blob: b1d643b5238de2f74ab3911d2bab4479c8745385 [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;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700423 unsigned long node_overflow;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800424 atomic_t allochit;
425 atomic_t allocmiss;
426 atomic_t freehit;
427 atomic_t freemiss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700428#endif
429#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800430 /*
431 * If debugging is enabled, then the allocator can add additional
432 * fields and/or padding to every object. buffer_size contains the total
433 * object size including these internal fields, the following two
434 * variables contain the offset to the user object and its size.
435 */
436 int obj_offset;
437 int obj_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700438#endif
439};
440
441#define CFLGS_OFF_SLAB (0x80000000UL)
442#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
443
444#define BATCHREFILL_LIMIT 16
Andrew Mortona737b3e2006-03-22 00:08:11 -0800445/*
446 * Optimization question: fewer reaps means less probability for unnessary
447 * cpucache drain/refill cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700448 *
Adrian Bunkdc6f3f22005-11-08 16:44:08 +0100449 * OTOH the cpuarrays can contain lots of objects,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700450 * which could lock up otherwise freeable slabs.
451 */
452#define REAPTIMEOUT_CPUC (2*HZ)
453#define REAPTIMEOUT_LIST3 (4*HZ)
454
455#if STATS
456#define STATS_INC_ACTIVE(x) ((x)->num_active++)
457#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
458#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
459#define STATS_INC_GROWN(x) ((x)->grown++)
460#define STATS_INC_REAPED(x) ((x)->reaped++)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800461#define STATS_SET_HIGH(x) \
462 do { \
463 if ((x)->num_active > (x)->high_mark) \
464 (x)->high_mark = (x)->num_active; \
465 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700466#define STATS_INC_ERR(x) ((x)->errors++)
467#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
Christoph Lametere498be72005-09-09 13:03:32 -0700468#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700469#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800470#define STATS_SET_FREEABLE(x, i) \
471 do { \
472 if ((x)->max_freeable < i) \
473 (x)->max_freeable = i; \
474 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700475#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
476#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
477#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
478#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
479#else
480#define STATS_INC_ACTIVE(x) do { } while (0)
481#define STATS_DEC_ACTIVE(x) do { } while (0)
482#define STATS_INC_ALLOCED(x) do { } while (0)
483#define STATS_INC_GROWN(x) do { } while (0)
484#define STATS_INC_REAPED(x) do { } while (0)
485#define STATS_SET_HIGH(x) do { } while (0)
486#define STATS_INC_ERR(x) do { } while (0)
487#define STATS_INC_NODEALLOCS(x) do { } while (0)
Christoph Lametere498be72005-09-09 13:03:32 -0700488#define STATS_INC_NODEFREES(x) do { } while (0)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700489#define STATS_INC_ACOVERFLOW(x) do { } while (0)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800490#define STATS_SET_FREEABLE(x, i) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700491#define STATS_INC_ALLOCHIT(x) do { } while (0)
492#define STATS_INC_ALLOCMISS(x) do { } while (0)
493#define STATS_INC_FREEHIT(x) do { } while (0)
494#define STATS_INC_FREEMISS(x) do { } while (0)
495#endif
496
497#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -0800498/*
499 * Magic nums for obj red zoning.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700500 * Placed in the first word before and the first word after an obj.
501 */
502#define RED_INACTIVE 0x5A2CF071UL /* when obj is inactive */
503#define RED_ACTIVE 0x170FC2A5UL /* when obj is active */
504
505/* ...and for poisoning */
506#define POISON_INUSE 0x5a /* for use-uninitialised poisoning */
507#define POISON_FREE 0x6b /* for use-after-free poisoning */
508#define POISON_END 0xa5 /* end-byte of poisoning */
509
Andrew Mortona737b3e2006-03-22 00:08:11 -0800510/*
511 * memory layout of objects:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700512 * 0 : objp
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800513 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
Linus Torvalds1da177e2005-04-16 15:20:36 -0700514 * the end of an object is aligned with the end of the real
515 * allocation. Catches writes behind the end of the allocation.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800516 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700517 * redzone word.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800518 * cachep->obj_offset: The real object.
519 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
Andrew Mortona737b3e2006-03-22 00:08:11 -0800520 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
521 * [BYTES_PER_WORD long]
Linus Torvalds1da177e2005-04-16 15:20:36 -0700522 */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800523static int obj_offset(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700524{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800525 return cachep->obj_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700526}
527
Pekka Enberg343e0d72006-02-01 03:05:50 -0800528static int obj_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700529{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800530 return cachep->obj_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700531}
532
Pekka Enberg343e0d72006-02-01 03:05:50 -0800533static unsigned long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700534{
535 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800536 return (unsigned long*) (objp+obj_offset(cachep)-BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700537}
538
Pekka Enberg343e0d72006-02-01 03:05:50 -0800539static unsigned long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700540{
541 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
542 if (cachep->flags & SLAB_STORE_USER)
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800543 return (unsigned long *)(objp + cachep->buffer_size -
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800544 2 * BYTES_PER_WORD);
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800545 return (unsigned long *)(objp + cachep->buffer_size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700546}
547
Pekka Enberg343e0d72006-02-01 03:05:50 -0800548static void **dbg_userword(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700549{
550 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800551 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700552}
553
554#else
555
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800556#define obj_offset(x) 0
557#define obj_size(cachep) (cachep->buffer_size)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700558#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;})
559#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;})
560#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
561
562#endif
563
564/*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800565 * Maximum size of an obj (in 2^order pages) and absolute limit for the gfp
566 * order.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700567 */
568#if defined(CONFIG_LARGE_ALLOCS)
569#define MAX_OBJ_ORDER 13 /* up to 32Mb */
570#define MAX_GFP_ORDER 13 /* up to 32Mb */
571#elif defined(CONFIG_MMU)
572#define MAX_OBJ_ORDER 5 /* 32 pages */
573#define MAX_GFP_ORDER 5 /* 32 pages */
574#else
575#define MAX_OBJ_ORDER 8 /* up to 1Mb */
576#define MAX_GFP_ORDER 8 /* up to 1Mb */
577#endif
578
579/*
580 * Do not go above this order unless 0 objects fit into the slab.
581 */
582#define BREAK_GFP_ORDER_HI 1
583#define BREAK_GFP_ORDER_LO 0
584static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
585
Andrew Mortona737b3e2006-03-22 00:08:11 -0800586/*
587 * Functions for storing/retrieving the cachep and or slab from the page
588 * allocator. These are used to find the slab an obj belongs to. With kfree(),
589 * these are used to find the cache which an obj belongs to.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700590 */
Pekka Enberg065d41c2005-11-13 16:06:46 -0800591static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
592{
593 page->lru.next = (struct list_head *)cache;
594}
595
596static inline struct kmem_cache *page_get_cache(struct page *page)
597{
Nick Piggin84097512006-03-22 00:08:34 -0800598 if (unlikely(PageCompound(page)))
599 page = (struct page *)page_private(page);
Pekka Enberg065d41c2005-11-13 16:06:46 -0800600 return (struct kmem_cache *)page->lru.next;
601}
602
603static inline void page_set_slab(struct page *page, struct slab *slab)
604{
605 page->lru.prev = (struct list_head *)slab;
606}
607
608static inline struct slab *page_get_slab(struct page *page)
609{
Nick Piggin84097512006-03-22 00:08:34 -0800610 if (unlikely(PageCompound(page)))
611 page = (struct page *)page_private(page);
Pekka Enberg065d41c2005-11-13 16:06:46 -0800612 return (struct slab *)page->lru.prev;
613}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700614
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800615static inline struct kmem_cache *virt_to_cache(const void *obj)
616{
617 struct page *page = virt_to_page(obj);
618 return page_get_cache(page);
619}
620
621static inline struct slab *virt_to_slab(const void *obj)
622{
623 struct page *page = virt_to_page(obj);
624 return page_get_slab(page);
625}
626
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800627static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
628 unsigned int idx)
629{
630 return slab->s_mem + cache->buffer_size * idx;
631}
632
633static inline unsigned int obj_to_index(struct kmem_cache *cache,
634 struct slab *slab, void *obj)
635{
636 return (unsigned)(obj - slab->s_mem) / cache->buffer_size;
637}
638
Andrew Mortona737b3e2006-03-22 00:08:11 -0800639/*
640 * These are the default caches for kmalloc. Custom caches can have other sizes.
641 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700642struct cache_sizes malloc_sizes[] = {
643#define CACHE(x) { .cs_size = (x) },
644#include <linux/kmalloc_sizes.h>
645 CACHE(ULONG_MAX)
646#undef CACHE
647};
648EXPORT_SYMBOL(malloc_sizes);
649
650/* Must match cache_sizes above. Out of line to keep cache footprint low. */
651struct cache_names {
652 char *name;
653 char *name_dma;
654};
655
656static struct cache_names __initdata cache_names[] = {
657#define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
658#include <linux/kmalloc_sizes.h>
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800659 {NULL,}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700660#undef CACHE
661};
662
663static struct arraycache_init initarray_cache __initdata =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800664 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700665static struct arraycache_init initarray_generic =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800666 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700667
668/* internal cache of cache description objs */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800669static struct kmem_cache cache_cache = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800670 .batchcount = 1,
671 .limit = BOOT_CPUCACHE_ENTRIES,
672 .shared = 1,
Pekka Enberg343e0d72006-02-01 03:05:50 -0800673 .buffer_size = sizeof(struct kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800674 .name = "kmem_cache",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700675#if DEBUG
Pekka Enberg343e0d72006-02-01 03:05:50 -0800676 .obj_size = sizeof(struct kmem_cache),
Linus Torvalds1da177e2005-04-16 15:20:36 -0700677#endif
678};
679
680/* Guard access to the cache-chain. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800681static DEFINE_MUTEX(cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700682static struct list_head cache_chain;
683
684/*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800685 * vm_enough_memory() looks at this to determine how many slab-allocated pages
686 * are possibly freeable under pressure
Linus Torvalds1da177e2005-04-16 15:20:36 -0700687 *
688 * SLAB_RECLAIM_ACCOUNT turns this on per-slab
689 */
690atomic_t slab_reclaim_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700691
692/*
693 * chicken and egg problem: delay the per-cpu array allocation
694 * until the general caches are up.
695 */
696static enum {
697 NONE,
Christoph Lametere498be72005-09-09 13:03:32 -0700698 PARTIAL_AC,
699 PARTIAL_L3,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700700 FULL
701} g_cpucache_up;
702
Mike Kravetz39d24e62006-05-15 09:44:13 -0700703/*
704 * used by boot code to determine if it can use slab based allocator
705 */
706int slab_is_available(void)
707{
708 return g_cpucache_up == FULL;
709}
710
Linus Torvalds1da177e2005-04-16 15:20:36 -0700711static DEFINE_PER_CPU(struct work_struct, reap_work);
712
Andrew Mortona737b3e2006-03-22 00:08:11 -0800713static void free_block(struct kmem_cache *cachep, void **objpp, int len,
714 int node);
Pekka Enberg343e0d72006-02-01 03:05:50 -0800715static void enable_cpucache(struct kmem_cache *cachep);
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800716static void cache_reap(void *unused);
Pekka Enberg343e0d72006-02-01 03:05:50 -0800717static int __node_shrink(struct kmem_cache *cachep, int node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700718
Pekka Enberg343e0d72006-02-01 03:05:50 -0800719static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700720{
721 return cachep->array[smp_processor_id()];
722}
723
Andrew Mortona737b3e2006-03-22 00:08:11 -0800724static inline struct kmem_cache *__find_general_cachep(size_t size,
725 gfp_t gfpflags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700726{
727 struct cache_sizes *csizep = malloc_sizes;
728
729#if DEBUG
730 /* This happens if someone tries to call
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800731 * kmem_cache_create(), or __kmalloc(), before
732 * the generic caches are initialized.
733 */
Alok Katariac7e43c72005-09-14 12:17:53 -0700734 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700735#endif
736 while (size > csizep->cs_size)
737 csizep++;
738
739 /*
Martin Hicks0abf40c2005-09-03 15:54:54 -0700740 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
Linus Torvalds1da177e2005-04-16 15:20:36 -0700741 * has cs_{dma,}cachep==NULL. Thus no special case
742 * for large kmalloc calls required.
743 */
744 if (unlikely(gfpflags & GFP_DMA))
745 return csizep->cs_dmacachep;
746 return csizep->cs_cachep;
747}
748
Pekka Enberg343e0d72006-02-01 03:05:50 -0800749struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700750{
751 return __find_general_cachep(size, gfpflags);
752}
753EXPORT_SYMBOL(kmem_find_general_cachep);
754
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800755static size_t slab_mgmt_size(size_t nr_objs, size_t align)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700756{
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800757 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
758}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700759
Andrew Mortona737b3e2006-03-22 00:08:11 -0800760/*
761 * Calculate the number of objects and left-over bytes for a given buffer size.
762 */
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800763static void cache_estimate(unsigned long gfporder, size_t buffer_size,
764 size_t align, int flags, size_t *left_over,
765 unsigned int *num)
766{
767 int nr_objs;
768 size_t mgmt_size;
769 size_t slab_size = PAGE_SIZE << gfporder;
770
771 /*
772 * The slab management structure can be either off the slab or
773 * on it. For the latter case, the memory allocated for a
774 * slab is used for:
775 *
776 * - The struct slab
777 * - One kmem_bufctl_t for each object
778 * - Padding to respect alignment of @align
779 * - @buffer_size bytes for each object
780 *
781 * If the slab management structure is off the slab, then the
782 * alignment will already be calculated into the size. Because
783 * the slabs are all pages aligned, the objects will be at the
784 * correct alignment when allocated.
785 */
786 if (flags & CFLGS_OFF_SLAB) {
787 mgmt_size = 0;
788 nr_objs = slab_size / buffer_size;
789
790 if (nr_objs > SLAB_LIMIT)
791 nr_objs = SLAB_LIMIT;
792 } else {
793 /*
794 * Ignore padding for the initial guess. The padding
795 * is at most @align-1 bytes, and @buffer_size is at
796 * least @align. In the worst case, this result will
797 * be one greater than the number of objects that fit
798 * into the memory allocation when taking the padding
799 * into account.
800 */
801 nr_objs = (slab_size - sizeof(struct slab)) /
802 (buffer_size + sizeof(kmem_bufctl_t));
803
804 /*
805 * This calculated number will be either the right
806 * amount, or one greater than what we want.
807 */
808 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
809 > slab_size)
810 nr_objs--;
811
812 if (nr_objs > SLAB_LIMIT)
813 nr_objs = SLAB_LIMIT;
814
815 mgmt_size = slab_mgmt_size(nr_objs, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700816 }
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800817 *num = nr_objs;
818 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700819}
820
821#define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
822
Andrew Mortona737b3e2006-03-22 00:08:11 -0800823static void __slab_error(const char *function, struct kmem_cache *cachep,
824 char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700825{
826 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800827 function, cachep->name, msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700828 dump_stack();
829}
830
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800831#ifdef CONFIG_NUMA
832/*
833 * Special reaping functions for NUMA systems called from cache_reap().
834 * These take care of doing round robin flushing of alien caches (containing
835 * objects freed on different nodes from which they were allocated) and the
836 * flushing of remote pcps by calling drain_node_pages.
837 */
838static DEFINE_PER_CPU(unsigned long, reap_node);
839
840static void init_reap_node(int cpu)
841{
842 int node;
843
844 node = next_node(cpu_to_node(cpu), node_online_map);
845 if (node == MAX_NUMNODES)
Paul Jackson442295c2006-03-22 00:09:11 -0800846 node = first_node(node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800847
848 __get_cpu_var(reap_node) = node;
849}
850
851static void next_reap_node(void)
852{
853 int node = __get_cpu_var(reap_node);
854
855 /*
856 * Also drain per cpu pages on remote zones
857 */
858 if (node != numa_node_id())
859 drain_node_pages(node);
860
861 node = next_node(node, node_online_map);
862 if (unlikely(node >= MAX_NUMNODES))
863 node = first_node(node_online_map);
864 __get_cpu_var(reap_node) = node;
865}
866
867#else
868#define init_reap_node(cpu) do { } while (0)
869#define next_reap_node(void) do { } while (0)
870#endif
871
Linus Torvalds1da177e2005-04-16 15:20:36 -0700872/*
873 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
874 * via the workqueue/eventd.
875 * Add the CPU number into the expiration time to minimize the possibility of
876 * the CPUs getting into lockstep and contending for the global cache chain
877 * lock.
878 */
879static void __devinit start_cpu_timer(int cpu)
880{
881 struct work_struct *reap_work = &per_cpu(reap_work, cpu);
882
883 /*
884 * When this gets called from do_initcalls via cpucache_init(),
885 * init_workqueues() has already run, so keventd will be setup
886 * at that time.
