blob: 6f5aeebd430683b7ecbcd1ca28f6efce2b991740 [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 Lametere498be72005-09-09 13:03:32 -0700901#ifdef CONFIG_NUMA
Pekka Enberg343e0d72006-02-01 03:05:50 -0800902static void *__cache_alloc_node(struct kmem_cache *, gfp_t, int);
Paul Jacksonc61afb12006-03-24 03:16:08 -0800903static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
Christoph Lameterdc85da12006-01-18 17:42:36 -0800904
Pekka Enberg5295a742006-02-01 03:05:48 -0800905static struct array_cache **alloc_alien_cache(int node, int limit)
Christoph Lametere498be72005-09-09 13:03:32 -0700906{
907 struct array_cache **ac_ptr;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800908 int memsize = sizeof(void *) * MAX_NUMNODES;
Christoph Lametere498be72005-09-09 13:03:32 -0700909 int i;
910
911 if (limit > 1)
912 limit = 12;
913 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
914 if (ac_ptr) {
915 for_each_node(i) {
916 if (i == node || !node_online(i)) {
917 ac_ptr[i] = NULL;
918 continue;
919 }
920 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
921 if (!ac_ptr[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800922 for (i--; i <= 0; i--)
Christoph Lametere498be72005-09-09 13:03:32 -0700923 kfree(ac_ptr[i]);
924 kfree(ac_ptr);
925 return NULL;
926 }
927 }
928 }
929 return ac_ptr;
930}
931
Pekka Enberg5295a742006-02-01 03:05:48 -0800932static void free_alien_cache(struct array_cache **ac_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -0700933{
934 int i;
935
936 if (!ac_ptr)
937 return;
Christoph Lametere498be72005-09-09 13:03:32 -0700938 for_each_node(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800939 kfree(ac_ptr[i]);
Christoph Lametere498be72005-09-09 13:03:32 -0700940 kfree(ac_ptr);
941}
942
Pekka Enberg343e0d72006-02-01 03:05:50 -0800943static void __drain_alien_cache(struct kmem_cache *cachep,
Pekka Enberg5295a742006-02-01 03:05:48 -0800944 struct array_cache *ac, int node)
Christoph Lametere498be72005-09-09 13:03:32 -0700945{
946 struct kmem_list3 *rl3 = cachep->nodelists[node];
947
948 if (ac->avail) {
949 spin_lock(&rl3->list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -0700950 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -0700951 ac->avail = 0;
952 spin_unlock(&rl3->list_lock);
953 }
954}
955
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800956/*
957 * Called from cache_reap() to regularly drain alien caches round robin.
958 */
959static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
960{
961 int node = __get_cpu_var(reap_node);
962
963 if (l3->alien) {
964 struct array_cache *ac = l3->alien[node];
965 if (ac && ac->avail) {
966 spin_lock_irq(&ac->lock);
967 __drain_alien_cache(cachep, ac, node);
968 spin_unlock_irq(&ac->lock);
969 }
970 }
971}
972
Andrew Mortona737b3e2006-03-22 00:08:11 -0800973static void drain_alien_cache(struct kmem_cache *cachep,
974 struct array_cache **alien)
Christoph Lametere498be72005-09-09 13:03:32 -0700975{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800976 int i = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700977 struct array_cache *ac;
978 unsigned long flags;
979
980 for_each_online_node(i) {
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -0800981 ac = alien[i];
Christoph Lametere498be72005-09-09 13:03:32 -0700982 if (ac) {
983 spin_lock_irqsave(&ac->lock, flags);
984 __drain_alien_cache(cachep, ac, i);
985 spin_unlock_irqrestore(&ac->lock, flags);
986 }
987 }
988}
989#else
Linus Torvalds7a21ef62006-02-05 11:26:38 -0800990
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -0800991#define drain_alien_cache(cachep, alien) do { } while (0)
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800992#define reap_alien(cachep, l3) do { } while (0)
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -0800993
Linus Torvalds7a21ef62006-02-05 11:26:38 -0800994static inline struct array_cache **alloc_alien_cache(int node, int limit)
995{
996 return (struct array_cache **) 0x01020304ul;
997}
998
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -0800999static inline void free_alien_cache(struct array_cache **ac_ptr)
1000{
1001}
Linus Torvalds7a21ef62006-02-05 11:26:38 -08001002
Christoph Lametere498be72005-09-09 13:03:32 -07001003#endif
1004
Linus Torvalds1da177e2005-04-16 15:20:36 -07001005static int __devinit cpuup_callback(struct notifier_block *nfb,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001006 unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001007{
1008 long cpu = (long)hcpu;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001009 struct kmem_cache *cachep;
Christoph Lametere498be72005-09-09 13:03:32 -07001010 struct kmem_list3 *l3 = NULL;
1011 int node = cpu_to_node(cpu);
1012 int memsize = sizeof(struct kmem_list3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001013
1014 switch (action) {
1015 case CPU_UP_PREPARE:
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001016 mutex_lock(&cache_chain_mutex);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001017 /*
1018 * We need to do this right in the beginning since
Christoph Lametere498be72005-09-09 13:03:32 -07001019 * alloc_arraycache's are going to use this list.
1020 * kmalloc_node allows us to add the slab to the right
1021 * kmem_list3 and not this cpu's kmem_list3
1022 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001023
Christoph Lametere498be72005-09-09 13:03:32 -07001024 list_for_each_entry(cachep, &cache_chain, next) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08001025 /*
1026 * Set up the size64 kmemlist for cpu before we can
Christoph Lametere498be72005-09-09 13:03:32 -07001027 * begin anything. Make sure some other cpu on this
1028 * node has not already allocated this
1029 */
1030 if (!cachep->nodelists[node]) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08001031 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1032 if (!l3)
Christoph Lametere498be72005-09-09 13:03:32 -07001033 goto bad;
1034 kmem_list3_init(l3);
1035 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001036 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -07001037
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001038 /*
1039 * The l3s don't come and go as CPUs come and
1040 * go. cache_chain_mutex is sufficient
1041 * protection here.
1042 */
Christoph Lametere498be72005-09-09 13:03:32 -07001043 cachep->nodelists[node] = l3;
1044 }
1045
1046 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1047 cachep->nodelists[node]->free_limit =
Andrew Mortona737b3e2006-03-22 00:08:11 -08001048 (1 + nr_cpus_node(node)) *
1049 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07001050 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1051 }
1052
Andrew Mortona737b3e2006-03-22 00:08:11 -08001053 /*
1054 * Now we can go ahead with allocating the shared arrays and
1055 * array caches
1056 */
Christoph Lametere498be72005-09-09 13:03:32 -07001057 list_for_each_entry(cachep, &cache_chain, next) {
Tobias Klausercd105df2006-01-08 01:00:59 -08001058 struct array_cache *nc;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001059 struct array_cache *shared;
1060 struct array_cache **alien;
Tobias Klausercd105df2006-01-08 01:00:59 -08001061
Christoph Lametere498be72005-09-09 13:03:32 -07001062 nc = alloc_arraycache(node, cachep->limit,
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001063 cachep->batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001064 if (!nc)
1065 goto bad;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001066 shared = alloc_arraycache(node,
1067 cachep->shared * cachep->batchcount,
1068 0xbaadf00d);
1069 if (!shared)
1070 goto bad;
Linus Torvalds7a21ef62006-02-05 11:26:38 -08001071
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001072 alien = alloc_alien_cache(node, cachep->limit);
1073 if (!alien)
1074 goto bad;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001075 cachep->array[cpu] = nc;
Christoph Lametere498be72005-09-09 13:03:32 -07001076 l3 = cachep->nodelists[node];
1077 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07001078
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001079 spin_lock_irq(&l3->list_lock);
1080 if (!l3->shared) {
1081 /*
1082 * We are serialised from CPU_DEAD or
1083 * CPU_UP_CANCELLED by the cpucontrol lock
1084 */
1085 l3->shared = shared;
1086 shared = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07001087 }
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001088#ifdef CONFIG_NUMA
1089 if (!l3->alien) {
1090 l3->alien = alien;
1091 alien = NULL;
1092 }
1093#endif
1094 spin_unlock_irq(&l3->list_lock);
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001095 kfree(shared);
1096 free_alien_cache(alien);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001097 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001098 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001099 break;
1100 case CPU_ONLINE:
1101 start_cpu_timer(cpu);
1102 break;
1103#ifdef CONFIG_HOTPLUG_CPU
1104 case CPU_DEAD:
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001105 /*
1106 * Even if all the cpus of a node are down, we don't free the
1107 * kmem_list3 of any cache. This to avoid a race between
1108 * cpu_down, and a kmalloc allocation from another cpu for
1109 * memory from the node of the cpu going down. The list3
1110 * structure is usually allocated from kmem_cache_create() and
1111 * gets destroyed at kmem_cache_destroy().
1112 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001113 /* fall thru */
1114 case CPU_UP_CANCELED:
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001115 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001116 list_for_each_entry(cachep, &cache_chain, next) {
1117 struct array_cache *nc;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001118 struct array_cache *shared;
1119 struct array_cache **alien;
Christoph Lametere498be72005-09-09 13:03:32 -07001120 cpumask_t mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001121
Christoph Lametere498be72005-09-09 13:03:32 -07001122 mask = node_to_cpumask(node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001123 /* cpu is dead; no one can alloc from it. */
1124 nc = cachep->array[cpu];
1125 cachep->array[cpu] = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07001126 l3 = cachep->nodelists[node];
1127
1128 if (!l3)
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001129 goto free_array_cache;
Christoph Lametere498be72005-09-09 13:03:32 -07001130
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08001131 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07001132
1133 /* Free limit for this kmem_list3 */
1134 l3->free_limit -= cachep->batchcount;
1135 if (nc)
Christoph Lameterff694162005-09-22 21:44:02 -07001136 free_block(cachep, nc->entry, nc->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001137
1138 if (!cpus_empty(mask)) {
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08001139 spin_unlock_irq(&l3->list_lock);
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001140 goto free_array_cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001141 }
Christoph Lametere498be72005-09-09 13:03:32 -07001142
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001143 shared = l3->shared;
1144 if (shared) {
Christoph Lametere498be72005-09-09 13:03:32 -07001145 free_block(cachep, l3->shared->entry,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001146 l3->shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001147 l3->shared = NULL;
1148 }
Christoph Lametere498be72005-09-09 13:03:32 -07001149
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001150 alien = l3->alien;
1151 l3->alien = NULL;
1152
1153 spin_unlock_irq(&l3->list_lock);
1154
1155 kfree(shared);
1156 if (alien) {
1157 drain_alien_cache(cachep, alien);
1158 free_alien_cache(alien);
Christoph Lametere498be72005-09-09 13:03:32 -07001159 }
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001160free_array_cache:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001161 kfree(nc);
1162 }
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001163 /*
1164 * In the previous loop, all the objects were freed to
1165 * the respective cache's slabs, now we can go ahead and
1166 * shrink each nodelist to its limit.
1167 */
1168 list_for_each_entry(cachep, &cache_chain, next) {
1169 l3 = cachep->nodelists[node];
1170 if (!l3)
1171 continue;
1172 spin_lock_irq(&l3->list_lock);
1173 /* free slabs belonging to this node */
1174 __node_shrink(cachep, node);
1175 spin_unlock_irq(&l3->list_lock);
1176 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001177 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001178 break;
1179#endif
1180 }
1181 return NOTIFY_OK;
Andrew Mortona737b3e2006-03-22 00:08:11 -08001182bad:
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001183 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001184 return NOTIFY_BAD;
1185}
1186
1187static struct notifier_block cpucache_notifier = { &cpuup_callback, NULL, 0 };
1188
Christoph Lametere498be72005-09-09 13:03:32 -07001189/*
1190 * swap the static kmem_list3 with kmalloced memory
1191 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001192static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1193 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001194{
1195 struct kmem_list3 *ptr;
1196
1197 BUG_ON(cachep->nodelists[nodeid] != list);
1198 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1199 BUG_ON(!ptr);
1200
1201 local_irq_disable();
1202 memcpy(ptr, list, sizeof(struct kmem_list3));
1203 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1204 cachep->nodelists[nodeid] = ptr;
1205 local_irq_enable();
1206}
1207
Andrew Mortona737b3e2006-03-22 00:08:11 -08001208/*
1209 * Initialisation. Called after the page allocator have been initialised and
1210 * before smp_init().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001211 */
1212void __init kmem_cache_init(void)
1213{
1214 size_t left_over;
1215 struct cache_sizes *sizes;
1216 struct cache_names *names;
Christoph Lametere498be72005-09-09 13:03:32 -07001217 int i;
Jack Steiner07ed76b2006-03-07 21:55:46 -08001218 int order;
Christoph Lametere498be72005-09-09 13:03:32 -07001219
1220 for (i = 0; i < NUM_INIT_LISTS; i++) {
1221 kmem_list3_init(&initkmem_list3[i]);
1222 if (i < MAX_NUMNODES)
1223 cache_cache.nodelists[i] = NULL;
1224 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001225
1226 /*
1227 * Fragmentation resistance on low memory - only use bigger
1228 * page orders on machines with more than 32MB of memory.
1229 */
1230 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1231 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1232
Linus Torvalds1da177e2005-04-16 15:20:36 -07001233 /* Bootstrap is tricky, because several objects are allocated
1234 * from caches that do not exist yet:
Andrew Mortona737b3e2006-03-22 00:08:11 -08001235 * 1) initialize the cache_cache cache: it contains the struct
1236 * kmem_cache structures of all caches, except cache_cache itself:
1237 * cache_cache is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001238 * Initially an __init data area is used for the head array and the
1239 * kmem_list3 structures, it's replaced with a kmalloc allocated
1240 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001241 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001242 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001243 * An __init data area is used for the head array.
1244 * 3) Create the remaining kmalloc caches, with minimally sized
1245 * head arrays.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001246 * 4) Replace the __init data head arrays for cache_cache and the first
1247 * kmalloc cache with kmalloc allocated arrays.
Christoph Lametere498be72005-09-09 13:03:32 -07001248 * 5) Replace the __init data for kmem_list3 for cache_cache and
1249 * the other cache's with kmalloc allocated memory.
