blob: 792bfe320a8b2c0896f8568f8d07de40203f8cf2 [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>
Randy Dunlapc9cf5522006-06-27 02:53:52 -070092#include <linux/poison.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070093#include <linux/swap.h>
94#include <linux/cache.h>
95#include <linux/interrupt.h>
96#include <linux/init.h>
97#include <linux/compiler.h>
Paul Jackson101a5002006-03-24 03:16:07 -080098#include <linux/cpuset.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070099#include <linux/seq_file.h>
100#include <linux/notifier.h>
101#include <linux/kallsyms.h>
102#include <linux/cpu.h>
103#include <linux/sysctl.h>
104#include <linux/module.h>
105#include <linux/rcupdate.h>
Paulo Marques543537b2005-06-23 00:09:02 -0700106#include <linux/string.h>
Christoph Lametere498be72005-09-09 13:03:32 -0700107#include <linux/nodemask.h>
Christoph Lameterdc85da12006-01-18 17:42:36 -0800108#include <linux/mempolicy.h>
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800109#include <linux/mutex.h>
Ingo Molnare7eebaf2006-06-27 02:54:55 -0700110#include <linux/rtmutex.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700111
112#include <asm/uaccess.h>
113#include <asm/cacheflush.h>
114#include <asm/tlbflush.h>
115#include <asm/page.h>
116
117/*
118 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL,
119 * SLAB_RED_ZONE & SLAB_POISON.
120 * 0 for faster, smaller code (especially in the critical paths).
121 *
122 * STATS - 1 to collect stats for /proc/slabinfo.
123 * 0 for faster, smaller code (especially in the critical paths).
124 *
125 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
126 */
127
128#ifdef CONFIG_DEBUG_SLAB
129#define DEBUG 1
130#define STATS 1
131#define FORCED_DEBUG 1
132#else
133#define DEBUG 0
134#define STATS 0
135#define FORCED_DEBUG 0
136#endif
137
Linus Torvalds1da177e2005-04-16 15:20:36 -0700138/* Shouldn't this be in a header file somewhere? */
139#define BYTES_PER_WORD sizeof(void *)
140
141#ifndef cache_line_size
142#define cache_line_size() L1_CACHE_BYTES
143#endif
144
145#ifndef ARCH_KMALLOC_MINALIGN
146/*
147 * Enforce a minimum alignment for the kmalloc caches.
148 * Usually, the kmalloc caches are cache_line_size() aligned, except when
149 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
150 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
151 * alignment larger than BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that.
152 * Note that this flag disables some debug features.
153 */
154#define ARCH_KMALLOC_MINALIGN 0
155#endif
156
157#ifndef ARCH_SLAB_MINALIGN
158/*
159 * Enforce a minimum alignment for all caches.
160 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
161 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
162 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
163 * some debug features.
164 */
165#define ARCH_SLAB_MINALIGN 0
166#endif
167
168#ifndef ARCH_KMALLOC_FLAGS
169#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
170#endif
171
172/* Legal flag mask for kmem_cache_create(). */
173#if DEBUG
174# define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \
175 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
Christoph Lameterac2b8982006-03-22 00:08:15 -0800176 SLAB_CACHE_DMA | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700177 SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \
178 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
Paul Jackson101a5002006-03-24 03:16:07 -0800179 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700180#else
Christoph Lameterac2b8982006-03-22 00:08:15 -0800181# define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700182 SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \
183 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
Paul Jackson101a5002006-03-24 03:16:07 -0800184 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700185#endif
186
187/*
188 * kmem_bufctl_t:
189 *
190 * Bufctl's are used for linking objs within a slab
191 * linked offsets.
192 *
193 * This implementation relies on "struct page" for locating the cache &
194 * slab an object belongs to.
195 * This allows the bufctl structure to be small (one int), but limits
196 * the number of objects a slab (not a cache) can contain when off-slab
197 * bufctls are used. The limit is the size of the largest general cache
198 * that does not use off-slab slabs.
199 * For 32bit archs with 4 kB pages, is this 56.
200 * This is not serious, as it is only for large objects, when it is unwise
201 * to have too many per slab.
202 * Note: This limit can be raised by introducing a general cache whose size
203 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
204 */
205
Kyle Moffettfa5b08d2005-09-03 15:55:03 -0700206typedef unsigned int kmem_bufctl_t;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700207#define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
208#define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
Al Viro871751e2006-03-25 03:06:39 -0800209#define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
210#define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700211
Linus Torvalds1da177e2005-04-16 15:20:36 -0700212/*
213 * struct slab
214 *
215 * Manages the objs in a slab. Placed either at the beginning of mem allocated
216 * for a slab, or allocated from an general cache.
217 * Slabs are chained into three list: fully used, partial, fully free slabs.
218 */
219struct slab {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800220 struct list_head list;
221 unsigned long colouroff;
222 void *s_mem; /* including colour offset */
223 unsigned int inuse; /* num of objs active in slab */
224 kmem_bufctl_t free;
225 unsigned short nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700226};
227
228/*
229 * struct slab_rcu
230 *
231 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
232 * arrange for kmem_freepages to be called via RCU. This is useful if
233 * we need to approach a kernel structure obliquely, from its address
234 * obtained without the usual locking. We can lock the structure to
235 * stabilize it and check it's still at the given address, only if we
236 * can be sure that the memory has not been meanwhile reused for some
237 * other kind of object (which our subsystem's lock might corrupt).
238 *
239 * rcu_read_lock before reading the address, then rcu_read_unlock after
240 * taking the spinlock within the structure expected at that address.
241 *
242 * We assume struct slab_rcu can overlay struct slab when destroying.
243 */
244struct slab_rcu {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800245 struct rcu_head head;
Pekka Enberg343e0d72006-02-01 03:05:50 -0800246 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800247 void *addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700248};
249
250/*
251 * struct array_cache
252 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700253 * Purpose:
254 * - LIFO ordering, to hand out cache-warm objects from _alloc
255 * - reduce the number of linked list operations
256 * - reduce spinlock operations
257 *
258 * The limit is stored in the per-cpu structure to reduce the data cache
259 * footprint.
260 *
261 */
262struct array_cache {
263 unsigned int avail;
264 unsigned int limit;
265 unsigned int batchcount;
266 unsigned int touched;
Christoph Lametere498be72005-09-09 13:03:32 -0700267 spinlock_t lock;
Andrew Mortona737b3e2006-03-22 00:08:11 -0800268 void *entry[0]; /*
269 * Must have this definition in here for the proper
270 * alignment of array_cache. Also simplifies accessing
271 * the entries.
272 * [0] is for gcc 2.95. It should really be [].
273 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700274};
275
Andrew Mortona737b3e2006-03-22 00:08:11 -0800276/*
277 * bootstrap: The caches do not work without cpuarrays anymore, but the
278 * cpuarrays are allocated from the generic caches...
Linus Torvalds1da177e2005-04-16 15:20:36 -0700279 */
280#define BOOT_CPUCACHE_ENTRIES 1
281struct arraycache_init {
282 struct array_cache cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800283 void *entries[BOOT_CPUCACHE_ENTRIES];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700284};
285
286/*
Christoph Lametere498be72005-09-09 13:03:32 -0700287 * The slab lists for all objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700288 */
289struct kmem_list3 {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800290 struct list_head slabs_partial; /* partial list first, better asm code */
291 struct list_head slabs_full;
292 struct list_head slabs_free;
293 unsigned long free_objects;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800294 unsigned int free_limit;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800295 unsigned int colour_next; /* Per-node cache coloring */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800296 spinlock_t list_lock;
297 struct array_cache *shared; /* shared per node */
298 struct array_cache **alien; /* on other nodes */
Christoph Lameter35386e32006-03-22 00:09:05 -0800299 unsigned long next_reap; /* updated without locking */
300 int free_touched; /* updated without locking */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700301};
302
Christoph Lametere498be72005-09-09 13:03:32 -0700303/*
304 * Need this for bootstrapping a per node allocator.
305 */
306#define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
307struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
308#define CACHE_CACHE 0
309#define SIZE_AC 1
310#define SIZE_L3 (1 + MAX_NUMNODES)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700311
Christoph Lametered11d9e2006-06-30 01:55:45 -0700312static int drain_freelist(struct kmem_cache *cache,
313 struct kmem_list3 *l3, int tofree);
314static void free_block(struct kmem_cache *cachep, void **objpp, int len,
315 int node);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -0700316static int enable_cpucache(struct kmem_cache *cachep);
Christoph Lametered11d9e2006-06-30 01:55:45 -0700317static void cache_reap(void *unused);
318
Christoph Lametere498be72005-09-09 13:03:32 -0700319/*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800320 * This function must be completely optimized away if a constant is passed to
321 * it. Mostly the same as what is in linux/slab.h except it returns an index.
Christoph Lametere498be72005-09-09 13:03:32 -0700322 */
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700323static __always_inline int index_of(const size_t size)
Christoph Lametere498be72005-09-09 13:03:32 -0700324{
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800325 extern void __bad_size(void);
326
Christoph Lametere498be72005-09-09 13:03:32 -0700327 if (__builtin_constant_p(size)) {
328 int i = 0;
329
330#define CACHE(x) \
331 if (size <=x) \
332 return i; \
333 else \
334 i++;
335#include "linux/kmalloc_sizes.h"
336#undef CACHE
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800337 __bad_size();
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700338 } else
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800339 __bad_size();
Christoph Lametere498be72005-09-09 13:03:32 -0700340 return 0;
341}
342
Ingo Molnare0a42722006-06-23 02:03:46 -0700343static int slab_early_init = 1;
344
Christoph Lametere498be72005-09-09 13:03:32 -0700345#define INDEX_AC index_of(sizeof(struct arraycache_init))
346#define INDEX_L3 index_of(sizeof(struct kmem_list3))
347
Pekka Enberg5295a742006-02-01 03:05:48 -0800348static void kmem_list3_init(struct kmem_list3 *parent)
Christoph Lametere498be72005-09-09 13:03:32 -0700349{
350 INIT_LIST_HEAD(&parent->slabs_full);
351 INIT_LIST_HEAD(&parent->slabs_partial);
352 INIT_LIST_HEAD(&parent->slabs_free);
353 parent->shared = NULL;
354 parent->alien = NULL;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800355 parent->colour_next = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700356 spin_lock_init(&parent->list_lock);
357 parent->free_objects = 0;
358 parent->free_touched = 0;
359}
360
Andrew Mortona737b3e2006-03-22 00:08:11 -0800361#define MAKE_LIST(cachep, listp, slab, nodeid) \
362 do { \
363 INIT_LIST_HEAD(listp); \
364 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
Christoph Lametere498be72005-09-09 13:03:32 -0700365 } while (0)
366
Andrew Mortona737b3e2006-03-22 00:08:11 -0800367#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
368 do { \
Christoph Lametere498be72005-09-09 13:03:32 -0700369 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
370 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
371 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
372 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700373
374/*
Pekka Enberg343e0d72006-02-01 03:05:50 -0800375 * struct kmem_cache
Linus Torvalds1da177e2005-04-16 15:20:36 -0700376 *
377 * manages a cache.
378 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800379
Pekka J Enberg2109a2d2005-11-07 00:58:01 -0800380struct kmem_cache {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700381/* 1) per-cpu data, touched during every alloc/free */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800382 struct array_cache *array[NR_CPUS];
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800383/* 2) Cache tunables. Protected by cache_chain_mutex */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800384 unsigned int batchcount;
385 unsigned int limit;
386 unsigned int shared;
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800387
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800388 unsigned int buffer_size;
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800389/* 3) touched by every alloc & free from the backend */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800390 struct kmem_list3 *nodelists[MAX_NUMNODES];
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800391
Andrew Mortona737b3e2006-03-22 00:08:11 -0800392 unsigned int flags; /* constant flags */
393 unsigned int num; /* # of objs per slab */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700394
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800395/* 4) cache_grow/shrink */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700396 /* order of pgs per slab (2^n) */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800397 unsigned int gfporder;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700398
399 /* force GFP flags, e.g. GFP_DMA */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800400 gfp_t gfpflags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700401
Andrew Mortona737b3e2006-03-22 00:08:11 -0800402 size_t colour; /* cache colouring range */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800403 unsigned int colour_off; /* colour offset */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800404 struct kmem_cache *slabp_cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800405 unsigned int slab_size;
Andrew Mortona737b3e2006-03-22 00:08:11 -0800406 unsigned int dflags; /* dynamic flags */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700407
408 /* constructor func */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800409 void (*ctor) (void *, struct kmem_cache *, unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700410
411 /* de-constructor func */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800412 void (*dtor) (void *, struct kmem_cache *, unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700413
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800414/* 5) cache creation/removal */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800415 const char *name;
416 struct list_head next;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700417
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800418/* 6) statistics */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700419#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800420 unsigned long num_active;
421 unsigned long num_allocations;
422 unsigned long high_mark;
423 unsigned long grown;
424 unsigned long reaped;
425 unsigned long errors;
426 unsigned long max_freeable;
427 unsigned long node_allocs;
428 unsigned long node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700429 unsigned long node_overflow;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800430 atomic_t allochit;
431 atomic_t allocmiss;
432 atomic_t freehit;
433 atomic_t freemiss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700434#endif
435#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800436 /*
437 * If debugging is enabled, then the allocator can add additional
438 * fields and/or padding to every object. buffer_size contains the total
439 * object size including these internal fields, the following two
440 * variables contain the offset to the user object and its size.