887 */
888 if (keventd_up() && reap_work->func == NULL) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800889 init_reap_node(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700890 INIT_WORK(reap_work, cache_reap, NULL);
891 schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu);
892 }
893}
894
Christoph Lametere498be72005-09-09 13:03:32 -0700895static struct array_cache *alloc_arraycache(int node, int entries,
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800896 int batchcount)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700897{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800898 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700899 struct array_cache *nc = NULL;
900
Christoph Lametere498be72005-09-09 13:03:32 -0700901 nc = kmalloc_node(memsize, GFP_KERNEL, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700902 if (nc) {
903 nc->avail = 0;
904 nc->limit = entries;
905 nc->batchcount = batchcount;
906 nc->touched = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700907 spin_lock_init(&nc->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700908 }
909 return nc;
910}
911
Christoph Lameter3ded1752006-03-25 03:06:44 -0800912/*
913 * Transfer objects in one arraycache to another.
914 * Locking must be handled by the caller.
915 *
916 * Return the number of entries transferred.
917 */
918static int transfer_objects(struct array_cache *to,
919 struct array_cache *from, unsigned int max)
920{
921 /* Figure out how many entries to transfer */
922 int nr = min(min(from->avail, max), to->limit - to->avail);
923
924 if (!nr)
925 return 0;
926
927 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
928 sizeof(void *) *nr);
929
930 from->avail -= nr;
931 to->avail += nr;
932 to->touched = 1;
933 return nr;
934}
935
Christoph Lametere498be72005-09-09 13:03:32 -0700936#ifdef CONFIG_NUMA
Pekka Enberg343e0d72006-02-01 03:05:50 -0800937static void *__cache_alloc_node(struct kmem_cache *, gfp_t, int);
Paul Jacksonc61afb12006-03-24 03:16:08 -0800938static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
Christoph Lameterdc85da12006-01-18 17:42:36 -0800939
Pekka Enberg5295a742006-02-01 03:05:48 -0800940static struct array_cache **alloc_alien_cache(int node, int limit)
Christoph Lametere498be72005-09-09 13:03:32 -0700941{
942 struct array_cache **ac_ptr;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800943 int memsize = sizeof(void *) * MAX_NUMNODES;
Christoph Lametere498be72005-09-09 13:03:32 -0700944 int i;
945
946 if (limit > 1)
947 limit = 12;
948 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
949 if (ac_ptr) {
950 for_each_node(i) {
951 if (i == node || !node_online(i)) {
952 ac_ptr[i] = NULL;
953 continue;
954 }
955 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
956 if (!ac_ptr[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800957 for (i--; i <= 0; i--)
Christoph Lametere498be72005-09-09 13:03:32 -0700958 kfree(ac_ptr[i]);
959 kfree(ac_ptr);
960 return NULL;
961 }
962 }
963 }
964 return ac_ptr;
965}
966
Pekka Enberg5295a742006-02-01 03:05:48 -0800967static void free_alien_cache(struct array_cache **ac_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -0700968{
969 int i;
970
971 if (!ac_ptr)
972 return;
Christoph Lametere498be72005-09-09 13:03:32 -0700973 for_each_node(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800974 kfree(ac_ptr[i]);
Christoph Lametere498be72005-09-09 13:03:32 -0700975 kfree(ac_ptr);
976}
977
Pekka Enberg343e0d72006-02-01 03:05:50 -0800978static void __drain_alien_cache(struct kmem_cache *cachep,
Pekka Enberg5295a742006-02-01 03:05:48 -0800979 struct array_cache *ac, int node)
Christoph Lametere498be72005-09-09 13:03:32 -0700980{
981 struct kmem_list3 *rl3 = cachep->nodelists[node];
982
983 if (ac->avail) {
984 spin_lock(&rl3->list_lock);
Christoph Lametere00946f2006-03-25 03:06:45 -0800985 /*
986 * Stuff objects into the remote nodes shared array first.
987 * That way we could avoid the overhead of putting the objects
988 * into the free lists and getting them back later.
989 */
shin, jacob693f7d32006-04-28 10:54:37 -0500990 if (rl3->shared)
991 transfer_objects(rl3->shared, ac, ac->limit);
Christoph Lametere00946f2006-03-25 03:06:45 -0800992
Christoph Lameterff694162005-09-22 21:44:02 -0700993 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -0700994 ac->avail = 0;
995 spin_unlock(&rl3->list_lock);
996 }
997}
998
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800999/*
1000 * Called from cache_reap() to regularly drain alien caches round robin.
1001 */
1002static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1003{
1004 int node = __get_cpu_var(reap_node);
1005
1006 if (l3->alien) {
1007 struct array_cache *ac = l3->alien[node];
Christoph Lametere00946f2006-03-25 03:06:45 -08001008
1009 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001010 __drain_alien_cache(cachep, ac, node);
1011 spin_unlock_irq(&ac->lock);
1012 }
1013 }
1014}
1015
Andrew Mortona737b3e2006-03-22 00:08:11 -08001016static void drain_alien_cache(struct kmem_cache *cachep,
1017 struct array_cache **alien)
Christoph Lametere498be72005-09-09 13:03:32 -07001018{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001019 int i = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07001020 struct array_cache *ac;
1021 unsigned long flags;
1022
1023 for_each_online_node(i) {
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001024 ac = alien[i];
Christoph Lametere498be72005-09-09 13:03:32 -07001025 if (ac) {
1026 spin_lock_irqsave(&ac->lock, flags);
1027 __drain_alien_cache(cachep, ac, i);
1028 spin_unlock_irqrestore(&ac->lock, flags);
1029 }
1030 }
1031}
1032#else
Linus Torvalds7a21ef62006-02-05 11:26:38 -08001033
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001034#define drain_alien_cache(cachep, alien) do { } while (0)
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001035#define reap_alien(cachep, l3) do { } while (0)
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001036
Linus Torvalds7a21ef62006-02-05 11:26:38 -08001037static inline struct array_cache **alloc_alien_cache(int node, int limit)
1038{
1039 return (struct array_cache **) 0x01020304ul;
1040}
1041
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001042static inline void free_alien_cache(struct array_cache **ac_ptr)
1043{
1044}
Linus Torvalds7a21ef62006-02-05 11:26:38 -08001045
Christoph Lametere498be72005-09-09 13:03:32 -07001046#endif
1047
Chandra Seetharaman83d722f2006-04-24 19:35:21 -07001048static int cpuup_callback(struct notifier_block *nfb,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001049 unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001050{
1051 long cpu = (long)hcpu;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001052 struct kmem_cache *cachep;
Christoph Lametere498be72005-09-09 13:03:32 -07001053 struct kmem_list3 *l3 = NULL;
1054 int node = cpu_to_node(cpu);
1055 int memsize = sizeof(struct kmem_list3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001056
1057 switch (action) {
1058 case CPU_UP_PREPARE:
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001059 mutex_lock(&cache_chain_mutex);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001060 /*
1061 * We need to do this right in the beginning since
Christoph Lametere498be72005-09-09 13:03:32 -07001062 * alloc_arraycache's are going to use this list.
1063 * kmalloc_node allows us to add the slab to the right
1064 * kmem_list3 and not this cpu's kmem_list3
1065 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001066
Christoph Lametere498be72005-09-09 13:03:32 -07001067 list_for_each_entry(cachep, &cache_chain, next) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08001068 /*
1069 * Set up the size64 kmemlist for cpu before we can
Christoph Lametere498be72005-09-09 13:03:32 -07001070 * begin anything. Make sure some other cpu on this
1071 * node has not already allocated this
1072 */
1073 if (!cachep->nodelists[node]) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08001074 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1075 if (!l3)
Christoph Lametere498be72005-09-09 13:03:32 -07001076 goto bad;
1077 kmem_list3_init(l3);
1078 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001079 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -07001080
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001081 /*
1082 * The l3s don't come and go as CPUs come and
1083 * go. cache_chain_mutex is sufficient
1084 * protection here.
1085 */
Christoph Lametere498be72005-09-09 13:03:32 -07001086 cachep->nodelists[node] = l3;
1087 }
1088
1089 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1090 cachep->nodelists[node]->free_limit =
Andrew Mortona737b3e2006-03-22 00:08:11 -08001091 (1 + nr_cpus_node(node)) *
1092 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07001093 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1094 }
1095
Andrew Mortona737b3e2006-03-22 00:08:11 -08001096 /*
1097 * Now we can go ahead with allocating the shared arrays and
1098 * array caches
1099 */
Christoph Lametere498be72005-09-09 13:03:32 -07001100 list_for_each_entry(cachep, &cache_chain, next) {
Tobias Klausercd105df2006-01-08 01:00:59 -08001101 struct array_cache *nc;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001102 struct array_cache *shared;
1103 struct array_cache **alien;
Tobias Klausercd105df2006-01-08 01:00:59 -08001104
Christoph Lametere498be72005-09-09 13:03:32 -07001105 nc = alloc_arraycache(node, cachep->limit,
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001106 cachep->batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001107 if (!nc)
1108 goto bad;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001109 shared = alloc_arraycache(node,
1110 cachep->shared * cachep->batchcount,
1111 0xbaadf00d);
1112 if (!shared)
1113 goto bad;
Linus Torvalds7a21ef62006-02-05 11:26:38 -08001114
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001115 alien = alloc_alien_cache(node, cachep->limit);
1116 if (!alien)
1117 goto bad;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001118 cachep->array[cpu] = nc;
Christoph Lametere498be72005-09-09 13:03:32 -07001119 l3 = cachep->nodelists[node];
1120 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07001121
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001122 spin_lock_irq(&l3->list_lock);
1123 if (!l3->shared) {
1124 /*
1125 * We are serialised from CPU_DEAD or
1126 * CPU_UP_CANCELLED by the cpucontrol lock
1127 */
1128 l3->shared = shared;
1129 shared = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07001130 }
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001131#ifdef CONFIG_NUMA
1132 if (!l3->alien) {
1133 l3->alien = alien;
1134 alien = NULL;
1135 }
1136#endif
1137 spin_unlock_irq(&l3->list_lock);
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001138 kfree(shared);
1139 free_alien_cache(alien);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001140 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001141 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001142 break;
1143 case CPU_ONLINE:
1144 start_cpu_timer(cpu);
1145 break;
1146#ifdef CONFIG_HOTPLUG_CPU
1147 case CPU_DEAD:
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001148 /*
1149 * Even if all the cpus of a node are down, we don't free the
1150 * kmem_list3 of any cache. This to avoid a race between
1151 * cpu_down, and a kmalloc allocation from another cpu for
1152 * memory from the node of the cpu going down. The list3
1153 * structure is usually allocated from kmem_cache_create() and
1154 * gets destroyed at kmem_cache_destroy().
1155 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001156 /* fall thru */
1157 case CPU_UP_CANCELED:
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001158 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001159 list_for_each_entry(cachep, &cache_chain, next) {
1160 struct array_cache *nc;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001161 struct array_cache *shared;
1162 struct array_cache **alien;
Christoph Lametere498be72005-09-09 13:03:32 -07001163 cpumask_t mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001164
Christoph Lametere498be72005-09-09 13:03:32 -07001165 mask = node_to_cpumask(node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001166 /* cpu is dead; no one can alloc from it. */
1167 nc = cachep->array[cpu];
1168 cachep->array[cpu] = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07001169 l3 = cachep->nodelists[node];
1170
1171 if (!l3)
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001172 goto free_array_cache;
Christoph Lametere498be72005-09-09 13:03:32 -07001173
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08001174 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07001175
1176 /* Free limit for this kmem_list3 */
1177 l3->free_limit -= cachep->batchcount;
1178 if (nc)
Christoph Lameterff694162005-09-22 21:44:02 -07001179 free_block(cachep, nc->entry, nc->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001180
1181 if (!cpus_empty(mask)) {
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08001182 spin_unlock_irq(&l3->list_lock);
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001183 goto free_array_cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001184 }
Christoph Lametere498be72005-09-09 13:03:32 -07001185
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001186 shared = l3->shared;
1187 if (shared) {
Christoph Lametere498be72005-09-09 13:03:32 -07001188 free_block(cachep, l3->shared->entry,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001189 l3->shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001190 l3->shared = NULL;
1191 }
Christoph Lametere498be72005-09-09 13:03:32 -07001192
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001193 alien = l3->alien;
1194 l3->alien = NULL;
1195
1196 spin_unlock_irq(&l3->list_lock);
1197
1198 kfree(shared);
1199 if (alien) {
1200 drain_alien_cache(cachep, alien);
1201 free_alien_cache(alien);
Christoph Lametere498be72005-09-09 13:03:32 -07001202 }
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001203free_array_cache:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001204 kfree(nc);
1205 }
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001206 /*
1207 * In the previous loop, all the objects were freed to
1208 * the respective cache's slabs, now we can go ahead and
1209 * shrink each nodelist to its limit.
1210 */
1211 list_for_each_entry(cachep, &cache_chain, next) {
1212 l3 = cachep->nodelists[node];
1213 if (!l3)
1214 continue;
1215 spin_lock_irq(&l3->list_lock);
1216 /* free slabs belonging to this node */
1217 __node_shrink(cachep, node);
1218 spin_unlock_irq(&l3->list_lock);
1219 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001220 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001221 break;
1222#endif
1223 }
1224 return NOTIFY_OK;
Andrew Mortona737b3e2006-03-22 00:08:11 -08001225bad:
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001226 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001227 return NOTIFY_BAD;
1228}
1229
1230static struct notifier_block cpucache_notifier = { &cpuup_callback, NULL, 0 };
1231
Christoph Lametere498be72005-09-09 13:03:32 -07001232/*
1233 * swap the static kmem_list3 with kmalloced memory
1234 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001235static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1236 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001237{
1238 struct kmem_list3 *ptr;
1239
1240 BUG_ON(cachep->nodelists[nodeid] != list);
1241 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1242 BUG_ON(!ptr);
1243
1244 local_irq_disable();
1245 memcpy(ptr, list, sizeof(struct kmem_list3));
1246 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1247 cachep->nodelists[nodeid] = ptr;
1248 local_irq_enable();
1249}
1250
Andrew Mortona737b3e2006-03-22 00:08:11 -08001251/*
1252 * Initialisation. Called after the page allocator have been initialised and
1253 * before smp_init().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001254 */
1255void __init kmem_cache_init(void)
1256{
1257 size_t left_over;
1258 struct cache_sizes *sizes;
1259 struct cache_names *names;
Christoph Lametere498be72005-09-09 13:03:32 -07001260 int i;
Jack Steiner07ed76b2006-03-07 21:55:46 -08001261 int order;
Christoph Lametere498be72005-09-09 13:03:32 -07001262
1263 for (i = 0; i < NUM_INIT_LISTS; i++) {
1264 kmem_list3_init(&initkmem_list3[i]);
1265 if (i < MAX_NUMNODES)
1266 cache_cache.nodelists[i] = NULL;
1267 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001268
1269 /*
1270 * Fragmentation resistance on low memory - only use bigger
1271 * page orders on machines with more than 32MB of memory.
1272 */
1273 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1274 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1275
Linus Torvalds1da177e2005-04-16 15:20:36 -07001276 /* Bootstrap is tricky, because several objects are allocated
1277 * from caches that do not exist yet:
Andrew Mortona737b3e2006-03-22 00:08:11 -08001278 * 1) initialize the cache_cache cache: it contains the struct
1279 * kmem_cache structures of all caches, except cache_cache itself:
1280 * cache_cache is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001281 * Initially an __init data area is used for the head array and the
1282 * kmem_list3 structures, it's replaced with a kmalloc allocated
1283 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001284 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001285 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001286 * An __init data area is used for the head array.
1287 * 3) Create the remaining kmalloc caches, with minimally sized
1288 * head arrays.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001289 * 4) Replace the __init data head arrays for cache_cache and the first
1290 * kmalloc cache with kmalloc allocated arrays.
Christoph Lametere498be72005-09-09 13:03:32 -07001291 * 5) Replace the __init data for kmem_list3 for cache_cache and
1292 * the other cache's with kmalloc allocated memory.
1293 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001294 */
1295
1296 /* 1) create the cache_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001297 INIT_LIST_HEAD(&cache_chain);
1298 list_add(&cache_cache.next, &cache_chain);
1299 cache_cache.colour_off = cache_line_size();
1300 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
Christoph Lametere498be72005-09-09 13:03:32 -07001301 cache_cache.nodelists[numa_node_id()] = &initkmem_list3[CACHE_CACHE];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001302
Andrew Mortona737b3e2006-03-22 00:08:11 -08001303 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1304 cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001305
Jack Steiner07ed76b2006-03-07 21:55:46 -08001306 for (order = 0; order < MAX_ORDER; order++) {
1307 cache_estimate(order, cache_cache.buffer_size,
1308 cache_line_size(), 0, &left_over, &cache_cache.num);
1309 if (cache_cache.num)
1310 break;
1311 }
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02001312 BUG_ON(!cache_cache.num);
Jack Steiner07ed76b2006-03-07 21:55:46 -08001313 cache_cache.gfporder = order;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001314 cache_cache.colour = left_over / cache_cache.colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001315 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1316 sizeof(struct slab), cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001317
1318 /* 2+3) create the kmalloc caches */
1319 sizes = malloc_sizes;
1320 names = cache_names;
1321
Andrew Mortona737b3e2006-03-22 00:08:11 -08001322 /*
1323 * Initialize the caches that provide memory for the array cache and the
1324 * kmem_list3 structures first. Without this, further allocations will
1325 * bug.
