* Always called with the cache_chain_mutex held */
Andrew Mortona737b3e2006-03-22 00:08:11 -08003521static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3522 int batchcount, int shared)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003523{
3524 struct ccupdate_struct new;
Christoph Lametere498be72005-09-09 13:03:32 -07003525 int i, err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003526
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003527 memset(&new.new, 0, sizeof(new.new));
Christoph Lametere498be72005-09-09 13:03:32 -07003528 for_each_online_cpu(i) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003529 new.new[i] = alloc_arraycache(cpu_to_node(i), limit,
3530 batchcount);
Christoph Lametere498be72005-09-09 13:03:32 -07003531 if (!new.new[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003532 for (i--; i >= 0; i--)
3533 kfree(new.new[i]);
Christoph Lametere498be72005-09-09 13:03:32 -07003534 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003535 }
3536 }
3537 new.cachep = cachep;
3538
Andrew Mortona07fa392006-03-22 00:08:17 -08003539 on_each_cpu(do_ccupdate_local, (void *)&new, 1, 1);
Christoph Lametere498be72005-09-09 13:03:32 -07003540
Linus Torvalds1da177e2005-04-16 15:20:36 -07003541 check_irq_on();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003542 cachep->batchcount = batchcount;
3543 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07003544 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003545
Christoph Lametere498be72005-09-09 13:03:32 -07003546 for_each_online_cpu(i) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003547 struct array_cache *ccold = new.new[i];
3548 if (!ccold)
3549 continue;
Christoph Lametere498be72005-09-09 13:03:32 -07003550 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -07003551 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
Christoph Lametere498be72005-09-09 13:03:32 -07003552 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003553 kfree(ccold);
3554 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003555
Christoph Lametere498be72005-09-09 13:03:32 -07003556 err = alloc_kmemlist(cachep);
3557 if (err) {
3558 printk(KERN_ERR "alloc_kmemlist failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003559 cachep->name, -err);
Christoph Lametere498be72005-09-09 13:03:32 -07003560 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003561 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003562 return 0;
3563}
3564
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003565/* Called with cache_chain_mutex held always */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003566static void enable_cpucache(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003567{
3568 int err;
3569 int limit, shared;
3570
Andrew Mortona737b3e2006-03-22 00:08:11 -08003571 /*
3572 * The head array serves three purposes:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003573 * - create a LIFO ordering, i.e. return objects that are cache-warm
3574 * - reduce the number of spinlock operations.
Andrew Mortona737b3e2006-03-22 00:08:11 -08003575 * - reduce the number of linked list operations on the slab and
Linus Torvalds1da177e2005-04-16 15:20:36 -07003576 * bufctl chains: array operations are cheaper.
3577 * The numbers are guessed, we should auto-tune as described by
3578 * Bonwick.
3579 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003580 if (cachep->buffer_size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003581 limit = 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003582 else if (cachep->buffer_size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003583 limit = 8;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003584 else if (cachep->buffer_size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003585 limit = 24;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003586 else if (cachep->buffer_size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003587 limit = 54;
3588 else
3589 limit = 120;
3590
Andrew Mortona737b3e2006-03-22 00:08:11 -08003591 /*
3592 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
Linus Torvalds1da177e2005-04-16 15:20:36 -07003593 * allocation behaviour: Most allocs on one cpu, most free operations
3594 * on another cpu. For these cases, an efficient object passing between
3595 * cpus is necessary. This is provided by a shared array. The array
3596 * replaces Bonwick's magazine layer.
3597 * On uniprocessor, it's functionally equivalent (but less efficient)
3598 * to a larger limit. Thus disabled by default.
3599 */
3600 shared = 0;
3601#ifdef CONFIG_SMP
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003602 if (cachep->buffer_size <= PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003603 shared = 8;
3604#endif
3605
3606#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003607 /*
3608 * With debugging enabled, large batchcount lead to excessively long
3609 * periods with disabled local interrupts. Limit the batchcount
Linus Torvalds1da177e2005-04-16 15:20:36 -07003610 */
3611 if (limit > 32)
3612 limit = 32;
3613#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003614 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003615 if (err)
3616 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003617 cachep->name, -err);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003618}
3619
Christoph Lameter1b552532006-03-22 00:09:07 -08003620/*
3621 * Drain an array if it contains any elements taking the l3 lock only if
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003622 * necessary. Note that the l3 listlock also protects the array_cache
3623 * if drain_array() is used on the shared array.