441 */
442 int obj_offset;
443 int obj_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700444#endif
445};
446
447#define CFLGS_OFF_SLAB (0x80000000UL)
448#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
449
450#define BATCHREFILL_LIMIT 16
Andrew Mortona737b3e2006-03-22 00:08:11 -0800451/*
452 * Optimization question: fewer reaps means less probability for unnessary
453 * cpucache drain/refill cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700454 *
Adrian Bunkdc6f3f22005-11-08 16:44:08 +0100455 * OTOH the cpuarrays can contain lots of objects,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700456 * which could lock up otherwise freeable slabs.
457 */
458#define REAPTIMEOUT_CPUC (2*HZ)
459#define REAPTIMEOUT_LIST3 (4*HZ)
460
461#if STATS
462#define STATS_INC_ACTIVE(x) ((x)->num_active++)
463#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
464#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
465#define STATS_INC_GROWN(x) ((x)->grown++)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700466#define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
Andrew Mortona737b3e2006-03-22 00:08:11 -0800467#define STATS_SET_HIGH(x) \
468 do { \
469 if ((x)->num_active > (x)->high_mark) \
470 (x)->high_mark = (x)->num_active; \
471 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700472#define STATS_INC_ERR(x) ((x)->errors++)
473#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
Christoph Lametere498be72005-09-09 13:03:32 -0700474#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700475#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800476#define STATS_SET_FREEABLE(x, i) \
477 do { \
478 if ((x)->max_freeable < i) \
479 (x)->max_freeable = i; \
480 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700481#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
482#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
483#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
484#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
485#else
486#define STATS_INC_ACTIVE(x) do { } while (0)
487#define STATS_DEC_ACTIVE(x) do { } while (0)
488#define STATS_INC_ALLOCED(x) do { } while (0)
489#define STATS_INC_GROWN(x) do { } while (0)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700490#define STATS_ADD_REAPED(x,y) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700491#define STATS_SET_HIGH(x) do { } while (0)
492#define STATS_INC_ERR(x) do { } while (0)
493#define STATS_INC_NODEALLOCS(x) do { } while (0)
Christoph Lametere498be72005-09-09 13:03:32 -0700494#define STATS_INC_NODEFREES(x) do { } while (0)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700495#define STATS_INC_ACOVERFLOW(x) do { } while (0)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800496#define STATS_SET_FREEABLE(x, i) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700497#define STATS_INC_ALLOCHIT(x) do { } while (0)
498#define STATS_INC_ALLOCMISS(x) do { } while (0)
499#define STATS_INC_FREEHIT(x) do { } while (0)
500#define STATS_INC_FREEMISS(x) do { } while (0)
501#endif
502
503#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -0700504
Andrew Mortona737b3e2006-03-22 00:08:11 -0800505/*
506 * memory layout of objects:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700507 * 0 : objp
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800508 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
Linus Torvalds1da177e2005-04-16 15:20:36 -0700509 * the end of an object is aligned with the end of the real
510 * allocation. Catches writes behind the end of the allocation.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800511 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700512 * redzone word.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800513 * cachep->obj_offset: The real object.
514 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
Andrew Mortona737b3e2006-03-22 00:08:11 -0800515 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
516 * [BYTES_PER_WORD long]
Linus Torvalds1da177e2005-04-16 15:20:36 -0700517 */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800518static int obj_offset(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700519{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800520 return cachep->obj_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700521}
522
Pekka Enberg343e0d72006-02-01 03:05:50 -0800523static int obj_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700524{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800525 return cachep->obj_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700526}
527
Pekka Enberg343e0d72006-02-01 03:05:50 -0800528static unsigned long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700529{
530 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800531 return (unsigned long*) (objp+obj_offset(cachep)-BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700532}
533
Pekka Enberg343e0d72006-02-01 03:05:50 -0800534static unsigned long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700535{
536 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
537 if (cachep->flags & SLAB_STORE_USER)
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800538 return (unsigned long *)(objp + cachep->buffer_size -
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800539 2 * BYTES_PER_WORD);
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800540 return (unsigned long *)(objp + cachep->buffer_size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700541}
542
Pekka Enberg343e0d72006-02-01 03:05:50 -0800543static void **dbg_userword(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700544{
545 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800546 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700547}
548
549#else
550
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800551#define obj_offset(x) 0
552#define obj_size(cachep) (cachep->buffer_size)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700553#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;})
554#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;})
555#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
556
557#endif
558
559/*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800560 * Maximum size of an obj (in 2^order pages) and absolute limit for the gfp
561 * order.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700562 */
563#if defined(CONFIG_LARGE_ALLOCS)
564#define MAX_OBJ_ORDER 13 /* up to 32Mb */
565#define MAX_GFP_ORDER 13 /* up to 32Mb */
566#elif defined(CONFIG_MMU)
567#define MAX_OBJ_ORDER 5 /* 32 pages */
568#define MAX_GFP_ORDER 5 /* 32 pages */
569#else
570#define MAX_OBJ_ORDER 8 /* up to 1Mb */
571#define MAX_GFP_ORDER 8 /* up to 1Mb */
572#endif
573
574/*
575 * Do not go above this order unless 0 objects fit into the slab.
576 */
577#define BREAK_GFP_ORDER_HI 1
578#define BREAK_GFP_ORDER_LO 0
579static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
580
Andrew Mortona737b3e2006-03-22 00:08:11 -0800581/*
582 * Functions for storing/retrieving the cachep and or slab from the page
583 * allocator. These are used to find the slab an obj belongs to. With kfree(),
584 * these are used to find the cache which an obj belongs to.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700585 */
Pekka Enberg065d41c2005-11-13 16:06:46 -0800586static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
587{
588 page->lru.next = (struct list_head *)cache;
589}
590
591static inline struct kmem_cache *page_get_cache(struct page *page)
592{
Nick Piggin84097512006-03-22 00:08:34 -0800593 if (unlikely(PageCompound(page)))
594 page = (struct page *)page_private(page);
Pekka Enbergddc2e812006-06-23 02:03:40 -0700595 BUG_ON(!PageSlab(page));
Pekka Enberg065d41c2005-11-13 16:06:46 -0800596 return (struct kmem_cache *)page->lru.next;
597}
598
599static inline void page_set_slab(struct page *page, struct slab *slab)
600{
601 page->lru.prev = (struct list_head *)slab;
602}
603
604static inline struct slab *page_get_slab(struct page *page)
605{
Nick Piggin84097512006-03-22 00:08:34 -0800606 if (unlikely(PageCompound(page)))
607 page = (struct page *)page_private(page);
Pekka Enbergddc2e812006-06-23 02:03:40 -0700608 BUG_ON(!PageSlab(page));
Pekka Enberg065d41c2005-11-13 16:06:46 -0800609 return (struct slab *)page->lru.prev;
610}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700611
Pekka Enberg6ed5eb22006-02-01 03:05:49 -0800612static inline struct kmem_cache *virt_to_cache(const void *obj)
613{
614 struct page *page = virt_to_page(obj);
615 return page_get_cache(page);
616}
617
618static inline struct slab *virt_to_slab(const void *obj)
619{
620 struct page *page = virt_to_page(obj);
621 return page_get_slab(page);
622}
623
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800624static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
625 unsigned int idx)
626{
627 return slab->s_mem + cache->buffer_size * idx;
628}
629
630static inline unsigned int obj_to_index(struct kmem_cache *cache,
631 struct slab *slab, void *obj)
632{
633 return (unsigned)(obj - slab->s_mem) / cache->buffer_size;
634}
635
Andrew Mortona737b3e2006-03-22 00:08:11 -0800636/*
637 * These are the default caches for kmalloc. Custom caches can have other sizes.
638 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700639struct cache_sizes malloc_sizes[] = {
640#define CACHE(x) { .cs_size = (x) },
641#include <linux/kmalloc_sizes.h>
642 CACHE(ULONG_MAX)
643#undef CACHE
644};
645EXPORT_SYMBOL(malloc_sizes);
646
647/* Must match cache_sizes above. Out of line to keep cache footprint low. */
648struct cache_names {
649 char *name;
650 char *name_dma;
651};
652
653static struct cache_names __initdata cache_names[] = {
654#define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
655#include <linux/kmalloc_sizes.h>
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800656 {NULL,}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700657#undef CACHE
658};
659
660static struct arraycache_init initarray_cache __initdata =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800661 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700662static struct arraycache_init initarray_generic =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800663 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700664
665/* internal cache of cache description objs */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800666static struct kmem_cache cache_cache = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800667 .batchcount = 1,
668 .limit = BOOT_CPUCACHE_ENTRIES,
669 .shared = 1,
Pekka Enberg343e0d72006-02-01 03:05:50 -0800670 .buffer_size = sizeof(struct kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800671 .name = "kmem_cache",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700672#if DEBUG
Pekka Enberg343e0d72006-02-01 03:05:50 -0800673 .obj_size = sizeof(struct kmem_cache),
Linus Torvalds1da177e2005-04-16 15:20:36 -0700674#endif
675};
676
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700677#define BAD_ALIEN_MAGIC 0x01020304ul
678
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200679#ifdef CONFIG_LOCKDEP
680
681/*
682 * Slab sometimes uses the kmalloc slabs to store the slab headers
683 * for other slabs "off slab".
684 * The locking for this is tricky in that it nests within the locks
685 * of all other slabs in a few places; to deal with this special
686 * locking we put on-slab caches into a separate lock-class.
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700687 *
688 * We set lock class for alien array caches which are up during init.
689 * The lock annotation will be lost if all cpus of a node goes down and
690 * then comes back up during hotplug
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200691 */
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700692static struct lock_class_key on_slab_l3_key;
693static struct lock_class_key on_slab_alc_key;
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200694
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700695static inline void init_lock_keys(void)
696
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200697{
698 int q;
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700699 struct cache_sizes *s = malloc_sizes;
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200700
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700701 while (s->cs_size != ULONG_MAX) {
702 for_each_node(q) {
703 struct array_cache **alc;
704 int r;
705 struct kmem_list3 *l3 = s->cs_cachep->nodelists[q];
706 if (!l3 || OFF_SLAB(s->cs_cachep))
707 continue;
708 lockdep_set_class(&l3->list_lock, &on_slab_l3_key);
709 alc = l3->alien;
710 /*
711 * FIXME: This check for BAD_ALIEN_MAGIC
712 * should go away when common slab code is taught to
713 * work even without alien caches.
714 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
715 * for alloc_alien_cache,
716 */
717 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
718 continue;
719 for_each_node(r) {
720 if (alc[r])
721 lockdep_set_class(&alc[r]->lock,
722 &on_slab_alc_key);
723 }
724 }
725 s++;
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200726 }
727}
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200728#else
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700729static inline void init_lock_keys(void)
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200730{
731}
732#endif
733
Linus Torvalds1da177e2005-04-16 15:20:36 -0700734/* Guard access to the cache-chain. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800735static DEFINE_MUTEX(cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700736static struct list_head cache_chain;
737
738/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700739 * chicken and egg problem: delay the per-cpu array allocation
740 * until the general caches are up.
741 */
742static enum {
743 NONE,
Christoph Lametere498be72005-09-09 13:03:32 -0700744 PARTIAL_AC,
745 PARTIAL_L3,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700746 FULL
747} g_cpucache_up;
748
Mike Kravetz39d24e62006-05-15 09:44:13 -0700749/*
750 * used by boot code to determine if it can use slab based allocator
751 */
752int slab_is_available(void)
753{
754 return g_cpucache_up == FULL;
755}
756
Linus Torvalds1da177e2005-04-16 15:20:36 -0700757static DEFINE_PER_CPU(struct work_struct, reap_work);
758
Pekka Enberg343e0d72006-02-01 03:05:50 -0800759static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700760{
761 return cachep->array[smp_processor_id()];
762}
763
Andrew Mortona737b3e2006-03-22 00:08:11 -0800764static inline struct kmem_cache *__find_general_cachep(size_t size,
765 gfp_t gfpflags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700766{
767 struct cache_sizes *csizep = malloc_sizes;
768
769#if DEBUG
770 /* This happens if someone tries to call
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800771 * kmem_cache_create(), or __kmalloc(), before
772 * the generic caches are initialized.
773 */
Alok Katariac7e43c72005-09-14 12:17:53 -0700774 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700775#endif
776 while (size > csizep->cs_size)
777 csizep++;
778
779 /*
Martin Hicks0abf40c2005-09-03 15:54:54 -0700780 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
Linus Torvalds1da177e2005-04-16 15:20:36 -0700781 * has cs_{dma,}cachep==NULL. Thus no special case
782 * for large kmalloc calls required.