1250 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001251 */
1252
1253 /* 1) create the cache_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001254 INIT_LIST_HEAD(&cache_chain);
1255 list_add(&cache_cache.next, &cache_chain);
1256 cache_cache.colour_off = cache_line_size();
1257 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
Christoph Lametere498be72005-09-09 13:03:32 -07001258 cache_cache.nodelists[numa_node_id()] = &initkmem_list3[CACHE_CACHE];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001259
Andrew Mortona737b3e2006-03-22 00:08:11 -08001260 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1261 cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001262
Jack Steiner07ed76b2006-03-07 21:55:46 -08001263 for (order = 0; order < MAX_ORDER; order++) {
1264 cache_estimate(order, cache_cache.buffer_size,
1265 cache_line_size(), 0, &left_over, &cache_cache.num);
1266 if (cache_cache.num)
1267 break;
1268 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001269 if (!cache_cache.num)
1270 BUG();
Jack Steiner07ed76b2006-03-07 21:55:46 -08001271 cache_cache.gfporder = order;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001272 cache_cache.colour = left_over / cache_cache.colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001273 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1274 sizeof(struct slab), cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001275
1276 /* 2+3) create the kmalloc caches */
1277 sizes = malloc_sizes;
1278 names = cache_names;
1279
Andrew Mortona737b3e2006-03-22 00:08:11 -08001280 /*
1281 * Initialize the caches that provide memory for the array cache and the
1282 * kmem_list3 structures first. Without this, further allocations will
1283 * bug.
Christoph Lametere498be72005-09-09 13:03:32 -07001284 */
1285
1286 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001287 sizes[INDEX_AC].cs_size,
1288 ARCH_KMALLOC_MINALIGN,
1289 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1290 NULL, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07001291
Andrew Mortona737b3e2006-03-22 00:08:11 -08001292 if (INDEX_AC != INDEX_L3) {
Christoph Lametere498be72005-09-09 13:03:32 -07001293 sizes[INDEX_L3].cs_cachep =
Andrew Mortona737b3e2006-03-22 00:08:11 -08001294 kmem_cache_create(names[INDEX_L3].name,
1295 sizes[INDEX_L3].cs_size,
1296 ARCH_KMALLOC_MINALIGN,
1297 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1298 NULL, NULL);
1299 }
Christoph Lametere498be72005-09-09 13:03:32 -07001300
Linus Torvalds1da177e2005-04-16 15:20:36 -07001301 while (sizes->cs_size != ULONG_MAX) {
Christoph Lametere498be72005-09-09 13:03:32 -07001302 /*
1303 * For performance, all the general caches are L1 aligned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001304 * This should be particularly beneficial on SMP boxes, as it
1305 * eliminates "false sharing".
1306 * Note for systems short on memory removing the alignment will
Christoph Lametere498be72005-09-09 13:03:32 -07001307 * allow tighter packing of the smaller caches.
1308 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001309 if (!sizes->cs_cachep) {
Christoph Lametere498be72005-09-09 13:03:32 -07001310 sizes->cs_cachep = kmem_cache_create(names->name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001311 sizes->cs_size,
1312 ARCH_KMALLOC_MINALIGN,
1313 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1314 NULL, NULL);
1315 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001316
1317 /* Inc off-slab bufctl limit until the ceiling is hit. */
1318 if (!(OFF_SLAB(sizes->cs_cachep))) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001319 offslab_limit = sizes->cs_size - sizeof(struct slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001320 offslab_limit /= sizeof(kmem_bufctl_t);
1321 }
1322
1323 sizes->cs_dmacachep = kmem_cache_create(names->name_dma,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001324 sizes->cs_size,
1325 ARCH_KMALLOC_MINALIGN,
1326 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1327 SLAB_PANIC,
1328 NULL, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001329 sizes++;
1330 names++;
1331 }
1332 /* 4) Replace the bootstrap head arrays */
1333 {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001334 void *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001335
Linus Torvalds1da177e2005-04-16 15:20:36 -07001336 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001337
Linus Torvalds1da177e2005-04-16 15:20:36 -07001338 local_irq_disable();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001339 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1340 memcpy(ptr, cpu_cache_get(&cache_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001341 sizeof(struct arraycache_init));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001342 cache_cache.array[smp_processor_id()] = ptr;
1343 local_irq_enable();
Christoph Lametere498be72005-09-09 13:03:32 -07001344
Linus Torvalds1da177e2005-04-16 15:20:36 -07001345 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001346
Linus Torvalds1da177e2005-04-16 15:20:36 -07001347 local_irq_disable();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001348 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001349 != &initarray_generic.cache);
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001350 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001351 sizeof(struct arraycache_init));
Christoph Lametere498be72005-09-09 13:03:32 -07001352 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001353 ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001354 local_irq_enable();
1355 }
Christoph Lametere498be72005-09-09 13:03:32 -07001356 /* 5) Replace the bootstrap kmem_list3's */
1357 {
1358 int node;
1359 /* Replace the static kmem_list3 structures for the boot cpu */
1360 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE],
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001361 numa_node_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001362
Christoph Lametere498be72005-09-09 13:03:32 -07001363 for_each_online_node(node) {
1364 init_list(malloc_sizes[INDEX_AC].cs_cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001365 &initkmem_list3[SIZE_AC + node], node);
Christoph Lametere498be72005-09-09 13:03:32 -07001366
1367 if (INDEX_AC != INDEX_L3) {
1368 init_list(malloc_sizes[INDEX_L3].cs_cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001369 &initkmem_list3[SIZE_L3 + node],
1370 node);
Christoph Lametere498be72005-09-09 13:03:32 -07001371 }
1372 }
1373 }
1374
1375 /* 6) resize the head arrays to their final sizes */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001376 {
Pekka Enberg343e0d72006-02-01 03:05:50 -08001377 struct kmem_cache *cachep;
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001378 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001379 list_for_each_entry(cachep, &cache_chain, next)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001380 enable_cpucache(cachep);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001381 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001382 }
1383
1384 /* Done! */
1385 g_cpucache_up = FULL;
1386
Andrew Mortona737b3e2006-03-22 00:08:11 -08001387 /*
1388 * Register a cpu startup notifier callback that initializes
1389 * cpu_cache_get for all new cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07001390 */
1391 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001392
Andrew Mortona737b3e2006-03-22 00:08:11 -08001393 /*
1394 * The reap timers are started later, with a module init call: That part
1395 * of the kernel is not yet operational.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001396 */
1397}
1398
1399static int __init cpucache_init(void)
1400{
1401 int cpu;
1402
Andrew Mortona737b3e2006-03-22 00:08:11 -08001403 /*
1404 * Register the timers that return unneeded pages to the page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -07001405 */
Christoph Lametere498be72005-09-09 13:03:32 -07001406 for_each_online_cpu(cpu)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001407 start_cpu_timer(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001408 return 0;
1409}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001410__initcall(cpucache_init);
1411
1412/*
1413 * Interface to system's page allocator. No need to hold the cache-lock.
1414 *
1415 * If we requested dmaable memory, we will get it. Even if we
1416 * did not request dmaable memory, we might get it, but that
1417 * would be relatively rare and ignorable.
1418 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001419static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001420{
1421 struct page *page;
1422 void *addr;
1423 int i;
1424
1425 flags |= cachep->gfpflags;
Christoph Lameter50c85a12005-11-13 16:06:47 -08001426 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001427 if (!page)
1428 return NULL;
1429 addr = page_address(page);
1430
1431 i = (1 << cachep->gfporder);
1432 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1433 atomic_add(i, &slab_reclaim_pages);
1434 add_page_state(nr_slab, i);
1435 while (i--) {
Nick Pigginf205b2f2006-03-22 00:08:02 -08001436 __SetPageSlab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001437 page++;
1438 }
1439 return addr;
1440}
1441
1442/*
1443 * Interface to system's page release.
1444 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001445static void kmem_freepages(struct kmem_cache *cachep, void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001446{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001447 unsigned long i = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001448 struct page *page = virt_to_page(addr);
1449 const unsigned long nr_freed = i;
1450
1451 while (i--) {
Nick Pigginf205b2f2006-03-22 00:08:02 -08001452 BUG_ON(!PageSlab(page));
1453 __ClearPageSlab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001454 page++;
1455 }
1456 sub_page_state(nr_slab, nr_freed);
1457 if (current->reclaim_state)
1458 current->reclaim_state->reclaimed_slab += nr_freed;
1459 free_pages((unsigned long)addr, cachep->gfporder);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001460 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1461 atomic_sub(1 << cachep->gfporder, &slab_reclaim_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001462}
1463
1464static void kmem_rcu_free(struct rcu_head *head)
1465{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001466 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001467 struct kmem_cache *cachep = slab_rcu->cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001468
1469 kmem_freepages(cachep, slab_rcu->addr);
1470 if (OFF_SLAB(cachep))
1471 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1472}
1473
1474#if DEBUG
1475
1476#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001477static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001478 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001479{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001480 int size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001481
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001482 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001483
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001484 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001485 return;
1486
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001487 *addr++ = 0x12345678;
1488 *addr++ = caller;
1489 *addr++ = smp_processor_id();
1490 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001491 {
1492 unsigned long *sptr = &caller;
1493 unsigned long svalue;
1494
1495 while (!kstack_end(sptr)) {
1496 svalue = *sptr++;
1497 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001498 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001499 size -= sizeof(unsigned long);
1500 if (size <= sizeof(unsigned long))
1501 break;
1502 }
1503 }
1504
1505 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001506 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001507}
1508#endif
1509
Pekka Enberg343e0d72006-02-01 03:05:50 -08001510static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001511{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001512 int size = obj_size(cachep);
1513 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001514
1515 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001516 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001517}
1518
1519static void dump_line(char *data, int offset, int limit)
1520{
1521 int i;
1522 printk(KERN_ERR "%03x:", offset);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001523 for (i = 0; i < limit; i++)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001524 printk(" %02x", (unsigned char)data[offset + i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001525 printk("\n");
1526}
1527#endif
1528
1529#if DEBUG
1530
Pekka Enberg343e0d72006-02-01 03:05:50 -08001531static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001532{
1533 int i, size;
1534 char *realobj;
1535
1536 if (cachep->flags & SLAB_RED_ZONE) {
1537 printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001538 *dbg_redzone1(cachep, objp),
1539 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001540 }
1541
1542 if (cachep->flags & SLAB_STORE_USER) {
1543 printk(KERN_ERR "Last user: [<%p>]",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001544 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001545 print_symbol("(%s)",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001546 (unsigned long)*dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001547 printk("\n");
1548 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001549 realobj = (char *)objp + obj_offset(cachep);
1550 size = obj_size(cachep);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001551 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001552 int limit;
1553 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001554 if (i + limit > size)
1555 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001556 dump_line(realobj, i, limit);
1557 }
1558}
1559
Pekka Enberg343e0d72006-02-01 03:05:50 -08001560static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001561{
1562 char *realobj;
1563 int size, i;
1564 int lines = 0;
1565
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001566 realobj = (char *)objp + obj_offset(cachep);
1567 size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001568
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001569 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001570 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001571 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001572 exp = POISON_END;
1573 if (realobj[i] != exp) {
1574 int limit;
1575 /* Mismatch ! */
1576 /* Print header */
1577 if (lines == 0) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001578 printk(KERN_ERR
Andrew Mortona737b3e2006-03-22 00:08:11 -08001579 "Slab corruption: start=%p, len=%d\n",
1580 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001581 print_objinfo(cachep, objp, 0);
1582 }
1583 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001584 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001585 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001586 if (i + limit > size)
1587 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001588 dump_line(realobj, i, limit);
1589 i += 16;
1590 lines++;
1591 /* Limit to 5 lines */
1592 if (lines > 5)
1593 break;
1594 }
1595 }
1596 if (lines != 0) {
1597 /* Print some data about the neighboring objects, if they
1598 * exist:
1599 */
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08001600 struct slab *slabp = virt_to_slab(objp);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001601 unsigned int objnr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001602
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001603 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001604 if (objnr) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001605 objp = index_to_obj(cachep, slabp, objnr - 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001606 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001607 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001608 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001609 print_objinfo(cachep, objp, 2);
1610 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001611 if (objnr + 1 < cachep->num) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001612 objp = index_to_obj(cachep, slabp, objnr + 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001613 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001614 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001615 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001616 print_objinfo(cachep, objp, 2);
1617 }
1618 }
1619}
1620#endif
1621
Linus Torvalds1da177e2005-04-16 15:20:36 -07001622#if DEBUG
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001623/**
Randy Dunlap911851e2006-03-22 00:08:14 -08001624 * slab_destroy_objs - destroy a slab and its objects
1625 * @cachep: cache pointer being destroyed
1626 * @slabp: slab pointer being destroyed
1627 *
1628 * Call the registered destructor for each object in a slab that is being
1629 * destroyed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001630 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001631static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001632{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001633 int i;
1634 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001635 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001636
1637 if (cachep->flags & SLAB_POISON) {
1638#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08001639 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1640 OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001641 kernel_map_pages(virt_to_page(objp),
Andrew Mortona737b3e2006-03-22 00:08:11 -08001642 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001643 else
1644 check_poison_obj(cachep, objp);
1645#else
1646 check_poison_obj(cachep, objp);
1647#endif
1648 }
1649 if (cachep->flags & SLAB_RED_ZONE) {
1650 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1651 slab_error(cachep, "start of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001652 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001653 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1654 slab_error(cachep, "end of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001655 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001656 }
1657 if (cachep->dtor && !(cachep->flags & SLAB_POISON))
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001658 (cachep->dtor) (objp + obj_offset(cachep), cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001659 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001660}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001661#else
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 if (cachep->dtor) {
1665 int i;
1666 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001667 void *objp = index_to_obj(cachep, slabp, i);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001668 (cachep->dtor) (objp, cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001669 }
1670 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001671}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001672#endif
1673
Randy Dunlap911851e2006-03-22 00:08:14 -08001674/**
1675 * slab_destroy - destroy and release all objects in a slab
1676 * @cachep: cache pointer being destroyed
1677 * @slabp: slab pointer being destroyed
1678 *
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001679 * Destroy all the objs in a slab, and release the mem back to the system.
Andrew Mortona737b3e2006-03-22 00:08:11 -08001680 * Before calling the slab must have been unlinked from the cache. The
1681 * cache-lock is not held/needed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001682 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001683static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001684{
1685 void *addr = slabp->s_mem - slabp->colouroff;
1686
1687 slab_destroy_objs(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001688 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1689 struct slab_rcu *slab_rcu;
1690
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001691 slab_rcu = (struct slab_rcu *)slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001692 slab_rcu->cachep = cachep;
1693 slab_rcu->addr = addr;
1694 call_rcu(&slab_rcu->head, kmem_rcu_free);
1695 } else {
1696 kmem_freepages(cachep, addr);
1697 if (OFF_SLAB(cachep))
1698 kmem_cache_free(cachep->slabp_cache, slabp);
1699 }
1700}
1701
Andrew Mortona737b3e2006-03-22 00:08:11 -08001702/*
1703 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1704 * size of kmem_list3.