Christoph Lameter1b552532006-03-22 00:09:07 -08003624 */
3625void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
3626 struct array_cache *ac, int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003627{
3628 int tofree;
3629
Christoph Lameter1b552532006-03-22 00:09:07 -08003630 if (!ac || !ac->avail)
3631 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003632 if (ac->touched && !force) {
3633 ac->touched = 0;
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003634 } else {
Christoph Lameter1b552532006-03-22 00:09:07 -08003635 spin_lock_irq(&l3->list_lock);
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003636 if (ac->avail) {
3637 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3638 if (tofree > ac->avail)
3639 tofree = (ac->avail + 1) / 2;
3640 free_block(cachep, ac->entry, tofree, node);
3641 ac->avail -= tofree;
3642 memmove(ac->entry, &(ac->entry[tofree]),
3643 sizeof(void *) * ac->avail);
3644 }
Christoph Lameter1b552532006-03-22 00:09:07 -08003645 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003646 }
3647}
3648
3649/**
3650 * cache_reap - Reclaim memory from caches.
Randy Dunlap1e5d5332005-11-07 01:01:06 -08003651 * @unused: unused parameter
Linus Torvalds1da177e2005-04-16 15:20:36 -07003652 *
3653 * Called from workqueue/eventd every few seconds.
3654 * Purpose:
3655 * - clear the per-cpu caches for this CPU.
3656 * - return freeable pages to the main free memory pool.
3657 *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003658 * If we cannot acquire the cache chain mutex then just give up - we'll try
3659 * again on the next iteration.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003660 */
3661static void cache_reap(void *unused)
3662{
3663 struct list_head *walk;
Christoph Lametere498be72005-09-09 13:03:32 -07003664 struct kmem_list3 *l3;
Christoph Lameteraab22072006-03-22 00:09:06 -08003665 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003666
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003667 if (!mutex_trylock(&cache_chain_mutex)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003668 /* Give up. Setup the next iteration. */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003669 schedule_delayed_work(&__get_cpu_var(reap_work),
3670 REAPTIMEOUT_CPUC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003671 return;
3672 }
3673
3674 list_for_each(walk, &cache_chain) {
Pekka Enberg343e0d72006-02-01 03:05:50 -08003675 struct kmem_cache *searchp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003676 struct list_head *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003677 int tofree;
3678 struct slab *slabp;
3679
Pekka Enberg343e0d72006-02-01 03:05:50 -08003680 searchp = list_entry(walk, struct kmem_cache, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003681 check_irq_on();
3682
Christoph Lameter35386e32006-03-22 00:09:05 -08003683 /*
3684 * We only take the l3 lock if absolutely necessary and we
3685 * have established with reasonable certainty that
3686 * we can do some work if the lock was obtained.
3687 */
Christoph Lameteraab22072006-03-22 00:09:06 -08003688 l3 = searchp->nodelists[node];
Christoph Lameter35386e32006-03-22 00:09:05 -08003689
Christoph Lameter8fce4d82006-03-09 17:33:54 -08003690 reap_alien(searchp, l3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003691
Christoph Lameteraab22072006-03-22 00:09:06 -08003692 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003693
Christoph Lameter35386e32006-03-22 00:09:05 -08003694 /*
3695 * These are racy checks but it does not matter
3696 * if we skip one check or scan twice.
3697 */
Christoph Lametere498be72005-09-09 13:03:32 -07003698 if (time_after(l3->next_reap, jiffies))
Christoph Lameter35386e32006-03-22 00:09:05 -08003699 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003700
Christoph Lametere498be72005-09-09 13:03:32 -07003701 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003702
Christoph Lameteraab22072006-03-22 00:09:06 -08003703 drain_array(searchp, l3, l3->shared, 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003704
Christoph Lametere498be72005-09-09 13:03:32 -07003705 if (l3->free_touched) {
3706 l3->free_touched = 0;
Christoph Lameter35386e32006-03-22 00:09:05 -08003707 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003708 }
3709
Andrew Mortona737b3e2006-03-22 00:08:11 -08003710 tofree = (l3->free_limit + 5 * searchp->num - 1) /
3711 (5 * searchp->num);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003712 do {
Christoph Lameter35386e32006-03-22 00:09:05 -08003713 /*
3714 * Do not lock if there are no free blocks.