783 */
784 if (unlikely(gfpflags & GFP_DMA))
785 return csizep->cs_dmacachep;
786 return csizep->cs_cachep;
787}
788
Adrian Bunkb2213852006-09-25 23:31:02 -0700789static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700790{
791 return __find_general_cachep(size, gfpflags);
792}
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700793
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800794static size_t slab_mgmt_size(size_t nr_objs, size_t align)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700795{
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800796 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
797}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700798
Andrew Mortona737b3e2006-03-22 00:08:11 -0800799/*
800 * Calculate the number of objects and left-over bytes for a given buffer size.
801 */
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800802static void cache_estimate(unsigned long gfporder, size_t buffer_size,
803 size_t align, int flags, size_t *left_over,
804 unsigned int *num)
805{
806 int nr_objs;
807 size_t mgmt_size;
808 size_t slab_size = PAGE_SIZE << gfporder;
809
810 /*
811 * The slab management structure can be either off the slab or
812 * on it. For the latter case, the memory allocated for a
813 * slab is used for:
814 *
815 * - The struct slab
816 * - One kmem_bufctl_t for each object
817 * - Padding to respect alignment of @align
818 * - @buffer_size bytes for each object
819 *
820 * If the slab management structure is off the slab, then the
821 * alignment will already be calculated into the size. Because
822 * the slabs are all pages aligned, the objects will be at the
823 * correct alignment when allocated.
824 */
825 if (flags & CFLGS_OFF_SLAB) {
826 mgmt_size = 0;
827 nr_objs = slab_size / buffer_size;
828
829 if (nr_objs > SLAB_LIMIT)
830 nr_objs = SLAB_LIMIT;
831 } else {
832 /*
833 * Ignore padding for the initial guess. The padding
834 * is at most @align-1 bytes, and @buffer_size is at
835 * least @align. In the worst case, this result will
836 * be one greater than the number of objects that fit
837 * into the memory allocation when taking the padding
838 * into account.
839 */
840 nr_objs = (slab_size - sizeof(struct slab)) /
841 (buffer_size + sizeof(kmem_bufctl_t));
842
843 /*
844 * This calculated number will be either the right
845 * amount, or one greater than what we want.
846 */
847 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
848 > slab_size)
849 nr_objs--;
850
851 if (nr_objs > SLAB_LIMIT)
852 nr_objs = SLAB_LIMIT;
853
854 mgmt_size = slab_mgmt_size(nr_objs, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700855 }
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800856 *num = nr_objs;
857 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700858}
859
860#define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
861
Andrew Mortona737b3e2006-03-22 00:08:11 -0800862static void __slab_error(const char *function, struct kmem_cache *cachep,
863 char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700864{
865 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800866 function, cachep->name, msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700867 dump_stack();
868}
869
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800870#ifdef CONFIG_NUMA
871/*
872 * Special reaping functions for NUMA systems called from cache_reap().
873 * These take care of doing round robin flushing of alien caches (containing
874 * objects freed on different nodes from which they were allocated) and the
875 * flushing of remote pcps by calling drain_node_pages.
876 */
877static DEFINE_PER_CPU(unsigned long, reap_node);
878
879static void init_reap_node(int cpu)
880{
881 int node;
882
883 node = next_node(cpu_to_node(cpu), node_online_map);
884 if (node == MAX_NUMNODES)
Paul Jackson442295c2006-03-22 00:09:11 -0800885 node = first_node(node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800886
887 __get_cpu_var(reap_node) = node;
888}
889
890static void next_reap_node(void)
891{
892 int node = __get_cpu_var(reap_node);
893
894 /*
895 * Also drain per cpu pages on remote zones
896 */
897 if (node != numa_node_id())
898 drain_node_pages(node);
899
900 node = next_node(node, node_online_map);
901 if (unlikely(node >= MAX_NUMNODES))
902 node = first_node(node_online_map);
903 __get_cpu_var(reap_node) = node;
904}
905
906#else
907#define init_reap_node(cpu) do { } while (0)
908#define next_reap_node(void) do { } while (0)
909#endif
910
Linus Torvalds1da177e2005-04-16 15:20:36 -0700911/*
912 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
913 * via the workqueue/eventd.
914 * Add the CPU number into the expiration time to minimize the possibility of
915 * the CPUs getting into lockstep and contending for the global cache chain
916 * lock.
917 */
918static void __devinit start_cpu_timer(int cpu)
919{
920 struct work_struct *reap_work = &per_cpu(reap_work, cpu);
921
922 /*
923 * When this gets called from do_initcalls via cpucache_init(),
924 * init_workqueues() has already run, so keventd will be setup
925 * at that time.
926 */
927 if (keventd_up() && reap_work->func == NULL) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800928 init_reap_node(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700929 INIT_WORK(reap_work, cache_reap, NULL);
930 schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu);
931 }
932}
933
Christoph Lametere498be72005-09-09 13:03:32 -0700934static struct array_cache *alloc_arraycache(int node, int entries,
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800935 int batchcount)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700936{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800937 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700938 struct array_cache *nc = NULL;
939
Christoph Lametere498be72005-09-09 13:03:32 -0700940 nc = kmalloc_node(memsize, GFP_KERNEL, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700941 if (nc) {
942 nc->avail = 0;
943 nc->limit = entries;
944 nc->batchcount = batchcount;
945 nc->touched = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700946 spin_lock_init(&nc->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700947 }
948 return nc;
949}
950
Christoph Lameter3ded1752006-03-25 03:06:44 -0800951/*
952 * Transfer objects in one arraycache to another.
953 * Locking must be handled by the caller.
954 *
955 * Return the number of entries transferred.
956 */
957static int transfer_objects(struct array_cache *to,
958 struct array_cache *from, unsigned int max)
959{
960 /* Figure out how many entries to transfer */
961 int nr = min(min(from->avail, max), to->limit - to->avail);
962
963 if (!nr)
964 return 0;
965
966 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
967 sizeof(void *) *nr);
968
969 from->avail -= nr;
970 to->avail += nr;
971 to->touched = 1;
972 return nr;
973}
974
Christoph Lameter765c4502006-09-27 01:50:08 -0700975#ifndef CONFIG_NUMA
976
977#define drain_alien_cache(cachep, alien) do { } while (0)
978#define reap_alien(cachep, l3) do { } while (0)
979
980static inline struct array_cache **alloc_alien_cache(int node, int limit)
981{
982 return (struct array_cache **)BAD_ALIEN_MAGIC;
983}
984
985static inline void free_alien_cache(struct array_cache **ac_ptr)
986{
987}
988
989static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
990{
991 return 0;
992}
993
994static inline void *alternate_node_alloc(struct kmem_cache *cachep,
995 gfp_t flags)
996{
997 return NULL;
998}
999
1000static inline void *__cache_alloc_node(struct kmem_cache *cachep,
1001 gfp_t flags, int nodeid)
1002{
1003 return NULL;
1004}
1005
1006#else /* CONFIG_NUMA */
1007
Pekka Enberg343e0d72006-02-01 03:05:50 -08001008static void *__cache_alloc_node(struct kmem_cache *, gfp_t, int);
Paul Jacksonc61afb12006-03-24 03:16:08 -08001009static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
Christoph Lameterdc85da12006-01-18 17:42:36 -08001010
Pekka Enberg5295a742006-02-01 03:05:48 -08001011static struct array_cache **alloc_alien_cache(int node, int limit)
Christoph Lametere498be72005-09-09 13:03:32 -07001012{
1013 struct array_cache **ac_ptr;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001014 int memsize = sizeof(void *) * MAX_NUMNODES;
Christoph Lametere498be72005-09-09 13:03:32 -07001015 int i;
1016
1017 if (limit > 1)
1018 limit = 12;
1019 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
1020 if (ac_ptr) {
1021 for_each_node(i) {
1022 if (i == node || !node_online(i)) {
1023 ac_ptr[i] = NULL;
1024 continue;
1025 }
1026 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
1027 if (!ac_ptr[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001028 for (i--; i <= 0; i--)
Christoph Lametere498be72005-09-09 13:03:32 -07001029 kfree(ac_ptr[i]);
1030 kfree(ac_ptr);
1031 return NULL;
1032 }
1033 }
1034 }
1035 return ac_ptr;
1036}
1037
Pekka Enberg5295a742006-02-01 03:05:48 -08001038static void free_alien_cache(struct array_cache **ac_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -07001039{
1040 int i;
1041
1042 if (!ac_ptr)
1043 return;
Christoph Lametere498be72005-09-09 13:03:32 -07001044 for_each_node(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001045 kfree(ac_ptr[i]);
Christoph Lametere498be72005-09-09 13:03:32 -07001046 kfree(ac_ptr);
1047}
1048
Pekka Enberg343e0d72006-02-01 03:05:50 -08001049static void __drain_alien_cache(struct kmem_cache *cachep,
Pekka Enberg5295a742006-02-01 03:05:48 -08001050 struct array_cache *ac, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07001051{
1052 struct kmem_list3 *rl3 = cachep->nodelists[node];
1053
1054 if (ac->avail) {
1055 spin_lock(&rl3->list_lock);
Christoph Lametere00946f2006-03-25 03:06:45 -08001056 /*
1057 * Stuff objects into the remote nodes shared array first.
1058 * That way we could avoid the overhead of putting the objects
1059 * into the free lists and getting them back later.
1060 */
shin, jacob693f7d32006-04-28 10:54:37 -05001061 if (rl3->shared)
1062 transfer_objects(rl3->shared, ac, ac->limit);
Christoph Lametere00946f2006-03-25 03:06:45 -08001063
Christoph Lameterff694162005-09-22 21:44:02 -07001064 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001065 ac->avail = 0;
1066 spin_unlock(&rl3->list_lock);
1067 }
1068}
1069
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001070/*
1071 * Called from cache_reap() to regularly drain alien caches round robin.
1072 */
1073static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1074{
1075 int node = __get_cpu_var(reap_node);
1076
1077 if (l3->alien) {
1078 struct array_cache *ac = l3->alien[node];
Christoph Lametere00946f2006-03-25 03:06:45 -08001079
1080 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001081 __drain_alien_cache(cachep, ac, node);
1082 spin_unlock_irq(&ac->lock);
1083 }
1084 }
1085}
1086
Andrew Mortona737b3e2006-03-22 00:08:11 -08001087static void drain_alien_cache(struct kmem_cache *cachep,
1088 struct array_cache **alien)
Christoph Lametere498be72005-09-09 13:03:32 -07001089{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001090 int i = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07001091 struct array_cache *ac;
1092 unsigned long flags;
1093
1094 for_each_online_node(i) {
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001095 ac = alien[i];
Christoph Lametere498be72005-09-09 13:03:32 -07001096 if (ac) {
1097 spin_lock_irqsave(&ac->lock, flags);
1098 __drain_alien_cache(cachep, ac, i);
1099 spin_unlock_irqrestore(&ac->lock, flags);
1100 }
1101 }
1102}
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001103
Ingo Molnar873623d2006-07-13 14:44:38 +02001104static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001105{
1106 struct slab *slabp = virt_to_slab(objp);
1107 int nodeid = slabp->nodeid;
1108 struct kmem_list3 *l3;
1109 struct array_cache *alien = NULL;
1110
1111 /*
1112 * Make sure we are not freeing a object from another node to the array
1113 * cache on this cpu.
1114 */
1115 if (likely(slabp->nodeid == numa_node_id()))
1116 return 0;
1117
1118 l3 = cachep->nodelists[numa_node_id()];
1119 STATS_INC_NODEFREES(cachep);
1120 if (l3->alien && l3->alien[nodeid]) {
1121 alien = l3->alien[nodeid];
Ingo Molnar873623d2006-07-13 14:44:38 +02001122 spin_lock(&alien->lock);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001123 if (unlikely(alien->avail == alien->limit)) {
1124 STATS_INC_ACOVERFLOW(cachep);
1125 __drain_alien_cache(cachep, alien, nodeid);
1126 }
1127 alien->entry[alien->avail++] = objp;
1128 spin_unlock(&alien->lock);
1129 } else {
1130 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1131 free_block(cachep, &objp, 1, nodeid);
1132 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1133 }
1134 return 1;
1135}
Christoph Lametere498be72005-09-09 13:03:32 -07001136#endif
1137
Chandra Seetharaman8c78f302006-07-30 03:03:35 -07001138static int __cpuinit cpuup_callback(struct notifier_block *nfb,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001139 unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001140{
1141 long cpu = (long)hcpu;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001142 struct kmem_cache *cachep;
Christoph Lametere498be72005-09-09 13:03:32 -07001143 struct kmem_list3 *l3 = NULL;
1144 int node = cpu_to_node(cpu);
1145 int memsize = sizeof(struct kmem_list3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001146
1147 switch (action) {
1148 case CPU_UP_PREPARE:
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001149 mutex_lock(&cache_chain_mutex);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001150 /*
1151 * We need to do this right in the beginning since
Christoph Lametere498be72005-09-09 13:03:32 -07001152 * alloc_arraycache's are going to use this list.