1705 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001706static void set_up_list3s(struct kmem_cache *cachep, int index)
Christoph Lametere498be72005-09-09 13:03:32 -07001707{
1708 int node;
1709
1710 for_each_online_node(node) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001711 cachep->nodelists[node] = &initkmem_list3[index + node];
Christoph Lametere498be72005-09-09 13:03:32 -07001712 cachep->nodelists[node]->next_reap = jiffies +
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001713 REAPTIMEOUT_LIST3 +
1714 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -07001715 }
1716}
1717
Linus Torvalds1da177e2005-04-16 15:20:36 -07001718/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08001719 * calculate_slab_order - calculate size (page order) of slabs
1720 * @cachep: pointer to the cache that is being created
1721 * @size: size of objects to be created in this cache.
1722 * @align: required alignment for the objects.
1723 * @flags: slab allocation flags
1724 *
1725 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001726 *
1727 * This could be made much more intelligent. For now, try to avoid using
1728 * high order pages for slabs. When the gfp() functions are more friendly
1729 * towards high-order requests, this should be changed.
1730 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001731static size_t calculate_slab_order(struct kmem_cache *cachep,
Randy Dunlapee13d782006-02-01 03:05:53 -08001732 size_t size, size_t align, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001733{
1734 size_t left_over = 0;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001735 int gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001736
Andrew Mortona737b3e2006-03-22 00:08:11 -08001737 for (gfporder = 0; gfporder <= MAX_GFP_ORDER; gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001738 unsigned int num;
1739 size_t remainder;
1740
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001741 cache_estimate(gfporder, size, align, flags, &remainder, &num);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001742 if (!num)
1743 continue;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001744
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001745 /* More than offslab_limit objects will cause problems */
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001746 if ((flags & CFLGS_OFF_SLAB) && num > offslab_limit)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001747 break;
1748
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001749 /* Found something acceptable - save it away */
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001750 cachep->num = num;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001751 cachep->gfporder = gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001752 left_over = remainder;
1753
1754 /*
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08001755 * A VFS-reclaimable slab tends to have most allocations
1756 * as GFP_NOFS and we really don't want to have to be allocating
1757 * higher-order pages when we are unable to shrink dcache.
1758 */
1759 if (flags & SLAB_RECLAIM_ACCOUNT)
1760 break;
1761
1762 /*
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001763 * Large number of objects is good, but very large slabs are
1764 * currently bad for the gfp()s.
1765 */
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001766 if (gfporder >= slab_break_gfp_order)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001767 break;
1768
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001769 /*
1770 * Acceptable internal fragmentation?
1771 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001772 if (left_over * 8 <= (PAGE_SIZE << gfporder))
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001773 break;
1774 }
1775 return left_over;
1776}
1777
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08001778static void setup_cpu_cache(struct kmem_cache *cachep)
1779{
1780 if (g_cpucache_up == FULL) {
1781 enable_cpucache(cachep);
1782 return;
1783 }
1784 if (g_cpucache_up == NONE) {
1785 /*
1786 * Note: the first kmem_cache_create must create the cache
1787 * that's used by kmalloc(24), otherwise the creation of
1788 * further caches will BUG().
1789 */
1790 cachep->array[smp_processor_id()] = &initarray_generic.cache;
1791
1792 /*
1793 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
1794 * the first cache, then we need to set up all its list3s,
1795 * otherwise the creation of further caches will BUG().
1796 */
1797 set_up_list3s(cachep, SIZE_AC);
1798 if (INDEX_AC == INDEX_L3)
1799 g_cpucache_up = PARTIAL_L3;
1800 else
1801 g_cpucache_up = PARTIAL_AC;
1802 } else {
1803 cachep->array[smp_processor_id()] =
1804 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1805
1806 if (g_cpucache_up == PARTIAL_AC) {
1807 set_up_list3s(cachep, SIZE_L3);
1808 g_cpucache_up = PARTIAL_L3;
1809 } else {
1810 int node;
1811 for_each_online_node(node) {
1812 cachep->nodelists[node] =
1813 kmalloc_node(sizeof(struct kmem_list3),
1814 GFP_KERNEL, node);
1815 BUG_ON(!cachep->nodelists[node]);
1816 kmem_list3_init(cachep->nodelists[node]);
1817 }
1818 }
1819 }
1820 cachep->nodelists[numa_node_id()]->next_reap =
1821 jiffies + REAPTIMEOUT_LIST3 +
1822 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1823
1824 cpu_cache_get(cachep)->avail = 0;
1825 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1826 cpu_cache_get(cachep)->batchcount = 1;
1827 cpu_cache_get(cachep)->touched = 0;
1828 cachep->batchcount = 1;
1829 cachep->limit = BOOT_CPUCACHE_ENTRIES;
1830}
1831
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001832/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07001833 * kmem_cache_create - Create a cache.
1834 * @name: A string which is used in /proc/slabinfo to identify this cache.
1835 * @size: The size of objects to be created in this cache.
1836 * @align: The required alignment for the objects.
1837 * @flags: SLAB flags
1838 * @ctor: A constructor for the objects.
1839 * @dtor: A destructor for the objects.
1840 *
1841 * Returns a ptr to the cache on success, NULL on failure.
1842 * Cannot be called within a int, but can be interrupted.
1843 * The @ctor is run when new pages are allocated by the cache
1844 * and the @dtor is run before the pages are handed back.
1845 *
1846 * @name must be valid until the cache is destroyed. This implies that
Andrew Mortona737b3e2006-03-22 00:08:11 -08001847 * the module calling this has to destroy the cache before getting unloaded.
1848 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001849 * The flags are
1850 *
1851 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1852 * to catch references to uninitialised memory.
1853 *
1854 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1855 * for buffer overruns.
1856 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001857 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1858 * cacheline. This can be beneficial if you're counting cycles as closely
1859 * as davem.
1860 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001861struct kmem_cache *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001862kmem_cache_create (const char *name, size_t size, size_t align,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001863 unsigned long flags,
1864 void (*ctor)(void*, struct kmem_cache *, unsigned long),
Pekka Enberg343e0d72006-02-01 03:05:50 -08001865 void (*dtor)(void*, struct kmem_cache *, unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001866{
1867 size_t left_over, slab_size, ralign;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001868 struct kmem_cache *cachep = NULL;
Andrew Morton4f12bb42005-11-07 00:58:00 -08001869 struct list_head *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001870
1871 /*
1872 * Sanity checks... these are all serious usage bugs.
1873 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001874 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001875 (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08001876 printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__,
1877 name);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001878 BUG();
1879 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001880
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08001881 /*
1882 * Prevent CPUs from coming and going.
1883 * lock_cpu_hotplug() nests outside cache_chain_mutex
1884 */
1885 lock_cpu_hotplug();
1886
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001887 mutex_lock(&cache_chain_mutex);
Andrew Morton4f12bb42005-11-07 00:58:00 -08001888
1889 list_for_each(p, &cache_chain) {
Pekka Enberg343e0d72006-02-01 03:05:50 -08001890 struct kmem_cache *pc = list_entry(p, struct kmem_cache, next);
Andrew Morton4f12bb42005-11-07 00:58:00 -08001891 mm_segment_t old_fs = get_fs();
1892 char tmp;
1893 int res;
1894
1895 /*
1896 * This happens when the module gets unloaded and doesn't
1897 * destroy its slab cache and no-one else reuses the vmalloc
1898 * area of the module. Print a warning.
1899 */
1900 set_fs(KERNEL_DS);
1901 res = __get_user(tmp, pc->name);
1902 set_fs(old_fs);
1903 if (res) {
1904 printk("SLAB: cache with size %d has lost its name\n",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001905 pc->buffer_size);
Andrew Morton4f12bb42005-11-07 00:58:00 -08001906 continue;
1907 }
1908
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001909 if (!strcmp(pc->name, name)) {
Andrew Morton4f12bb42005-11-07 00:58:00 -08001910 printk("kmem_cache_create: duplicate cache %s\n", name);
1911 dump_stack();
1912 goto oops;
1913 }
1914 }
1915
Linus Torvalds1da177e2005-04-16 15:20:36 -07001916#if DEBUG
1917 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
1918 if ((flags & SLAB_DEBUG_INITIAL) && !ctor) {
1919 /* No constructor, but inital state check requested */
1920 printk(KERN_ERR "%s: No con, but init state check "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001921 "requested - %s\n", __FUNCTION__, name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001922 flags &= ~SLAB_DEBUG_INITIAL;
1923 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001924#if FORCED_DEBUG
1925 /*
1926 * Enable redzoning and last user accounting, except for caches with
1927 * large objects, if the increased size would increase the object size
1928 * above the next power of two: caches with object sizes just above a
1929 * power of two have a significant amount of internal fragmentation.
1930 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001931 if (size < 4096 || fls(size - 1) == fls(size-1 + 3 * BYTES_PER_WORD))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001932 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001933 if (!(flags & SLAB_DESTROY_BY_RCU))
1934 flags |= SLAB_POISON;
1935#endif
1936 if (flags & SLAB_DESTROY_BY_RCU)
1937 BUG_ON(flags & SLAB_POISON);
1938#endif
1939 if (flags & SLAB_DESTROY_BY_RCU)
1940 BUG_ON(dtor);
1941
1942 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08001943 * Always checks flags, a caller might be expecting debug support which
1944 * isn't available.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001945 */
1946 if (flags & ~CREATE_MASK)
1947 BUG();
1948
Andrew Mortona737b3e2006-03-22 00:08:11 -08001949 /*
1950 * Check that size is in terms of words. This is needed to avoid
Linus Torvalds1da177e2005-04-16 15:20:36 -07001951 * unaligned accesses for some archs when redzoning is used, and makes
1952 * sure any on-slab bufctl's are also correctly aligned.
1953 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001954 if (size & (BYTES_PER_WORD - 1)) {
1955 size += (BYTES_PER_WORD - 1);
1956 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001957 }
1958
Andrew Mortona737b3e2006-03-22 00:08:11 -08001959 /* calculate the final buffer alignment: */
1960
Linus Torvalds1da177e2005-04-16 15:20:36 -07001961 /* 1) arch recommendation: can be overridden for debug */
1962 if (flags & SLAB_HWCACHE_ALIGN) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08001963 /*
1964 * Default alignment: as specified by the arch code. Except if
1965 * an object is really small, then squeeze multiple objects into
1966 * one cacheline.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001967 */
1968 ralign = cache_line_size();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001969 while (size <= ralign / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001970 ralign /= 2;
1971 } else {
1972 ralign = BYTES_PER_WORD;
1973 }
1974 /* 2) arch mandated alignment: disables debug if necessary */
1975 if (ralign < ARCH_SLAB_MINALIGN) {
1976 ralign = ARCH_SLAB_MINALIGN;
1977 if (ralign > BYTES_PER_WORD)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001978 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001979 }
1980 /* 3) caller mandated alignment: disables debug if necessary */
1981 if (ralign < align) {
1982 ralign = align;
1983 if (ralign > BYTES_PER_WORD)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001984 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001985 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08001986 /*
1987 * 4) Store it. Note that the debug code below can reduce
Linus Torvalds1da177e2005-04-16 15:20:36 -07001988 * the alignment to BYTES_PER_WORD.
1989 */
1990 align = ralign;
1991
1992 /* Get cache's description obj. */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001993 cachep = kmem_cache_alloc(&cache_cache, SLAB_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001994 if (!cachep)
Andrew Morton4f12bb42005-11-07 00:58:00 -08001995 goto oops;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001996 memset(cachep, 0, sizeof(struct kmem_cache));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001997
1998#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001999 cachep->obj_size = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002000
2001 if (flags & SLAB_RED_ZONE) {
2002 /* redzoning only works with word aligned caches */
2003 align = BYTES_PER_WORD;
2004
2005 /* add space for red zone words */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002006 cachep->obj_offset += BYTES_PER_WORD;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002007 size += 2 * BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002008 }
2009 if (flags & SLAB_STORE_USER) {
2010 /* user store requires word alignment and
2011 * one word storage behind the end of the real
2012 * object.
2013 */
2014 align = BYTES_PER_WORD;
2015 size += BYTES_PER_WORD;
2016 }
2017#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002018 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002019 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2020 cachep->obj_offset += PAGE_SIZE - size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002021 size = PAGE_SIZE;
2022 }
2023#endif
2024#endif
2025
2026 /* Determine if the slab management is 'on' or 'off' slab. */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002027 if (size >= (PAGE_SIZE >> 3))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002028 /*
2029 * Size is large, assume best to place the slab management obj
2030 * off-slab (should allow better packing of objs).
2031 */
2032 flags |= CFLGS_OFF_SLAB;
2033
2034 size = ALIGN(size, align);
2035
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002036 left_over = calculate_slab_order(cachep, size, align, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002037
2038 if (!cachep->num) {
2039 printk("kmem_cache_create: couldn't create cache %s.\n", name);
2040 kmem_cache_free(&cache_cache, cachep);
2041 cachep = NULL;
Andrew Morton4f12bb42005-11-07 00:58:00 -08002042 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002043 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002044 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2045 + sizeof(struct slab), align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002046
2047 /*
2048 * If the slab has been placed off-slab, and we have enough space then
2049 * move it on-slab. This is at the expense of any extra colouring.