Christoph Lametere498be72005-09-09 13:03:32 -07001326 */
1327
1328 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001329 sizes[INDEX_AC].cs_size,
1330 ARCH_KMALLOC_MINALIGN,
1331 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1332 NULL, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07001333
Andrew Mortona737b3e2006-03-22 00:08:11 -08001334 if (INDEX_AC != INDEX_L3) {
Christoph Lametere498be72005-09-09 13:03:32 -07001335 sizes[INDEX_L3].cs_cachep =
Andrew Mortona737b3e2006-03-22 00:08:11 -08001336 kmem_cache_create(names[INDEX_L3].name,
1337 sizes[INDEX_L3].cs_size,
1338 ARCH_KMALLOC_MINALIGN,
1339 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1340 NULL, NULL);
1341 }
Christoph Lametere498be72005-09-09 13:03:32 -07001342
Linus Torvalds1da177e2005-04-16 15:20:36 -07001343 while (sizes->cs_size != ULONG_MAX) {
Christoph Lametere498be72005-09-09 13:03:32 -07001344 /*
1345 * For performance, all the general caches are L1 aligned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001346 * This should be particularly beneficial on SMP boxes, as it
1347 * eliminates "false sharing".
1348 * Note for systems short on memory removing the alignment will
Christoph Lametere498be72005-09-09 13:03:32 -07001349 * allow tighter packing of the smaller caches.
1350 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001351 if (!sizes->cs_cachep) {
Christoph Lametere498be72005-09-09 13:03:32 -07001352 sizes->cs_cachep = kmem_cache_create(names->name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001353 sizes->cs_size,
1354 ARCH_KMALLOC_MINALIGN,
1355 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1356 NULL, NULL);
1357 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001358
1359 /* Inc off-slab bufctl limit until the ceiling is hit. */
1360 if (!(OFF_SLAB(sizes->cs_cachep))) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001361 offslab_limit = sizes->cs_size - sizeof(struct slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001362 offslab_limit /= sizeof(kmem_bufctl_t);
1363 }
1364
1365 sizes->cs_dmacachep = kmem_cache_create(names->name_dma,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001366 sizes->cs_size,
1367 ARCH_KMALLOC_MINALIGN,
1368 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1369 SLAB_PANIC,
1370 NULL, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001371 sizes++;
1372 names++;
1373 }
1374 /* 4) Replace the bootstrap head arrays */
1375 {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001376 void *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001377
Linus Torvalds1da177e2005-04-16 15:20:36 -07001378 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001379
Linus Torvalds1da177e2005-04-16 15:20:36 -07001380 local_irq_disable();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001381 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1382 memcpy(ptr, cpu_cache_get(&cache_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001383 sizeof(struct arraycache_init));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001384 cache_cache.array[smp_processor_id()] = ptr;
1385 local_irq_enable();
Christoph Lametere498be72005-09-09 13:03:32 -07001386
Linus Torvalds1da177e2005-04-16 15:20:36 -07001387 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001388
Linus Torvalds1da177e2005-04-16 15:20:36 -07001389 local_irq_disable();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001390 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001391 != &initarray_generic.cache);
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001392 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001393 sizeof(struct arraycache_init));
Christoph Lametere498be72005-09-09 13:03:32 -07001394 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001395 ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001396 local_irq_enable();
1397 }
Christoph Lametere498be72005-09-09 13:03:32 -07001398 /* 5) Replace the bootstrap kmem_list3's */
1399 {
1400 int node;
1401 /* Replace the static kmem_list3 structures for the boot cpu */
1402 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE],
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001403 numa_node_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001404
Christoph Lametere498be72005-09-09 13:03:32 -07001405 for_each_online_node(node) {
1406 init_list(malloc_sizes[INDEX_AC].cs_cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001407 &initkmem_list3[SIZE_AC + node], node);
Christoph Lametere498be72005-09-09 13:03:32 -07001408
1409 if (INDEX_AC != INDEX_L3) {
1410 init_list(malloc_sizes[INDEX_L3].cs_cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001411 &initkmem_list3[SIZE_L3 + node],
1412 node);
Christoph Lametere498be72005-09-09 13:03:32 -07001413 }
1414 }
1415 }
1416
1417 /* 6) resize the head arrays to their final sizes */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001418 {
Pekka Enberg343e0d72006-02-01 03:05:50 -08001419 struct kmem_cache *cachep;
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001420 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001421 list_for_each_entry(cachep, &cache_chain, next)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001422 enable_cpucache(cachep);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001423 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001424 }
1425
1426 /* Done! */
1427 g_cpucache_up = FULL;
1428
Andrew Mortona737b3e2006-03-22 00:08:11 -08001429 /*
1430 * Register a cpu startup notifier callback that initializes
1431 * cpu_cache_get for all new cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07001432 */
1433 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001434
Andrew Mortona737b3e2006-03-22 00:08:11 -08001435 /*
1436 * The reap timers are started later, with a module init call: That part
1437 * of the kernel is not yet operational.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001438 */
1439}
1440
1441static int __init cpucache_init(void)
1442{
1443 int cpu;
1444
Andrew Mortona737b3e2006-03-22 00:08:11 -08001445 /*
1446 * Register the timers that return unneeded pages to the page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -07001447 */
Christoph Lametere498be72005-09-09 13:03:32 -07001448 for_each_online_cpu(cpu)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001449 start_cpu_timer(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001450 return 0;
1451}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001452__initcall(cpucache_init);
1453
1454/*
1455 * Interface to system's page allocator. No need to hold the cache-lock.
1456 *
1457 * If we requested dmaable memory, we will get it. Even if we
1458 * did not request dmaable memory, we might get it, but that
1459 * would be relatively rare and ignorable.
1460 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001461static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001462{
1463 struct page *page;
1464 void *addr;
1465 int i;
1466
1467 flags |= cachep->gfpflags;
Luke Yangd6fef9d2006-04-10 22:52:56 -07001468#ifndef CONFIG_MMU
1469 /* nommu uses slab's for process anonymous memory allocations, so
1470 * requires __GFP_COMP to properly refcount higher order allocations"
1471 */
1472 page = alloc_pages_node(nodeid, (flags | __GFP_COMP), cachep->gfporder);
1473#else
Christoph Lameter50c85a12005-11-13 16:06:47 -08001474 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
Luke Yangd6fef9d2006-04-10 22:52:56 -07001475#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001476 if (!page)
1477 return NULL;
1478 addr = page_address(page);
1479
1480 i = (1 << cachep->gfporder);
1481 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1482 atomic_add(i, &slab_reclaim_pages);
1483 add_page_state(nr_slab, i);
1484 while (i--) {
Nick Pigginf205b2f2006-03-22 00:08:02 -08001485 __SetPageSlab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001486 page++;
1487 }
1488 return addr;
1489}
1490
1491/*
1492 * Interface to system's page release.
1493 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001494static void kmem_freepages(struct kmem_cache *cachep, void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001495{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001496 unsigned long i = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001497 struct page *page = virt_to_page(addr);
1498 const unsigned long nr_freed = i;
1499
1500 while (i--) {
Nick Pigginf205b2f2006-03-22 00:08:02 -08001501 BUG_ON(!PageSlab(page));
1502 __ClearPageSlab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001503 page++;
1504 }
1505 sub_page_state(nr_slab, nr_freed);
1506 if (current->reclaim_state)
1507 current->reclaim_state->reclaimed_slab += nr_freed;
1508 free_pages((unsigned long)addr, cachep->gfporder);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001509 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1510 atomic_sub(1 << cachep->gfporder, &slab_reclaim_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001511}
1512
1513static void kmem_rcu_free(struct rcu_head *head)
1514{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001515 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001516 struct kmem_cache *cachep = slab_rcu->cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001517
1518 kmem_freepages(cachep, slab_rcu->addr);
1519 if (OFF_SLAB(cachep))
1520 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1521}
1522
1523#if DEBUG
1524
1525#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001526static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001527 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001528{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001529 int size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001530
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001531 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001532
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001533 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001534 return;
1535
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001536 *addr++ = 0x12345678;
1537 *addr++ = caller;
1538 *addr++ = smp_processor_id();
1539 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001540 {
1541 unsigned long *sptr = &caller;
1542 unsigned long svalue;
1543
1544 while (!kstack_end(sptr)) {
1545 svalue = *sptr++;
1546 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001547 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001548 size -= sizeof(unsigned long);
1549 if (size <= sizeof(unsigned long))
1550 break;
1551 }
1552 }
1553
1554 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001555 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001556}
1557#endif
1558
Pekka Enberg343e0d72006-02-01 03:05:50 -08001559static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001560{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001561 int size = obj_size(cachep);
1562 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001563
1564 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001565 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001566}
1567
1568static void dump_line(char *data, int offset, int limit)
1569{
1570 int i;
1571 printk(KERN_ERR "%03x:", offset);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001572 for (i = 0; i < limit; i++)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001573 printk(" %02x", (unsigned char)data[offset + i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001574 printk("\n");
1575}
1576#endif
1577
1578#if DEBUG
1579
Pekka Enberg343e0d72006-02-01 03:05:50 -08001580static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001581{
1582 int i, size;
1583 char *realobj;
1584
1585 if (cachep->flags & SLAB_RED_ZONE) {
1586 printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001587 *dbg_redzone1(cachep, objp),
1588 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001589 }
1590
1591 if (cachep->flags & SLAB_STORE_USER) {
1592 printk(KERN_ERR "Last user: [<%p>]",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001593 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001594 print_symbol("(%s)",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001595 (unsigned long)*dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001596 printk("\n");
1597 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001598 realobj = (char *)objp + obj_offset(cachep);
1599 size = obj_size(cachep);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001600 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001601 int limit;
1602 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001603 if (i + limit > size)
1604 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001605 dump_line(realobj, i, limit);
1606 }
1607}
1608
Pekka Enberg343e0d72006-02-01 03:05:50 -08001609static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001610{
1611 char *realobj;
1612 int size, i;
1613 int lines = 0;
1614
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001615 realobj = (char *)objp + obj_offset(cachep);
1616 size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001617
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001618 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001619 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001620 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001621 exp = POISON_END;
1622 if (realobj[i] != exp) {
1623 int limit;
1624 /* Mismatch ! */
1625 /* Print header */
1626 if (lines == 0) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001627 printk(KERN_ERR
Andrew Mortona737b3e2006-03-22 00:08:11 -08001628 "Slab corruption: start=%p, len=%d\n",
1629 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001630 print_objinfo(cachep, objp, 0);
1631 }
1632 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001633 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001634 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001635 if (i + limit > size)
1636 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001637 dump_line(realobj, i, limit);
1638 i += 16;
1639 lines++;
1640 /* Limit to 5 lines */
1641 if (lines > 5)
1642 break;
1643 }
1644 }
1645 if (lines != 0) {
1646 /* Print some data about the neighboring objects, if they
1647 * exist:
1648 */
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08001649 struct slab *slabp = virt_to_slab(objp);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001650 unsigned int objnr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001651
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001652 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001653 if (objnr) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001654 objp = index_to_obj(cachep, slabp, objnr - 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001655 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001656 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001657 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001658 print_objinfo(cachep, objp, 2);
1659 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001660 if (objnr + 1 < cachep->num) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001661 objp = index_to_obj(cachep, slabp, objnr + 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001662 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001663 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001664 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001665 print_objinfo(cachep, objp, 2);
1666 }
1667 }
1668}
1669#endif
1670
Linus Torvalds1da177e2005-04-16 15:20:36 -07001671#if DEBUG
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001672/**
Randy Dunlap911851e2006-03-22 00:08:14 -08001673 * slab_destroy_objs - destroy a slab and its objects
1674 * @cachep: cache pointer being destroyed
1675 * @slabp: slab pointer being destroyed
1676 *
1677 * Call the registered destructor for each object in a slab that is being
1678 * destroyed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001679 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001680static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001681{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001682 int i;
1683 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001684 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001685
1686 if (cachep->flags & SLAB_POISON) {
1687#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08001688 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1689 OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001690 kernel_map_pages(virt_to_page(objp),
Andrew Mortona737b3e2006-03-22 00:08:11 -08001691 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001692 else
1693 check_poison_obj(cachep, objp);
1694#else
1695 check_poison_obj(cachep, objp);
1696#endif
1697 }
1698 if (cachep->flags & SLAB_RED_ZONE) {
1699 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1700 slab_error(cachep, "start of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001701 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001702 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1703 slab_error(cachep, "end of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001704 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001705 }
1706 if (cachep->dtor && !(cachep->flags & SLAB_POISON))
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001707 (cachep->dtor) (objp + obj_offset(cachep), cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001708 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001709}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001710#else
Pekka Enberg343e0d72006-02-01 03:05:50 -08001711static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001712{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001713 if (cachep->dtor) {
1714 int i;
1715 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001716 void *objp = index_to_obj(cachep, slabp, i);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001717 (cachep->dtor) (objp, cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001718 }
1719 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001720}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001721#endif
1722
Randy Dunlap911851e2006-03-22 00:08:14 -08001723/**
1724 * slab_destroy - destroy and release all objects in a slab
1725 * @cachep: cache pointer being destroyed
1726 * @slabp: slab pointer being destroyed
1727 *
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001728 * Destroy all the objs in a slab, and release the mem back to the system.
Andrew Mortona737b3e2006-03-22 00:08:11 -08001729 * Before calling the slab must have been unlinked from the cache. The
1730 * cache-lock is not held/needed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001731 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001732static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001733{
1734 void *addr = slabp->s_mem - slabp->colouroff;
1735
1736 slab_destroy_objs(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001737 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1738 struct slab_rcu *slab_rcu;
1739
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001740 slab_rcu = (struct slab_rcu *)slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001741 slab_rcu->cachep = cachep;
1742 slab_rcu->addr = addr;
1743 call_rcu(&slab_rcu->head, kmem_rcu_free);
1744 } else {
1745 kmem_freepages(cachep, addr);
1746 if (OFF_SLAB(cachep))
1747 kmem_cache_free(cachep->slabp_cache, slabp);
1748 }
1749}
1750
Andrew Mortona737b3e2006-03-22 00:08:11 -08001751/*
1752 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1753 * size of kmem_list3.
1754 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001755static void set_up_list3s(struct kmem_cache *cachep, int index)
Christoph Lametere498be72005-09-09 13:03:32 -07001756{
1757 int node;
1758
1759 for_each_online_node(node) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001760 cachep->nodelists[node] = &initkmem_list3[index + node];
Christoph Lametere498be72005-09-09 13:03:32 -07001761 cachep->nodelists[node]->next_reap = jiffies +
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001762 REAPTIMEOUT_LIST3 +
1763 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -07001764 }
1765}
1766
Linus Torvalds1da177e2005-04-16 15:20:36 -07001767/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08001768 * calculate_slab_order - calculate size (page order) of slabs
1769 * @cachep: pointer to the cache that is being created
1770 * @size: size of objects to be created in this cache.
1771 * @align: required alignment for the objects.
1772 * @flags: slab allocation flags
1773 *
1774 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001775 *
1776 * This could be made much more intelligent. For now, try to avoid using
1777 * high order pages for slabs. When the gfp() functions are more friendly
1778 * towards high-order requests, this should be changed.
1779 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001780static size_t calculate_slab_order(struct kmem_cache *cachep,
Randy Dunlapee13d782006-02-01 03:05:53 -08001781 size_t size, size_t align, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001782{
1783 size_t left_over = 0;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001784 int gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001785
Andrew Mortona737b3e2006-03-22 00:08:11 -08001786 for (gfporder = 0; gfporder <= MAX_GFP_ORDER; gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001787 unsigned int num;
1788 size_t remainder;
1789
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001790 cache_estimate(gfporder, size, align, flags, &remainder, &num);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001791 if (!num)
1792 continue;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001793
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001794 /* More than offslab_limit objects will cause problems */
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001795 if ((flags & CFLGS_OFF_SLAB) && num > offslab_limit)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001796 break;
1797
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001798 /* Found something acceptable - save it away */
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001799 cachep->num = num;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001800 cachep->gfporder = gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001801 left_over = remainder;
1802
1803 /*
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08001804 * A VFS-reclaimable slab tends to have most allocations
1805 * as GFP_NOFS and we really don't want to have to be allocating
1806 * higher-order pages when we are unable to shrink dcache.
1807 */
1808 if (flags & SLAB_RECLAIM_ACCOUNT)
1809 break;
1810
1811 /*
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001812 * Large number of objects is good, but very large slabs are
1813 * currently bad for the gfp()s.
1814 */
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001815 if (gfporder >= slab_break_gfp_order)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001816 break;
1817
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001818 /*
1819 * Acceptable internal fragmentation?
1820 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001821 if (left_over * 8 <= (PAGE_SIZE << gfporder))
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001822 break;
1823 }
1824 return left_over;
1825}
1826
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08001827static void setup_cpu_cache(struct kmem_cache *cachep)
1828{
1829 if (g_cpucache_up == FULL) {
1830 enable_cpucache(cachep);
1831 return;
1832 }
1833 if (g_cpucache_up == NONE) {
1834 /*
1835 * Note: the first kmem_cache_create must create the cache
1836 * that's used by kmalloc(24), otherwise the creation of
1837 * further caches will BUG().