1153 * kmalloc_node allows us to add the slab to the right
1154 * kmem_list3 and not this cpu's kmem_list3
1155 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001156
Christoph Lametere498be72005-09-09 13:03:32 -07001157 list_for_each_entry(cachep, &cache_chain, next) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08001158 /*
1159 * Set up the size64 kmemlist for cpu before we can
Christoph Lametere498be72005-09-09 13:03:32 -07001160 * begin anything. Make sure some other cpu on this
1161 * node has not already allocated this
1162 */
1163 if (!cachep->nodelists[node]) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08001164 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1165 if (!l3)
Christoph Lametere498be72005-09-09 13:03:32 -07001166 goto bad;
1167 kmem_list3_init(l3);
1168 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001169 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -07001170
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001171 /*
1172 * The l3s don't come and go as CPUs come and
1173 * go. cache_chain_mutex is sufficient
1174 * protection here.
1175 */
Christoph Lametere498be72005-09-09 13:03:32 -07001176 cachep->nodelists[node] = l3;
1177 }
1178
1179 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1180 cachep->nodelists[node]->free_limit =
Andrew Mortona737b3e2006-03-22 00:08:11 -08001181 (1 + nr_cpus_node(node)) *
1182 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07001183 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1184 }
1185
Andrew Mortona737b3e2006-03-22 00:08:11 -08001186 /*
1187 * Now we can go ahead with allocating the shared arrays and
1188 * array caches
1189 */
Christoph Lametere498be72005-09-09 13:03:32 -07001190 list_for_each_entry(cachep, &cache_chain, next) {
Tobias Klausercd105df2006-01-08 01:00:59 -08001191 struct array_cache *nc;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001192 struct array_cache *shared;
1193 struct array_cache **alien;
Tobias Klausercd105df2006-01-08 01:00:59 -08001194
Christoph Lametere498be72005-09-09 13:03:32 -07001195 nc = alloc_arraycache(node, cachep->limit,
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001196 cachep->batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001197 if (!nc)
1198 goto bad;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001199 shared = alloc_arraycache(node,
1200 cachep->shared * cachep->batchcount,
1201 0xbaadf00d);
1202 if (!shared)
1203 goto bad;
Linus Torvalds7a21ef62006-02-05 11:26:38 -08001204
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001205 alien = alloc_alien_cache(node, cachep->limit);
1206 if (!alien)
1207 goto bad;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001208 cachep->array[cpu] = nc;
Christoph Lametere498be72005-09-09 13:03:32 -07001209 l3 = cachep->nodelists[node];
1210 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07001211
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001212 spin_lock_irq(&l3->list_lock);
1213 if (!l3->shared) {
1214 /*
1215 * We are serialised from CPU_DEAD or
1216 * CPU_UP_CANCELLED by the cpucontrol lock
1217 */
1218 l3->shared = shared;
1219 shared = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07001220 }
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001221#ifdef CONFIG_NUMA
1222 if (!l3->alien) {
1223 l3->alien = alien;
1224 alien = NULL;
1225 }
1226#endif
1227 spin_unlock_irq(&l3->list_lock);
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001228 kfree(shared);
1229 free_alien_cache(alien);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001230 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001231 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001232 break;
1233 case CPU_ONLINE:
1234 start_cpu_timer(cpu);
1235 break;
1236#ifdef CONFIG_HOTPLUG_CPU
1237 case CPU_DEAD:
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001238 /*
1239 * Even if all the cpus of a node are down, we don't free the
1240 * kmem_list3 of any cache. This to avoid a race between
1241 * cpu_down, and a kmalloc allocation from another cpu for
1242 * memory from the node of the cpu going down. The list3
1243 * structure is usually allocated from kmem_cache_create() and
1244 * gets destroyed at kmem_cache_destroy().
1245 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001246 /* fall thru */
1247 case CPU_UP_CANCELED:
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001248 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001249 list_for_each_entry(cachep, &cache_chain, next) {
1250 struct array_cache *nc;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001251 struct array_cache *shared;
1252 struct array_cache **alien;
Christoph Lametere498be72005-09-09 13:03:32 -07001253 cpumask_t mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001254
Christoph Lametere498be72005-09-09 13:03:32 -07001255 mask = node_to_cpumask(node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001256 /* cpu is dead; no one can alloc from it. */
1257 nc = cachep->array[cpu];
1258 cachep->array[cpu] = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07001259 l3 = cachep->nodelists[node];
1260
1261 if (!l3)
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001262 goto free_array_cache;
Christoph Lametere498be72005-09-09 13:03:32 -07001263
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08001264 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07001265
1266 /* Free limit for this kmem_list3 */
1267 l3->free_limit -= cachep->batchcount;
1268 if (nc)
Christoph Lameterff694162005-09-22 21:44:02 -07001269 free_block(cachep, nc->entry, nc->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001270
1271 if (!cpus_empty(mask)) {
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08001272 spin_unlock_irq(&l3->list_lock);
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001273 goto free_array_cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001274 }
Christoph Lametere498be72005-09-09 13:03:32 -07001275
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001276 shared = l3->shared;
1277 if (shared) {
Christoph Lametere498be72005-09-09 13:03:32 -07001278 free_block(cachep, l3->shared->entry,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001279 l3->shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001280 l3->shared = NULL;
1281 }
Christoph Lametere498be72005-09-09 13:03:32 -07001282
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001283 alien = l3->alien;
1284 l3->alien = NULL;
1285
1286 spin_unlock_irq(&l3->list_lock);
1287
1288 kfree(shared);
1289 if (alien) {
1290 drain_alien_cache(cachep, alien);
1291 free_alien_cache(alien);
Christoph Lametere498be72005-09-09 13:03:32 -07001292 }
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001293free_array_cache:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001294 kfree(nc);
1295 }
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001296 /*
1297 * In the previous loop, all the objects were freed to
1298 * the respective cache's slabs, now we can go ahead and
1299 * shrink each nodelist to its limit.
1300 */
1301 list_for_each_entry(cachep, &cache_chain, next) {
1302 l3 = cachep->nodelists[node];
1303 if (!l3)
1304 continue;
Christoph Lametered11d9e2006-06-30 01:55:45 -07001305 drain_freelist(cachep, l3, l3->free_objects);
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001306 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001307 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001308 break;
1309#endif
1310 }
1311 return NOTIFY_OK;
Andrew Mortona737b3e2006-03-22 00:08:11 -08001312bad:
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001313 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001314 return NOTIFY_BAD;
1315}
1316
Chandra Seetharaman74b85f32006-06-27 02:54:09 -07001317static struct notifier_block __cpuinitdata cpucache_notifier = {
1318 &cpuup_callback, NULL, 0
1319};
Linus Torvalds1da177e2005-04-16 15:20:36 -07001320
Christoph Lametere498be72005-09-09 13:03:32 -07001321/*
1322 * swap the static kmem_list3 with kmalloced memory
1323 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001324static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1325 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001326{
1327 struct kmem_list3 *ptr;
1328
1329 BUG_ON(cachep->nodelists[nodeid] != list);
1330 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1331 BUG_ON(!ptr);
1332
1333 local_irq_disable();
1334 memcpy(ptr, list, sizeof(struct kmem_list3));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001335 /*
1336 * Do not assume that spinlocks can be initialized via memcpy:
1337 */
1338 spin_lock_init(&ptr->list_lock);
1339
Christoph Lametere498be72005-09-09 13:03:32 -07001340 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1341 cachep->nodelists[nodeid] = ptr;
1342 local_irq_enable();
1343}
1344
Andrew Mortona737b3e2006-03-22 00:08:11 -08001345/*
1346 * Initialisation. Called after the page allocator have been initialised and
1347 * before smp_init().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001348 */
1349void __init kmem_cache_init(void)
1350{
1351 size_t left_over;
1352 struct cache_sizes *sizes;
1353 struct cache_names *names;
Christoph Lametere498be72005-09-09 13:03:32 -07001354 int i;
Jack Steiner07ed76b2006-03-07 21:55:46 -08001355 int order;
Christoph Lametere498be72005-09-09 13:03:32 -07001356
1357 for (i = 0; i < NUM_INIT_LISTS; i++) {
1358 kmem_list3_init(&initkmem_list3[i]);
1359 if (i < MAX_NUMNODES)
1360 cache_cache.nodelists[i] = NULL;
1361 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001362
1363 /*
1364 * Fragmentation resistance on low memory - only use bigger
1365 * page orders on machines with more than 32MB of memory.
1366 */
1367 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1368 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1369
Linus Torvalds1da177e2005-04-16 15:20:36 -07001370 /* Bootstrap is tricky, because several objects are allocated
1371 * from caches that do not exist yet:
Andrew Mortona737b3e2006-03-22 00:08:11 -08001372 * 1) initialize the cache_cache cache: it contains the struct
1373 * kmem_cache structures of all caches, except cache_cache itself:
1374 * cache_cache is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001375 * Initially an __init data area is used for the head array and the
1376 * kmem_list3 structures, it's replaced with a kmalloc allocated
1377 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001378 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001379 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001380 * An __init data area is used for the head array.
1381 * 3) Create the remaining kmalloc caches, with minimally sized
1382 * head arrays.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001383 * 4) Replace the __init data head arrays for cache_cache and the first
1384 * kmalloc cache with kmalloc allocated arrays.
Christoph Lametere498be72005-09-09 13:03:32 -07001385 * 5) Replace the __init data for kmem_list3 for cache_cache and
1386 * the other cache's with kmalloc allocated memory.
1387 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001388 */
1389
1390 /* 1) create the cache_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001391 INIT_LIST_HEAD(&cache_chain);
1392 list_add(&cache_cache.next, &cache_chain);
1393 cache_cache.colour_off = cache_line_size();
1394 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
Christoph Lametere498be72005-09-09 13:03:32 -07001395 cache_cache.nodelists[numa_node_id()] = &initkmem_list3[CACHE_CACHE];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001396
Andrew Mortona737b3e2006-03-22 00:08:11 -08001397 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1398 cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001399
Jack Steiner07ed76b2006-03-07 21:55:46 -08001400 for (order = 0; order < MAX_ORDER; order++) {
1401 cache_estimate(order, cache_cache.buffer_size,
1402 cache_line_size(), 0, &left_over, &cache_cache.num);
1403 if (cache_cache.num)
1404 break;
1405 }
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02001406 BUG_ON(!cache_cache.num);
Jack Steiner07ed76b2006-03-07 21:55:46 -08001407 cache_cache.gfporder = order;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001408 cache_cache.colour = left_over / cache_cache.colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001409 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1410 sizeof(struct slab), cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001411
1412 /* 2+3) create the kmalloc caches */
1413 sizes = malloc_sizes;
1414 names = cache_names;
1415
Andrew Mortona737b3e2006-03-22 00:08:11 -08001416 /*
1417 * Initialize the caches that provide memory for the array cache and the
1418 * kmem_list3 structures first. Without this, further allocations will
1419 * bug.
Christoph Lametere498be72005-09-09 13:03:32 -07001420 */
1421
1422 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001423 sizes[INDEX_AC].cs_size,
1424 ARCH_KMALLOC_MINALIGN,
1425 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1426 NULL, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07001427
Andrew Mortona737b3e2006-03-22 00:08:11 -08001428 if (INDEX_AC != INDEX_L3) {
Christoph Lametere498be72005-09-09 13:03:32 -07001429 sizes[INDEX_L3].cs_cachep =
Andrew Mortona737b3e2006-03-22 00:08:11 -08001430 kmem_cache_create(names[INDEX_L3].name,
1431 sizes[INDEX_L3].cs_size,
1432 ARCH_KMALLOC_MINALIGN,
1433 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1434 NULL, NULL);
1435 }
Christoph Lametere498be72005-09-09 13:03:32 -07001436
Ingo Molnare0a42722006-06-23 02:03:46 -07001437 slab_early_init = 0;
1438
Linus Torvalds1da177e2005-04-16 15:20:36 -07001439 while (sizes->cs_size != ULONG_MAX) {
Christoph Lametere498be72005-09-09 13:03:32 -07001440 /*
1441 * For performance, all the general caches are L1 aligned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001442 * This should be particularly beneficial on SMP boxes, as it
1443 * eliminates "false sharing".
1444 * Note for systems short on memory removing the alignment will
Christoph Lametere498be72005-09-09 13:03:32 -07001445 * allow tighter packing of the smaller caches.