2050 */
2051 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2052 flags &= ~CFLGS_OFF_SLAB;
2053 left_over -= slab_size;
2054 }
2055
2056 if (flags & CFLGS_OFF_SLAB) {
2057 /* really off slab. No need for manual alignment */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002058 slab_size =
2059 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002060 }
2061
2062 cachep->colour_off = cache_line_size();
2063 /* Offset must be a multiple of the alignment. */
2064 if (cachep->colour_off < align)
2065 cachep->colour_off = align;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002066 cachep->colour = left_over / cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002067 cachep->slab_size = slab_size;
2068 cachep->flags = flags;
2069 cachep->gfpflags = 0;
2070 if (flags & SLAB_CACHE_DMA)
2071 cachep->gfpflags |= GFP_DMA;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002072 cachep->buffer_size = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002073
2074 if (flags & CFLGS_OFF_SLAB)
Victor Fuscob2d55072005-09-10 00:26:36 -07002075 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002076 cachep->ctor = ctor;
2077 cachep->dtor = dtor;
2078 cachep->name = name;
2079
Linus Torvalds1da177e2005-04-16 15:20:36 -07002080
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002081 setup_cpu_cache(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002082
Linus Torvalds1da177e2005-04-16 15:20:36 -07002083 /* cache setup completed, link it into the list */
2084 list_add(&cachep->next, &cache_chain);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002085oops:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002086 if (!cachep && (flags & SLAB_PANIC))
2087 panic("kmem_cache_create(): failed to create slab `%s'\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002088 name);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002089 mutex_unlock(&cache_chain_mutex);
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002090 unlock_cpu_hotplug();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002091 return cachep;
2092}
2093EXPORT_SYMBOL(kmem_cache_create);
2094
2095#if DEBUG
2096static void check_irq_off(void)
2097{
2098 BUG_ON(!irqs_disabled());
2099}
2100
2101static void check_irq_on(void)
2102{
2103 BUG_ON(irqs_disabled());
2104}
2105
Pekka Enberg343e0d72006-02-01 03:05:50 -08002106static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002107{
2108#ifdef CONFIG_SMP
2109 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002110 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002111#endif
2112}
Christoph Lametere498be72005-09-09 13:03:32 -07002113
Pekka Enberg343e0d72006-02-01 03:05:50 -08002114static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002115{
2116#ifdef CONFIG_SMP
2117 check_irq_off();
2118 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2119#endif
2120}
2121
Linus Torvalds1da177e2005-04-16 15:20:36 -07002122#else
2123#define check_irq_off() do { } while(0)
2124#define check_irq_on() do { } while(0)
2125#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002126#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002127#endif
2128
Christoph Lameteraab22072006-03-22 00:09:06 -08002129static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2130 struct array_cache *ac,
2131 int force, int node);
2132
Linus Torvalds1da177e2005-04-16 15:20:36 -07002133static void do_drain(void *arg)
2134{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002135 struct kmem_cache *cachep = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002136 struct array_cache *ac;
Christoph Lameterff694162005-09-22 21:44:02 -07002137 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002138
2139 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002140 ac = cpu_cache_get(cachep);
Christoph Lameterff694162005-09-22 21:44:02 -07002141 spin_lock(&cachep->nodelists[node]->list_lock);
2142 free_block(cachep, ac->entry, ac->avail, node);
2143 spin_unlock(&cachep->nodelists[node]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002144 ac->avail = 0;
2145}
2146
Pekka Enberg343e0d72006-02-01 03:05:50 -08002147static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002148{
Christoph Lametere498be72005-09-09 13:03:32 -07002149 struct kmem_list3 *l3;
2150 int node;
2151
Andrew Mortona07fa392006-03-22 00:08:17 -08002152 on_each_cpu(do_drain, cachep, 1, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002153 check_irq_on();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002154 for_each_online_node(node) {
Christoph Lametere498be72005-09-09 13:03:32 -07002155 l3 = cachep->nodelists[node];
2156 if (l3) {
Christoph Lameteraab22072006-03-22 00:09:06 -08002157 drain_array(cachep, l3, l3->shared, 1, node);
Christoph Lametere498be72005-09-09 13:03:32 -07002158 if (l3->alien)
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08002159 drain_alien_cache(cachep, l3->alien);
Christoph Lametere498be72005-09-09 13:03:32 -07002160 }
2161 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002162}
2163
Pekka Enberg343e0d72006-02-01 03:05:50 -08002164static int __node_shrink(struct kmem_cache *cachep, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002165{
2166 struct slab *slabp;
Christoph Lametere498be72005-09-09 13:03:32 -07002167 struct kmem_list3 *l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002168 int ret;
2169
Christoph Lametere498be72005-09-09 13:03:32 -07002170 for (;;) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002171 struct list_head *p;
2172
Christoph Lametere498be72005-09-09 13:03:32 -07002173 p = l3->slabs_free.prev;
2174 if (p == &l3->slabs_free)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002175 break;
2176
Christoph Lametere498be72005-09-09 13:03:32 -07002177 slabp = list_entry(l3->slabs_free.prev, struct slab, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002178#if DEBUG
2179 if (slabp->inuse)
2180 BUG();
2181#endif
2182 list_del(&slabp->list);
2183
Christoph Lametere498be72005-09-09 13:03:32 -07002184 l3->free_objects -= cachep->num;
2185 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002186 slab_destroy(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07002187 spin_lock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002188 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002189 ret = !list_empty(&l3->slabs_full) || !list_empty(&l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002190 return ret;
2191}
2192
Pekka Enberg343e0d72006-02-01 03:05:50 -08002193static int __cache_shrink(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07002194{
2195 int ret = 0, i = 0;
2196 struct kmem_list3 *l3;
2197
2198 drain_cpu_caches(cachep);
2199
2200 check_irq_on();
2201 for_each_online_node(i) {
2202 l3 = cachep->nodelists[i];
2203 if (l3) {
2204 spin_lock_irq(&l3->list_lock);
2205 ret += __node_shrink(cachep, i);
2206 spin_unlock_irq(&l3->list_lock);
2207 }
2208 }
2209 return (ret ? 1 : 0);
2210}
2211
Linus Torvalds1da177e2005-04-16 15:20:36 -07002212/**
2213 * kmem_cache_shrink - Shrink a cache.
2214 * @cachep: The cache to shrink.
2215 *
2216 * Releases as many slabs as possible for a cache.
2217 * To help debugging, a zero exit status indicates all slabs were released.
2218 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002219int kmem_cache_shrink(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002220{
2221 if (!cachep || in_interrupt())
2222 BUG();
2223
2224 return __cache_shrink(cachep);
2225}
2226EXPORT_SYMBOL(kmem_cache_shrink);
2227
2228/**
2229 * kmem_cache_destroy - delete a cache
2230 * @cachep: the cache to destroy
2231 *
Pekka Enberg343e0d72006-02-01 03:05:50 -08002232 * Remove a struct kmem_cache object from the slab cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002233 * Returns 0 on success.
2234 *
2235 * It is expected this function will be called by a module when it is
2236 * unloaded. This will remove the cache completely, and avoid a duplicate
2237 * cache being allocated each time a module is loaded and unloaded, if the
2238 * module doesn't have persistent in-kernel storage across loads and unloads.
2239 *
2240 * The cache must be empty before calling this function.
2241 *
2242 * The caller must guarantee that noone will allocate memory from the cache
2243 * during the kmem_cache_destroy().
2244 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002245int kmem_cache_destroy(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002246{
2247 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07002248 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002249
2250 if (!cachep || in_interrupt())
2251 BUG();
2252
2253 /* Don't let CPUs to come and go */
2254 lock_cpu_hotplug();
2255
2256 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002257 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002258 /*
2259 * the chain is never empty, cache_cache is never destroyed
2260 */
2261 list_del(&cachep->next);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002262 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002263
2264 if (__cache_shrink(cachep)) {
2265 slab_error(cachep, "Can't free all objects");
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002266 mutex_lock(&cache_chain_mutex);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002267 list_add(&cachep->next, &cache_chain);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002268 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002269 unlock_cpu_hotplug();
2270 return 1;
2271 }
2272
2273 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
Paul E. McKenneyfbd568a3e2005-05-01 08:59:04 -07002274 synchronize_rcu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002275
Christoph Lametere498be72005-09-09 13:03:32 -07002276 for_each_online_cpu(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002277 kfree(cachep->array[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002278
2279 /* NUMA: free the list3 structures */
Christoph Lametere498be72005-09-09 13:03:32 -07002280 for_each_online_node(i) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002281 l3 = cachep->nodelists[i];
2282 if (l3) {
Christoph Lametere498be72005-09-09 13:03:32 -07002283 kfree(l3->shared);
2284 free_alien_cache(l3->alien);
2285 kfree(l3);
2286 }
2287 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002288 kmem_cache_free(&cache_cache, cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002289 unlock_cpu_hotplug();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002290 return 0;
2291}
2292EXPORT_SYMBOL(kmem_cache_destroy);
2293
2294/* Get the memory for a slab management obj. */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002295static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002296 int colour_off, gfp_t local_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002297{
2298 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002299
Linus Torvalds1da177e2005-04-16 15:20:36 -07002300 if (OFF_SLAB(cachep)) {
2301 /* Slab management obj is off-slab. */
2302 slabp = kmem_cache_alloc(cachep->slabp_cache, local_flags);
2303 if (!slabp)
2304 return NULL;
2305 } else {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002306 slabp = objp + colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002307 colour_off += cachep->slab_size;
2308 }
2309 slabp->inuse = 0;
2310 slabp->colouroff = colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002311 slabp->s_mem = objp + colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002312 return slabp;
2313}
2314
2315static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2316{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002317 return (kmem_bufctl_t *) (slabp + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002318}
2319
Pekka Enberg343e0d72006-02-01 03:05:50 -08002320static void cache_init_objs(struct kmem_cache *cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002321 struct slab *slabp, unsigned long ctor_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002322{
2323 int i;
2324
2325 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002326 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002327#if DEBUG
2328 /* need to poison the objs? */
2329 if (cachep->flags & SLAB_POISON)
2330 poison_obj(cachep, objp, POISON_FREE);
2331 if (cachep->flags & SLAB_STORE_USER)
2332 *dbg_userword(cachep, objp) = NULL;
2333
2334 if (cachep->flags & SLAB_RED_ZONE) {
2335 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2336 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2337 }
2338 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002339 * Constructors are not allowed to allocate memory from the same
2340 * cache which they are a constructor for. Otherwise, deadlock.
2341 * They must also be threaded.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002342 */
2343 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002344 cachep->ctor(objp + obj_offset(cachep), cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002345 ctor_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002346
2347 if (cachep->flags & SLAB_RED_ZONE) {
2348 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2349 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002350 " end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002351 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2352 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002353 " start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002354 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08002355 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2356 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002357 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002358 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002359#else
2360 if (cachep->ctor)
2361 cachep->ctor(objp, cachep, ctor_flags);
2362#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002363 slab_bufctl(slabp)[i] = i + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002364 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002365 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002366 slabp->free = 0;
2367}
2368
Pekka Enberg343e0d72006-02-01 03:05:50 -08002369static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002370{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002371 if (flags & SLAB_DMA)
2372 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2373 else
2374 BUG_ON(cachep->gfpflags & GFP_DMA);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002375}
2376
Andrew Mortona737b3e2006-03-22 00:08:11 -08002377static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2378 int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002379{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002380 void *objp = index_to_obj(cachep, slabp, slabp->free);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002381 kmem_bufctl_t next;
2382
2383 slabp->inuse++;
2384 next = slab_bufctl(slabp)[slabp->free];
2385#if DEBUG
2386 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2387 WARN_ON(slabp->nodeid != nodeid);
2388#endif
2389 slabp->free = next;
2390
2391 return objp;
2392}
2393
Andrew Mortona737b3e2006-03-22 00:08:11 -08002394static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2395 void *objp, int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002396{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002397 unsigned int objnr = obj_to_index(cachep, slabp, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002398
2399#if DEBUG
2400 /* Verify that the slab belongs to the intended node */
2401 WARN_ON(slabp->nodeid != nodeid);
2402
Al Viro871751e2006-03-25 03:06:39 -08002403 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
Matthew Dobson78d382d2006-02-01 03:05:47 -08002404 printk(KERN_ERR "slab: double free detected in cache "
Andrew Mortona737b3e2006-03-22 00:08:11 -08002405 "'%s', objp %p\n", cachep->name, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002406 BUG();
2407 }
2408#endif
2409 slab_bufctl(slabp)[objnr] = slabp->free;
2410 slabp->free = objnr;
2411 slabp->inuse--;
2412}
2413
Andrew Mortona737b3e2006-03-22 00:08:11 -08002414static void set_slab_attr(struct kmem_cache *cachep, struct slab *slabp,
2415 void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002416{
2417 int i;
2418 struct page *page;
2419
2420 /* Nasty!!!!!! I hope this is OK. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002421 page = virt_to_page(objp);
Nick Piggin84097512006-03-22 00:08:34 -08002422
2423 i = 1;
2424 if (likely(!PageCompound(page)))
2425 i <<= cachep->gfporder;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002426 do {
Pekka Enberg065d41c2005-11-13 16:06:46 -08002427 page_set_cache(page, cachep);
2428 page_set_slab(page, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002429 page++;
2430 } while (--i);
2431}
2432
2433/*
2434 * Grow (by 1) the number of slabs within a cache. This is called by
2435 * kmem_cache_alloc() when there are no active objs left in a cache.
2436 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002437static int cache_grow(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002438{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002439 struct slab *slabp;
2440 void *objp;
2441 size_t offset;
2442 gfp_t local_flags;
2443 unsigned long ctor_flags;
Christoph Lametere498be72005-09-09 13:03:32 -07002444 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002445
Andrew Mortona737b3e2006-03-22 00:08:11 -08002446 /*
2447 * Be lazy and only check for valid flags here, keeping it out of the
2448 * critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002449 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002450 if (flags & ~(SLAB_DMA | SLAB_LEVEL_MASK | SLAB_NO_GROW))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002451 BUG();
2452 if (flags & SLAB_NO_GROW)
2453 return 0;
2454
2455 ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2456 local_flags = (flags & SLAB_LEVEL_MASK);
2457 if (!(local_flags & __GFP_WAIT))
2458 /*
2459 * Not allowed to sleep. Need to tell a constructor about
2460 * this - it might need to know...
2461 */
2462 ctor_flags |= SLAB_CTOR_ATOMIC;
2463
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002464 /* Take the l3 list lock to change the colour_next on this node */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002465 check_irq_off();
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002466 l3 = cachep->nodelists[nodeid];
2467 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002468
2469 /* Get colour for the slab, and cal the next value. */
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002470 offset = l3->colour_next;
2471 l3->colour_next++;
2472 if (l3->colour_next >= cachep->colour)
2473 l3->colour_next = 0;
2474 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002475
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002476 offset *= cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002477
2478 if (local_flags & __GFP_WAIT)
2479 local_irq_enable();
2480
2481 /*
2482 * The test for missing atomic flag is performed here, rather than
2483 * the more obvious place, simply to reduce the critical path length
2484 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2485 * will eventually be caught here (where it matters).
2486 */
2487 kmem_flagcheck(cachep, flags);
2488
Andrew Mortona737b3e2006-03-22 00:08:11 -08002489 /*
2490 * Get mem for the objs. Attempt to allocate a physical page from
2491 * 'nodeid'.