1838 */
1839 cachep->array[smp_processor_id()] = &initarray_generic.cache;
1840
1841 /*
1842 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
1843 * the first cache, then we need to set up all its list3s,
1844 * otherwise the creation of further caches will BUG().
1845 */
1846 set_up_list3s(cachep, SIZE_AC);
1847 if (INDEX_AC == INDEX_L3)
1848 g_cpucache_up = PARTIAL_L3;
1849 else
1850 g_cpucache_up = PARTIAL_AC;
1851 } else {
1852 cachep->array[smp_processor_id()] =
1853 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1854
1855 if (g_cpucache_up == PARTIAL_AC) {
1856 set_up_list3s(cachep, SIZE_L3);
1857 g_cpucache_up = PARTIAL_L3;
1858 } else {
1859 int node;
1860 for_each_online_node(node) {
1861 cachep->nodelists[node] =
1862 kmalloc_node(sizeof(struct kmem_list3),
1863 GFP_KERNEL, node);
1864 BUG_ON(!cachep->nodelists[node]);
1865 kmem_list3_init(cachep->nodelists[node]);
1866 }
1867 }
1868 }
1869 cachep->nodelists[numa_node_id()]->next_reap =
1870 jiffies + REAPTIMEOUT_LIST3 +
1871 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1872
1873 cpu_cache_get(cachep)->avail = 0;
1874 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1875 cpu_cache_get(cachep)->batchcount = 1;
1876 cpu_cache_get(cachep)->touched = 0;
1877 cachep->batchcount = 1;
1878 cachep->limit = BOOT_CPUCACHE_ENTRIES;
1879}
1880
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001881/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07001882 * kmem_cache_create - Create a cache.
1883 * @name: A string which is used in /proc/slabinfo to identify this cache.
1884 * @size: The size of objects to be created in this cache.
1885 * @align: The required alignment for the objects.
1886 * @flags: SLAB flags
1887 * @ctor: A constructor for the objects.
1888 * @dtor: A destructor for the objects.
1889 *
1890 * Returns a ptr to the cache on success, NULL on failure.
1891 * Cannot be called within a int, but can be interrupted.
1892 * The @ctor is run when new pages are allocated by the cache
1893 * and the @dtor is run before the pages are handed back.
1894 *
1895 * @name must be valid until the cache is destroyed. This implies that
Andrew Mortona737b3e2006-03-22 00:08:11 -08001896 * the module calling this has to destroy the cache before getting unloaded.
1897 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001898 * The flags are
1899 *
1900 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1901 * to catch references to uninitialised memory.
1902 *
1903 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1904 * for buffer overruns.
1905 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001906 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1907 * cacheline. This can be beneficial if you're counting cycles as closely
1908 * as davem.
1909 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001910struct kmem_cache *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001911kmem_cache_create (const char *name, size_t size, size_t align,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001912 unsigned long flags,
1913 void (*ctor)(void*, struct kmem_cache *, unsigned long),
Pekka Enberg343e0d72006-02-01 03:05:50 -08001914 void (*dtor)(void*, struct kmem_cache *, unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001915{
1916 size_t left_over, slab_size, ralign;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001917 struct kmem_cache *cachep = NULL;
Andrew Morton4f12bb42005-11-07 00:58:00 -08001918 struct list_head *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001919
1920 /*
1921 * Sanity checks... these are all serious usage bugs.
1922 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001923 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001924 (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08001925 printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__,
1926 name);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001927 BUG();
1928 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001929
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08001930 /*
1931 * Prevent CPUs from coming and going.
1932 * lock_cpu_hotplug() nests outside cache_chain_mutex
1933 */
1934 lock_cpu_hotplug();
1935
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001936 mutex_lock(&cache_chain_mutex);
Andrew Morton4f12bb42005-11-07 00:58:00 -08001937
1938 list_for_each(p, &cache_chain) {
Pekka Enberg343e0d72006-02-01 03:05:50 -08001939 struct kmem_cache *pc = list_entry(p, struct kmem_cache, next);
Andrew Morton4f12bb42005-11-07 00:58:00 -08001940 mm_segment_t old_fs = get_fs();
1941 char tmp;
1942 int res;
1943
1944 /*
1945 * This happens when the module gets unloaded and doesn't
1946 * destroy its slab cache and no-one else reuses the vmalloc
1947 * area of the module. Print a warning.
1948 */
1949 set_fs(KERNEL_DS);
1950 res = __get_user(tmp, pc->name);
1951 set_fs(old_fs);
1952 if (res) {
1953 printk("SLAB: cache with size %d has lost its name\n",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001954 pc->buffer_size);
Andrew Morton4f12bb42005-11-07 00:58:00 -08001955 continue;
1956 }
1957
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001958 if (!strcmp(pc->name, name)) {
Andrew Morton4f12bb42005-11-07 00:58:00 -08001959 printk("kmem_cache_create: duplicate cache %s\n", name);
1960 dump_stack();
1961 goto oops;
1962 }
1963 }
1964
Linus Torvalds1da177e2005-04-16 15:20:36 -07001965#if DEBUG
1966 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
1967 if ((flags & SLAB_DEBUG_INITIAL) && !ctor) {
1968 /* No constructor, but inital state check requested */
1969 printk(KERN_ERR "%s: No con, but init state check "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001970 "requested - %s\n", __FUNCTION__, name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001971 flags &= ~SLAB_DEBUG_INITIAL;
1972 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001973#if FORCED_DEBUG
1974 /*
1975 * Enable redzoning and last user accounting, except for caches with
1976 * large objects, if the increased size would increase the object size
1977 * above the next power of two: caches with object sizes just above a
1978 * power of two have a significant amount of internal fragmentation.
1979 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001980 if (size < 4096 || fls(size - 1) == fls(size-1 + 3 * BYTES_PER_WORD))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001981 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001982 if (!(flags & SLAB_DESTROY_BY_RCU))
1983 flags |= SLAB_POISON;
1984#endif
1985 if (flags & SLAB_DESTROY_BY_RCU)
1986 BUG_ON(flags & SLAB_POISON);
1987#endif
1988 if (flags & SLAB_DESTROY_BY_RCU)
1989 BUG_ON(dtor);
1990
1991 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08001992 * Always checks flags, a caller might be expecting debug support which
1993 * isn't available.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001994 */
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02001995 BUG_ON(flags & ~CREATE_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001996
Andrew Mortona737b3e2006-03-22 00:08:11 -08001997 /*
1998 * Check that size is in terms of words. This is needed to avoid
Linus Torvalds1da177e2005-04-16 15:20:36 -07001999 * unaligned accesses for some archs when redzoning is used, and makes
2000 * sure any on-slab bufctl's are also correctly aligned.
2001 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002002 if (size & (BYTES_PER_WORD - 1)) {
2003 size += (BYTES_PER_WORD - 1);
2004 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002005 }
2006
Andrew Mortona737b3e2006-03-22 00:08:11 -08002007 /* calculate the final buffer alignment: */
2008
Linus Torvalds1da177e2005-04-16 15:20:36 -07002009 /* 1) arch recommendation: can be overridden for debug */
2010 if (flags & SLAB_HWCACHE_ALIGN) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002011 /*
2012 * Default alignment: as specified by the arch code. Except if
2013 * an object is really small, then squeeze multiple objects into
2014 * one cacheline.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002015 */
2016 ralign = cache_line_size();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002017 while (size <= ralign / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002018 ralign /= 2;
2019 } else {
2020 ralign = BYTES_PER_WORD;
2021 }
2022 /* 2) arch mandated alignment: disables debug if necessary */
2023 if (ralign < ARCH_SLAB_MINALIGN) {
2024 ralign = ARCH_SLAB_MINALIGN;
2025 if (ralign > BYTES_PER_WORD)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002026 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002027 }
2028 /* 3) caller mandated alignment: disables debug if necessary */
2029 if (ralign < align) {
2030 ralign = align;
2031 if (ralign > BYTES_PER_WORD)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002032 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002033 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08002034 /*
2035 * 4) Store it. Note that the debug code below can reduce
Linus Torvalds1da177e2005-04-16 15:20:36 -07002036 * the alignment to BYTES_PER_WORD.
2037 */
2038 align = ralign;
2039
2040 /* Get cache's description obj. */
Pekka Enbergc5e3b832006-03-25 03:06:43 -08002041 cachep = kmem_cache_zalloc(&cache_cache, SLAB_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002042 if (!cachep)
Andrew Morton4f12bb42005-11-07 00:58:00 -08002043 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002044
2045#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002046 cachep->obj_size = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002047
2048 if (flags & SLAB_RED_ZONE) {
2049 /* redzoning only works with word aligned caches */
2050 align = BYTES_PER_WORD;
2051
2052 /* add space for red zone words */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002053 cachep->obj_offset += BYTES_PER_WORD;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002054 size += 2 * BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002055 }
2056 if (flags & SLAB_STORE_USER) {
2057 /* user store requires word alignment and
2058 * one word storage behind the end of the real
2059 * object.
2060 */
2061 align = BYTES_PER_WORD;
2062 size += BYTES_PER_WORD;
2063 }
2064#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002065 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002066 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2067 cachep->obj_offset += PAGE_SIZE - size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002068 size = PAGE_SIZE;
2069 }
2070#endif
2071#endif
2072
2073 /* Determine if the slab management is 'on' or 'off' slab. */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002074 if (size >= (PAGE_SIZE >> 3))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002075 /*
2076 * Size is large, assume best to place the slab management obj
2077 * off-slab (should allow better packing of objs).
2078 */
2079 flags |= CFLGS_OFF_SLAB;
2080
2081 size = ALIGN(size, align);
2082
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002083 left_over = calculate_slab_order(cachep, size, align, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002084
2085 if (!cachep->num) {
2086 printk("kmem_cache_create: couldn't create cache %s.\n", name);
2087 kmem_cache_free(&cache_cache, cachep);
2088 cachep = NULL;
Andrew Morton4f12bb42005-11-07 00:58:00 -08002089 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002090 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002091 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2092 + sizeof(struct slab), align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002093
2094 /*
2095 * If the slab has been placed off-slab, and we have enough space then
2096 * move it on-slab. This is at the expense of any extra colouring.
2097 */
2098 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2099 flags &= ~CFLGS_OFF_SLAB;
2100 left_over -= slab_size;
2101 }
2102
2103 if (flags & CFLGS_OFF_SLAB) {
2104 /* really off slab. No need for manual alignment */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002105 slab_size =
2106 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002107 }
2108
2109 cachep->colour_off = cache_line_size();
2110 /* Offset must be a multiple of the alignment. */
2111 if (cachep->colour_off < align)
2112 cachep->colour_off = align;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002113 cachep->colour = left_over / cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002114 cachep->slab_size = slab_size;
2115 cachep->flags = flags;
2116 cachep->gfpflags = 0;
2117 if (flags & SLAB_CACHE_DMA)
2118 cachep->gfpflags |= GFP_DMA;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002119 cachep->buffer_size = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002120
2121 if (flags & CFLGS_OFF_SLAB)
Victor Fuscob2d55072005-09-10 00:26:36 -07002122 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002123 cachep->ctor = ctor;
2124 cachep->dtor = dtor;
2125 cachep->name = name;
2126
Linus Torvalds1da177e2005-04-16 15:20:36 -07002127
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002128 setup_cpu_cache(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002129
Linus Torvalds1da177e2005-04-16 15:20:36 -07002130 /* cache setup completed, link it into the list */
2131 list_add(&cachep->next, &cache_chain);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002132oops:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002133 if (!cachep && (flags & SLAB_PANIC))
2134 panic("kmem_cache_create(): failed to create slab `%s'\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002135 name);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002136 mutex_unlock(&cache_chain_mutex);
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002137 unlock_cpu_hotplug();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002138 return cachep;
2139}
2140EXPORT_SYMBOL(kmem_cache_create);
2141
2142#if DEBUG
2143static void check_irq_off(void)
2144{
2145 BUG_ON(!irqs_disabled());
2146}
2147
2148static void check_irq_on(void)
2149{
2150 BUG_ON(irqs_disabled());
2151}
2152
Pekka Enberg343e0d72006-02-01 03:05:50 -08002153static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002154{
2155#ifdef CONFIG_SMP
2156 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002157 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002158#endif
2159}
Christoph Lametere498be72005-09-09 13:03:32 -07002160
Pekka Enberg343e0d72006-02-01 03:05:50 -08002161static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002162{
2163#ifdef CONFIG_SMP
2164 check_irq_off();
2165 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2166#endif
2167}
2168
Linus Torvalds1da177e2005-04-16 15:20:36 -07002169#else
2170#define check_irq_off() do { } while(0)
2171#define check_irq_on() do { } while(0)
2172#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002173#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002174#endif
2175
Christoph Lameteraab22072006-03-22 00:09:06 -08002176static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2177 struct array_cache *ac,
2178 int force, int node);
2179
Linus Torvalds1da177e2005-04-16 15:20:36 -07002180static void do_drain(void *arg)
2181{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002182 struct kmem_cache *cachep = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002183 struct array_cache *ac;
Christoph Lameterff694162005-09-22 21:44:02 -07002184 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002185
2186 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002187 ac = cpu_cache_get(cachep);
Christoph Lameterff694162005-09-22 21:44:02 -07002188 spin_lock(&cachep->nodelists[node]->list_lock);
2189 free_block(cachep, ac->entry, ac->avail, node);
2190 spin_unlock(&cachep->nodelists[node]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002191 ac->avail = 0;
2192}
2193
Pekka Enberg343e0d72006-02-01 03:05:50 -08002194static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002195{
Christoph Lametere498be72005-09-09 13:03:32 -07002196 struct kmem_list3 *l3;
2197 int node;
2198
Andrew Mortona07fa392006-03-22 00:08:17 -08002199 on_each_cpu(do_drain, cachep, 1, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002200 check_irq_on();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002201 for_each_online_node(node) {
Christoph Lametere498be72005-09-09 13:03:32 -07002202 l3 = cachep->nodelists[node];
2203 if (l3) {
Christoph Lameteraab22072006-03-22 00:09:06 -08002204 drain_array(cachep, l3, l3->shared, 1, node);
Christoph Lametere498be72005-09-09 13:03:32 -07002205 if (l3->alien)
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08002206 drain_alien_cache(cachep, l3->alien);
Christoph Lametere498be72005-09-09 13:03:32 -07002207 }
2208 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002209}
2210
Pekka Enberg343e0d72006-02-01 03:05:50 -08002211static int __node_shrink(struct kmem_cache *cachep, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002212{
2213 struct slab *slabp;
Christoph Lametere498be72005-09-09 13:03:32 -07002214 struct kmem_list3 *l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002215 int ret;
2216
Christoph Lametere498be72005-09-09 13:03:32 -07002217 for (;;) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002218 struct list_head *p;
2219
Christoph Lametere498be72005-09-09 13:03:32 -07002220 p = l3->slabs_free.prev;
2221 if (p == &l3->slabs_free)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002222 break;
2223
Christoph Lametere498be72005-09-09 13:03:32 -07002224 slabp = list_entry(l3->slabs_free.prev, struct slab, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002225#if DEBUG
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002226 BUG_ON(slabp->inuse);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002227#endif
2228 list_del(&slabp->list);
2229
Christoph Lametere498be72005-09-09 13:03:32 -07002230 l3->free_objects -= cachep->num;
2231 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002232 slab_destroy(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07002233 spin_lock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002234 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002235 ret = !list_empty(&l3->slabs_full) || !list_empty(&l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002236 return ret;
2237}
2238
Pekka Enberg343e0d72006-02-01 03:05:50 -08002239static int __cache_shrink(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07002240{
2241 int ret = 0, i = 0;
2242 struct kmem_list3 *l3;
2243
2244 drain_cpu_caches(cachep);
2245
2246 check_irq_on();
2247 for_each_online_node(i) {
2248 l3 = cachep->nodelists[i];
2249 if (l3) {
2250 spin_lock_irq(&l3->list_lock);
2251 ret += __node_shrink(cachep, i);
2252 spin_unlock_irq(&l3->list_lock);
2253 }
2254 }
2255 return (ret ? 1 : 0);
2256}
2257
Linus Torvalds1da177e2005-04-16 15:20:36 -07002258/**
2259 * kmem_cache_shrink - Shrink a cache.
2260 * @cachep: The cache to shrink.
2261 *
2262 * Releases as many slabs as possible for a cache.
2263 * To help debugging, a zero exit status indicates all slabs were released.
2264 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002265int kmem_cache_shrink(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002266{
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002267 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002268
2269 return __cache_shrink(cachep);
2270}
2271EXPORT_SYMBOL(kmem_cache_shrink);
2272
2273/**
2274 * kmem_cache_destroy - delete a cache
2275 * @cachep: the cache to destroy
2276 *
Pekka Enberg343e0d72006-02-01 03:05:50 -08002277 * Remove a struct kmem_cache object from the slab cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002278 * Returns 0 on success.