1446 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001447 if (!sizes->cs_cachep) {
Christoph Lametere498be72005-09-09 13:03:32 -07001448 sizes->cs_cachep = kmem_cache_create(names->name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001449 sizes->cs_size,
1450 ARCH_KMALLOC_MINALIGN,
1451 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1452 NULL, NULL);
1453 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001454
Linus Torvalds1da177e2005-04-16 15:20:36 -07001455 sizes->cs_dmacachep = kmem_cache_create(names->name_dma,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001456 sizes->cs_size,
1457 ARCH_KMALLOC_MINALIGN,
1458 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1459 SLAB_PANIC,
1460 NULL, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001461 sizes++;
1462 names++;
1463 }
1464 /* 4) Replace the bootstrap head arrays */
1465 {
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001466 struct array_cache *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001467
Linus Torvalds1da177e2005-04-16 15:20:36 -07001468 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001469
Linus Torvalds1da177e2005-04-16 15:20:36 -07001470 local_irq_disable();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001471 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1472 memcpy(ptr, cpu_cache_get(&cache_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001473 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001474 /*
1475 * Do not assume that spinlocks can be initialized via memcpy:
1476 */
1477 spin_lock_init(&ptr->lock);
1478
Linus Torvalds1da177e2005-04-16 15:20:36 -07001479 cache_cache.array[smp_processor_id()] = ptr;
1480 local_irq_enable();
Christoph Lametere498be72005-09-09 13:03:32 -07001481
Linus Torvalds1da177e2005-04-16 15:20:36 -07001482 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001483
Linus Torvalds1da177e2005-04-16 15:20:36 -07001484 local_irq_disable();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001485 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001486 != &initarray_generic.cache);
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001487 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001488 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001489 /*
1490 * Do not assume that spinlocks can be initialized via memcpy:
1491 */
1492 spin_lock_init(&ptr->lock);
1493
Christoph Lametere498be72005-09-09 13:03:32 -07001494 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001495 ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001496 local_irq_enable();
1497 }
Christoph Lametere498be72005-09-09 13:03:32 -07001498 /* 5) Replace the bootstrap kmem_list3's */
1499 {
1500 int node;
1501 /* Replace the static kmem_list3 structures for the boot cpu */
1502 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE],
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001503 numa_node_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001504
Christoph Lametere498be72005-09-09 13:03:32 -07001505 for_each_online_node(node) {
1506 init_list(malloc_sizes[INDEX_AC].cs_cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001507 &initkmem_list3[SIZE_AC + node], node);
Christoph Lametere498be72005-09-09 13:03:32 -07001508
1509 if (INDEX_AC != INDEX_L3) {
1510 init_list(malloc_sizes[INDEX_L3].cs_cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001511 &initkmem_list3[SIZE_L3 + node],
1512 node);
Christoph Lametere498be72005-09-09 13:03:32 -07001513 }
1514 }
1515 }
1516
1517 /* 6) resize the head arrays to their final sizes */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001518 {
Pekka Enberg343e0d72006-02-01 03:05:50 -08001519 struct kmem_cache *cachep;
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001520 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001521 list_for_each_entry(cachep, &cache_chain, next)
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07001522 if (enable_cpucache(cachep))
1523 BUG();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001524 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001525 }
1526
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001527 /* Annotate slab for lockdep -- annotate the malloc caches */
1528 init_lock_keys();
1529
1530
Linus Torvalds1da177e2005-04-16 15:20:36 -07001531 /* Done! */
1532 g_cpucache_up = FULL;
1533
Andrew Mortona737b3e2006-03-22 00:08:11 -08001534 /*
1535 * Register a cpu startup notifier callback that initializes
1536 * cpu_cache_get for all new cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07001537 */
1538 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001539
Andrew Mortona737b3e2006-03-22 00:08:11 -08001540 /*
1541 * The reap timers are started later, with a module init call: That part
1542 * of the kernel is not yet operational.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001543 */
1544}
1545
1546static int __init cpucache_init(void)
1547{
1548 int cpu;
1549
Andrew Mortona737b3e2006-03-22 00:08:11 -08001550 /*
1551 * Register the timers that return unneeded pages to the page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -07001552 */
Christoph Lametere498be72005-09-09 13:03:32 -07001553 for_each_online_cpu(cpu)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001554 start_cpu_timer(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001555 return 0;
1556}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001557__initcall(cpucache_init);
1558
1559/*
1560 * Interface to system's page allocator. No need to hold the cache-lock.
1561 *
1562 * If we requested dmaable memory, we will get it. Even if we
1563 * did not request dmaable memory, we might get it, but that
1564 * would be relatively rare and ignorable.
1565 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001566static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001567{
1568 struct page *page;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001569 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001570 int i;
1571
Luke Yangd6fef9d2006-04-10 22:52:56 -07001572#ifndef CONFIG_MMU
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001573 /*
1574 * Nommu uses slab's for process anonymous memory allocations, and thus
1575 * requires __GFP_COMP to properly refcount higher order allocations
Luke Yangd6fef9d2006-04-10 22:52:56 -07001576 */
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001577 flags |= __GFP_COMP;
Luke Yangd6fef9d2006-04-10 22:52:56 -07001578#endif
Christoph Lameter765c4502006-09-27 01:50:08 -07001579
1580 /*
1581 * Under NUMA we want memory on the indicated node. We will handle
1582 * the needed fallback ourselves since we want to serve from our
1583 * per node object lists first for other nodes.
1584 */
1585 flags |= cachep->gfpflags | GFP_THISNODE;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001586
1587 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001588 if (!page)
1589 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001590
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001591 nr_pages = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001592 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
Christoph Lameter972d1a72006-09-25 23:31:51 -07001593 add_zone_page_state(page_zone(page),
1594 NR_SLAB_RECLAIMABLE, nr_pages);
1595 else
1596 add_zone_page_state(page_zone(page),
1597 NR_SLAB_UNRECLAIMABLE, nr_pages);
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001598 for (i = 0; i < nr_pages; i++)
1599 __SetPageSlab(page + i);
1600 return page_address(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001601}
1602
1603/*
1604 * Interface to system's page release.
1605 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001606static void kmem_freepages(struct kmem_cache *cachep, void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001607{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001608 unsigned long i = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001609 struct page *page = virt_to_page(addr);
1610 const unsigned long nr_freed = i;
1611
Christoph Lameter972d1a72006-09-25 23:31:51 -07001612 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1613 sub_zone_page_state(page_zone(page),
1614 NR_SLAB_RECLAIMABLE, nr_freed);
1615 else
1616 sub_zone_page_state(page_zone(page),
1617 NR_SLAB_UNRECLAIMABLE, nr_freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001618 while (i--) {
Nick Pigginf205b2f2006-03-22 00:08:02 -08001619 BUG_ON(!PageSlab(page));
1620 __ClearPageSlab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001621 page++;
1622 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001623 if (current->reclaim_state)
1624 current->reclaim_state->reclaimed_slab += nr_freed;
1625 free_pages((unsigned long)addr, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001626}
1627
1628static void kmem_rcu_free(struct rcu_head *head)
1629{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001630 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001631 struct kmem_cache *cachep = slab_rcu->cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001632
1633 kmem_freepages(cachep, slab_rcu->addr);
1634 if (OFF_SLAB(cachep))
1635 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1636}
1637
1638#if DEBUG
1639
1640#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001641static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001642 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001643{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001644 int size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001645
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001646 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001647
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001648 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001649 return;
1650
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001651 *addr++ = 0x12345678;
1652 *addr++ = caller;
1653 *addr++ = smp_processor_id();
1654 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001655 {
1656 unsigned long *sptr = &caller;
1657 unsigned long svalue;
1658
1659 while (!kstack_end(sptr)) {
1660 svalue = *sptr++;
1661 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001662 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001663 size -= sizeof(unsigned long);
1664 if (size <= sizeof(unsigned long))
1665 break;
1666 }
1667 }
1668
1669 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001670 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001671}
1672#endif
1673
Pekka Enberg343e0d72006-02-01 03:05:50 -08001674static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001675{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001676 int size = obj_size(cachep);
1677 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001678
1679 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001680 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001681}
1682
1683static void dump_line(char *data, int offset, int limit)
1684{
1685 int i;
1686 printk(KERN_ERR "%03x:", offset);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001687 for (i = 0; i < limit; i++)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001688 printk(" %02x", (unsigned char)data[offset + i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001689 printk("\n");
1690}
1691#endif
1692
1693#if DEBUG
1694
Pekka Enberg343e0d72006-02-01 03:05:50 -08001695static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001696{
1697 int i, size;
1698 char *realobj;
1699
1700 if (cachep->flags & SLAB_RED_ZONE) {
1701 printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001702 *dbg_redzone1(cachep, objp),
1703 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001704 }
1705
1706 if (cachep->flags & SLAB_STORE_USER) {
1707 printk(KERN_ERR "Last user: [<%p>]",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001708 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001709 print_symbol("(%s)",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001710 (unsigned long)*dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001711 printk("\n");
1712 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001713 realobj = (char *)objp + obj_offset(cachep);
1714 size = obj_size(cachep);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001715 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001716 int limit;
1717 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001718 if (i + limit > size)
1719 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001720 dump_line(realobj, i, limit);
1721 }
1722}
1723
Pekka Enberg343e0d72006-02-01 03:05:50 -08001724static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001725{
1726 char *realobj;
1727 int size, i;
1728 int lines = 0;
1729
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001730 realobj = (char *)objp + obj_offset(cachep);
1731 size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001732
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001733 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001734 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001735 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001736 exp = POISON_END;
1737 if (realobj[i] != exp) {
1738 int limit;
1739 /* Mismatch ! */
1740 /* Print header */
1741 if (lines == 0) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001742 printk(KERN_ERR
Andrew Mortona737b3e2006-03-22 00:08:11 -08001743 "Slab corruption: start=%p, len=%d\n",
1744 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001745 print_objinfo(cachep, objp, 0);
1746 }
1747 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001748 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001749 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001750 if (i + limit > size)
1751 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001752 dump_line(realobj, i, limit);
1753 i += 16;
1754 lines++;
1755 /* Limit to 5 lines */
1756 if (lines > 5)
1757 break;
1758 }
1759 }
1760 if (lines != 0) {
1761 /* Print some data about the neighboring objects, if they
1762 * exist:
1763 */
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08001764 struct slab *slabp = virt_to_slab(objp);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001765 unsigned int objnr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001766
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001767 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001768 if (objnr) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001769 objp = index_to_obj(cachep, slabp, objnr - 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001770 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001771 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001772 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001773 print_objinfo(cachep, objp, 2);
1774 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001775 if (objnr + 1 < cachep->num) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001776 objp = index_to_obj(cachep, slabp, objnr + 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001777 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001778 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001779 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001780 print_objinfo(cachep, objp, 2);
1781 }
1782 }
1783}
1784#endif
1785
Linus Torvalds1da177e2005-04-16 15:20:36 -07001786#if DEBUG
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001787/**
Randy Dunlap911851e2006-03-22 00:08:14 -08001788 * slab_destroy_objs - destroy a slab and its objects
1789 * @cachep: cache pointer being destroyed
1790 * @slabp: slab pointer being destroyed
1791 *
1792 * Call the registered destructor for each object in a slab that is being
1793 * destroyed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001794 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001795static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001796{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001797 int i;
1798 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001799 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001800
1801 if (cachep->flags & SLAB_POISON) {
1802#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08001803 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1804 OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001805 kernel_map_pages(virt_to_page(objp),
Andrew Mortona737b3e2006-03-22 00:08:11 -08001806 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001807 else
1808 check_poison_obj(cachep, objp);
1809#else
1810 check_poison_obj(cachep, objp);
1811#endif
1812 }
1813 if (cachep->flags & SLAB_RED_ZONE) {
1814 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1815 slab_error(cachep, "start of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001816 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001817 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1818 slab_error(cachep, "end of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001819 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001820 }
1821 if (cachep->dtor && !(cachep->flags & SLAB_POISON))
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001822 (cachep->dtor) (objp + obj_offset(cachep), cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001823 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001824}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001825#else
Pekka Enberg343e0d72006-02-01 03:05:50 -08001826static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001827{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001828 if (cachep->dtor) {
1829 int i;
1830 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001831 void *objp = index_to_obj(cachep, slabp, i);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001832 (cachep->dtor) (objp, cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001833 }
1834 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001835}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001836#endif
1837
Randy Dunlap911851e2006-03-22 00:08:14 -08001838/**
1839 * slab_destroy - destroy and release all objects in a slab
1840 * @cachep: cache pointer being destroyed
1841 * @slabp: slab pointer being destroyed
1842 *
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001843 * Destroy all the objs in a slab, and release the mem back to the system.
Andrew Mortona737b3e2006-03-22 00:08:11 -08001844 * Before calling the slab must have been unlinked from the cache. The
1845 * cache-lock is not held/needed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001846 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001847static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001848{
1849 void *addr = slabp->s_mem - slabp->colouroff;
1850
1851 slab_destroy_objs(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001852 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1853 struct slab_rcu *slab_rcu;
1854
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001855 slab_rcu = (struct slab_rcu *)slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001856 slab_rcu->cachep = cachep;
1857 slab_rcu->addr = addr;
1858 call_rcu(&slab_rcu->head, kmem_rcu_free);
1859 } else {
1860 kmem_freepages(cachep, addr);
Ingo Molnar873623d2006-07-13 14:44:38 +02001861 if (OFF_SLAB(cachep))
1862 kmem_cache_free(cachep->slabp_cache, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001863 }
1864}
1865
Andrew Mortona737b3e2006-03-22 00:08:11 -08001866/*
1867 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1868 * size of kmem_list3.