Christoph Lametere498be72005-09-09 13:03:32 -07002492 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002493 objp = kmem_getpages(cachep, flags, nodeid);
2494 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002495 goto failed;
2496
2497 /* Get slab management. */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002498 slabp = alloc_slabmgmt(cachep, objp, offset, local_flags);
2499 if (!slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002500 goto opps1;
2501
Christoph Lametere498be72005-09-09 13:03:32 -07002502 slabp->nodeid = nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002503 set_slab_attr(cachep, slabp, objp);
2504
2505 cache_init_objs(cachep, slabp, ctor_flags);
2506
2507 if (local_flags & __GFP_WAIT)
2508 local_irq_disable();
2509 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002510 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002511
2512 /* Make slab active. */
Christoph Lametere498be72005-09-09 13:03:32 -07002513 list_add_tail(&slabp->list, &(l3->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002514 STATS_INC_GROWN(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002515 l3->free_objects += cachep->num;
2516 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002517 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002518opps1:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002519 kmem_freepages(cachep, objp);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002520failed:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002521 if (local_flags & __GFP_WAIT)
2522 local_irq_disable();
2523 return 0;
2524}
2525
2526#if DEBUG
2527
2528/*
2529 * Perform extra freeing checks:
2530 * - detect bad pointers.
2531 * - POISON/RED_ZONE checking
2532 * - destructor calls, for caches with POISON+dtor
2533 */
2534static void kfree_debugcheck(const void *objp)
2535{
2536 struct page *page;
2537
2538 if (!virt_addr_valid(objp)) {
2539 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002540 (unsigned long)objp);
2541 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002542 }
2543 page = virt_to_page(objp);
2544 if (!PageSlab(page)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002545 printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n",
2546 (unsigned long)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002547 BUG();
2548 }
2549}
2550
Pekka Enberg343e0d72006-02-01 03:05:50 -08002551static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002552 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002553{
2554 struct page *page;
2555 unsigned int objnr;
2556 struct slab *slabp;
2557
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002558 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002559 kfree_debugcheck(objp);
2560 page = virt_to_page(objp);
2561
Pekka Enberg065d41c2005-11-13 16:06:46 -08002562 if (page_get_cache(page) != cachep) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002563 printk(KERN_ERR "mismatch in kmem_cache_free: expected "
2564 "cache %p, got %p\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002565 page_get_cache(page), cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002566 printk(KERN_ERR "%p is %s.\n", cachep, cachep->name);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002567 printk(KERN_ERR "%p is %s.\n", page_get_cache(page),
2568 page_get_cache(page)->name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002569 WARN_ON(1);
2570 }
Pekka Enberg065d41c2005-11-13 16:06:46 -08002571 slabp = page_get_slab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002572
2573 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002574 if (*dbg_redzone1(cachep, objp) != RED_ACTIVE ||
2575 *dbg_redzone2(cachep, objp) != RED_ACTIVE) {
2576 slab_error(cachep, "double free, or memory outside"
2577 " object was overwritten");
2578 printk(KERN_ERR "%p: redzone 1:0x%lx, "
2579 "redzone 2:0x%lx.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002580 objp, *dbg_redzone1(cachep, objp),
2581 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002582 }
2583 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2584 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2585 }
2586 if (cachep->flags & SLAB_STORE_USER)
2587 *dbg_userword(cachep, objp) = caller;
2588
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002589 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002590
2591 BUG_ON(objnr >= cachep->num);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002592 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002593
2594 if (cachep->flags & SLAB_DEBUG_INITIAL) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002595 /*
2596 * Need to call the slab's constructor so the caller can
2597 * perform a verify of its state (debugging). Called without
2598 * the cache-lock held.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002599 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002600 cachep->ctor(objp + obj_offset(cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002601 cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002602 }
2603 if (cachep->flags & SLAB_POISON && cachep->dtor) {
2604 /* we want to cache poison the object,
2605 * call the destruction callback
2606 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002607 cachep->dtor(objp + obj_offset(cachep), cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002608 }
Al Viro871751e2006-03-25 03:06:39 -08002609#ifdef CONFIG_DEBUG_SLAB_LEAK
2610 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2611#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002612 if (cachep->flags & SLAB_POISON) {
2613#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08002614 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002615 store_stackinfo(cachep, objp, (unsigned long)caller);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002616 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002617 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002618 } else {
2619 poison_obj(cachep, objp, POISON_FREE);
2620 }
2621#else
2622 poison_obj(cachep, objp, POISON_FREE);
2623#endif
2624 }
2625 return objp;
2626}
2627
Pekka Enberg343e0d72006-02-01 03:05:50 -08002628static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002629{
2630 kmem_bufctl_t i;
2631 int entries = 0;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002632
Linus Torvalds1da177e2005-04-16 15:20:36 -07002633 /* Check slab's freelist to see if this obj is there. */
2634 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2635 entries++;
2636 if (entries > cachep->num || i >= cachep->num)
2637 goto bad;
2638 }
2639 if (entries != cachep->num - slabp->inuse) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002640bad:
2641 printk(KERN_ERR "slab: Internal list corruption detected in "
2642 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2643 cachep->name, cachep->num, slabp, slabp->inuse);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002644 for (i = 0;
Linus Torvalds264132b2006-03-06 12:10:07 -08002645 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002646 i++) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002647 if (i % 16 == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002648 printk("\n%03x:", i);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002649 printk(" %02x", ((unsigned char *)slabp)[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002650 }
2651 printk("\n");
2652 BUG();
2653 }
2654}
2655#else
2656#define kfree_debugcheck(x) do { } while(0)
2657#define cache_free_debugcheck(x,objp,z) (objp)
2658#define check_slabp(x,y) do { } while(0)
2659#endif
2660
Pekka Enberg343e0d72006-02-01 03:05:50 -08002661static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002662{
2663 int batchcount;
2664 struct kmem_list3 *l3;
2665 struct array_cache *ac;
2666
2667 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002668 ac = cpu_cache_get(cachep);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002669retry:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002670 batchcount = ac->batchcount;
2671 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002672 /*
2673 * If there was little recent activity on this cache, then
2674 * perform only a partial refill. Otherwise we could generate
2675 * refill bouncing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002676 */
2677 batchcount = BATCHREFILL_LIMIT;
2678 }
Christoph Lametere498be72005-09-09 13:03:32 -07002679 l3 = cachep->nodelists[numa_node_id()];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002680
Christoph Lametere498be72005-09-09 13:03:32 -07002681 BUG_ON(ac->avail > 0 || !l3);
2682 spin_lock(&l3->list_lock);
2683
Linus Torvalds1da177e2005-04-16 15:20:36 -07002684 if (l3->shared) {
2685 struct array_cache *shared_array = l3->shared;
2686 if (shared_array->avail) {
2687 if (batchcount > shared_array->avail)
2688 batchcount = shared_array->avail;
2689 shared_array->avail -= batchcount;
2690 ac->avail = batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07002691 memcpy(ac->entry,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002692 &(shared_array->entry[shared_array->avail]),
2693 sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002694 shared_array->touched = 1;
2695 goto alloc_done;
2696 }
2697 }
2698 while (batchcount > 0) {
2699 struct list_head *entry;
2700 struct slab *slabp;
2701 /* Get slab alloc is to come from. */
2702 entry = l3->slabs_partial.next;
2703 if (entry == &l3->slabs_partial) {
2704 l3->free_touched = 1;
2705 entry = l3->slabs_free.next;
2706 if (entry == &l3->slabs_free)
2707 goto must_grow;
2708 }
2709
2710 slabp = list_entry(entry, struct slab, list);
2711 check_slabp(cachep, slabp);
2712 check_spinlock_acquired(cachep);
2713 while (slabp->inuse < cachep->num && batchcount--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002714 STATS_INC_ALLOCED(cachep);
2715 STATS_INC_ACTIVE(cachep);
2716 STATS_SET_HIGH(cachep);
2717
Matthew Dobson78d382d2006-02-01 03:05:47 -08002718 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
2719 numa_node_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002720 }
2721 check_slabp(cachep, slabp);
2722
2723 /* move slabp to correct slabp list: */
2724 list_del(&slabp->list);
2725 if (slabp->free == BUFCTL_END)
2726 list_add(&slabp->list, &l3->slabs_full);
2727 else
2728 list_add(&slabp->list, &l3->slabs_partial);
2729 }
2730
Andrew Mortona737b3e2006-03-22 00:08:11 -08002731must_grow:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002732 l3->free_objects -= ac->avail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002733alloc_done:
Christoph Lametere498be72005-09-09 13:03:32 -07002734 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002735
2736 if (unlikely(!ac->avail)) {
2737 int x;
Christoph Lametere498be72005-09-09 13:03:32 -07002738 x = cache_grow(cachep, flags, numa_node_id());
2739
Andrew Mortona737b3e2006-03-22 00:08:11 -08002740 /* cache_grow can reenable interrupts, then ac could change. */
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002741 ac = cpu_cache_get(cachep);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002742 if (!x && ac->avail == 0) /* no objects in sight? abort */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002743 return NULL;
2744
Andrew Mortona737b3e2006-03-22 00:08:11 -08002745 if (!ac->avail) /* objects refilled by interrupt? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002746 goto retry;
2747 }
2748 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07002749 return ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002750}
2751
Andrew Mortona737b3e2006-03-22 00:08:11 -08002752static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
2753 gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002754{
2755 might_sleep_if(flags & __GFP_WAIT);
2756#if DEBUG
2757 kmem_flagcheck(cachep, flags);
2758#endif
2759}
2760
2761#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08002762static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
2763 gfp_t flags, void *objp, void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002764{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002765 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002766 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002767 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002768#ifdef CONFIG_DEBUG_PAGEALLOC
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002769 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002770 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002771 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002772 else
2773 check_poison_obj(cachep, objp);
2774#else
2775 check_poison_obj(cachep, objp);
2776#endif
2777 poison_obj(cachep, objp, POISON_INUSE);
2778 }
2779 if (cachep->flags & SLAB_STORE_USER)
2780 *dbg_userword(cachep, objp) = caller;
2781
2782 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002783 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
2784 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2785 slab_error(cachep, "double free, or memory outside"
2786 " object was overwritten");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002787 printk(KERN_ERR
Andrew Mortona737b3e2006-03-22 00:08:11 -08002788 "%p: redzone 1:0x%lx, redzone 2:0x%lx\n",
2789 objp, *dbg_redzone1(cachep, objp),
2790 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002791 }
2792 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2793 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2794 }
Al Viro871751e2006-03-25 03:06:39 -08002795#ifdef CONFIG_DEBUG_SLAB_LEAK
2796 {
2797 struct slab *slabp;
2798 unsigned objnr;
2799
2800 slabp = page_get_slab(virt_to_page(objp));
2801 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
2802 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
2803 }
2804#endif
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002805 objp += obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002806 if (cachep->ctor && cachep->flags & SLAB_POISON) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002807 unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002808
2809 if (!(flags & __GFP_WAIT))
2810 ctor_flags |= SLAB_CTOR_ATOMIC;
2811
2812 cachep->ctor(objp, cachep, ctor_flags);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002813 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002814 return objp;
2815}
2816#else
2817#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2818#endif
2819
Pekka Enberg343e0d72006-02-01 03:05:50 -08002820static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002821{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002822 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002823 struct array_cache *ac;
2824
Christoph Lameterdc85da12006-01-18 17:42:36 -08002825#ifdef CONFIG_NUMA
Paul Jacksonb2455392006-03-24 03:16:12 -08002826 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
Paul Jacksonc61afb12006-03-24 03:16:08 -08002827 objp = alternate_node_alloc(cachep, flags);
2828 if (objp != NULL)
2829 return objp;
Paul Jackson101a5002006-03-24 03:16:07 -08002830 }
Christoph Lameterdc85da12006-01-18 17:42:36 -08002831#endif
2832
Alok N Kataria5c382302005-09-27 21:45:46 -07002833 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002834 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002835 if (likely(ac->avail)) {
2836 STATS_INC_ALLOCHIT(cachep);
2837 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07002838 objp = ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002839 } else {
2840 STATS_INC_ALLOCMISS(cachep);
2841 objp = cache_alloc_refill(cachep, flags);
2842 }
Alok N Kataria5c382302005-09-27 21:45:46 -07002843 return objp;
2844}
2845
Andrew Mortona737b3e2006-03-22 00:08:11 -08002846static __always_inline void *__cache_alloc(struct kmem_cache *cachep,
2847 gfp_t flags, void *caller)
Alok N Kataria5c382302005-09-27 21:45:46 -07002848{
2849 unsigned long save_flags;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002850 void *objp;
Alok N Kataria5c382302005-09-27 21:45:46 -07002851
2852 cache_alloc_debugcheck_before(cachep, flags);
2853
2854 local_irq_save(save_flags);
2855 objp = ____cache_alloc(cachep, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002856 local_irq_restore(save_flags);
Eric Dumazet34342e82005-09-03 15:55:06 -07002857 objp = cache_alloc_debugcheck_after(cachep, flags, objp,
Pekka Enberg7fd6b142006-02-01 03:05:52 -08002858 caller);
Eric Dumazet34342e82005-09-03 15:55:06 -07002859 prefetchw(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002860 return objp;
2861}
2862
Christoph Lametere498be72005-09-09 13:03:32 -07002863#ifdef CONFIG_NUMA
2864/*
Paul Jacksonb2455392006-03-24 03:16:12 -08002865 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
Paul Jacksonc61afb12006-03-24 03:16:08 -08002866 *
2867 * If we are in_interrupt, then process context, including cpusets and
2868 * mempolicy, may not apply and should not be used for allocation policy.