2279 *
2280 * It is expected this function will be called by a module when it is
2281 * unloaded. This will remove the cache completely, and avoid a duplicate
2282 * cache being allocated each time a module is loaded and unloaded, if the
2283 * module doesn't have persistent in-kernel storage across loads and unloads.
2284 *
2285 * The cache must be empty before calling this function.
2286 *
2287 * The caller must guarantee that noone will allocate memory from the cache
2288 * during the kmem_cache_destroy().
2289 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002290int kmem_cache_destroy(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002291{
2292 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07002293 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002294
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002295 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002296
2297 /* Don't let CPUs to come and go */
2298 lock_cpu_hotplug();
2299
2300 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002301 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002302 /*
2303 * the chain is never empty, cache_cache is never destroyed
2304 */
2305 list_del(&cachep->next);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002306 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002307
2308 if (__cache_shrink(cachep)) {
2309 slab_error(cachep, "Can't free all objects");
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002310 mutex_lock(&cache_chain_mutex);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002311 list_add(&cachep->next, &cache_chain);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002312 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002313 unlock_cpu_hotplug();
2314 return 1;
2315 }
2316
2317 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
Paul E. McKenneyfbd568a3e2005-05-01 08:59:04 -07002318 synchronize_rcu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002319
Christoph Lametere498be72005-09-09 13:03:32 -07002320 for_each_online_cpu(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002321 kfree(cachep->array[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002322
2323 /* NUMA: free the list3 structures */
Christoph Lametere498be72005-09-09 13:03:32 -07002324 for_each_online_node(i) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002325 l3 = cachep->nodelists[i];
2326 if (l3) {
Christoph Lametere498be72005-09-09 13:03:32 -07002327 kfree(l3->shared);
2328 free_alien_cache(l3->alien);
2329 kfree(l3);
2330 }
2331 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002332 kmem_cache_free(&cache_cache, cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002333 unlock_cpu_hotplug();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002334 return 0;
2335}
2336EXPORT_SYMBOL(kmem_cache_destroy);
2337
2338/* Get the memory for a slab management obj. */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002339static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002340 int colour_off, gfp_t local_flags,
2341 int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002342{
2343 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002344
Linus Torvalds1da177e2005-04-16 15:20:36 -07002345 if (OFF_SLAB(cachep)) {
2346 /* Slab management obj is off-slab. */
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002347 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2348 local_flags, nodeid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002349 if (!slabp)
2350 return NULL;
2351 } else {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002352 slabp = objp + colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002353 colour_off += cachep->slab_size;
2354 }
2355 slabp->inuse = 0;
2356 slabp->colouroff = colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002357 slabp->s_mem = objp + colour_off;
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002358 slabp->nodeid = nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002359 return slabp;
2360}
2361
2362static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2363{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002364 return (kmem_bufctl_t *) (slabp + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002365}
2366
Pekka Enberg343e0d72006-02-01 03:05:50 -08002367static void cache_init_objs(struct kmem_cache *cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002368 struct slab *slabp, unsigned long ctor_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002369{
2370 int i;
2371
2372 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002373 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002374#if DEBUG
2375 /* need to poison the objs? */
2376 if (cachep->flags & SLAB_POISON)
2377 poison_obj(cachep, objp, POISON_FREE);
2378 if (cachep->flags & SLAB_STORE_USER)
2379 *dbg_userword(cachep, objp) = NULL;
2380
2381 if (cachep->flags & SLAB_RED_ZONE) {
2382 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2383 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2384 }
2385 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002386 * Constructors are not allowed to allocate memory from the same
2387 * cache which they are a constructor for. Otherwise, deadlock.
2388 * They must also be threaded.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002389 */
2390 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002391 cachep->ctor(objp + obj_offset(cachep), cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002392 ctor_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002393
2394 if (cachep->flags & SLAB_RED_ZONE) {
2395 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2396 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002397 " end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002398 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2399 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002400 " start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002401 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08002402 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2403 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002404 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002405 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002406#else
2407 if (cachep->ctor)
2408 cachep->ctor(objp, cachep, ctor_flags);
2409#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002410 slab_bufctl(slabp)[i] = i + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002411 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002412 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002413 slabp->free = 0;
2414}
2415
Pekka Enberg343e0d72006-02-01 03:05:50 -08002416static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002417{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002418 if (flags & SLAB_DMA)
2419 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2420 else
2421 BUG_ON(cachep->gfpflags & GFP_DMA);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002422}
2423
Andrew Mortona737b3e2006-03-22 00:08:11 -08002424static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2425 int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002426{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002427 void *objp = index_to_obj(cachep, slabp, slabp->free);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002428 kmem_bufctl_t next;
2429
2430 slabp->inuse++;
2431 next = slab_bufctl(slabp)[slabp->free];
2432#if DEBUG
2433 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2434 WARN_ON(slabp->nodeid != nodeid);
2435#endif
2436 slabp->free = next;
2437
2438 return objp;
2439}
2440
Andrew Mortona737b3e2006-03-22 00:08:11 -08002441static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2442 void *objp, int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002443{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002444 unsigned int objnr = obj_to_index(cachep, slabp, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002445
2446#if DEBUG
2447 /* Verify that the slab belongs to the intended node */
2448 WARN_ON(slabp->nodeid != nodeid);
2449
Al Viro871751e2006-03-25 03:06:39 -08002450 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
Matthew Dobson78d382d2006-02-01 03:05:47 -08002451 printk(KERN_ERR "slab: double free detected in cache "
Andrew Mortona737b3e2006-03-22 00:08:11 -08002452 "'%s', objp %p\n", cachep->name, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002453 BUG();
2454 }
2455#endif
2456 slab_bufctl(slabp)[objnr] = slabp->free;
2457 slabp->free = objnr;
2458 slabp->inuse--;
2459}
2460
Andrew Mortona737b3e2006-03-22 00:08:11 -08002461static void set_slab_attr(struct kmem_cache *cachep, struct slab *slabp,
2462 void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002463{
2464 int i;
2465 struct page *page;
2466
2467 /* Nasty!!!!!! I hope this is OK. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002468 page = virt_to_page(objp);
Nick Piggin84097512006-03-22 00:08:34 -08002469
2470 i = 1;
2471 if (likely(!PageCompound(page)))
2472 i <<= cachep->gfporder;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002473 do {
Pekka Enberg065d41c2005-11-13 16:06:46 -08002474 page_set_cache(page, cachep);
2475 page_set_slab(page, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002476 page++;
2477 } while (--i);
2478}
2479
2480/*
2481 * Grow (by 1) the number of slabs within a cache. This is called by
2482 * kmem_cache_alloc() when there are no active objs left in a cache.
2483 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002484static int cache_grow(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002485{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002486 struct slab *slabp;
2487 void *objp;
2488 size_t offset;
2489 gfp_t local_flags;
2490 unsigned long ctor_flags;
Christoph Lametere498be72005-09-09 13:03:32 -07002491 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002492
Andrew Mortona737b3e2006-03-22 00:08:11 -08002493 /*
2494 * Be lazy and only check for valid flags here, keeping it out of the
2495 * critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002496 */
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002497 BUG_ON(flags & ~(SLAB_DMA | SLAB_LEVEL_MASK | SLAB_NO_GROW));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002498 if (flags & SLAB_NO_GROW)
2499 return 0;
2500
2501 ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2502 local_flags = (flags & SLAB_LEVEL_MASK);
2503 if (!(local_flags & __GFP_WAIT))
2504 /*
2505 * Not allowed to sleep. Need to tell a constructor about
2506 * this - it might need to know...
2507 */
2508 ctor_flags |= SLAB_CTOR_ATOMIC;
2509
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002510 /* Take the l3 list lock to change the colour_next on this node */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002511 check_irq_off();
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002512 l3 = cachep->nodelists[nodeid];
2513 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002514
2515 /* Get colour for the slab, and cal the next value. */
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002516 offset = l3->colour_next;
2517 l3->colour_next++;
2518 if (l3->colour_next >= cachep->colour)
2519 l3->colour_next = 0;
2520 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002521
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002522 offset *= cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002523
2524 if (local_flags & __GFP_WAIT)
2525 local_irq_enable();
2526
2527 /*
2528 * The test for missing atomic flag is performed here, rather than
2529 * the more obvious place, simply to reduce the critical path length
2530 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2531 * will eventually be caught here (where it matters).
2532 */
2533 kmem_flagcheck(cachep, flags);
2534
Andrew Mortona737b3e2006-03-22 00:08:11 -08002535 /*
2536 * Get mem for the objs. Attempt to allocate a physical page from
2537 * 'nodeid'.
Christoph Lametere498be72005-09-09 13:03:32 -07002538 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002539 objp = kmem_getpages(cachep, flags, nodeid);
2540 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002541 goto failed;
2542
2543 /* Get slab management. */
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002544 slabp = alloc_slabmgmt(cachep, objp, offset, local_flags, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002545 if (!slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002546 goto opps1;
2547
Christoph Lametere498be72005-09-09 13:03:32 -07002548 slabp->nodeid = nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002549 set_slab_attr(cachep, slabp, objp);
2550
2551 cache_init_objs(cachep, slabp, ctor_flags);
2552
2553 if (local_flags & __GFP_WAIT)
2554 local_irq_disable();
2555 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002556 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002557
2558 /* Make slab active. */
Christoph Lametere498be72005-09-09 13:03:32 -07002559 list_add_tail(&slabp->list, &(l3->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002560 STATS_INC_GROWN(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002561 l3->free_objects += cachep->num;
2562 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002563 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002564opps1:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002565 kmem_freepages(cachep, objp);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002566failed:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002567 if (local_flags & __GFP_WAIT)
2568 local_irq_disable();
2569 return 0;
2570}
2571
2572#if DEBUG
2573
2574/*
2575 * Perform extra freeing checks:
2576 * - detect bad pointers.
2577 * - POISON/RED_ZONE checking
2578 * - destructor calls, for caches with POISON+dtor
2579 */
2580static void kfree_debugcheck(const void *objp)
2581{
2582 struct page *page;
2583
2584 if (!virt_addr_valid(objp)) {
2585 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002586 (unsigned long)objp);
2587 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002588 }
2589 page = virt_to_page(objp);
2590 if (!PageSlab(page)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002591 printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n",
2592 (unsigned long)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002593 BUG();
2594 }
2595}
2596
Pekka Enberg343e0d72006-02-01 03:05:50 -08002597static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002598 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002599{
2600 struct page *page;
2601 unsigned int objnr;
2602 struct slab *slabp;
2603
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002604 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002605 kfree_debugcheck(objp);
2606 page = virt_to_page(objp);
2607
Pekka Enberg065d41c2005-11-13 16:06:46 -08002608 if (page_get_cache(page) != cachep) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002609 printk(KERN_ERR "mismatch in kmem_cache_free: expected "
2610 "cache %p, got %p\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002611 page_get_cache(page), cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002612 printk(KERN_ERR "%p is %s.\n", cachep, cachep->name);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002613 printk(KERN_ERR "%p is %s.\n", page_get_cache(page),
2614 page_get_cache(page)->name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002615 WARN_ON(1);
2616 }
Pekka Enberg065d41c2005-11-13 16:06:46 -08002617 slabp = page_get_slab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002618
2619 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002620 if (*dbg_redzone1(cachep, objp) != RED_ACTIVE ||
2621 *dbg_redzone2(cachep, objp) != RED_ACTIVE) {
2622 slab_error(cachep, "double free, or memory outside"
2623 " object was overwritten");
2624 printk(KERN_ERR "%p: redzone 1:0x%lx, "
2625 "redzone 2:0x%lx.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002626 objp, *dbg_redzone1(cachep, objp),
2627 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002628 }
2629 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2630 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2631 }
2632 if (cachep->flags & SLAB_STORE_USER)
2633 *dbg_userword(cachep, objp) = caller;
2634
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002635 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002636
2637 BUG_ON(objnr >= cachep->num);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002638 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002639
2640 if (cachep->flags & SLAB_DEBUG_INITIAL) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002641 /*
2642 * Need to call the slab's constructor so the caller can
2643 * perform a verify of its state (debugging). Called without
2644 * the cache-lock held.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002645 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002646 cachep->ctor(objp + obj_offset(cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002647 cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002648 }
2649 if (cachep->flags & SLAB_POISON && cachep->dtor) {
2650 /* we want to cache poison the object,
2651 * call the destruction callback
2652 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002653 cachep->dtor(objp + obj_offset(cachep), cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002654 }
Al Viro871751e2006-03-25 03:06:39 -08002655#ifdef CONFIG_DEBUG_SLAB_LEAK
2656 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2657#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002658 if (cachep->flags & SLAB_POISON) {
2659#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08002660 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002661 store_stackinfo(cachep, objp, (unsigned long)caller);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002662 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002663 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002664 } else {
2665 poison_obj(cachep, objp, POISON_FREE);
2666 }
2667#else
2668 poison_obj(cachep, objp, POISON_FREE);
2669#endif
2670 }
2671 return objp;
2672}
2673
Pekka Enberg343e0d72006-02-01 03:05:50 -08002674static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002675{
2676 kmem_bufctl_t i;
2677 int entries = 0;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002678
Linus Torvalds1da177e2005-04-16 15:20:36 -07002679 /* Check slab's freelist to see if this obj is there. */
2680 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2681 entries++;
2682 if (entries > cachep->num || i >= cachep->num)
2683 goto bad;
2684 }
2685 if (entries != cachep->num - slabp->inuse) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002686bad:
2687 printk(KERN_ERR "slab: Internal list corruption detected in "
2688 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2689 cachep->name, cachep->num, slabp, slabp->inuse);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002690 for (i = 0;
Linus Torvalds264132b2006-03-06 12:10:07 -08002691 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002692 i++) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002693 if (i % 16 == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002694 printk("\n%03x:", i);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002695 printk(" %02x", ((unsigned char *)slabp)[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002696 }
2697 printk("\n");
2698 BUG();
2699 }
2700}
2701#else
2702#define kfree_debugcheck(x) do { } while(0)
2703#define cache_free_debugcheck(x,objp,z) (objp)
2704#define check_slabp(x,y) do { } while(0)
2705#endif
2706
Pekka Enberg343e0d72006-02-01 03:05:50 -08002707static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002708{
2709 int batchcount;
2710 struct kmem_list3 *l3;
2711 struct array_cache *ac;
2712
2713 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002714 ac = cpu_cache_get(cachep);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002715retry:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002716 batchcount = ac->batchcount;
2717 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002718 /*
2719 * If there was little recent activity on this cache, then
2720 * perform only a partial refill. Otherwise we could generate
2721 * refill bouncing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002722 */
2723 batchcount = BATCHREFILL_LIMIT;
2724 }
Christoph Lametere498be72005-09-09 13:03:32 -07002725 l3 = cachep->nodelists[numa_node_id()];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002726
Christoph Lametere498be72005-09-09 13:03:32 -07002727 BUG_ON(ac->avail > 0 || !l3);
2728 spin_lock(&l3->list_lock);
2729
Christoph Lameter3ded1752006-03-25 03:06:44 -08002730 /* See if we can refill from the shared array */
2731 if (l3->shared && transfer_objects(ac, l3->shared, batchcount))
2732 goto alloc_done;
2733
Linus Torvalds1da177e2005-04-16 15:20:36 -07002734 while (batchcount > 0) {
2735 struct list_head *entry;
2736 struct slab *slabp;
2737 /* Get slab alloc is to come from. */
2738 entry = l3->slabs_partial.next;
2739 if (entry == &l3->slabs_partial) {
2740 l3->free_touched = 1;
2741 entry = l3->slabs_free.next;
2742 if (entry == &l3->slabs_free)
2743 goto must_grow;
2744 }
2745
2746 slabp = list_entry(entry, struct slab, list);
2747 check_slabp(cachep, slabp);
2748 check_spinlock_acquired(cachep);
2749 while (slabp->inuse < cachep->num && batchcount--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002750 STATS_INC_ALLOCED(cachep);
2751 STATS_INC_ACTIVE(cachep);
2752 STATS_SET_HIGH(cachep);
2753
Matthew Dobson78d382d2006-02-01 03:05:47 -08002754 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
2755 numa_node_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002756 }
2757 check_slabp(cachep, slabp);
2758
2759 /* move slabp to correct slabp list: */
2760 list_del(&slabp->list);
2761 if (slabp->free == BUFCTL_END)
2762 list_add(&slabp->list, &l3->slabs_full);
2763 else
2764 list_add(&slabp->list, &l3->slabs_partial);
2765 }
2766
Andrew Mortona737b3e2006-03-22 00:08:11 -08002767must_grow:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002768 l3->free_objects -= ac->avail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002769alloc_done:
Christoph Lametere498be72005-09-09 13:03:32 -07002770 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002771
2772 if (unlikely(!ac->avail)) {
2773 int x;
Christoph Lametere498be72005-09-09 13:03:32 -07002774 x = cache_grow(cachep, flags, numa_node_id());
2775
Andrew Mortona737b3e2006-03-22 00:08:11 -08002776 /* cache_grow can reenable interrupts, then ac could change. */
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002777 ac = cpu_cache_get(cachep);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002778 if (!x && ac->avail == 0) /* no objects in sight? abort */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002779 return NULL;
2780
Andrew Mortona737b3e2006-03-22 00:08:11 -08002781 if (!ac->avail) /* objects refilled by interrupt? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002782 goto retry;
2783 }
2784 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07002785 return ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002786}
2787
Andrew Mortona737b3e2006-03-22 00:08:11 -08002788static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
2789 gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002790{
2791 might_sleep_if(flags & __GFP_WAIT);
2792#if DEBUG
2793 kmem_flagcheck(cachep, flags);
2794#endif
2795}
2796
2797#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08002798static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
2799 gfp_t flags, void *objp, void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002800{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002801 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002802 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002803 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002804#ifdef CONFIG_DEBUG_PAGEALLOC
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002805 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002806 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002807 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002808 else
2809 check_poison_obj(cachep, objp);
2810#else
2811 check_poison_obj(cachep, objp);
2812#endif
2813 poison_obj(cachep, objp, POISON_INUSE);
2814 }
2815 if (cachep->flags & SLAB_STORE_USER)
2816 *dbg_userword(cachep, objp) = caller;
2817
2818 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002819 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
2820 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2821 slab_error(cachep, "double free, or memory outside"
2822 " object was overwritten");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002823 printk(KERN_ERR
Andrew Mortona737b3e2006-03-22 00:08:11 -08002824 "%p: redzone 1:0x%lx, redzone 2:0x%lx\n",
2825 objp, *dbg_redzone1(cachep, objp),
2826 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002827 }
2828 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2829 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2830 }
Al Viro871751e2006-03-25 03:06:39 -08002831#ifdef CONFIG_DEBUG_SLAB_LEAK
2832 {
2833 struct slab *slabp;
2834 unsigned objnr;
2835
2836 slabp = page_get_slab(virt_to_page(objp));
2837 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
2838 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
2839 }
2840#endif
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002841 objp += obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002842 if (cachep->ctor && cachep->flags & SLAB_POISON) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002843 unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002844
2845 if (!(flags & __GFP_WAIT))
2846 ctor_flags |= SLAB_CTOR_ATOMIC;
2847
2848 cachep->ctor(objp, cachep, ctor_flags);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002849 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002850 return objp;
2851}
2852#else
2853#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2854#endif
2855
Pekka Enberg343e0d72006-02-01 03:05:50 -08002856static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002857{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002858 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002859 struct array_cache *ac;
2860
Christoph Lameterdc85da12006-01-18 17:42:36 -08002861#ifdef CONFIG_NUMA
Paul Jacksonb2455392006-03-24 03:16:12 -08002862 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
Paul Jacksonc61afb12006-03-24 03:16:08 -08002863 objp = alternate_node_alloc(cachep, flags);
2864 if (objp != NULL)
2865 return objp;
Paul Jackson101a5002006-03-24 03:16:07 -08002866 }
Christoph Lameterdc85da12006-01-18 17:42:36 -08002867#endif
2868
Alok N Kataria5c382302005-09-27 21:45:46 -07002869 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002870 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002871 if (likely(ac->avail)) {
2872 STATS_INC_ALLOCHIT(cachep);
2873 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07002874 objp = ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002875 } else {
2876 STATS_INC_ALLOCMISS(cachep);
2877 objp = cache_alloc_refill(cachep, flags);
2878 }
Alok N Kataria5c382302005-09-27 21:45:46 -07002879 return objp;
2880}
2881
Andrew Mortona737b3e2006-03-22 00:08:11 -08002882static __always_inline void *__cache_alloc(struct kmem_cache *cachep,
2883 gfp_t flags, void *caller)
Alok N Kataria5c382302005-09-27 21:45:46 -07002884{
2885 unsigned long save_flags;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002886 void *objp;
Alok N Kataria5c382302005-09-27 21:45:46 -07002887
2888 cache_alloc_debugcheck_before(cachep, flags);
2889
2890 local_irq_save(save_flags);
2891 objp = ____cache_alloc(cachep, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002892 local_irq_restore(save_flags);
Eric Dumazet34342e82005-09-03 15:55:06 -07002893 objp = cache_alloc_debugcheck_after(cachep, flags, objp,
Pekka Enberg7fd6b142006-02-01 03:05:52 -08002894 caller);
Eric Dumazet34342e82005-09-03 15:55:06 -07002895 prefetchw(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002896 return objp;
2897}
2898
Christoph Lametere498be72005-09-09 13:03:32 -07002899#ifdef CONFIG_NUMA
2900/*
Paul Jacksonb2455392006-03-24 03:16:12 -08002901 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
Paul Jacksonc61afb12006-03-24 03:16:08 -08002902 *
2903 * If we are in_interrupt, then process context, including cpusets and
2904 * mempolicy, may not apply and should not be used for allocation policy.