1869 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001870static void set_up_list3s(struct kmem_cache *cachep, int index)
Christoph Lametere498be72005-09-09 13:03:32 -07001871{
1872 int node;
1873
1874 for_each_online_node(node) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001875 cachep->nodelists[node] = &initkmem_list3[index + node];
Christoph Lametere498be72005-09-09 13:03:32 -07001876 cachep->nodelists[node]->next_reap = jiffies +
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001877 REAPTIMEOUT_LIST3 +
1878 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -07001879 }
1880}
1881
Christoph Lameter117f6eb2006-09-25 23:31:37 -07001882static void __kmem_cache_destroy(struct kmem_cache *cachep)
1883{
1884 int i;
1885 struct kmem_list3 *l3;
1886
1887 for_each_online_cpu(i)
1888 kfree(cachep->array[i]);
1889
1890 /* NUMA: free the list3 structures */
1891 for_each_online_node(i) {
1892 l3 = cachep->nodelists[i];
1893 if (l3) {
1894 kfree(l3->shared);
1895 free_alien_cache(l3->alien);
1896 kfree(l3);
1897 }
1898 }
1899 kmem_cache_free(&cache_cache, cachep);
1900}
1901
1902
Linus Torvalds1da177e2005-04-16 15:20:36 -07001903/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08001904 * calculate_slab_order - calculate size (page order) of slabs
1905 * @cachep: pointer to the cache that is being created
1906 * @size: size of objects to be created in this cache.
1907 * @align: required alignment for the objects.
1908 * @flags: slab allocation flags
1909 *
1910 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001911 *
1912 * This could be made much more intelligent. For now, try to avoid using
1913 * high order pages for slabs. When the gfp() functions are more friendly
1914 * towards high-order requests, this should be changed.
1915 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001916static size_t calculate_slab_order(struct kmem_cache *cachep,
Randy Dunlapee13d782006-02-01 03:05:53 -08001917 size_t size, size_t align, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001918{
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001919 unsigned long offslab_limit;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001920 size_t left_over = 0;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001921 int gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001922
Andrew Mortona737b3e2006-03-22 00:08:11 -08001923 for (gfporder = 0; gfporder <= MAX_GFP_ORDER; gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001924 unsigned int num;
1925 size_t remainder;
1926
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001927 cache_estimate(gfporder, size, align, flags, &remainder, &num);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001928 if (!num)
1929 continue;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001930
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001931 if (flags & CFLGS_OFF_SLAB) {
1932 /*
1933 * Max number of objs-per-slab for caches which
1934 * use off-slab slabs. Needed to avoid a possible
1935 * looping condition in cache_grow().
1936 */
1937 offslab_limit = size - sizeof(struct slab);
1938 offslab_limit /= sizeof(kmem_bufctl_t);
1939
1940 if (num > offslab_limit)
1941 break;
1942 }
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001943
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001944 /* Found something acceptable - save it away */
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001945 cachep->num = num;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001946 cachep->gfporder = gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001947 left_over = remainder;
1948
1949 /*
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08001950 * A VFS-reclaimable slab tends to have most allocations
1951 * as GFP_NOFS and we really don't want to have to be allocating
1952 * higher-order pages when we are unable to shrink dcache.
1953 */
1954 if (flags & SLAB_RECLAIM_ACCOUNT)
1955 break;
1956
1957 /*
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001958 * Large number of objects is good, but very large slabs are
1959 * currently bad for the gfp()s.
1960 */
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001961 if (gfporder >= slab_break_gfp_order)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001962 break;
1963
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001964 /*
1965 * Acceptable internal fragmentation?
1966 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001967 if (left_over * 8 <= (PAGE_SIZE << gfporder))
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001968 break;
1969 }
1970 return left_over;
1971}
1972
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07001973static int setup_cpu_cache(struct kmem_cache *cachep)
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08001974{
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07001975 if (g_cpucache_up == FULL)
1976 return enable_cpucache(cachep);
1977
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08001978 if (g_cpucache_up == NONE) {
1979 /*
1980 * Note: the first kmem_cache_create must create the cache
1981 * that's used by kmalloc(24), otherwise the creation of
1982 * further caches will BUG().
1983 */
1984 cachep->array[smp_processor_id()] = &initarray_generic.cache;
1985
1986 /*
1987 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
1988 * the first cache, then we need to set up all its list3s,
1989 * otherwise the creation of further caches will BUG().
1990 */
1991 set_up_list3s(cachep, SIZE_AC);
1992 if (INDEX_AC == INDEX_L3)
1993 g_cpucache_up = PARTIAL_L3;
1994 else
1995 g_cpucache_up = PARTIAL_AC;
1996 } else {
1997 cachep->array[smp_processor_id()] =
1998 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1999
2000 if (g_cpucache_up == PARTIAL_AC) {
2001 set_up_list3s(cachep, SIZE_L3);
2002 g_cpucache_up = PARTIAL_L3;
2003 } else {
2004 int node;
2005 for_each_online_node(node) {
2006 cachep->nodelists[node] =
2007 kmalloc_node(sizeof(struct kmem_list3),
2008 GFP_KERNEL, node);
2009 BUG_ON(!cachep->nodelists[node]);
2010 kmem_list3_init(cachep->nodelists[node]);
2011 }
2012 }
2013 }
2014 cachep->nodelists[numa_node_id()]->next_reap =
2015 jiffies + REAPTIMEOUT_LIST3 +
2016 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2017
2018 cpu_cache_get(cachep)->avail = 0;
2019 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2020 cpu_cache_get(cachep)->batchcount = 1;
2021 cpu_cache_get(cachep)->touched = 0;
2022 cachep->batchcount = 1;
2023 cachep->limit = BOOT_CPUCACHE_ENTRIES;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002024 return 0;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002025}
2026
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002027/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07002028 * kmem_cache_create - Create a cache.
2029 * @name: A string which is used in /proc/slabinfo to identify this cache.
2030 * @size: The size of objects to be created in this cache.
2031 * @align: The required alignment for the objects.
2032 * @flags: SLAB flags
2033 * @ctor: A constructor for the objects.
2034 * @dtor: A destructor for the objects.
2035 *
2036 * Returns a ptr to the cache on success, NULL on failure.
2037 * Cannot be called within a int, but can be interrupted.
2038 * The @ctor is run when new pages are allocated by the cache
2039 * and the @dtor is run before the pages are handed back.
2040 *
2041 * @name must be valid until the cache is destroyed. This implies that
Andrew Mortona737b3e2006-03-22 00:08:11 -08002042 * the module calling this has to destroy the cache before getting unloaded.
2043 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002044 * The flags are
2045 *
2046 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2047 * to catch references to uninitialised memory.
2048 *
2049 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2050 * for buffer overruns.
2051 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002052 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2053 * cacheline. This can be beneficial if you're counting cycles as closely
2054 * as davem.
2055 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002056struct kmem_cache *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002057kmem_cache_create (const char *name, size_t size, size_t align,
Andrew Mortona737b3e2006-03-22 00:08:11 -08002058 unsigned long flags,
2059 void (*ctor)(void*, struct kmem_cache *, unsigned long),
Pekka Enberg343e0d72006-02-01 03:05:50 -08002060 void (*dtor)(void*, struct kmem_cache *, unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002061{
2062 size_t left_over, slab_size, ralign;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07002063 struct kmem_cache *cachep = NULL, *pc;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002064
2065 /*
2066 * Sanity checks... these are all serious usage bugs.
2067 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002068 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002069 (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002070 printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__,
2071 name);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002072 BUG();
2073 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002074
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002075 /*
2076 * Prevent CPUs from coming and going.
2077 * lock_cpu_hotplug() nests outside cache_chain_mutex
2078 */
2079 lock_cpu_hotplug();
2080
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002081 mutex_lock(&cache_chain_mutex);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002082
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07002083 list_for_each_entry(pc, &cache_chain, next) {
Andrew Morton4f12bb42005-11-07 00:58:00 -08002084 mm_segment_t old_fs = get_fs();
2085 char tmp;
2086 int res;
2087
2088 /*
2089 * This happens when the module gets unloaded and doesn't
2090 * destroy its slab cache and no-one else reuses the vmalloc
2091 * area of the module. Print a warning.
2092 */
2093 set_fs(KERNEL_DS);
2094 res = __get_user(tmp, pc->name);
2095 set_fs(old_fs);
2096 if (res) {
2097 printk("SLAB: cache with size %d has lost its name\n",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002098 pc->buffer_size);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002099 continue;
2100 }
2101
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002102 if (!strcmp(pc->name, name)) {
Andrew Morton4f12bb42005-11-07 00:58:00 -08002103 printk("kmem_cache_create: duplicate cache %s\n", name);
2104 dump_stack();
2105 goto oops;
2106 }
2107 }
2108
Linus Torvalds1da177e2005-04-16 15:20:36 -07002109#if DEBUG
2110 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
2111 if ((flags & SLAB_DEBUG_INITIAL) && !ctor) {
2112 /* No constructor, but inital state check requested */
2113 printk(KERN_ERR "%s: No con, but init state check "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002114 "requested - %s\n", __FUNCTION__, name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002115 flags &= ~SLAB_DEBUG_INITIAL;
2116 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002117#if FORCED_DEBUG
2118 /*
2119 * Enable redzoning and last user accounting, except for caches with
2120 * large objects, if the increased size would increase the object size
2121 * above the next power of two: caches with object sizes just above a
2122 * power of two have a significant amount of internal fragmentation.
2123 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002124 if (size < 4096 || fls(size - 1) == fls(size-1 + 3 * BYTES_PER_WORD))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002125 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002126 if (!(flags & SLAB_DESTROY_BY_RCU))
2127 flags |= SLAB_POISON;
2128#endif
2129 if (flags & SLAB_DESTROY_BY_RCU)
2130 BUG_ON(flags & SLAB_POISON);
2131#endif
2132 if (flags & SLAB_DESTROY_BY_RCU)
2133 BUG_ON(dtor);
2134
2135 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002136 * Always checks flags, a caller might be expecting debug support which
2137 * isn't available.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002138 */
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002139 BUG_ON(flags & ~CREATE_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002140
Andrew Mortona737b3e2006-03-22 00:08:11 -08002141 /*
2142 * Check that size is in terms of words. This is needed to avoid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002143 * unaligned accesses for some archs when redzoning is used, and makes
2144 * sure any on-slab bufctl's are also correctly aligned.
2145 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002146 if (size & (BYTES_PER_WORD - 1)) {
2147 size += (BYTES_PER_WORD - 1);
2148 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002149 }
2150
Andrew Mortona737b3e2006-03-22 00:08:11 -08002151 /* calculate the final buffer alignment: */
2152
Linus Torvalds1da177e2005-04-16 15:20:36 -07002153 /* 1) arch recommendation: can be overridden for debug */
2154 if (flags & SLAB_HWCACHE_ALIGN) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002155 /*
2156 * Default alignment: as specified by the arch code. Except if
2157 * an object is really small, then squeeze multiple objects into
2158 * one cacheline.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002159 */
2160 ralign = cache_line_size();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002161 while (size <= ralign / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002162 ralign /= 2;
2163 } else {
2164 ralign = BYTES_PER_WORD;
2165 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002166
2167 /*
2168 * Redzoning and user store require word alignment. Note this will be
2169 * overridden by architecture or caller mandated alignment if either
2170 * is greater than BYTES_PER_WORD.
2171 */
2172 if (flags & SLAB_RED_ZONE || flags & SLAB_STORE_USER)
2173 ralign = BYTES_PER_WORD;
2174
Linus Torvalds1da177e2005-04-16 15:20:36 -07002175 /* 2) arch mandated alignment: disables debug if necessary */
2176 if (ralign < ARCH_SLAB_MINALIGN) {
2177 ralign = ARCH_SLAB_MINALIGN;
2178 if (ralign > BYTES_PER_WORD)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002179 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002180 }
2181 /* 3) caller mandated alignment: disables debug if necessary */
2182 if (ralign < align) {
2183 ralign = align;
2184 if (ralign > BYTES_PER_WORD)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002185 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002186 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08002187 /*
Pekka Enbergca5f9702006-09-25 23:31:25 -07002188 * 4) Store it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002189 */
2190 align = ralign;
2191
2192 /* Get cache's description obj. */
Pekka Enbergc5e3b832006-03-25 03:06:43 -08002193 cachep = kmem_cache_zalloc(&cache_cache, SLAB_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002194 if (!cachep)
Andrew Morton4f12bb42005-11-07 00:58:00 -08002195 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002196
2197#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002198 cachep->obj_size = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002199
Pekka Enbergca5f9702006-09-25 23:31:25 -07002200 /*
2201 * Both debugging options require word-alignment which is calculated
2202 * into align above.
2203 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002204 if (flags & SLAB_RED_ZONE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002205 /* add space for red zone words */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002206 cachep->obj_offset += BYTES_PER_WORD;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002207 size += 2 * BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002208 }
2209 if (flags & SLAB_STORE_USER) {
Pekka Enbergca5f9702006-09-25 23:31:25 -07002210 /* user store requires one word storage behind the end of
2211 * the real object.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002212 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002213 size += BYTES_PER_WORD;
2214 }
2215#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002216 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002217 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2218 cachep->obj_offset += PAGE_SIZE - size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002219 size = PAGE_SIZE;
2220 }
2221#endif
2222#endif
2223
Ingo Molnare0a42722006-06-23 02:03:46 -07002224 /*
2225 * Determine if the slab management is 'on' or 'off' slab.
2226 * (bootstrapping cannot cope with offslab caches so don't do
2227 * it too early on.)
2228 */
2229 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002230 /*
2231 * Size is large, assume best to place the slab management obj
2232 * off-slab (should allow better packing of objs).