3715 */
3716 if (list_empty(&l3->slabs_free))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003717 break;
3718
Christoph Lameter35386e32006-03-22 00:09:05 -08003719 spin_lock_irq(&l3->list_lock);
3720 p = l3->slabs_free.next;
3721 if (p == &(l3->slabs_free)) {
3722 spin_unlock_irq(&l3->list_lock);
3723 break;
3724 }
3725
Linus Torvalds1da177e2005-04-16 15:20:36 -07003726 slabp = list_entry(p, struct slab, list);
3727 BUG_ON(slabp->inuse);
3728 list_del(&slabp->list);
3729 STATS_INC_REAPED(searchp);
3730
Andrew Mortona737b3e2006-03-22 00:08:11 -08003731 /*
3732 * Safe to drop the lock. The slab is no longer linked
3733 * to the cache. searchp cannot disappear, we hold
Linus Torvalds1da177e2005-04-16 15:20:36 -07003734 * cache_chain_lock
3735 */
Christoph Lametere498be72005-09-09 13:03:32 -07003736 l3->free_objects -= searchp->num;
3737 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003738 slab_destroy(searchp, slabp);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003739 } while (--tofree > 0);
Christoph Lameter35386e32006-03-22 00:09:05 -08003740next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003741 cond_resched();
3742 }
3743 check_irq_on();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003744 mutex_unlock(&cache_chain_mutex);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08003745 next_reap_node();
Andrew Mortona737b3e2006-03-22 00:08:11 -08003746 /* Set up the next iteration */
Manfred Spraulcd61ef62005-11-07 00:58:02 -08003747 schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003748}
3749
3750#ifdef CONFIG_PROC_FS
3751
Pekka Enberg85289f92006-01-08 01:00:36 -08003752static void print_slabinfo_header(struct seq_file *m)
3753{
3754 /*
3755 * Output format version, so at least we can change it
3756 * without _too_ many complaints.
3757 */
3758#if STATS
3759 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
3760#else
3761 seq_puts(m, "slabinfo - version: 2.1\n");
3762#endif
3763 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
3764 "<objperslab> <pagesperslab>");
3765 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
3766 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
3767#if STATS
3768 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
3769 "<error> <maxfreeable> <nodeallocs> <remotefrees>");
3770 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
3771#endif
3772 seq_putc(m, '\n');
3773}
3774
Linus Torvalds1da177e2005-04-16 15:20:36 -07003775static void *s_start(struct seq_file *m, loff_t *pos)
3776{
3777 loff_t n = *pos;
3778 struct list_head *p;
3779
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003780 mutex_lock(&cache_chain_mutex);
Pekka Enberg85289f92006-01-08 01:00:36 -08003781 if (!n)
3782 print_slabinfo_header(m);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003783 p = cache_chain.next;
3784 while (n--) {
3785 p = p->next;
3786 if (p == &cache_chain)
3787 return NULL;
3788 }
Pekka Enberg343e0d72006-02-01 03:05:50 -08003789 return list_entry(p, struct kmem_cache, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003790}
3791
3792static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3793{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003794 struct kmem_cache *cachep = p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003795 ++*pos;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003796 return cachep->next.next == &cache_chain ?
3797 NULL : list_entry(cachep->next.next, struct kmem_cache, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003798}
3799
3800static void s_stop(struct seq_file *m, void *p)
3801{
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003802 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003803}
3804
3805static int s_show(struct seq_file *m, void *p)
3806{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003807 struct kmem_cache *cachep = p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003808 struct list_head *q;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003809 struct slab *slabp;
3810 unsigned long active_objs;
3811 unsigned long num_objs;
3812 unsigned long active_slabs = 0;
3813 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07003814 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003815 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003816 int node;
3817 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003818
Linus Torvalds1da177e2005-04-16 15:20:36 -07003819 active_objs = 0;
3820 num_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07003821 for_each_online_node(node) {
3822 l3 = cachep->nodelists[node];
3823 if (!l3)
3824 continue;
3825
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003826 check_irq_on();
3827 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003828
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003829 list_for_each(q, &l3->slabs_full) {
Christoph Lametere498be72005-09-09 13:03:32 -07003830 slabp = list_entry(q, struct slab, list);
3831 if (slabp->inuse != cachep->num && !error)
3832 error = "slabs_full accounting error";
3833 active_objs += cachep->num;
3834 active_slabs++;
3835 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003836 list_for_each(q, &l3->slabs_partial) {
Christoph Lametere498be72005-09-09 13:03:32 -07003837 slabp = list_entry(q, struct slab, list);
3838 if (slabp->inuse == cachep->num && !error)
3839 error = "slabs_partial inuse accounting error";
3840 if (!slabp->inuse && !error)