2869 */
2870static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
2871{
2872 int nid_alloc, nid_here;
2873
2874 if (in_interrupt())
2875 return NULL;
2876 nid_alloc = nid_here = numa_node_id();
2877 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
2878 nid_alloc = cpuset_mem_spread_node();
2879 else if (current->mempolicy)
2880 nid_alloc = slab_node(current->mempolicy);
2881 if (nid_alloc != nid_here)
2882 return __cache_alloc_node(cachep, flags, nid_alloc);
2883 return NULL;
2884}
2885
2886/*
Christoph Lametere498be72005-09-09 13:03:32 -07002887 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002888 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002889static void *__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
2890 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07002891{
2892 struct list_head *entry;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002893 struct slab *slabp;
2894 struct kmem_list3 *l3;
2895 void *obj;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002896 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002897
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002898 l3 = cachep->nodelists[nodeid];
2899 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07002900
Andrew Mortona737b3e2006-03-22 00:08:11 -08002901retry:
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08002902 check_irq_off();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002903 spin_lock(&l3->list_lock);
2904 entry = l3->slabs_partial.next;
2905 if (entry == &l3->slabs_partial) {
2906 l3->free_touched = 1;
2907 entry = l3->slabs_free.next;
2908 if (entry == &l3->slabs_free)
2909 goto must_grow;
2910 }
Christoph Lametere498be72005-09-09 13:03:32 -07002911
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002912 slabp = list_entry(entry, struct slab, list);
2913 check_spinlock_acquired_node(cachep, nodeid);
2914 check_slabp(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07002915
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002916 STATS_INC_NODEALLOCS(cachep);
2917 STATS_INC_ACTIVE(cachep);
2918 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002919
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002920 BUG_ON(slabp->inuse == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07002921
Matthew Dobson78d382d2006-02-01 03:05:47 -08002922 obj = slab_get_obj(cachep, slabp, nodeid);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002923 check_slabp(cachep, slabp);
2924 l3->free_objects--;
2925 /* move slabp to correct slabp list: */
2926 list_del(&slabp->list);
Christoph Lametere498be72005-09-09 13:03:32 -07002927
Andrew Mortona737b3e2006-03-22 00:08:11 -08002928 if (slabp->free == BUFCTL_END)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002929 list_add(&slabp->list, &l3->slabs_full);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002930 else
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002931 list_add(&slabp->list, &l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07002932
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002933 spin_unlock(&l3->list_lock);
2934 goto done;
Christoph Lametere498be72005-09-09 13:03:32 -07002935
Andrew Mortona737b3e2006-03-22 00:08:11 -08002936must_grow:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002937 spin_unlock(&l3->list_lock);
2938 x = cache_grow(cachep, flags, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07002939
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002940 if (!x)
2941 return NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07002942
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002943 goto retry;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002944done:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002945 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07002946}
2947#endif
2948
2949/*
2950 * Caller needs to acquire correct kmem_list's list_lock
2951 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002952static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002953 int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002954{
2955 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07002956 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002957
2958 for (i = 0; i < nr_objects; i++) {
2959 void *objp = objpp[i];
2960 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002961
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08002962 slabp = virt_to_slab(objp);
Christoph Lameterff694162005-09-22 21:44:02 -07002963 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002964 list_del(&slabp->list);
Christoph Lameterff694162005-09-22 21:44:02 -07002965 check_spinlock_acquired_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002966 check_slabp(cachep, slabp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002967 slab_put_obj(cachep, slabp, objp, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002968 STATS_DEC_ACTIVE(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002969 l3->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002970 check_slabp(cachep, slabp);
2971
2972 /* fixup slab chains */
2973 if (slabp->inuse == 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07002974 if (l3->free_objects > l3->free_limit) {
2975 l3->free_objects -= cachep->num;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002976 slab_destroy(cachep, slabp);
2977 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07002978 list_add(&slabp->list, &l3->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002979 }
2980 } else {
2981 /* Unconditionally move a slab to the end of the
2982 * partial list on free - maximum time for the
2983 * other objects to be freed, too.
2984 */
Christoph Lametere498be72005-09-09 13:03:32 -07002985 list_add_tail(&slabp->list, &l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002986 }
2987 }
2988}
2989
Pekka Enberg343e0d72006-02-01 03:05:50 -08002990static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002991{
2992 int batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07002993 struct kmem_list3 *l3;
Christoph Lameterff694162005-09-22 21:44:02 -07002994 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002995
2996 batchcount = ac->batchcount;
2997#if DEBUG
2998 BUG_ON(!batchcount || batchcount > ac->avail);
2999#endif
3000 check_irq_off();
Christoph Lameterff694162005-09-22 21:44:02 -07003001 l3 = cachep->nodelists[node];
Christoph Lametere498be72005-09-09 13:03:32 -07003002 spin_lock(&l3->list_lock);
3003 if (l3->shared) {
3004 struct array_cache *shared_array = l3->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003005 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003006 if (max) {
3007 if (batchcount > max)
3008 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07003009 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003010 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003011 shared_array->avail += batchcount;
3012 goto free_done;
3013 }
3014 }
3015
Christoph Lameterff694162005-09-22 21:44:02 -07003016 free_block(cachep, ac->entry, batchcount, node);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003017free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003018#if STATS
3019 {
3020 int i = 0;
3021 struct list_head *p;
3022
Christoph Lametere498be72005-09-09 13:03:32 -07003023 p = l3->slabs_free.next;
3024 while (p != &(l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003025 struct slab *slabp;
3026
3027 slabp = list_entry(p, struct slab, list);
3028 BUG_ON(slabp->inuse);
3029
3030 i++;
3031 p = p->next;
3032 }
3033 STATS_SET_FREEABLE(cachep, i);
3034 }
3035#endif
Christoph Lametere498be72005-09-09 13:03:32 -07003036 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003037 ac->avail -= batchcount;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003038 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003039}
3040
3041/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08003042 * Release an obj back to its cache. If the obj has a constructed state, it must
3043 * be in this state _before_ it is released. Called with disabled ints.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003044 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003045static inline void __cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003046{
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003047 struct array_cache *ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003048
3049 check_irq_off();
3050 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3051
Christoph Lametere498be72005-09-09 13:03:32 -07003052 /* Make sure we are not freeing a object from another
3053 * node to the array cache on this cpu.
3054 */
3055#ifdef CONFIG_NUMA
3056 {
3057 struct slab *slabp;
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003058 slabp = virt_to_slab(objp);
Christoph Lametere498be72005-09-09 13:03:32 -07003059 if (unlikely(slabp->nodeid != numa_node_id())) {
3060 struct array_cache *alien = NULL;
3061 int nodeid = slabp->nodeid;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003062 struct kmem_list3 *l3;
Christoph Lametere498be72005-09-09 13:03:32 -07003063
Andrew Mortona737b3e2006-03-22 00:08:11 -08003064 l3 = cachep->nodelists[numa_node_id()];
Christoph Lametere498be72005-09-09 13:03:32 -07003065 STATS_INC_NODEFREES(cachep);
3066 if (l3->alien && l3->alien[nodeid]) {
3067 alien = l3->alien[nodeid];
3068 spin_lock(&alien->lock);
3069 if (unlikely(alien->avail == alien->limit))
3070 __drain_alien_cache(cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003071 alien, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07003072 alien->entry[alien->avail++] = objp;
3073 spin_unlock(&alien->lock);
3074 } else {
3075 spin_lock(&(cachep->nodelists[nodeid])->
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003076 list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -07003077 free_block(cachep, &objp, 1, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07003078 spin_unlock(&(cachep->nodelists[nodeid])->
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003079 list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003080 }
3081 return;
3082 }
3083 }
3084#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003085 if (likely(ac->avail < ac->limit)) {
3086 STATS_INC_FREEHIT(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003087 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003088 return;
3089 } else {
3090 STATS_INC_FREEMISS(cachep);
3091 cache_flusharray(cachep, ac);
Christoph Lametere498be72005-09-09 13:03:32 -07003092 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003093 }
3094}
3095
3096/**
3097 * kmem_cache_alloc - Allocate an object
3098 * @cachep: The cache to allocate from.
3099 * @flags: See kmalloc().
3100 *
3101 * Allocate an object from this cache. The flags are only relevant
3102 * if the cache has no available objects.
3103 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003104void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003105{
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003106 return __cache_alloc(cachep, flags, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003107}
3108EXPORT_SYMBOL(kmem_cache_alloc);
3109
3110/**
Pekka Enberga8c0f9a2006-03-25 03:06:42 -08003111 * kmem_cache_alloc - Allocate an object. The memory is set to zero.
3112 * @cache: The cache to allocate from.
3113 * @flags: See kmalloc().
3114 *
3115 * Allocate an object from this cache and set the allocated memory to zero.
3116 * The flags are only relevant if the cache has no available objects.
3117 */
3118void *kmem_cache_zalloc(struct kmem_cache *cache, gfp_t flags)
3119{
3120 void *ret = __cache_alloc(cache, flags, __builtin_return_address(0));
3121 if (ret)
3122 memset(ret, 0, obj_size(cache));
3123 return ret;
3124}
3125EXPORT_SYMBOL(kmem_cache_zalloc);
3126
3127/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003128 * kmem_ptr_validate - check if an untrusted pointer might
3129 * be a slab entry.
3130 * @cachep: the cache we're checking against
3131 * @ptr: pointer to validate
3132 *
3133 * This verifies that the untrusted pointer looks sane:
3134 * it is _not_ a guarantee that the pointer is actually
3135 * part of the slab cache in question, but it at least
3136 * validates that the pointer can be dereferenced and
3137 * looks half-way sane.
3138 *
3139 * Currently only used for dentry validation.
3140 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003141int fastcall kmem_ptr_validate(struct kmem_cache *cachep, void *ptr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003142{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003143 unsigned long addr = (unsigned long)ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003144 unsigned long min_addr = PAGE_OFFSET;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003145 unsigned long align_mask = BYTES_PER_WORD - 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003146 unsigned long size = cachep->buffer_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003147 struct page *page;
3148
3149 if (unlikely(addr < min_addr))
3150 goto out;
3151 if (unlikely(addr > (unsigned long)high_memory - size))
3152 goto out;
3153 if (unlikely(addr & align_mask))
3154 goto out;
3155 if (unlikely(!kern_addr_valid(addr)))
3156 goto out;
3157 if (unlikely(!kern_addr_valid(addr + size - 1)))
3158 goto out;
3159 page = virt_to_page(ptr);
3160 if (unlikely(!PageSlab(page)))
3161 goto out;
Pekka Enberg065d41c2005-11-13 16:06:46 -08003162 if (unlikely(page_get_cache(page) != cachep))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003163 goto out;
3164 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003165out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003166 return 0;
3167}
3168
3169#ifdef CONFIG_NUMA
3170/**
3171 * kmem_cache_alloc_node - Allocate an object on the specified node
3172 * @cachep: The cache to allocate from.
3173 * @flags: See kmalloc().
3174 * @nodeid: node number of the target node.
3175 *
3176 * Identical to kmem_cache_alloc, except that this function is slow
3177 * and can sleep. And it will allocate memory on the given node, which
3178 * can improve the performance for cpu bound structures.
Christoph Lametere498be72005-09-09 13:03:32 -07003179 * New and improved: it will now make sure that the object gets
3180 * put on the correct node list so that there is no false sharing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003181 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003182void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003183{
Christoph Lametere498be72005-09-09 13:03:32 -07003184 unsigned long save_flags;
3185 void *ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003186
Christoph Lametere498be72005-09-09 13:03:32 -07003187 cache_alloc_debugcheck_before(cachep, flags);
3188 local_irq_save(save_flags);
Christoph Lameter18f820f2006-02-01 03:05:43 -08003189
3190 if (nodeid == -1 || nodeid == numa_node_id() ||
Andrew Mortona737b3e2006-03-22 00:08:11 -08003191 !cachep->nodelists[nodeid])
Alok N Kataria5c382302005-09-27 21:45:46 -07003192 ptr = ____cache_alloc(cachep, flags);
3193 else
3194 ptr = __cache_alloc_node(cachep, flags, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07003195 local_irq_restore(save_flags);
Christoph Lameter18f820f2006-02-01 03:05:43 -08003196
3197 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr,
3198 __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003199
Christoph Lametere498be72005-09-09 13:03:32 -07003200 return ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003201}
3202EXPORT_SYMBOL(kmem_cache_alloc_node);
3203
Al Virodd0fc662005-10-07 07:46:04 +01003204void *kmalloc_node(size_t size, gfp_t flags, int node)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003205{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003206 struct kmem_cache *cachep;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003207
3208 cachep = kmem_find_general_cachep(size, flags);
3209 if (unlikely(cachep == NULL))
3210 return NULL;
3211 return kmem_cache_alloc_node(cachep, flags, node);
3212}
3213EXPORT_SYMBOL(kmalloc_node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003214#endif
3215
3216/**
3217 * kmalloc - allocate memory
3218 * @size: how many bytes of memory are required.
3219 * @flags: the type of memory to allocate.
Randy Dunlap911851e2006-03-22 00:08:14 -08003220 * @caller: function caller for debug tracking of the caller
Linus Torvalds1da177e2005-04-16 15:20:36 -07003221 *
3222 * kmalloc is the normal method of allocating memory
3223 * in the kernel.
3224 *
3225 * The @flags argument may be one of:
3226 *
3227 * %GFP_USER - Allocate memory on behalf of user. May sleep.
3228 *
3229 * %GFP_KERNEL - Allocate normal kernel ram. May sleep.
3230 *
3231 * %GFP_ATOMIC - Allocation will not sleep. Use inside interrupt handlers.
3232 *
3233 * Additionally, the %GFP_DMA flag may be set to indicate the memory
3234 * must be suitable for DMA. This can mean different things on different
3235 * platforms. For example, on i386, it means that the memory must come
3236 * from the first 16MB.
3237 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003238static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3239 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003240{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003241 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003242
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003243 /* If you want to save a few bytes .text space: replace
3244 * __ with kmem_.
3245 * Then kmalloc uses the uninlined functions instead of the inline
3246 * functions.
3247 */
3248 cachep = __find_general_cachep(size, flags);
Andrew Mortondbdb9042005-09-23 13:24:10 -07003249 if (unlikely(cachep == NULL))
3250 return NULL;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003251 return __cache_alloc(cachep, flags, caller);
3252}
3253
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003254
3255void *__kmalloc(size_t size, gfp_t flags)
3256{
Al Viro871751e2006-03-25 03:06:39 -08003257#ifndef CONFIG_DEBUG_SLAB
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003258 return __do_kmalloc(size, flags, NULL);
Al Viro871751e2006-03-25 03:06:39 -08003259#else
3260 return __do_kmalloc(size, flags, __builtin_return_address(0));
3261#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003262}
3263EXPORT_SYMBOL(__kmalloc);
3264
Al Viro871751e2006-03-25 03:06:39 -08003265#ifdef CONFIG_DEBUG_SLAB
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003266void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller)
3267{
3268 return __do_kmalloc(size, flags, caller);
3269}
3270EXPORT_SYMBOL(__kmalloc_track_caller);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003271#endif
3272
Linus Torvalds1da177e2005-04-16 15:20:36 -07003273#ifdef CONFIG_SMP
3274/**
3275 * __alloc_percpu - allocate one copy of the object for every present
3276 * cpu in the system, zeroing them.
3277 * Objects should be dereferenced using the per_cpu_ptr macro only.
3278 *
3279 * @size: how many bytes of memory are required.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003280 */
Pekka Enbergf9f75002006-01-08 01:00:33 -08003281void *__alloc_percpu(size_t size)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003282{
3283 int i;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003284 struct percpu_data *pdata = kmalloc(sizeof(*pdata), GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003285
3286 if (!pdata)
3287 return NULL;
3288
Christoph Lametere498be72005-09-09 13:03:32 -07003289 /*
3290 * Cannot use for_each_online_cpu since a cpu may come online
3291 * and we have no way of figuring out how to fix the array
3292 * that we have allocated then....