2233 */
2234 flags |= CFLGS_OFF_SLAB;
2235
2236 size = ALIGN(size, align);
2237
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002238 left_over = calculate_slab_order(cachep, size, align, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002239
2240 if (!cachep->num) {
2241 printk("kmem_cache_create: couldn't create cache %s.\n", name);
2242 kmem_cache_free(&cache_cache, cachep);
2243 cachep = NULL;
Andrew Morton4f12bb42005-11-07 00:58:00 -08002244 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002245 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002246 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2247 + sizeof(struct slab), align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002248
2249 /*
2250 * If the slab has been placed off-slab, and we have enough space then
2251 * move it on-slab. This is at the expense of any extra colouring.
2252 */
2253 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2254 flags &= ~CFLGS_OFF_SLAB;
2255 left_over -= slab_size;
2256 }
2257
2258 if (flags & CFLGS_OFF_SLAB) {
2259 /* really off slab. No need for manual alignment */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002260 slab_size =
2261 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002262 }
2263
2264 cachep->colour_off = cache_line_size();
2265 /* Offset must be a multiple of the alignment. */
2266 if (cachep->colour_off < align)
2267 cachep->colour_off = align;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002268 cachep->colour = left_over / cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002269 cachep->slab_size = slab_size;
2270 cachep->flags = flags;
2271 cachep->gfpflags = 0;
2272 if (flags & SLAB_CACHE_DMA)
2273 cachep->gfpflags |= GFP_DMA;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002274 cachep->buffer_size = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002275
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002276 if (flags & CFLGS_OFF_SLAB) {
Victor Fuscob2d55072005-09-10 00:26:36 -07002277 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002278 /*
2279 * This is a possibility for one of the malloc_sizes caches.
2280 * But since we go off slab only for object size greater than
2281 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2282 * this should not happen at all.
2283 * But leave a BUG_ON for some lucky dude.
2284 */
2285 BUG_ON(!cachep->slabp_cache);
2286 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002287 cachep->ctor = ctor;
2288 cachep->dtor = dtor;
2289 cachep->name = name;
2290
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002291 if (setup_cpu_cache(cachep)) {
2292 __kmem_cache_destroy(cachep);
2293 cachep = NULL;
2294 goto oops;
2295 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002296
Linus Torvalds1da177e2005-04-16 15:20:36 -07002297 /* cache setup completed, link it into the list */
2298 list_add(&cachep->next, &cache_chain);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002299oops:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002300 if (!cachep && (flags & SLAB_PANIC))
2301 panic("kmem_cache_create(): failed to create slab `%s'\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002302 name);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002303 mutex_unlock(&cache_chain_mutex);
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002304 unlock_cpu_hotplug();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002305 return cachep;
2306}
2307EXPORT_SYMBOL(kmem_cache_create);
2308
2309#if DEBUG
2310static void check_irq_off(void)
2311{
2312 BUG_ON(!irqs_disabled());
2313}
2314
2315static void check_irq_on(void)
2316{
2317 BUG_ON(irqs_disabled());
2318}
2319
Pekka Enberg343e0d72006-02-01 03:05:50 -08002320static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002321{
2322#ifdef CONFIG_SMP
2323 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002324 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002325#endif
2326}
Christoph Lametere498be72005-09-09 13:03:32 -07002327
Pekka Enberg343e0d72006-02-01 03:05:50 -08002328static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002329{
2330#ifdef CONFIG_SMP
2331 check_irq_off();
2332 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2333#endif
2334}
2335
Linus Torvalds1da177e2005-04-16 15:20:36 -07002336#else
2337#define check_irq_off() do { } while(0)
2338#define check_irq_on() do { } while(0)
2339#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002340#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002341#endif
2342
Christoph Lameteraab22072006-03-22 00:09:06 -08002343static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2344 struct array_cache *ac,
2345 int force, int node);
2346
Linus Torvalds1da177e2005-04-16 15:20:36 -07002347static void do_drain(void *arg)
2348{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002349 struct kmem_cache *cachep = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002350 struct array_cache *ac;
Christoph Lameterff694162005-09-22 21:44:02 -07002351 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002352
2353 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002354 ac = cpu_cache_get(cachep);
Christoph Lameterff694162005-09-22 21:44:02 -07002355 spin_lock(&cachep->nodelists[node]->list_lock);
2356 free_block(cachep, ac->entry, ac->avail, node);
2357 spin_unlock(&cachep->nodelists[node]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002358 ac->avail = 0;
2359}
2360
Pekka Enberg343e0d72006-02-01 03:05:50 -08002361static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002362{
Christoph Lametere498be72005-09-09 13:03:32 -07002363 struct kmem_list3 *l3;
2364 int node;
2365
Andrew Mortona07fa392006-03-22 00:08:17 -08002366 on_each_cpu(do_drain, cachep, 1, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002367 check_irq_on();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002368 for_each_online_node(node) {
Christoph Lametere498be72005-09-09 13:03:32 -07002369 l3 = cachep->nodelists[node];
Roland Dreiera4523a82006-05-15 11:41:00 -07002370 if (l3 && l3->alien)
2371 drain_alien_cache(cachep, l3->alien);
2372 }
2373
2374 for_each_online_node(node) {
2375 l3 = cachep->nodelists[node];
2376 if (l3)
Christoph Lameteraab22072006-03-22 00:09:06 -08002377 drain_array(cachep, l3, l3->shared, 1, node);
Christoph Lametere498be72005-09-09 13:03:32 -07002378 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002379}
2380
Christoph Lametered11d9e2006-06-30 01:55:45 -07002381/*
2382 * Remove slabs from the list of free slabs.
2383 * Specify the number of slabs to drain in tofree.
2384 *
2385 * Returns the actual number of slabs released.
2386 */
2387static int drain_freelist(struct kmem_cache *cache,
2388 struct kmem_list3 *l3, int tofree)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002389{
Christoph Lametered11d9e2006-06-30 01:55:45 -07002390 struct list_head *p;
2391 int nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002392 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002393
Christoph Lametered11d9e2006-06-30 01:55:45 -07002394 nr_freed = 0;
2395 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002396
Christoph Lametered11d9e2006-06-30 01:55:45 -07002397 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002398 p = l3->slabs_free.prev;
Christoph Lametered11d9e2006-06-30 01:55:45 -07002399 if (p == &l3->slabs_free) {
2400 spin_unlock_irq(&l3->list_lock);
2401 goto out;
2402 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002403
Christoph Lametered11d9e2006-06-30 01:55:45 -07002404 slabp = list_entry(p, struct slab, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002405#if DEBUG
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002406 BUG_ON(slabp->inuse);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002407#endif
2408 list_del(&slabp->list);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002409 /*
2410 * Safe to drop the lock. The slab is no longer linked
2411 * to the cache.
2412 */
2413 l3->free_objects -= cache->num;
Christoph Lametere498be72005-09-09 13:03:32 -07002414 spin_unlock_irq(&l3->list_lock);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002415 slab_destroy(cache, slabp);
2416 nr_freed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002417 }
Christoph Lametered11d9e2006-06-30 01:55:45 -07002418out:
2419 return nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002420}
2421
Pekka Enberg343e0d72006-02-01 03:05:50 -08002422static int __cache_shrink(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07002423{
2424 int ret = 0, i = 0;
2425 struct kmem_list3 *l3;
2426
2427 drain_cpu_caches(cachep);
2428
2429 check_irq_on();
2430 for_each_online_node(i) {
2431 l3 = cachep->nodelists[i];
Christoph Lametered11d9e2006-06-30 01:55:45 -07002432 if (!l3)
2433 continue;
2434
2435 drain_freelist(cachep, l3, l3->free_objects);
2436
2437 ret += !list_empty(&l3->slabs_full) ||
2438 !list_empty(&l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07002439 }
2440 return (ret ? 1 : 0);
2441}
2442
Linus Torvalds1da177e2005-04-16 15:20:36 -07002443/**
2444 * kmem_cache_shrink - Shrink a cache.
2445 * @cachep: The cache to shrink.
2446 *
2447 * Releases as many slabs as possible for a cache.
2448 * To help debugging, a zero exit status indicates all slabs were released.
2449 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002450int kmem_cache_shrink(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002451{
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002452 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002453
2454 return __cache_shrink(cachep);
2455}
2456EXPORT_SYMBOL(kmem_cache_shrink);
2457
2458/**
2459 * kmem_cache_destroy - delete a cache
2460 * @cachep: the cache to destroy
2461 *
Pekka Enberg343e0d72006-02-01 03:05:50 -08002462 * Remove a struct kmem_cache object from the slab cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002463 *
2464 * It is expected this function will be called by a module when it is
2465 * unloaded. This will remove the cache completely, and avoid a duplicate
2466 * cache being allocated each time a module is loaded and unloaded, if the
2467 * module doesn't have persistent in-kernel storage across loads and unloads.
2468 *
2469 * The cache must be empty before calling this function.
2470 *
2471 * The caller must guarantee that noone will allocate memory from the cache
2472 * during the kmem_cache_destroy().
2473 */
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002474void kmem_cache_destroy(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002475{
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002476 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002477
2478 /* Don't let CPUs to come and go */
2479 lock_cpu_hotplug();
2480
2481 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002482 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002483 /*
2484 * the chain is never empty, cache_cache is never destroyed
2485 */
2486 list_del(&cachep->next);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002487 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002488
2489 if (__cache_shrink(cachep)) {
2490 slab_error(cachep, "Can't free all objects");
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002491 mutex_lock(&cache_chain_mutex);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002492 list_add(&cachep->next, &cache_chain);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002493 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002494 unlock_cpu_hotplug();
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002495 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002496 }
2497
2498 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
Paul E. McKenneyfbd568a3e2005-05-01 08:59:04 -07002499 synchronize_rcu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002500
Christoph Lameter117f6eb2006-09-25 23:31:37 -07002501 __kmem_cache_destroy(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002502 unlock_cpu_hotplug();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002503}
2504EXPORT_SYMBOL(kmem_cache_destroy);
2505
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002506/*
2507 * Get the memory for a slab management obj.
2508 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2509 * always come from malloc_sizes caches. The slab descriptor cannot
2510 * come from the same cache which is getting created because,
2511 * when we are searching for an appropriate cache for these
2512 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2513 * If we are creating a malloc_sizes cache here it would not be visible to
2514 * kmem_find_general_cachep till the initialization is complete.
2515 * Hence we cannot have slabp_cache same as the original cache.
2516 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002517static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002518 int colour_off, gfp_t local_flags,
2519 int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002520{
2521 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002522
Linus Torvalds1da177e2005-04-16 15:20:36 -07002523 if (OFF_SLAB(cachep)) {
2524 /* Slab management obj is off-slab. */
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002525 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2526 local_flags, nodeid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002527 if (!slabp)
2528 return NULL;
2529 } else {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002530 slabp = objp + colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002531 colour_off += cachep->slab_size;
2532 }
2533 slabp->inuse = 0;
2534 slabp->colouroff = colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002535 slabp->s_mem = objp + colour_off;
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002536 slabp->nodeid = nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002537 return slabp;
2538}
2539
2540static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2541{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002542 return (kmem_bufctl_t *) (slabp + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002543}
2544
Pekka Enberg343e0d72006-02-01 03:05:50 -08002545static void cache_init_objs(struct kmem_cache *cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002546 struct slab *slabp, unsigned long ctor_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002547{
2548 int i;
2549
2550 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002551 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002552#if DEBUG
2553 /* need to poison the objs? */
2554 if (cachep->flags & SLAB_POISON)
2555 poison_obj(cachep, objp, POISON_FREE);
2556 if (cachep->flags & SLAB_STORE_USER)
2557 *dbg_userword(cachep, objp) = NULL;
2558
2559 if (cachep->flags & SLAB_RED_ZONE) {
2560 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2561 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2562 }
2563 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002564 * Constructors are not allowed to allocate memory from the same
2565 * cache which they are a constructor for. Otherwise, deadlock.