2905 */
2906static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
2907{
2908 int nid_alloc, nid_here;
2909
2910 if (in_interrupt())
2911 return NULL;
2912 nid_alloc = nid_here = numa_node_id();
2913 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
2914 nid_alloc = cpuset_mem_spread_node();
2915 else if (current->mempolicy)
2916 nid_alloc = slab_node(current->mempolicy);
2917 if (nid_alloc != nid_here)
2918 return __cache_alloc_node(cachep, flags, nid_alloc);
2919 return NULL;
2920}
2921
2922/*
Christoph Lametere498be72005-09-09 13:03:32 -07002923 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002924 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002925static void *__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
2926 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07002927{
2928 struct list_head *entry;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002929 struct slab *slabp;
2930 struct kmem_list3 *l3;
2931 void *obj;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002932 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002933
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002934 l3 = cachep->nodelists[nodeid];
2935 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07002936
Andrew Mortona737b3e2006-03-22 00:08:11 -08002937retry:
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08002938 check_irq_off();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002939 spin_lock(&l3->list_lock);
2940 entry = l3->slabs_partial.next;
2941 if (entry == &l3->slabs_partial) {
2942 l3->free_touched = 1;
2943 entry = l3->slabs_free.next;
2944 if (entry == &l3->slabs_free)
2945 goto must_grow;
2946 }
Christoph Lametere498be72005-09-09 13:03:32 -07002947
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002948 slabp = list_entry(entry, struct slab, list);
2949 check_spinlock_acquired_node(cachep, nodeid);
2950 check_slabp(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07002951
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002952 STATS_INC_NODEALLOCS(cachep);
2953 STATS_INC_ACTIVE(cachep);
2954 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002955
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002956 BUG_ON(slabp->inuse == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07002957
Matthew Dobson78d382d2006-02-01 03:05:47 -08002958 obj = slab_get_obj(cachep, slabp, nodeid);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002959 check_slabp(cachep, slabp);
2960 l3->free_objects--;
2961 /* move slabp to correct slabp list: */
2962 list_del(&slabp->list);
Christoph Lametere498be72005-09-09 13:03:32 -07002963
Andrew Mortona737b3e2006-03-22 00:08:11 -08002964 if (slabp->free == BUFCTL_END)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002965 list_add(&slabp->list, &l3->slabs_full);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002966 else
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002967 list_add(&slabp->list, &l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07002968
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002969 spin_unlock(&l3->list_lock);
2970 goto done;
Christoph Lametere498be72005-09-09 13:03:32 -07002971
Andrew Mortona737b3e2006-03-22 00:08:11 -08002972must_grow:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002973 spin_unlock(&l3->list_lock);
2974 x = cache_grow(cachep, flags, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07002975
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002976 if (!x)
2977 return NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07002978
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002979 goto retry;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002980done:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002981 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07002982}
2983#endif
2984
2985/*
2986 * Caller needs to acquire correct kmem_list's list_lock
2987 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002988static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002989 int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002990{
2991 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07002992 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002993
2994 for (i = 0; i < nr_objects; i++) {
2995 void *objp = objpp[i];
2996 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002997
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08002998 slabp = virt_to_slab(objp);
Christoph Lameterff694162005-09-22 21:44:02 -07002999 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003000 list_del(&slabp->list);
Christoph Lameterff694162005-09-22 21:44:02 -07003001 check_spinlock_acquired_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003002 check_slabp(cachep, slabp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08003003 slab_put_obj(cachep, slabp, objp, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003004 STATS_DEC_ACTIVE(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003005 l3->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003006 check_slabp(cachep, slabp);
3007
3008 /* fixup slab chains */
3009 if (slabp->inuse == 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07003010 if (l3->free_objects > l3->free_limit) {
3011 l3->free_objects -= cachep->num;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003012 slab_destroy(cachep, slabp);
3013 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07003014 list_add(&slabp->list, &l3->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003015 }
3016 } else {
3017 /* Unconditionally move a slab to the end of the
3018 * partial list on free - maximum time for the
3019 * other objects to be freed, too.
3020 */
Christoph Lametere498be72005-09-09 13:03:32 -07003021 list_add_tail(&slabp->list, &l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003022 }
3023 }
3024}
3025
Pekka Enberg343e0d72006-02-01 03:05:50 -08003026static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003027{
3028 int batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07003029 struct kmem_list3 *l3;
Christoph Lameterff694162005-09-22 21:44:02 -07003030 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003031
3032 batchcount = ac->batchcount;
3033#if DEBUG
3034 BUG_ON(!batchcount || batchcount > ac->avail);
3035#endif
3036 check_irq_off();
Christoph Lameterff694162005-09-22 21:44:02 -07003037 l3 = cachep->nodelists[node];
Christoph Lametere498be72005-09-09 13:03:32 -07003038 spin_lock(&l3->list_lock);
3039 if (l3->shared) {
3040 struct array_cache *shared_array = l3->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003041 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003042 if (max) {
3043 if (batchcount > max)
3044 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07003045 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003046 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003047 shared_array->avail += batchcount;
3048 goto free_done;
3049 }
3050 }
3051
Christoph Lameterff694162005-09-22 21:44:02 -07003052 free_block(cachep, ac->entry, batchcount, node);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003053free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003054#if STATS
3055 {
3056 int i = 0;
3057 struct list_head *p;
3058
Christoph Lametere498be72005-09-09 13:03:32 -07003059 p = l3->slabs_free.next;
3060 while (p != &(l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003061 struct slab *slabp;
3062
3063 slabp = list_entry(p, struct slab, list);
3064 BUG_ON(slabp->inuse);
3065
3066 i++;
3067 p = p->next;
3068 }
3069 STATS_SET_FREEABLE(cachep, i);
3070 }
3071#endif
Christoph Lametere498be72005-09-09 13:03:32 -07003072 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003073 ac->avail -= batchcount;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003074 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003075}
3076
3077/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08003078 * Release an obj back to its cache. If the obj has a constructed state, it must
3079 * be in this state _before_ it is released. Called with disabled ints.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003080 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003081static inline void __cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003082{
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003083 struct array_cache *ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003084
3085 check_irq_off();
3086 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3087
Christoph Lametere498be72005-09-09 13:03:32 -07003088 /* Make sure we are not freeing a object from another
3089 * node to the array cache on this cpu.
3090 */
3091#ifdef CONFIG_NUMA
3092 {
3093 struct slab *slabp;
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003094 slabp = virt_to_slab(objp);
Christoph Lametere498be72005-09-09 13:03:32 -07003095 if (unlikely(slabp->nodeid != numa_node_id())) {
3096 struct array_cache *alien = NULL;
3097 int nodeid = slabp->nodeid;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003098 struct kmem_list3 *l3;
Christoph Lametere498be72005-09-09 13:03:32 -07003099
Andrew Mortona737b3e2006-03-22 00:08:11 -08003100 l3 = cachep->nodelists[numa_node_id()];
Christoph Lametere498be72005-09-09 13:03:32 -07003101 STATS_INC_NODEFREES(cachep);
3102 if (l3->alien && l3->alien[nodeid]) {
3103 alien = l3->alien[nodeid];
3104 spin_lock(&alien->lock);
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07003105 if (unlikely(alien->avail == alien->limit)) {
3106 STATS_INC_ACOVERFLOW(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003107 __drain_alien_cache(cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003108 alien, nodeid);
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07003109 }
Christoph Lametere498be72005-09-09 13:03:32 -07003110 alien->entry[alien->avail++] = objp;
3111 spin_unlock(&alien->lock);
3112 } else {
3113 spin_lock(&(cachep->nodelists[nodeid])->
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003114 list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -07003115 free_block(cachep, &objp, 1, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07003116 spin_unlock(&(cachep->nodelists[nodeid])->
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003117 list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003118 }
3119 return;
3120 }
3121 }
3122#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003123 if (likely(ac->avail < ac->limit)) {
3124 STATS_INC_FREEHIT(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003125 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003126 return;
3127 } else {
3128 STATS_INC_FREEMISS(cachep);
3129 cache_flusharray(cachep, ac);
Christoph Lametere498be72005-09-09 13:03:32 -07003130 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003131 }
3132}
3133
3134/**
3135 * kmem_cache_alloc - Allocate an object
3136 * @cachep: The cache to allocate from.
3137 * @flags: See kmalloc().
3138 *
3139 * Allocate an object from this cache. The flags are only relevant
3140 * if the cache has no available objects.
3141 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003142void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003143{
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003144 return __cache_alloc(cachep, flags, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003145}
3146EXPORT_SYMBOL(kmem_cache_alloc);
3147
3148/**
Pekka Enberga8c0f9a2006-03-25 03:06:42 -08003149 * kmem_cache_alloc - Allocate an object. The memory is set to zero.
3150 * @cache: The cache to allocate from.
3151 * @flags: See kmalloc().
3152 *
3153 * Allocate an object from this cache and set the allocated memory to zero.
3154 * The flags are only relevant if the cache has no available objects.
3155 */
3156void *kmem_cache_zalloc(struct kmem_cache *cache, gfp_t flags)
3157{
3158 void *ret = __cache_alloc(cache, flags, __builtin_return_address(0));
3159 if (ret)
3160 memset(ret, 0, obj_size(cache));
3161 return ret;
3162}
3163EXPORT_SYMBOL(kmem_cache_zalloc);
3164
3165/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003166 * kmem_ptr_validate - check if an untrusted pointer might
3167 * be a slab entry.
3168 * @cachep: the cache we're checking against
3169 * @ptr: pointer to validate
3170 *
3171 * This verifies that the untrusted pointer looks sane:
3172 * it is _not_ a guarantee that the pointer is actually
3173 * part of the slab cache in question, but it at least
3174 * validates that the pointer can be dereferenced and
3175 * looks half-way sane.
3176 *
3177 * Currently only used for dentry validation.
3178 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003179int fastcall kmem_ptr_validate(struct kmem_cache *cachep, void *ptr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003180{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003181 unsigned long addr = (unsigned long)ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003182 unsigned long min_addr = PAGE_OFFSET;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003183 unsigned long align_mask = BYTES_PER_WORD - 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003184 unsigned long size = cachep->buffer_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003185 struct page *page;
3186
3187 if (unlikely(addr < min_addr))
3188 goto out;
3189 if (unlikely(addr > (unsigned long)high_memory - size))
3190 goto out;
3191 if (unlikely(addr & align_mask))
3192 goto out;
3193 if (unlikely(!kern_addr_valid(addr)))
3194 goto out;
3195 if (unlikely(!kern_addr_valid(addr + size - 1)))
3196 goto out;
3197 page = virt_to_page(ptr);
3198 if (unlikely(!PageSlab(page)))
3199 goto out;
Pekka Enberg065d41c2005-11-13 16:06:46 -08003200 if (unlikely(page_get_cache(page) != cachep))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003201 goto out;
3202 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003203out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003204 return 0;
3205}
3206
3207#ifdef CONFIG_NUMA
3208/**
3209 * kmem_cache_alloc_node - Allocate an object on the specified node
3210 * @cachep: The cache to allocate from.
3211 * @flags: See kmalloc().
3212 * @nodeid: node number of the target node.
3213 *
3214 * Identical to kmem_cache_alloc, except that this function is slow
3215 * and can sleep. And it will allocate memory on the given node, which
3216 * can improve the performance for cpu bound structures.
Christoph Lametere498be72005-09-09 13:03:32 -07003217 * New and improved: it will now make sure that the object gets
3218 * put on the correct node list so that there is no false sharing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003219 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003220void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003221{
Christoph Lametere498be72005-09-09 13:03:32 -07003222 unsigned long save_flags;
3223 void *ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003224
Christoph Lametere498be72005-09-09 13:03:32 -07003225 cache_alloc_debugcheck_before(cachep, flags);
3226 local_irq_save(save_flags);
Christoph Lameter18f820f2006-02-01 03:05:43 -08003227
3228 if (nodeid == -1 || nodeid == numa_node_id() ||
Andrew Mortona737b3e2006-03-22 00:08:11 -08003229 !cachep->nodelists[nodeid])
Alok N Kataria5c382302005-09-27 21:45:46 -07003230 ptr = ____cache_alloc(cachep, flags);
3231 else
3232 ptr = __cache_alloc_node(cachep, flags, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07003233 local_irq_restore(save_flags);
Christoph Lameter18f820f2006-02-01 03:05:43 -08003234
3235 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr,
3236 __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003237
Christoph Lametere498be72005-09-09 13:03:32 -07003238 return ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003239}
3240EXPORT_SYMBOL(kmem_cache_alloc_node);
3241
Al Virodd0fc662005-10-07 07:46:04 +01003242void *kmalloc_node(size_t size, gfp_t flags, int node)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003243{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003244 struct kmem_cache *cachep;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003245
3246 cachep = kmem_find_general_cachep(size, flags);
3247 if (unlikely(cachep == NULL))
3248 return NULL;
3249 return kmem_cache_alloc_node(cachep, flags, node);
3250}
3251EXPORT_SYMBOL(kmalloc_node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003252#endif
3253
3254/**
3255 * kmalloc - allocate memory
3256 * @size: how many bytes of memory are required.