3293 */
3294 for_each_cpu(i) {
3295 int node = cpu_to_node(i);
3296
3297 if (node_online(node))
3298 pdata->ptrs[i] = kmalloc_node(size, GFP_KERNEL, node);
3299 else
3300 pdata->ptrs[i] = kmalloc(size, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003301
3302 if (!pdata->ptrs[i])
3303 goto unwind_oom;
3304 memset(pdata->ptrs[i], 0, size);
3305 }
3306
3307 /* Catch derefs w/o wrappers */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003308 return (void *)(~(unsigned long)pdata);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003309
Andrew Mortona737b3e2006-03-22 00:08:11 -08003310unwind_oom:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003311 while (--i >= 0) {
3312 if (!cpu_possible(i))
3313 continue;
3314 kfree(pdata->ptrs[i]);
3315 }
3316 kfree(pdata);
3317 return NULL;
3318}
3319EXPORT_SYMBOL(__alloc_percpu);
3320#endif
3321
3322/**
3323 * kmem_cache_free - Deallocate an object
3324 * @cachep: The cache the allocation was from.
3325 * @objp: The previously allocated object.
3326 *
3327 * Free an object which was previously allocated from this
3328 * cache.
3329 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003330void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003331{
3332 unsigned long flags;
3333
3334 local_irq_save(flags);
3335 __cache_free(cachep, objp);
3336 local_irq_restore(flags);
3337}
3338EXPORT_SYMBOL(kmem_cache_free);
3339
3340/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003341 * kfree - free previously allocated memory
3342 * @objp: pointer returned by kmalloc.
3343 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003344 * If @objp is NULL, no operation is performed.
3345 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003346 * Don't free memory not originally allocated by kmalloc()
3347 * or you will run into trouble.
3348 */
3349void kfree(const void *objp)
3350{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003351 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003352 unsigned long flags;
3353
3354 if (unlikely(!objp))
3355 return;
3356 local_irq_save(flags);
3357 kfree_debugcheck(objp);
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003358 c = virt_to_cache(objp);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003359 mutex_debug_check_no_locks_freed(objp, obj_size(c));
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003360 __cache_free(c, (void *)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003361 local_irq_restore(flags);
3362}
3363EXPORT_SYMBOL(kfree);
3364
3365#ifdef CONFIG_SMP
3366/**
3367 * free_percpu - free previously allocated percpu memory
3368 * @objp: pointer returned by alloc_percpu.
3369 *
3370 * Don't free memory not originally allocated by alloc_percpu()
3371 * The complemented objp is to check for that.
3372 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003373void free_percpu(const void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003374{
3375 int i;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003376 struct percpu_data *p = (struct percpu_data *)(~(unsigned long)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003377
Christoph Lametere498be72005-09-09 13:03:32 -07003378 /*
3379 * We allocate for all cpus so we cannot use for online cpu here.
3380 */
3381 for_each_cpu(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003382 kfree(p->ptrs[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003383 kfree(p);
3384}
3385EXPORT_SYMBOL(free_percpu);
3386#endif
3387
Pekka Enberg343e0d72006-02-01 03:05:50 -08003388unsigned int kmem_cache_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003389{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003390 return obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003391}
3392EXPORT_SYMBOL(kmem_cache_size);
3393
Pekka Enberg343e0d72006-02-01 03:05:50 -08003394const char *kmem_cache_name(struct kmem_cache *cachep)
Arnaldo Carvalho de Melo19449722005-06-18 22:46:19 -07003395{
3396 return cachep->name;
3397}
3398EXPORT_SYMBOL_GPL(kmem_cache_name);
3399
Christoph Lametere498be72005-09-09 13:03:32 -07003400/*
3401 * This initializes kmem_list3 for all nodes.
3402 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003403static int alloc_kmemlist(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07003404{
3405 int node;
3406 struct kmem_list3 *l3;
3407 int err = 0;
3408
3409 for_each_online_node(node) {
3410 struct array_cache *nc = NULL, *new;
3411 struct array_cache **new_alien = NULL;
3412#ifdef CONFIG_NUMA
Andrew Mortona737b3e2006-03-22 00:08:11 -08003413 new_alien = alloc_alien_cache(node, cachep->limit);
3414 if (!new_alien)
Christoph Lametere498be72005-09-09 13:03:32 -07003415 goto fail;
3416#endif
Andrew Mortona737b3e2006-03-22 00:08:11 -08003417 new = alloc_arraycache(node, cachep->shared*cachep->batchcount,
3418 0xbaadf00d);
3419 if (!new)
Christoph Lametere498be72005-09-09 13:03:32 -07003420 goto fail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003421 l3 = cachep->nodelists[node];
3422 if (l3) {
Christoph Lametere498be72005-09-09 13:03:32 -07003423 spin_lock_irq(&l3->list_lock);
3424
Andrew Mortona737b3e2006-03-22 00:08:11 -08003425 nc = cachep->nodelists[node]->shared;
3426 if (nc)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003427 free_block(cachep, nc->entry, nc->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07003428
3429 l3->shared = new;
3430 if (!cachep->nodelists[node]->alien) {
3431 l3->alien = new_alien;
3432 new_alien = NULL;
3433 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003434 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003435 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003436 spin_unlock_irq(&l3->list_lock);
3437 kfree(nc);
3438 free_alien_cache(new_alien);
3439 continue;
3440 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08003441 l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node);
3442 if (!l3)
Christoph Lametere498be72005-09-09 13:03:32 -07003443 goto fail;
3444
3445 kmem_list3_init(l3);
3446 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Andrew Mortona737b3e2006-03-22 00:08:11 -08003447 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -07003448 l3->shared = new;
3449 l3->alien = new_alien;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003450 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003451 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003452 cachep->nodelists[node] = l3;
3453 }
3454 return err;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003455fail:
Christoph Lametere498be72005-09-09 13:03:32 -07003456 err = -ENOMEM;
3457 return err;
3458}
3459
Linus Torvalds1da177e2005-04-16 15:20:36 -07003460struct ccupdate_struct {
Pekka Enberg343e0d72006-02-01 03:05:50 -08003461 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003462 struct array_cache *new[NR_CPUS];
3463};
3464
3465static void do_ccupdate_local(void *info)
3466{
Andrew Mortona737b3e2006-03-22 00:08:11 -08003467 struct ccupdate_struct *new = info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003468 struct array_cache *old;
3469
3470 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003471 old = cpu_cache_get(new->cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003472
Linus Torvalds1da177e2005-04-16 15:20:36 -07003473 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3474 new->new[smp_processor_id()] = old;
3475}
3476
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003477/* Always called with the cache_chain_mutex held */
Andrew Mortona737b3e2006-03-22 00:08:11 -08003478static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3479 int batchcount, int shared)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003480{
3481 struct ccupdate_struct new;
Christoph Lametere498be72005-09-09 13:03:32 -07003482 int i, err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003483
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003484 memset(&new.new, 0, sizeof(new.new));
Christoph Lametere498be72005-09-09 13:03:32 -07003485 for_each_online_cpu(i) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003486 new.new[i] = alloc_arraycache(cpu_to_node(i), limit,
3487 batchcount);
Christoph Lametere498be72005-09-09 13:03:32 -07003488 if (!new.new[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003489 for (i--; i >= 0; i--)
3490 kfree(new.new[i]);
Christoph Lametere498be72005-09-09 13:03:32 -07003491 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003492 }
3493 }
3494 new.cachep = cachep;
3495
Andrew Mortona07fa392006-03-22 00:08:17 -08003496 on_each_cpu(do_ccupdate_local, (void *)&new, 1, 1);
Christoph Lametere498be72005-09-09 13:03:32 -07003497
Linus Torvalds1da177e2005-04-16 15:20:36 -07003498 check_irq_on();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003499 cachep->batchcount = batchcount;
3500 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07003501 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003502
Christoph Lametere498be72005-09-09 13:03:32 -07003503 for_each_online_cpu(i) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003504 struct array_cache *ccold = new.new[i];
3505 if (!ccold)
3506 continue;
Christoph Lametere498be72005-09-09 13:03:32 -07003507 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -07003508 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
Christoph Lametere498be72005-09-09 13:03:32 -07003509 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003510 kfree(ccold);
3511 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003512
Christoph Lametere498be72005-09-09 13:03:32 -07003513 err = alloc_kmemlist(cachep);
3514 if (err) {
3515 printk(KERN_ERR "alloc_kmemlist failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003516 cachep->name, -err);
Christoph Lametere498be72005-09-09 13:03:32 -07003517 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003518 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003519 return 0;
3520}
3521
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003522/* Called with cache_chain_mutex held always */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003523static void enable_cpucache(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003524{
3525 int err;
3526 int limit, shared;
3527
Andrew Mortona737b3e2006-03-22 00:08:11 -08003528 /*
3529 * The head array serves three purposes:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003530 * - create a LIFO ordering, i.e. return objects that are cache-warm
3531 * - reduce the number of spinlock operations.
Andrew Mortona737b3e2006-03-22 00:08:11 -08003532 * - reduce the number of linked list operations on the slab and
Linus Torvalds1da177e2005-04-16 15:20:36 -07003533 * bufctl chains: array operations are cheaper.
3534 * The numbers are guessed, we should auto-tune as described by
3535 * Bonwick.
3536 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003537 if (cachep->buffer_size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003538 limit = 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003539 else if (cachep->buffer_size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003540 limit = 8;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003541 else if (cachep->buffer_size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003542 limit = 24;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003543 else if (cachep->buffer_size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003544 limit = 54;
3545 else
3546 limit = 120;
3547
Andrew Mortona737b3e2006-03-22 00:08:11 -08003548 /*
3549 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
Linus Torvalds1da177e2005-04-16 15:20:36 -07003550 * allocation behaviour: Most allocs on one cpu, most free operations
3551 * on another cpu. For these cases, an efficient object passing between
3552 * cpus is necessary. This is provided by a shared array. The array
3553 * replaces Bonwick's magazine layer.
3554 * On uniprocessor, it's functionally equivalent (but less efficient)
3555 * to a larger limit. Thus disabled by default.
3556 */
3557 shared = 0;
3558#ifdef CONFIG_SMP
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003559 if (cachep->buffer_size <= PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003560 shared = 8;
3561#endif
3562
3563#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003564 /*
3565 * With debugging enabled, large batchcount lead to excessively long
3566 * periods with disabled local interrupts. Limit the batchcount
Linus Torvalds1da177e2005-04-16 15:20:36 -07003567 */
3568 if (limit > 32)
3569 limit = 32;
3570#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003571 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003572 if (err)
3573 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003574 cachep->name, -err);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003575}
3576
Christoph Lameter1b552532006-03-22 00:09:07 -08003577/*
3578 * Drain an array if it contains any elements taking the l3 lock only if
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003579 * necessary. Note that the l3 listlock also protects the array_cache
3580 * if drain_array() is used on the shared array.
Christoph Lameter1b552532006-03-22 00:09:07 -08003581 */
3582void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
3583 struct array_cache *ac, int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003584{
3585 int tofree;
3586
Christoph Lameter1b552532006-03-22 00:09:07 -08003587 if (!ac || !ac->avail)
3588 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003589 if (ac->touched && !force) {
3590 ac->touched = 0;
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003591 } else {
Christoph Lameter1b552532006-03-22 00:09:07 -08003592 spin_lock_irq(&l3->list_lock);
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003593 if (ac->avail) {
3594 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3595 if (tofree > ac->avail)
3596 tofree = (ac->avail + 1) / 2;
3597 free_block(cachep, ac->entry, tofree, node);
3598 ac->avail -= tofree;
3599 memmove(ac->entry, &(ac->entry[tofree]),
3600 sizeof(void *) * ac->avail);
3601 }
Christoph Lameter1b552532006-03-22 00:09:07 -08003602 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003603 }
3604}
3605
3606/**
3607 * cache_reap - Reclaim memory from caches.
Randy Dunlap1e5d5332005-11-07 01:01:06 -08003608 * @unused: unused parameter
Linus Torvalds1da177e2005-04-16 15:20:36 -07003609 *
3610 * Called from workqueue/eventd every few seconds.
3611 * Purpose:
3612 * - clear the per-cpu caches for this CPU.
3613 * - return freeable pages to the main free memory pool.
3614 *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003615 * If we cannot acquire the cache chain mutex then just give up - we'll try
3616 * again on the next iteration.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003617 */
3618static void cache_reap(void *unused)
3619{
3620 struct list_head *walk;
Christoph Lametere498be72005-09-09 13:03:32 -07003621 struct kmem_list3 *l3;
Christoph Lameteraab22072006-03-22 00:09:06 -08003622 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003623
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003624 if (!mutex_trylock(&cache_chain_mutex)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003625 /* Give up. Setup the next iteration. */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003626 schedule_delayed_work(&__get_cpu_var(reap_work),
3627 REAPTIMEOUT_CPUC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003628 return;
3629 }
3630
3631 list_for_each(walk, &cache_chain) {
Pekka Enberg343e0d72006-02-01 03:05:50 -08003632 struct kmem_cache *searchp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003633 struct list_head *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003634 int tofree;
3635 struct slab *slabp;
3636
Pekka Enberg343e0d72006-02-01 03:05:50 -08003637 searchp = list_entry(walk, struct kmem_cache, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003638 check_irq_on();
3639
Christoph Lameter35386e32006-03-22 00:09:05 -08003640 /*
3641 * We only take the l3 lock if absolutely necessary and we
3642 * have established with reasonable certainty that
3643 * we can do some work if the lock was obtained.
3644 */
Christoph Lameteraab22072006-03-22 00:09:06 -08003645 l3 = searchp->nodelists[node];
Christoph Lameter35386e32006-03-22 00:09:05 -08003646
Christoph Lameter8fce4d82006-03-09 17:33:54 -08003647 reap_alien(searchp, l3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003648
Christoph Lameteraab22072006-03-22 00:09:06 -08003649 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003650
Christoph Lameter35386e32006-03-22 00:09:05 -08003651 /*
3652 * These are racy checks but it does not matter
3653 * if we skip one check or scan twice.