3841 error = "slabs_partial/inuse accounting error";
3842 active_objs += slabp->inuse;
3843 active_slabs++;
3844 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003845 list_for_each(q, &l3->slabs_free) {
Christoph Lametere498be72005-09-09 13:03:32 -07003846 slabp = list_entry(q, struct slab, list);
3847 if (slabp->inuse && !error)
3848 error = "slabs_free/inuse accounting error";
3849 num_slabs++;
3850 }
3851 free_objects += l3->free_objects;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08003852 if (l3->shared)
3853 shared_avail += l3->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07003854
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003855 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003856 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003857 num_slabs += active_slabs;
3858 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003859 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003860 error = "free_objects accounting error";
3861
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003862 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003863 if (error)
3864 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
3865
3866 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003867 name, active_objs, num_objs, cachep->buffer_size,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003868 cachep->num, (1 << cachep->gfporder));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003869 seq_printf(m, " : tunables %4u %4u %4u",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003870 cachep->limit, cachep->batchcount, cachep->shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003871 seq_printf(m, " : slabdata %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003872 active_slabs, num_slabs, shared_avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003873#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003874 { /* list3 stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003875 unsigned long high = cachep->high_mark;
3876 unsigned long allocs = cachep->num_allocations;
3877 unsigned long grown = cachep->grown;
3878 unsigned long reaped = cachep->reaped;
3879 unsigned long errors = cachep->errors;
3880 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003881 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07003882 unsigned long node_frees = cachep->node_frees;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003883
Christoph Lametere498be72005-09-09 13:03:32 -07003884 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
Andrew Mortona737b3e2006-03-22 00:08:11 -08003885 %4lu %4lu %4lu %4lu", allocs, high, grown,
3886 reaped, errors, max_freeable, node_allocs,
3887 node_frees);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003888 }
3889 /* cpu stats */
3890 {
3891 unsigned long allochit = atomic_read(&cachep->allochit);
3892 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
3893 unsigned long freehit = atomic_read(&cachep->freehit);
3894 unsigned long freemiss = atomic_read(&cachep->freemiss);
3895
3896 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003897 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003898 }
3899#endif
3900 seq_putc(m, '\n');
Linus Torvalds1da177e2005-04-16 15:20:36 -07003901 return 0;
3902}
3903
3904/*
3905 * slabinfo_op - iterator that generates /proc/slabinfo
3906 *
3907 * Output layout:
3908 * cache-name
3909 * num-active-objs
3910 * total-objs
3911 * object size
3912 * num-active-slabs
3913 * total-slabs
3914 * num-pages-per-slab
3915 * + further values on SMP and with statistics enabled
3916 */
3917
3918struct seq_operations slabinfo_op = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003919 .start = s_start,
3920 .next = s_next,
3921 .stop = s_stop,
3922 .show = s_show,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003923};
3924
3925#define MAX_SLABINFO_WRITE 128
3926/**
3927 * slabinfo_write - Tuning for the slab allocator
3928 * @file: unused
3929 * @buffer: user buffer
3930 * @count: data length
3931 * @ppos: unused
3932 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003933ssize_t slabinfo_write(struct file *file, const char __user * buffer,
3934 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003935{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003936 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003937 int limit, batchcount, shared, res;
3938 struct list_head *p;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003939
Linus Torvalds1da177e2005-04-16 15:20:36 -07003940 if (count > MAX_SLABINFO_WRITE)
3941 return -EINVAL;
3942 if (copy_from_user(&kbuf, buffer, count))
3943 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003944 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07003945
3946 tmp = strchr(kbuf, ' ');
3947 if (!tmp)
3948 return -EINVAL;
3949 *tmp = '\0';
3950 tmp++;
3951 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
3952 return -EINVAL;
3953
3954 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003955 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003956 res = -EINVAL;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003957 list_for_each(p, &cache_chain) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003958 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003959
Andrew Mortona737b3e2006-03-22 00:08:11 -08003960 cachep = list_entry(p, struct kmem_cache, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003961 if (!strcmp(cachep->name, kbuf)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003962 if (limit < 1 || batchcount < 1 ||