3257 * @flags: the type of memory to allocate.
Randy Dunlap911851e2006-03-22 00:08:14 -08003258 * @caller: function caller for debug tracking of the caller
Linus Torvalds1da177e2005-04-16 15:20:36 -07003259 *
3260 * kmalloc is the normal method of allocating memory
3261 * in the kernel.
3262 *
3263 * The @flags argument may be one of:
3264 *
3265 * %GFP_USER - Allocate memory on behalf of user. May sleep.
3266 *
3267 * %GFP_KERNEL - Allocate normal kernel ram. May sleep.
3268 *
3269 * %GFP_ATOMIC - Allocation will not sleep. Use inside interrupt handlers.
3270 *
3271 * Additionally, the %GFP_DMA flag may be set to indicate the memory
3272 * must be suitable for DMA. This can mean different things on different
3273 * platforms. For example, on i386, it means that the memory must come
3274 * from the first 16MB.
3275 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003276static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3277 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003278{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003279 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003280
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003281 /* If you want to save a few bytes .text space: replace
3282 * __ with kmem_.
3283 * Then kmalloc uses the uninlined functions instead of the inline
3284 * functions.
3285 */
3286 cachep = __find_general_cachep(size, flags);
Andrew Mortondbdb9042005-09-23 13:24:10 -07003287 if (unlikely(cachep == NULL))
3288 return NULL;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003289 return __cache_alloc(cachep, flags, caller);
3290}
3291
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003292
3293void *__kmalloc(size_t size, gfp_t flags)
3294{
Al Viro871751e2006-03-25 03:06:39 -08003295#ifndef CONFIG_DEBUG_SLAB
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003296 return __do_kmalloc(size, flags, NULL);
Al Viro871751e2006-03-25 03:06:39 -08003297#else
3298 return __do_kmalloc(size, flags, __builtin_return_address(0));
3299#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003300}
3301EXPORT_SYMBOL(__kmalloc);
3302
Al Viro871751e2006-03-25 03:06:39 -08003303#ifdef CONFIG_DEBUG_SLAB
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003304void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller)
3305{
3306 return __do_kmalloc(size, flags, caller);
3307}
3308EXPORT_SYMBOL(__kmalloc_track_caller);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003309#endif
3310
Linus Torvalds1da177e2005-04-16 15:20:36 -07003311#ifdef CONFIG_SMP
3312/**
3313 * __alloc_percpu - allocate one copy of the object for every present
3314 * cpu in the system, zeroing them.
3315 * Objects should be dereferenced using the per_cpu_ptr macro only.
3316 *
3317 * @size: how many bytes of memory are required.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003318 */
Pekka Enbergf9f75002006-01-08 01:00:33 -08003319void *__alloc_percpu(size_t size)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003320{
3321 int i;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003322 struct percpu_data *pdata = kmalloc(sizeof(*pdata), GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003323
3324 if (!pdata)
3325 return NULL;
3326
Christoph Lametere498be72005-09-09 13:03:32 -07003327 /*
3328 * Cannot use for_each_online_cpu since a cpu may come online
3329 * and we have no way of figuring out how to fix the array
3330 * that we have allocated then....
3331 */
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08003332 for_each_possible_cpu(i) {
Christoph Lametere498be72005-09-09 13:03:32 -07003333 int node = cpu_to_node(i);
3334
3335 if (node_online(node))
3336 pdata->ptrs[i] = kmalloc_node(size, GFP_KERNEL, node);
3337 else
3338 pdata->ptrs[i] = kmalloc(size, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003339
3340 if (!pdata->ptrs[i])
3341 goto unwind_oom;
3342 memset(pdata->ptrs[i], 0, size);
3343 }
3344
3345 /* Catch derefs w/o wrappers */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003346 return (void *)(~(unsigned long)pdata);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003347
Andrew Mortona737b3e2006-03-22 00:08:11 -08003348unwind_oom:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003349 while (--i >= 0) {
3350 if (!cpu_possible(i))
3351 continue;
3352 kfree(pdata->ptrs[i]);
3353 }
3354 kfree(pdata);
3355 return NULL;
3356}
3357EXPORT_SYMBOL(__alloc_percpu);
3358#endif
3359
3360/**
3361 * kmem_cache_free - Deallocate an object
3362 * @cachep: The cache the allocation was from.
3363 * @objp: The previously allocated object.
3364 *
3365 * Free an object which was previously allocated from this
3366 * cache.
3367 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003368void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003369{
3370 unsigned long flags;
3371
3372 local_irq_save(flags);
3373 __cache_free(cachep, objp);
3374 local_irq_restore(flags);
3375}
3376EXPORT_SYMBOL(kmem_cache_free);
3377
3378/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003379 * kfree - free previously allocated memory
3380 * @objp: pointer returned by kmalloc.
3381 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003382 * If @objp is NULL, no operation is performed.
3383 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003384 * Don't free memory not originally allocated by kmalloc()
3385 * or you will run into trouble.
3386 */
3387void kfree(const void *objp)
3388{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003389 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003390 unsigned long flags;
3391
3392 if (unlikely(!objp))
3393 return;
3394 local_irq_save(flags);
3395 kfree_debugcheck(objp);
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003396 c = virt_to_cache(objp);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003397 mutex_debug_check_no_locks_freed(objp, obj_size(c));
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003398 __cache_free(c, (void *)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003399 local_irq_restore(flags);
3400}
3401EXPORT_SYMBOL(kfree);
3402
3403#ifdef CONFIG_SMP
3404/**
3405 * free_percpu - free previously allocated percpu memory
3406 * @objp: pointer returned by alloc_percpu.
3407 *
3408 * Don't free memory not originally allocated by alloc_percpu()
3409 * The complemented objp is to check for that.
3410 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003411void free_percpu(const void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003412{
3413 int i;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003414 struct percpu_data *p = (struct percpu_data *)(~(unsigned long)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003415
Christoph Lametere498be72005-09-09 13:03:32 -07003416 /*
3417 * We allocate for all cpus so we cannot use for online cpu here.
3418 */
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08003419 for_each_possible_cpu(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003420 kfree(p->ptrs[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003421 kfree(p);
3422}
3423EXPORT_SYMBOL(free_percpu);
3424#endif
3425
Pekka Enberg343e0d72006-02-01 03:05:50 -08003426unsigned int kmem_cache_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003427{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003428 return obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003429}
3430EXPORT_SYMBOL(kmem_cache_size);
3431
Pekka Enberg343e0d72006-02-01 03:05:50 -08003432const char *kmem_cache_name(struct kmem_cache *cachep)
Arnaldo Carvalho de Melo19449722005-06-18 22:46:19 -07003433{
3434 return cachep->name;
3435}
3436EXPORT_SYMBOL_GPL(kmem_cache_name);
3437
Christoph Lametere498be72005-09-09 13:03:32 -07003438/*
Christoph Lameter0718dc22006-03-25 03:06:47 -08003439 * This initializes kmem_list3 or resizes varioius caches for all nodes.
Christoph Lametere498be72005-09-09 13:03:32 -07003440 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003441static int alloc_kmemlist(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07003442{
3443 int node;
3444 struct kmem_list3 *l3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003445 struct array_cache *new_shared;
3446 struct array_cache **new_alien;
Christoph Lametere498be72005-09-09 13:03:32 -07003447
3448 for_each_online_node(node) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003449
Andrew Mortona737b3e2006-03-22 00:08:11 -08003450 new_alien = alloc_alien_cache(node, cachep->limit);
3451 if (!new_alien)
Christoph Lametere498be72005-09-09 13:03:32 -07003452 goto fail;
Christoph Lametercafeb022006-03-25 03:06:46 -08003453
Christoph Lameter0718dc22006-03-25 03:06:47 -08003454 new_shared = alloc_arraycache(node,
3455 cachep->shared*cachep->batchcount,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003456 0xbaadf00d);
Christoph Lameter0718dc22006-03-25 03:06:47 -08003457 if (!new_shared) {
3458 free_alien_cache(new_alien);
Christoph Lametere498be72005-09-09 13:03:32 -07003459 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003460 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003461
Andrew Mortona737b3e2006-03-22 00:08:11 -08003462 l3 = cachep->nodelists[node];
3463 if (l3) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003464 struct array_cache *shared = l3->shared;
3465
Christoph Lametere498be72005-09-09 13:03:32 -07003466 spin_lock_irq(&l3->list_lock);
3467
Christoph Lametercafeb022006-03-25 03:06:46 -08003468 if (shared)
Christoph Lameter0718dc22006-03-25 03:06:47 -08003469 free_block(cachep, shared->entry,
3470 shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07003471
Christoph Lametercafeb022006-03-25 03:06:46 -08003472 l3->shared = new_shared;
3473 if (!l3->alien) {
Christoph Lametere498be72005-09-09 13:03:32 -07003474 l3->alien = new_alien;
3475 new_alien = NULL;
3476 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003477 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003478 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003479 spin_unlock_irq(&l3->list_lock);
Christoph Lametercafeb022006-03-25 03:06:46 -08003480 kfree(shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003481 free_alien_cache(new_alien);
3482 continue;
3483 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08003484 l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node);
Christoph Lameter0718dc22006-03-25 03:06:47 -08003485 if (!l3) {
3486 free_alien_cache(new_alien);
3487 kfree(new_shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003488 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003489 }
Christoph Lametere498be72005-09-09 13:03:32 -07003490
3491 kmem_list3_init(l3);
3492 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Andrew Mortona737b3e2006-03-22 00:08:11 -08003493 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003494 l3->shared = new_shared;
Christoph Lametere498be72005-09-09 13:03:32 -07003495 l3->alien = new_alien;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003496 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003497 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003498 cachep->nodelists[node] = l3;
3499 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003500 return 0;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003501
Andrew Mortona737b3e2006-03-22 00:08:11 -08003502fail:
Christoph Lameter0718dc22006-03-25 03:06:47 -08003503 if (!cachep->next.next) {
3504 /* Cache is not active yet. Roll back what we did */
3505 node--;
3506 while (node >= 0) {
3507 if (cachep->nodelists[node]) {
3508 l3 = cachep->nodelists[node];
3509
3510 kfree(l3->shared);
3511 free_alien_cache(l3->alien);
3512 kfree(l3);
3513 cachep->nodelists[node] = NULL;
3514 }
3515 node--;
3516 }
3517 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003518 return -ENOMEM;
Christoph Lametere498be72005-09-09 13:03:32 -07003519}
3520
Linus Torvalds1da177e2005-04-16 15:20:36 -07003521struct ccupdate_struct {
Pekka Enberg343e0d72006-02-01 03:05:50 -08003522 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003523 struct array_cache *new[NR_CPUS];
3524};
3525
3526static void do_ccupdate_local(void *info)
3527{
Andrew Mortona737b3e2006-03-22 00:08:11 -08003528 struct ccupdate_struct *new = info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003529 struct array_cache *old;
3530
3531 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003532 old = cpu_cache_get(new->cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003533
Linus Torvalds1da177e2005-04-16 15:20:36 -07003534 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3535 new->new[smp_processor_id()] = old;
3536}
3537
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003538/* Always called with the cache_chain_mutex held */
Andrew Mortona737b3e2006-03-22 00:08:11 -08003539static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3540 int batchcount, int shared)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003541{
3542 struct ccupdate_struct new;
Christoph Lametere498be72005-09-09 13:03:32 -07003543 int i, err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003544
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003545 memset(&new.new, 0, sizeof(new.new));
Christoph Lametere498be72005-09-09 13:03:32 -07003546 for_each_online_cpu(i) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003547 new.new[i] = alloc_arraycache(cpu_to_node(i), limit,
3548 batchcount);
Christoph Lametere498be72005-09-09 13:03:32 -07003549 if (!new.new[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003550 for (i--; i >= 0; i--)
3551 kfree(new.new[i]);
Christoph Lametere498be72005-09-09 13:03:32 -07003552 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003553 }
3554 }
3555 new.cachep = cachep;
3556
Andrew Mortona07fa392006-03-22 00:08:17 -08003557 on_each_cpu(do_ccupdate_local, (void *)&new, 1, 1);
Christoph Lametere498be72005-09-09 13:03:32 -07003558
Linus Torvalds1da177e2005-04-16 15:20:36 -07003559 check_irq_on();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003560 cachep->batchcount = batchcount;
3561 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07003562 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003563
Christoph Lametere498be72005-09-09 13:03:32 -07003564 for_each_online_cpu(i) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003565 struct array_cache *ccold = new.new[i];
3566 if (!ccold)
3567 continue;
Christoph Lametere498be72005-09-09 13:03:32 -07003568 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -07003569 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
Christoph Lametere498be72005-09-09 13:03:32 -07003570 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003571 kfree(ccold);
3572 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003573
Christoph Lametere498be72005-09-09 13:03:32 -07003574 err = alloc_kmemlist(cachep);
3575 if (err) {
3576 printk(KERN_ERR "alloc_kmemlist failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003577 cachep->name, -err);
Christoph Lametere498be72005-09-09 13:03:32 -07003578 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003579 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003580 return 0;
3581}
3582
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003583/* Called with cache_chain_mutex held always */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003584static void enable_cpucache(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003585{
3586 int err;
3587 int limit, shared;
3588
Andrew Mortona737b3e2006-03-22 00:08:11 -08003589 /*
3590 * The head array serves three purposes:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003591 * - create a LIFO ordering, i.e. return objects that are cache-warm
3592 * - reduce the number of spinlock operations.
Andrew Mortona737b3e2006-03-22 00:08:11 -08003593 * - reduce the number of linked list operations on the slab and
Linus Torvalds1da177e2005-04-16 15:20:36 -07003594 * bufctl chains: array operations are cheaper.
3595 * The numbers are guessed, we should auto-tune as described by
3596 * Bonwick.
3597 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003598 if (cachep->buffer_size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003599 limit = 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003600 else if (cachep->buffer_size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003601 limit = 8;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003602 else if (cachep->buffer_size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003603 limit = 24;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003604 else if (cachep->buffer_size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003605 limit = 54;
3606 else
3607 limit = 120;
3608
Andrew Mortona737b3e2006-03-22 00:08:11 -08003609 /*
3610 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
Linus Torvalds1da177e2005-04-16 15:20:36 -07003611 * allocation behaviour: Most allocs on one cpu, most free operations
3612 * on another cpu. For these cases, an efficient object passing between
3613 * cpus is necessary. This is provided by a shared array. The array
3614 * replaces Bonwick's magazine layer.
3615 * On uniprocessor, it's functionally equivalent (but less efficient)
3616 * to a larger limit. Thus disabled by default.
3617 */
3618 shared = 0;
3619#ifdef CONFIG_SMP
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003620 if (cachep->buffer_size <= PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003621 shared = 8;
3622#endif
3623
3624#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003625 /*
3626 * With debugging enabled, large batchcount lead to excessively long
3627 * periods with disabled local interrupts. Limit the batchcount
Linus Torvalds1da177e2005-04-16 15:20:36 -07003628 */
3629 if (limit > 32)
3630 limit = 32;
3631#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003632 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003633 if (err)
3634 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003635 cachep->name, -err);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003636}
3637
Christoph Lameter1b552532006-03-22 00:09:07 -08003638/*
3639 * Drain an array if it contains any elements taking the l3 lock only if
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003640 * necessary. Note that the l3 listlock also protects the array_cache
3641 * if drain_array() is used on the shared array.
Christoph Lameter1b552532006-03-22 00:09:07 -08003642 */
3643void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
3644 struct array_cache *ac, int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003645{
3646 int tofree;
3647
Christoph Lameter1b552532006-03-22 00:09:07 -08003648 if (!ac || !ac->avail)
3649 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003650 if (ac->touched && !force) {
3651 ac->touched = 0;
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003652 } else {
Christoph Lameter1b552532006-03-22 00:09:07 -08003653 spin_lock_irq(&l3->list_lock);
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003654 if (ac->avail) {
3655 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3656 if (tofree > ac->avail)
3657 tofree = (ac->avail + 1) / 2;
3658 free_block(cachep, ac->entry, tofree, node);
3659 ac->avail -= tofree;
3660 memmove(ac->entry, &(ac->entry[tofree]),
3661 sizeof(void *) * ac->avail);
3662 }
Christoph Lameter1b552532006-03-22 00:09:07 -08003663 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003664 }
3665}
3666
3667/**
3668 * cache_reap - Reclaim memory from caches.