3654 */
Christoph Lametere498be72005-09-09 13:03:32 -07003655 if (time_after(l3->next_reap, jiffies))
Christoph Lameter35386e32006-03-22 00:09:05 -08003656 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003657
Christoph Lametere498be72005-09-09 13:03:32 -07003658 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003659
Christoph Lameteraab22072006-03-22 00:09:06 -08003660 drain_array(searchp, l3, l3->shared, 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003661
Christoph Lametere498be72005-09-09 13:03:32 -07003662 if (l3->free_touched) {
3663 l3->free_touched = 0;
Christoph Lameter35386e32006-03-22 00:09:05 -08003664 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003665 }
3666
Andrew Mortona737b3e2006-03-22 00:08:11 -08003667 tofree = (l3->free_limit + 5 * searchp->num - 1) /
3668 (5 * searchp->num);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003669 do {
Christoph Lameter35386e32006-03-22 00:09:05 -08003670 /*
3671 * Do not lock if there are no free blocks.
3672 */
3673 if (list_empty(&l3->slabs_free))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003674 break;
3675
Christoph Lameter35386e32006-03-22 00:09:05 -08003676 spin_lock_irq(&l3->list_lock);
3677 p = l3->slabs_free.next;
3678 if (p == &(l3->slabs_free)) {
3679 spin_unlock_irq(&l3->list_lock);
3680 break;
3681 }
3682
Linus Torvalds1da177e2005-04-16 15:20:36 -07003683 slabp = list_entry(p, struct slab, list);
3684 BUG_ON(slabp->inuse);
3685 list_del(&slabp->list);
3686 STATS_INC_REAPED(searchp);
3687
Andrew Mortona737b3e2006-03-22 00:08:11 -08003688 /*
3689 * Safe to drop the lock. The slab is no longer linked
3690 * to the cache. searchp cannot disappear, we hold
Linus Torvalds1da177e2005-04-16 15:20:36 -07003691 * cache_chain_lock
3692 */
Christoph Lametere498be72005-09-09 13:03:32 -07003693 l3->free_objects -= searchp->num;
3694 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003695 slab_destroy(searchp, slabp);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003696 } while (--tofree > 0);
Christoph Lameter35386e32006-03-22 00:09:05 -08003697next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003698 cond_resched();
3699 }
3700 check_irq_on();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003701 mutex_unlock(&cache_chain_mutex);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08003702 next_reap_node();
Andrew Mortona737b3e2006-03-22 00:08:11 -08003703 /* Set up the next iteration */
Manfred Spraulcd61ef62005-11-07 00:58:02 -08003704 schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003705}
3706
3707#ifdef CONFIG_PROC_FS
3708
Pekka Enberg85289f92006-01-08 01:00:36 -08003709static void print_slabinfo_header(struct seq_file *m)
3710{
3711 /*
3712 * Output format version, so at least we can change it
3713 * without _too_ many complaints.
3714 */
3715#if STATS
3716 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
3717#else
3718 seq_puts(m, "slabinfo - version: 2.1\n");
3719#endif
3720 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
3721 "<objperslab> <pagesperslab>");
3722 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
3723 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
3724#if STATS
3725 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
3726 "<error> <maxfreeable> <nodeallocs> <remotefrees>");
3727 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
3728#endif
3729 seq_putc(m, '\n');
3730}
3731
Linus Torvalds1da177e2005-04-16 15:20:36 -07003732static void *s_start(struct seq_file *m, loff_t *pos)
3733{
3734 loff_t n = *pos;
3735 struct list_head *p;
3736
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003737 mutex_lock(&cache_chain_mutex);
Pekka Enberg85289f92006-01-08 01:00:36 -08003738 if (!n)
3739 print_slabinfo_header(m);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003740 p = cache_chain.next;
3741 while (n--) {
3742 p = p->next;
3743 if (p == &cache_chain)
3744 return NULL;
3745 }
Pekka Enberg343e0d72006-02-01 03:05:50 -08003746 return list_entry(p, struct kmem_cache, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003747}
3748
3749static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3750{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003751 struct kmem_cache *cachep = p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003752 ++*pos;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003753 return cachep->next.next == &cache_chain ?
3754 NULL : list_entry(cachep->next.next, struct kmem_cache, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003755}
3756
3757static void s_stop(struct seq_file *m, void *p)
3758{
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003759 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003760}
3761
3762static int s_show(struct seq_file *m, void *p)
3763{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003764 struct kmem_cache *cachep = p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003765 struct list_head *q;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003766 struct slab *slabp;
3767 unsigned long active_objs;
3768 unsigned long num_objs;
3769 unsigned long active_slabs = 0;
3770 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07003771 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003772 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003773 int node;
3774 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003775
Linus Torvalds1da177e2005-04-16 15:20:36 -07003776 active_objs = 0;
3777 num_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07003778 for_each_online_node(node) {
3779 l3 = cachep->nodelists[node];
3780 if (!l3)
3781 continue;
3782
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003783 check_irq_on();
3784 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003785
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003786 list_for_each(q, &l3->slabs_full) {
Christoph Lametere498be72005-09-09 13:03:32 -07003787 slabp = list_entry(q, struct slab, list);
3788 if (slabp->inuse != cachep->num && !error)
3789 error = "slabs_full accounting error";
3790 active_objs += cachep->num;
3791 active_slabs++;
3792 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003793 list_for_each(q, &l3->slabs_partial) {
Christoph Lametere498be72005-09-09 13:03:32 -07003794 slabp = list_entry(q, struct slab, list);
3795 if (slabp->inuse == cachep->num && !error)
3796 error = "slabs_partial inuse accounting error";
3797 if (!slabp->inuse && !error)
3798 error = "slabs_partial/inuse accounting error";
3799 active_objs += slabp->inuse;
3800 active_slabs++;
3801 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003802 list_for_each(q, &l3->slabs_free) {
Christoph Lametere498be72005-09-09 13:03:32 -07003803 slabp = list_entry(q, struct slab, list);
3804 if (slabp->inuse && !error)
3805 error = "slabs_free/inuse accounting error";
3806 num_slabs++;
3807 }
3808 free_objects += l3->free_objects;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08003809 if (l3->shared)
3810 shared_avail += l3->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07003811
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003812 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003813 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003814 num_slabs += active_slabs;
3815 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003816 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003817 error = "free_objects accounting error";
3818
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003819 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003820 if (error)
3821 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
3822
3823 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003824 name, active_objs, num_objs, cachep->buffer_size,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003825 cachep->num, (1 << cachep->gfporder));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003826 seq_printf(m, " : tunables %4u %4u %4u",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003827 cachep->limit, cachep->batchcount, cachep->shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003828 seq_printf(m, " : slabdata %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003829 active_slabs, num_slabs, shared_avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003830#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003831 { /* list3 stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003832 unsigned long high = cachep->high_mark;
3833 unsigned long allocs = cachep->num_allocations;
3834 unsigned long grown = cachep->grown;
3835 unsigned long reaped = cachep->reaped;
3836 unsigned long errors = cachep->errors;
3837 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003838 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07003839 unsigned long node_frees = cachep->node_frees;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003840
Christoph Lametere498be72005-09-09 13:03:32 -07003841 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
Andrew Mortona737b3e2006-03-22 00:08:11 -08003842 %4lu %4lu %4lu %4lu", allocs, high, grown,
3843 reaped, errors, max_freeable, node_allocs,
3844 node_frees);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003845 }
3846 /* cpu stats */
3847 {
3848 unsigned long allochit = atomic_read(&cachep->allochit);
3849 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
3850 unsigned long freehit = atomic_read(&cachep->freehit);
3851 unsigned long freemiss = atomic_read(&cachep->freemiss);
3852
3853 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003854 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003855 }
3856#endif
3857 seq_putc(m, '\n');
Linus Torvalds1da177e2005-04-16 15:20:36 -07003858 return 0;
3859}
3860
3861/*
3862 * slabinfo_op - iterator that generates /proc/slabinfo
3863 *
3864 * Output layout:
3865 * cache-name
3866 * num-active-objs
3867 * total-objs
3868 * object size
3869 * num-active-slabs
3870 * total-slabs
3871 * num-pages-per-slab
3872 * + further values on SMP and with statistics enabled
3873 */
3874
3875struct seq_operations slabinfo_op = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003876 .start = s_start,
3877 .next = s_next,
3878 .stop = s_stop,
3879 .show = s_show,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003880};
3881
3882#define MAX_SLABINFO_WRITE 128
3883/**
3884 * slabinfo_write - Tuning for the slab allocator
3885 * @file: unused
3886 * @buffer: user buffer
3887 * @count: data length
3888 * @ppos: unused
3889 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003890ssize_t slabinfo_write(struct file *file, const char __user * buffer,
3891 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003892{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003893 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003894 int limit, batchcount, shared, res;
3895 struct list_head *p;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003896
Linus Torvalds1da177e2005-04-16 15:20:36 -07003897 if (count > MAX_SLABINFO_WRITE)
3898 return -EINVAL;
3899 if (copy_from_user(&kbuf, buffer, count))
3900 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003901 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07003902
3903 tmp = strchr(kbuf, ' ');
3904 if (!tmp)
3905 return -EINVAL;
3906 *tmp = '\0';
3907 tmp++;
3908 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
3909 return -EINVAL;
3910
3911 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003912 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003913 res = -EINVAL;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003914 list_for_each(p, &cache_chain) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003915 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003916
Andrew Mortona737b3e2006-03-22 00:08:11 -08003917 cachep = list_entry(p, struct kmem_cache, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003918 if (!strcmp(cachep->name, kbuf)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003919 if (limit < 1 || batchcount < 1 ||
3920 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07003921 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003922 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07003923 res = do_tune_cpucache(cachep, limit,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003924 batchcount, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003925 }
3926 break;
3927 }
3928 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003929 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003930 if (res >= 0)
3931 res = count;
3932 return res;
3933}
Al Viro871751e2006-03-25 03:06:39 -08003934
3935#ifdef CONFIG_DEBUG_SLAB_LEAK
3936
3937static void *leaks_start(struct seq_file *m, loff_t *pos)
3938{
3939 loff_t n = *pos;
3940 struct list_head *p;
3941
3942 mutex_lock(&cache_chain_mutex);
3943 p = cache_chain.next;
3944 while (n--) {
3945 p = p->next;
3946 if (p == &cache_chain)
3947 return NULL;
3948 }
3949 return list_entry(p, struct kmem_cache, next);
3950}
3951
3952static inline int add_caller(unsigned long *n, unsigned long v)
3953{
3954 unsigned long *p;
3955 int l;
3956 if (!v)
3957 return 1;
3958 l = n[1];
3959 p = n + 2;
3960 while (l) {
3961 int i = l/2;
3962 unsigned long *q = p + 2 * i;
3963 if (*q == v) {
3964 q[1]++;
3965 return 1;
3966 }
3967 if (*q > v) {
3968 l = i;
3969 } else {
3970 p = q + 2;
3971 l -= i + 1;
3972 }
3973 }
3974 if (++n[1] == n[0])
3975 return 0;
3976 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
3977 p[0] = v;
3978 p[1] = 1;
3979 return 1;
3980}
3981
3982static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
3983{
3984 void *p;
3985 int i;
3986 if (n[0] == n[1])
3987 return;
3988 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
3989 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
3990 continue;
3991 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
3992 return;
3993 }
3994}
3995
3996static void show_symbol(struct seq_file *m, unsigned long address)
3997{
3998#ifdef CONFIG_KALLSYMS
3999 char *modname;
4000 const char *name;
4001 unsigned long offset, size;
4002 char namebuf[KSYM_NAME_LEN+1];
4003
4004 name = kallsyms_lookup(address, &size, &offset, &modname, namebuf);
4005
4006 if (name) {
4007 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4008 if (modname)
4009 seq_printf(m, " [%s]", modname);
4010 return;
4011 }
4012#endif
4013 seq_printf(m, "%p", (void *)address);
4014}
4015
4016static int leaks_show(struct seq_file *m, void *p)
4017{
4018 struct kmem_cache *cachep = p;
4019 struct list_head *q;
4020 struct slab *slabp;
4021 struct kmem_list3 *l3;
4022 const char *name;
4023 unsigned long *n = m->private;
4024 int node;
4025 int i;
4026
4027 if (!(cachep->flags & SLAB_STORE_USER))
4028 return 0;
4029 if (!(cachep->flags & SLAB_RED_ZONE))
4030 return 0;
4031
4032 /* OK, we can do it */
4033
4034 n[1] = 0;
4035
4036 for_each_online_node(node) {
4037 l3 = cachep->nodelists[node];
4038 if (!l3)
4039 continue;
4040
4041 check_irq_on();
4042 spin_lock_irq(&l3->list_lock);
4043
4044 list_for_each(q, &l3->slabs_full) {
4045 slabp = list_entry(q, struct slab, list);
4046 handle_slab(n, cachep, slabp);
4047 }
4048 list_for_each(q, &l3->slabs_partial) {
4049 slabp = list_entry(q, struct slab, list);
4050 handle_slab(n, cachep, slabp);
4051 }
4052 spin_unlock_irq(&l3->list_lock);
4053 }
4054 name = cachep->name;
4055 if (n[0] == n[1]) {
4056 /* Increase the buffer size */
4057 mutex_unlock(&cache_chain_mutex);
4058 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4059 if (!m->private) {
4060 /* Too bad, we are really out */
4061 m->private = n;
4062 mutex_lock(&cache_chain_mutex);
4063 return -ENOMEM;
4064 }
4065 *(unsigned long *)m->private = n[0] * 2;
4066 kfree(n);
4067 mutex_lock(&cache_chain_mutex);
4068 /* Now make sure this entry will be retried */
4069 m->count = m->size;
4070 return 0;
4071 }
4072 for (i = 0; i < n[1]; i++) {
4073 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4074 show_symbol(m, n[2*i+2]);
4075 seq_putc(m, '\n');
4076 }
4077 return 0;
4078}
4079
4080struct seq_operations slabstats_op = {
4081 .start = leaks_start,
4082 .next = s_next,
4083 .stop = s_stop,
4084 .show = leaks_show,
4085};
4086#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004087#endif
4088
Manfred Spraul00e145b2005-09-03 15:55:07 -07004089/**
4090 * ksize - get the actual amount of memory allocated for a given object
4091 * @objp: Pointer to the object
4092 *
4093 * kmalloc may internally round up allocations and return more memory
4094 * than requested. ksize() can be used to determine the actual amount of
4095 * memory allocated. The caller may use this additional memory, even though
4096 * a smaller amount of memory was initially specified with the kmalloc call.
4097 * The caller must guarantee that objp points to a valid object previously
4098 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4099 * must not be freed during the duration of the call.
4100 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004101unsigned int ksize(const void *objp)
4102{
Manfred Spraul00e145b2005-09-03 15:55:07 -07004103 if (unlikely(objp == NULL))
4104 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004105
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08004106 return obj_size(virt_to_cache(objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004107}