3963 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07003964 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003965 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07003966 res = do_tune_cpucache(cachep, limit,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003967 batchcount, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003968 }
3969 break;
3970 }
3971 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003972 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003973 if (res >= 0)
3974 res = count;
3975 return res;
3976}
Al Viro871751e2006-03-25 03:06:39 -08003977
3978#ifdef CONFIG_DEBUG_SLAB_LEAK
3979
3980static void *leaks_start(struct seq_file *m, loff_t *pos)
3981{
3982 loff_t n = *pos;
3983 struct list_head *p;
3984
3985 mutex_lock(&cache_chain_mutex);
3986 p = cache_chain.next;
3987 while (n--) {
3988 p = p->next;
3989 if (p == &cache_chain)
3990 return NULL;
3991 }
3992 return list_entry(p, struct kmem_cache, next);
3993}
3994
3995static inline int add_caller(unsigned long *n, unsigned long v)
3996{
3997 unsigned long *p;
3998 int l;
3999 if (!v)
4000 return 1;
4001 l = n[1];
4002 p = n + 2;
4003 while (l) {
4004 int i = l/2;
4005 unsigned long *q = p + 2 * i;
4006 if (*q == v) {
4007 q[1]++;
4008 return 1;
4009 }
4010 if (*q > v) {
4011 l = i;
4012 } else {
4013 p = q + 2;
4014 l -= i + 1;
4015 }
4016 }
4017 if (++n[1] == n[0])
4018 return 0;
4019 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4020 p[0] = v;
4021 p[1] = 1;
4022 return 1;
4023}
4024
4025static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4026{
4027 void *p;
4028 int i;
4029 if (n[0] == n[1])
4030 return;
4031 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4032 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4033 continue;
4034 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4035 return;
4036 }
4037}
4038
4039static void show_symbol(struct seq_file *m, unsigned long address)
4040{
4041#ifdef CONFIG_KALLSYMS
4042 char *modname;
4043 const char *name;
4044 unsigned long offset, size;
4045 char namebuf[KSYM_NAME_LEN+1];
4046
4047 name = kallsyms_lookup(address, &size, &offset, &modname, namebuf);
4048
4049 if (name) {
4050 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4051 if (modname)
4052 seq_printf(m, " [%s]", modname);
4053 return;
4054 }
4055#endif
4056 seq_printf(m, "%p", (void *)address);
4057}
4058
4059static int leaks_show(struct seq_file *m, void *p)
4060{
4061 struct kmem_cache *cachep = p;
4062 struct list_head *q;
4063 struct slab *slabp;
4064 struct kmem_list3 *l3;
4065 const char *name;
4066 unsigned long *n = m->private;
4067 int node;
4068 int i;
4069
4070 if (!(cachep->flags & SLAB_STORE_USER))
4071 return 0;
4072 if (!(cachep->flags & SLAB_RED_ZONE))
4073 return 0;
4074
4075 /* OK, we can do it */
4076
4077 n[1] = 0;
4078
4079 for_each_online_node(node) {
4080 l3 = cachep->nodelists[node];
4081 if (!l3)
4082 continue;
4083
4084 check_irq_on();
4085 spin_lock_irq(&l3->list_lock);
4086
4087 list_for_each(q, &l3->slabs_full) {
4088 slabp = list_entry(q, struct slab, list);
4089 handle_slab(n, cachep, slabp);
4090 }
4091 list_for_each(q, &l3->slabs_partial) {
4092 slabp = list_entry(q, struct slab, list);
4093 handle_slab(n, cachep, slabp);
4094 }
4095 spin_unlock_irq(&l3->list_lock);
4096 }
4097 name = cachep->name;
4098 if (n[0] == n[1]) {
4099 /* Increase the buffer size */
4100 mutex_unlock(&cache_chain_mutex);
4101 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4102 if (!m->private) {
4103 /* Too bad, we are really out */
4104 m->private = n;
4105 mutex_lock(&cache_chain_mutex);
4106 return -ENOMEM;
4107 }
4108 *(unsigned long *)m->private = n[0] * 2;
4109 kfree(n);
4110 mutex_lock(&cache_chain_mutex);
4111 /* Now make sure this entry will be retried */
4112 m->count = m->size;
4113 return 0;
4114 }
4115 for (i = 0; i < n[1]; i++) {
4116 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4117 show_symbol(m, n[2*i+2]);
4118 seq_putc(m, '\n');
4119 }
4120 return 0;
4121}
4122
4123struct seq_operations slabstats_op = {
4124 .start = leaks_start,
4125 .next = s_next,
4126 .stop = s_stop,
4127 .show = leaks_show,
4128};
4129#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004130#endif
4131
Manfred Spraul00e145b2005-09-03 15:55:07 -07004132/**
4133 * ksize - get the actual amount of memory allocated for a given object
4134 * @objp: Pointer to the object
4135 *
4136 * kmalloc may internally round up allocations and return more memory
4137 * than requested. ksize() can be used to determine the actual amount of
4138 * memory allocated. The caller may use this additional memory, even though
4139 * a smaller amount of memory was initially specified with the kmalloc call.
4140 * The caller must guarantee that objp points to a valid object previously
4141 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4142 * must not be freed during the duration of the call.
4143 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004144unsigned int ksize(const void *objp)
4145{
Manfred Spraul00e145b2005-09-03 15:55:07 -07004146 if (unlikely(objp == NULL))
4147 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004148
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08004149 return obj_size(virt_to_cache(objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004150}