Randy Dunlap1e5d5332005-11-07 01:01:06 -08003669 * @unused: unused parameter
Linus Torvalds1da177e2005-04-16 15:20:36 -07003670 *
3671 * Called from workqueue/eventd every few seconds.
3672 * Purpose:
3673 * - clear the per-cpu caches for this CPU.
3674 * - return freeable pages to the main free memory pool.
3675 *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003676 * If we cannot acquire the cache chain mutex then just give up - we'll try
3677 * again on the next iteration.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003678 */
3679static void cache_reap(void *unused)
3680{
3681 struct list_head *walk;
Christoph Lametere498be72005-09-09 13:03:32 -07003682 struct kmem_list3 *l3;
Christoph Lameteraab22072006-03-22 00:09:06 -08003683 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003684
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003685 if (!mutex_trylock(&cache_chain_mutex)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003686 /* Give up. Setup the next iteration. */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003687 schedule_delayed_work(&__get_cpu_var(reap_work),
3688 REAPTIMEOUT_CPUC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003689 return;
3690 }
3691
3692 list_for_each(walk, &cache_chain) {
Pekka Enberg343e0d72006-02-01 03:05:50 -08003693 struct kmem_cache *searchp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003694 struct list_head *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003695 int tofree;
3696 struct slab *slabp;
3697
Pekka Enberg343e0d72006-02-01 03:05:50 -08003698 searchp = list_entry(walk, struct kmem_cache, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003699 check_irq_on();
3700
Christoph Lameter35386e32006-03-22 00:09:05 -08003701 /*
3702 * We only take the l3 lock if absolutely necessary and we
3703 * have established with reasonable certainty that
3704 * we can do some work if the lock was obtained.
3705 */
Christoph Lameteraab22072006-03-22 00:09:06 -08003706 l3 = searchp->nodelists[node];
Christoph Lameter35386e32006-03-22 00:09:05 -08003707
Christoph Lameter8fce4d82006-03-09 17:33:54 -08003708 reap_alien(searchp, l3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003709
Christoph Lameteraab22072006-03-22 00:09:06 -08003710 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003711
Christoph Lameter35386e32006-03-22 00:09:05 -08003712 /*
3713 * These are racy checks but it does not matter
3714 * if we skip one check or scan twice.
3715 */
Christoph Lametere498be72005-09-09 13:03:32 -07003716 if (time_after(l3->next_reap, jiffies))
Christoph Lameter35386e32006-03-22 00:09:05 -08003717 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003718
Christoph Lametere498be72005-09-09 13:03:32 -07003719 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003720
Christoph Lameteraab22072006-03-22 00:09:06 -08003721 drain_array(searchp, l3, l3->shared, 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003722
Christoph Lametere498be72005-09-09 13:03:32 -07003723 if (l3->free_touched) {
3724 l3->free_touched = 0;
Christoph Lameter35386e32006-03-22 00:09:05 -08003725 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003726 }
3727
Andrew Mortona737b3e2006-03-22 00:08:11 -08003728 tofree = (l3->free_limit + 5 * searchp->num - 1) /
3729 (5 * searchp->num);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003730 do {
Christoph Lameter35386e32006-03-22 00:09:05 -08003731 /*
3732 * Do not lock if there are no free blocks.
3733 */
3734 if (list_empty(&l3->slabs_free))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003735 break;
3736
Christoph Lameter35386e32006-03-22 00:09:05 -08003737 spin_lock_irq(&l3->list_lock);
3738 p = l3->slabs_free.next;
3739 if (p == &(l3->slabs_free)) {
3740 spin_unlock_irq(&l3->list_lock);
3741 break;
3742 }
3743
Linus Torvalds1da177e2005-04-16 15:20:36 -07003744 slabp = list_entry(p, struct slab, list);
3745 BUG_ON(slabp->inuse);
3746 list_del(&slabp->list);
3747 STATS_INC_REAPED(searchp);
3748
Andrew Mortona737b3e2006-03-22 00:08:11 -08003749 /*
3750 * Safe to drop the lock. The slab is no longer linked
3751 * to the cache. searchp cannot disappear, we hold
Linus Torvalds1da177e2005-04-16 15:20:36 -07003752 * cache_chain_lock
3753 */
Christoph Lametere498be72005-09-09 13:03:32 -07003754 l3->free_objects -= searchp->num;
3755 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003756 slab_destroy(searchp, slabp);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003757 } while (--tofree > 0);
Christoph Lameter35386e32006-03-22 00:09:05 -08003758next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003759 cond_resched();
3760 }
3761 check_irq_on();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003762 mutex_unlock(&cache_chain_mutex);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08003763 next_reap_node();
Andrew Mortona737b3e2006-03-22 00:08:11 -08003764 /* Set up the next iteration */
Manfred Spraulcd61ef62005-11-07 00:58:02 -08003765 schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003766}
3767
3768#ifdef CONFIG_PROC_FS
3769
Pekka Enberg85289f92006-01-08 01:00:36 -08003770static void print_slabinfo_header(struct seq_file *m)
3771{
3772 /*
3773 * Output format version, so at least we can change it
3774 * without _too_ many complaints.
3775 */
3776#if STATS
3777 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
3778#else
3779 seq_puts(m, "slabinfo - version: 2.1\n");
3780#endif
3781 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
3782 "<objperslab> <pagesperslab>");
3783 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
3784 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
3785#if STATS
3786 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07003787 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
Pekka Enberg85289f92006-01-08 01:00:36 -08003788 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
3789#endif
3790 seq_putc(m, '\n');
3791}
3792
Linus Torvalds1da177e2005-04-16 15:20:36 -07003793static void *s_start(struct seq_file *m, loff_t *pos)
3794{
3795 loff_t n = *pos;
3796 struct list_head *p;
3797
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003798 mutex_lock(&cache_chain_mutex);
Pekka Enberg85289f92006-01-08 01:00:36 -08003799 if (!n)
3800 print_slabinfo_header(m);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003801 p = cache_chain.next;
3802 while (n--) {
3803 p = p->next;
3804 if (p == &cache_chain)
3805 return NULL;
3806 }
Pekka Enberg343e0d72006-02-01 03:05:50 -08003807 return list_entry(p, struct kmem_cache, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003808}
3809
3810static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3811{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003812 struct kmem_cache *cachep = p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003813 ++*pos;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003814 return cachep->next.next == &cache_chain ?
3815 NULL : list_entry(cachep->next.next, struct kmem_cache, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003816}
3817
3818static void s_stop(struct seq_file *m, void *p)
3819{
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003820 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003821}
3822
3823static int s_show(struct seq_file *m, void *p)
3824{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003825 struct kmem_cache *cachep = p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003826 struct list_head *q;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003827 struct slab *slabp;
3828 unsigned long active_objs;
3829 unsigned long num_objs;
3830 unsigned long active_slabs = 0;
3831 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07003832 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003833 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003834 int node;
3835 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003836
Linus Torvalds1da177e2005-04-16 15:20:36 -07003837 active_objs = 0;
3838 num_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07003839 for_each_online_node(node) {
3840 l3 = cachep->nodelists[node];
3841 if (!l3)
3842 continue;
3843
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003844 check_irq_on();
3845 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003846
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003847 list_for_each(q, &l3->slabs_full) {
Christoph Lametere498be72005-09-09 13:03:32 -07003848 slabp = list_entry(q, struct slab, list);
3849 if (slabp->inuse != cachep->num && !error)
3850 error = "slabs_full accounting error";
3851 active_objs += cachep->num;
3852 active_slabs++;
3853 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003854 list_for_each(q, &l3->slabs_partial) {
Christoph Lametere498be72005-09-09 13:03:32 -07003855 slabp = list_entry(q, struct slab, list);
3856 if (slabp->inuse == cachep->num && !error)
3857 error = "slabs_partial inuse accounting error";
3858 if (!slabp->inuse && !error)
3859 error = "slabs_partial/inuse accounting error";
3860 active_objs += slabp->inuse;
3861 active_slabs++;
3862 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003863 list_for_each(q, &l3->slabs_free) {
Christoph Lametere498be72005-09-09 13:03:32 -07003864 slabp = list_entry(q, struct slab, list);
3865 if (slabp->inuse && !error)
3866 error = "slabs_free/inuse accounting error";
3867 num_slabs++;
3868 }
3869 free_objects += l3->free_objects;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08003870 if (l3->shared)
3871 shared_avail += l3->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07003872
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003873 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003874 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003875 num_slabs += active_slabs;
3876 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003877 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003878 error = "free_objects accounting error";
3879
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003880 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003881 if (error)
3882 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
3883
3884 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003885 name, active_objs, num_objs, cachep->buffer_size,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003886 cachep->num, (1 << cachep->gfporder));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003887 seq_printf(m, " : tunables %4u %4u %4u",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003888 cachep->limit, cachep->batchcount, cachep->shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003889 seq_printf(m, " : slabdata %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003890 active_slabs, num_slabs, shared_avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003891#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003892 { /* list3 stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003893 unsigned long high = cachep->high_mark;
3894 unsigned long allocs = cachep->num_allocations;
3895 unsigned long grown = cachep->grown;
3896 unsigned long reaped = cachep->reaped;
3897 unsigned long errors = cachep->errors;
3898 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003899 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07003900 unsigned long node_frees = cachep->node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07003901 unsigned long overflows = cachep->node_overflow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003902
Christoph Lametere498be72005-09-09 13:03:32 -07003903 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07003904 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003905 reaped, errors, max_freeable, node_allocs,
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07003906 node_frees, overflows);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003907 }
3908 /* cpu stats */
3909 {
3910 unsigned long allochit = atomic_read(&cachep->allochit);
3911 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
3912 unsigned long freehit = atomic_read(&cachep->freehit);
3913 unsigned long freemiss = atomic_read(&cachep->freemiss);
3914
3915 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003916 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003917 }
3918#endif
3919 seq_putc(m, '\n');
Linus Torvalds1da177e2005-04-16 15:20:36 -07003920 return 0;
3921}
3922
3923/*
3924 * slabinfo_op - iterator that generates /proc/slabinfo
3925 *
3926 * Output layout:
3927 * cache-name
3928 * num-active-objs
3929 * total-objs
3930 * object size
3931 * num-active-slabs
3932 * total-slabs
3933 * num-pages-per-slab
3934 * + further values on SMP and with statistics enabled
3935 */
3936
3937struct seq_operations slabinfo_op = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003938 .start = s_start,
3939 .next = s_next,
3940 .stop = s_stop,
3941 .show = s_show,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003942};
3943
3944#define MAX_SLABINFO_WRITE 128
3945/**
3946 * slabinfo_write - Tuning for the slab allocator
3947 * @file: unused
3948 * @buffer: user buffer
3949 * @count: data length
3950 * @ppos: unused
3951 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003952ssize_t slabinfo_write(struct file *file, const char __user * buffer,
3953 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003954{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003955 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003956 int limit, batchcount, shared, res;
3957 struct list_head *p;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003958
Linus Torvalds1da177e2005-04-16 15:20:36 -07003959 if (count > MAX_SLABINFO_WRITE)
3960 return -EINVAL;
3961 if (copy_from_user(&kbuf, buffer, count))
3962 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003963 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07003964
3965 tmp = strchr(kbuf, ' ');
3966 if (!tmp)
3967 return -EINVAL;
3968 *tmp = '\0';
3969 tmp++;
3970 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
3971 return -EINVAL;
3972
3973 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003974 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003975 res = -EINVAL;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003976 list_for_each(p, &cache_chain) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003977 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003978
Andrew Mortona737b3e2006-03-22 00:08:11 -08003979 cachep = list_entry(p, struct kmem_cache, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003980 if (!strcmp(cachep->name, kbuf)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003981 if (limit < 1 || batchcount < 1 ||
3982 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07003983 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003984 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07003985 res = do_tune_cpucache(cachep, limit,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003986 batchcount, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003987 }
3988 break;
3989 }
3990 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003991 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003992 if (res >= 0)
3993 res = count;
3994 return res;
3995}
Al Viro871751e2006-03-25 03:06:39 -08003996
3997#ifdef CONFIG_DEBUG_SLAB_LEAK
3998
3999static void *leaks_start(struct seq_file *m, loff_t *pos)
4000{
4001 loff_t n = *pos;
4002 struct list_head *p;
4003
4004 mutex_lock(&cache_chain_mutex);
4005 p = cache_chain.next;
4006 while (n--) {
4007 p = p->next;
4008 if (p == &cache_chain)
4009 return NULL;
4010 }
4011 return list_entry(p, struct kmem_cache, next);
4012}
4013
4014static inline int add_caller(unsigned long *n, unsigned long v)
4015{
4016 unsigned long *p;
4017 int l;
4018 if (!v)
4019 return 1;
4020 l = n[1];
4021 p = n + 2;
4022 while (l) {
4023 int i = l/2;
4024 unsigned long *q = p + 2 * i;
4025 if (*q == v) {
4026 q[1]++;
4027 return 1;
4028 }
4029 if (*q > v) {
4030 l = i;
4031 } else {
4032 p = q + 2;
4033 l -= i + 1;
4034 }
4035 }
4036 if (++n[1] == n[0])
4037 return 0;
4038 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4039 p[0] = v;
4040 p[1] = 1;
4041 return 1;
4042}
4043
4044static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4045{
4046 void *p;
4047 int i;
4048 if (n[0] == n[1])
4049 return;
4050 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4051 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4052 continue;
4053 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4054 return;
4055 }
4056}
4057
4058static void show_symbol(struct seq_file *m, unsigned long address)
4059{
4060#ifdef CONFIG_KALLSYMS
4061 char *modname;
4062 const char *name;
4063 unsigned long offset, size;
4064 char namebuf[KSYM_NAME_LEN+1];
4065
4066 name = kallsyms_lookup(address, &size, &offset, &modname, namebuf);
4067
4068 if (name) {
4069 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4070 if (modname)
4071 seq_printf(m, " [%s]", modname);
4072 return;
4073 }
4074#endif
4075 seq_printf(m, "%p", (void *)address);
4076}
4077
4078static int leaks_show(struct seq_file *m, void *p)
4079{
4080 struct kmem_cache *cachep = p;
4081 struct list_head *q;
4082 struct slab *slabp;
4083 struct kmem_list3 *l3;
4084 const char *name;
4085 unsigned long *n = m->private;
4086 int node;
4087 int i;
4088
4089 if (!(cachep->flags & SLAB_STORE_USER))
4090 return 0;
4091 if (!(cachep->flags & SLAB_RED_ZONE))
4092 return 0;
4093
4094 /* OK, we can do it */
4095
4096 n[1] = 0;
4097
4098 for_each_online_node(node) {
4099 l3 = cachep->nodelists[node];
4100 if (!l3)
4101 continue;
4102
4103 check_irq_on();
4104 spin_lock_irq(&l3->list_lock);
4105
4106 list_for_each(q, &l3->slabs_full) {
4107 slabp = list_entry(q, struct slab, list);
4108 handle_slab(n, cachep, slabp);
4109 }
4110 list_for_each(q, &l3->slabs_partial) {
4111 slabp = list_entry(q, struct slab, list);
4112 handle_slab(n, cachep, slabp);
4113 }
4114 spin_unlock_irq(&l3->list_lock);
4115 }
4116 name = cachep->name;
4117 if (n[0] == n[1]) {
4118 /* Increase the buffer size */
4119 mutex_unlock(&cache_chain_mutex);
4120 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4121 if (!m->private) {
4122 /* Too bad, we are really out */
4123 m->private = n;
4124 mutex_lock(&cache_chain_mutex);
4125 return -ENOMEM;
4126 }
4127 *(unsigned long *)m->private = n[0] * 2;
4128 kfree(n);
4129 mutex_lock(&cache_chain_mutex);
4130 /* Now make sure this entry will be retried */
4131 m->count = m->size;
4132 return 0;
4133 }
4134 for (i = 0; i < n[1]; i++) {
4135 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4136 show_symbol(m, n[2*i+2]);
4137 seq_putc(m, '\n');
4138 }
4139 return 0;
4140}
4141
4142struct seq_operations slabstats_op = {
4143 .start = leaks_start,
4144 .next = s_next,
4145 .stop = s_stop,
4146 .show = leaks_show,
4147};
4148#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004149#endif
4150
Manfred Spraul00e145b2005-09-03 15:55:07 -07004151/**
4152 * ksize - get the actual amount of memory allocated for a given object
4153 * @objp: Pointer to the object
4154 *
4155 * kmalloc may internally round up allocations and return more memory
4156 * than requested. ksize() can be used to determine the actual amount of
4157 * memory allocated. The caller may use this additional memory, even though
4158 * a smaller amount of memory was initially specified with the kmalloc call.
4159 * The caller must guarantee that objp points to a valid object previously
4160 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4161 * must not be freed during the duration of the call.
4162 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004163unsigned int ksize(const void *objp)
4164{
Manfred Spraul00e145b2005-09-03 15:55:07 -07004165 if (unlikely(objp == NULL))
4166 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004167
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08004168 return obj_size(virt_to_cache(objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004169}