2566 * They must also be threaded.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002567 */
2568 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002569 cachep->ctor(objp + obj_offset(cachep), cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002570 ctor_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002571
2572 if (cachep->flags & SLAB_RED_ZONE) {
2573 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2574 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002575 " end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002576 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2577 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002578 " start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002579 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08002580 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2581 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002582 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002583 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002584#else
2585 if (cachep->ctor)
2586 cachep->ctor(objp, cachep, ctor_flags);
2587#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002588 slab_bufctl(slabp)[i] = i + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002589 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002590 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002591 slabp->free = 0;
2592}
2593
Pekka Enberg343e0d72006-02-01 03:05:50 -08002594static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002595{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002596 if (flags & SLAB_DMA)
2597 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2598 else
2599 BUG_ON(cachep->gfpflags & GFP_DMA);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002600}
2601
Andrew Mortona737b3e2006-03-22 00:08:11 -08002602static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2603 int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002604{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002605 void *objp = index_to_obj(cachep, slabp, slabp->free);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002606 kmem_bufctl_t next;
2607
2608 slabp->inuse++;
2609 next = slab_bufctl(slabp)[slabp->free];
2610#if DEBUG
2611 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2612 WARN_ON(slabp->nodeid != nodeid);
2613#endif
2614 slabp->free = next;
2615
2616 return objp;
2617}
2618
Andrew Mortona737b3e2006-03-22 00:08:11 -08002619static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2620 void *objp, int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002621{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002622 unsigned int objnr = obj_to_index(cachep, slabp, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002623
2624#if DEBUG
2625 /* Verify that the slab belongs to the intended node */
2626 WARN_ON(slabp->nodeid != nodeid);
2627
Al Viro871751e2006-03-25 03:06:39 -08002628 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
Matthew Dobson78d382d2006-02-01 03:05:47 -08002629 printk(KERN_ERR "slab: double free detected in cache "
Andrew Mortona737b3e2006-03-22 00:08:11 -08002630 "'%s', objp %p\n", cachep->name, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002631 BUG();
2632 }
2633#endif
2634 slab_bufctl(slabp)[objnr] = slabp->free;
2635 slabp->free = objnr;
2636 slabp->inuse--;
2637}
2638
Pekka Enberg47768742006-06-23 02:03:07 -07002639/*
2640 * Map pages beginning at addr to the given cache and slab. This is required
2641 * for the slab allocator to be able to lookup the cache and slab of a
2642 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2643 */
2644static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2645 void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002646{
Pekka Enberg47768742006-06-23 02:03:07 -07002647 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002648 struct page *page;
2649
Pekka Enberg47768742006-06-23 02:03:07 -07002650 page = virt_to_page(addr);
Nick Piggin84097512006-03-22 00:08:34 -08002651
Pekka Enberg47768742006-06-23 02:03:07 -07002652 nr_pages = 1;
Nick Piggin84097512006-03-22 00:08:34 -08002653 if (likely(!PageCompound(page)))
Pekka Enberg47768742006-06-23 02:03:07 -07002654 nr_pages <<= cache->gfporder;
2655
Linus Torvalds1da177e2005-04-16 15:20:36 -07002656 do {
Pekka Enberg47768742006-06-23 02:03:07 -07002657 page_set_cache(page, cache);
2658 page_set_slab(page, slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002659 page++;
Pekka Enberg47768742006-06-23 02:03:07 -07002660 } while (--nr_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002661}
2662
2663/*
2664 * Grow (by 1) the number of slabs within a cache. This is called by
2665 * kmem_cache_alloc() when there are no active objs left in a cache.
2666 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002667static int cache_grow(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002668{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002669 struct slab *slabp;
2670 void *objp;
2671 size_t offset;
2672 gfp_t local_flags;
2673 unsigned long ctor_flags;
Christoph Lametere498be72005-09-09 13:03:32 -07002674 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002675
Andrew Mortona737b3e2006-03-22 00:08:11 -08002676 /*
2677 * Be lazy and only check for valid flags here, keeping it out of the
2678 * critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002679 */
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002680 BUG_ON(flags & ~(SLAB_DMA | SLAB_LEVEL_MASK | SLAB_NO_GROW));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002681 if (flags & SLAB_NO_GROW)
2682 return 0;
2683
2684 ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2685 local_flags = (flags & SLAB_LEVEL_MASK);
2686 if (!(local_flags & __GFP_WAIT))
2687 /*
2688 * Not allowed to sleep. Need to tell a constructor about
2689 * this - it might need to know...
2690 */
2691 ctor_flags |= SLAB_CTOR_ATOMIC;
2692
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002693 /* Take the l3 list lock to change the colour_next on this node */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002694 check_irq_off();
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002695 l3 = cachep->nodelists[nodeid];
2696 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002697
2698 /* Get colour for the slab, and cal the next value. */
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002699 offset = l3->colour_next;
2700 l3->colour_next++;
2701 if (l3->colour_next >= cachep->colour)
2702 l3->colour_next = 0;
2703 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002704
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002705 offset *= cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002706
2707 if (local_flags & __GFP_WAIT)
2708 local_irq_enable();
2709
2710 /*
2711 * The test for missing atomic flag is performed here, rather than
2712 * the more obvious place, simply to reduce the critical path length
2713 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2714 * will eventually be caught here (where it matters).
2715 */
2716 kmem_flagcheck(cachep, flags);
2717
Andrew Mortona737b3e2006-03-22 00:08:11 -08002718 /*
2719 * Get mem for the objs. Attempt to allocate a physical page from
2720 * 'nodeid'.
Christoph Lametere498be72005-09-09 13:03:32 -07002721 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002722 objp = kmem_getpages(cachep, flags, nodeid);
2723 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002724 goto failed;
2725
2726 /* Get slab management. */
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002727 slabp = alloc_slabmgmt(cachep, objp, offset, local_flags, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002728 if (!slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002729 goto opps1;
2730
Christoph Lametere498be72005-09-09 13:03:32 -07002731 slabp->nodeid = nodeid;
Pekka Enberg47768742006-06-23 02:03:07 -07002732 slab_map_pages(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002733
2734 cache_init_objs(cachep, slabp, ctor_flags);
2735
2736 if (local_flags & __GFP_WAIT)
2737 local_irq_disable();
2738 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002739 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002740
2741 /* Make slab active. */
Christoph Lametere498be72005-09-09 13:03:32 -07002742 list_add_tail(&slabp->list, &(l3->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002743 STATS_INC_GROWN(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002744 l3->free_objects += cachep->num;
2745 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002746 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002747opps1:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002748 kmem_freepages(cachep, objp);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002749failed:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002750 if (local_flags & __GFP_WAIT)
2751 local_irq_disable();
2752 return 0;
2753}
2754
2755#if DEBUG
2756
2757/*
2758 * Perform extra freeing checks:
2759 * - detect bad pointers.
2760 * - POISON/RED_ZONE checking
2761 * - destructor calls, for caches with POISON+dtor
2762 */
2763static void kfree_debugcheck(const void *objp)
2764{
2765 struct page *page;
2766
2767 if (!virt_addr_valid(objp)) {
2768 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002769 (unsigned long)objp);
2770 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002771 }
2772 page = virt_to_page(objp);
2773 if (!PageSlab(page)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002774 printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n",
2775 (unsigned long)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002776 BUG();
2777 }
2778}
2779
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002780static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2781{
2782 unsigned long redzone1, redzone2;
2783
2784 redzone1 = *dbg_redzone1(cache, obj);
2785 redzone2 = *dbg_redzone2(cache, obj);
2786
2787 /*
2788 * Redzone is ok.
2789 */
2790 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2791 return;
2792
2793 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2794 slab_error(cache, "double free detected");
2795 else
2796 slab_error(cache, "memory outside object was overwritten");
2797
2798 printk(KERN_ERR "%p: redzone 1:0x%lx, redzone 2:0x%lx.\n",
2799 obj, redzone1, redzone2);
2800}
2801
Pekka Enberg343e0d72006-02-01 03:05:50 -08002802static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002803 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002804{
2805 struct page *page;
2806 unsigned int objnr;
2807 struct slab *slabp;
2808
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002809 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002810 kfree_debugcheck(objp);
2811 page = virt_to_page(objp);
2812
Pekka Enberg065d41c2005-11-13 16:06:46 -08002813 slabp = page_get_slab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002814
2815 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002816 verify_redzone_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002817 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2818 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2819 }
2820 if (cachep->flags & SLAB_STORE_USER)
2821 *dbg_userword(cachep, objp) = caller;
2822
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002823 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002824
2825 BUG_ON(objnr >= cachep->num);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002826 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002827
2828 if (cachep->flags & SLAB_DEBUG_INITIAL) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002829 /*
2830 * Need to call the slab's constructor so the caller can
2831 * perform a verify of its state (debugging). Called without
2832 * the cache-lock held.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002833 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002834 cachep->ctor(objp + obj_offset(cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002835 cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002836 }
2837 if (cachep->flags & SLAB_POISON && cachep->dtor) {
2838 /* we want to cache poison the object,
2839 * call the destruction callback
2840 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002841 cachep->dtor(objp + obj_offset(cachep), cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002842 }
Al Viro871751e2006-03-25 03:06:39 -08002843#ifdef CONFIG_DEBUG_SLAB_LEAK
2844 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2845#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002846 if (cachep->flags & SLAB_POISON) {
2847#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08002848 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002849 store_stackinfo(cachep, objp, (unsigned long)caller);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002850 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002851 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002852 } else {
2853 poison_obj(cachep, objp, POISON_FREE);
2854 }
2855#else
2856 poison_obj(cachep, objp, POISON_FREE);
2857#endif
2858 }
2859 return objp;
2860}
2861
Pekka Enberg343e0d72006-02-01 03:05:50 -08002862static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002863{
2864 kmem_bufctl_t i;
2865 int entries = 0;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002866
Linus Torvalds1da177e2005-04-16 15:20:36 -07002867 /* Check slab's freelist to see if this obj is there. */
2868 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2869 entries++;
2870 if (entries > cachep->num || i >= cachep->num)
2871 goto bad;
2872 }
2873 if (entries != cachep->num - slabp->inuse) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002874bad:
2875 printk(KERN_ERR "slab: Internal list corruption detected in "
2876 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2877 cachep->name, cachep->num, slabp, slabp->inuse);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002878 for (i = 0;
Linus Torvalds264132b2006-03-06 12:10:07 -08002879 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002880 i++) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002881 if (i % 16 == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002882 printk("\n%03x:", i);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002883 printk(" %02x", ((unsigned char *)slabp)[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002884 }
2885 printk("\n");
2886 BUG();
2887 }
2888}
2889#else
2890#define kfree_debugcheck(x) do { } while(0)
2891#define cache_free_debugcheck(x,objp,z) (objp)
2892#define check_slabp(x,y) do { } while(0)
2893#endif
2894
Pekka Enberg343e0d72006-02-01 03:05:50 -08002895static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002896{
2897 int batchcount;
2898 struct kmem_list3 *l3;
2899 struct array_cache *ac;
2900
2901 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002902 ac = cpu_cache_get(cachep);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002903retry:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002904 batchcount = ac->batchcount;
2905 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002906 /*
2907 * If there was little recent activity on this cache, then
2908 * perform only a partial refill. Otherwise we could generate
2909 * refill bouncing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002910 */
2911 batchcount = BATCHREFILL_LIMIT;
2912 }
Christoph Lametere498be72005-09-09 13:03:32 -07002913 l3 = cachep->nodelists[numa_node_id()];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002914
Christoph Lametere498be72005-09-09 13:03:32 -07002915 BUG_ON(ac->avail > 0 || !l3);
2916 spin_lock(&l3->list_lock);
2917
Christoph Lameter3ded1752006-03-25 03:06:44 -08002918 /* See if we can refill from the shared array */
2919 if (l3->shared && transfer_objects(ac, l3->shared, batchcount))
2920 goto alloc_done;
2921
Linus Torvalds1da177e2005-04-16 15:20:36 -07002922 while (batchcount > 0) {
2923 struct list_head *entry;
2924 struct slab *slabp;
2925 /* Get slab alloc is to come from. */
2926 entry = l3->slabs_partial.next;
2927 if (entry == &l3->slabs_partial) {
2928 l3->free_touched = 1;
2929 entry = l3->slabs_free.next;
2930 if (entry == &l3->slabs_free)
2931 goto must_grow;
2932 }
2933
2934 slabp = list_entry(entry, struct slab, list);
2935 check_slabp(cachep, slabp);
2936 check_spinlock_acquired(cachep);
2937 while (slabp->inuse < cachep->num && batchcount--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002938 STATS_INC_ALLOCED(cachep);
2939 STATS_INC_ACTIVE(cachep);
2940 STATS_SET_HIGH(cachep);
2941
Matthew Dobson78d382d2006-02-01 03:05:47 -08002942 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
2943 numa_node_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002944 }
2945 check_slabp(cachep, slabp);
2946
2947 /* move slabp to correct slabp list: */
2948 list_del(&slabp->list);
2949 if (slabp->free == BUFCTL_END)
2950 list_add(&slabp->list, &l3->slabs_full);
2951 else
2952 list_add(&slabp->list, &l3->slabs_partial);
2953 }
2954
Andrew Mortona737b3e2006-03-22 00:08:11 -08002955must_grow:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002956 l3->free_objects -= ac->avail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002957alloc_done:
Christoph Lametere498be72005-09-09 13:03:32 -07002958 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002959
2960 if (unlikely(!ac->avail)) {
2961 int x;
Christoph Lametere498be72005-09-09 13:03:32 -07002962 x = cache_grow(cachep, flags, numa_node_id());
2963
Andrew Mortona737b3e2006-03-22 00:08:11 -08002964 /* cache_grow can reenable interrupts, then ac could change. */
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002965 ac = cpu_cache_get(cachep);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002966 if (!x && ac->avail == 0) /* no objects in sight? abort */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002967 return NULL;
2968
Andrew Mortona737b3e2006-03-22 00:08:11 -08002969 if (!ac->avail) /* objects refilled by interrupt? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002970 goto retry;
2971 }
2972 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07002973 return ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002974}
2975
Andrew Mortona737b3e2006-03-22 00:08:11 -08002976static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
2977 gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002978{
2979 might_sleep_if(flags & __GFP_WAIT);
2980#if DEBUG
2981 kmem_flagcheck(cachep, flags);
2982#endif
2983}
2984
2985#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08002986static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
2987 gfp_t flags, void *objp, void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002988{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002989 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002990 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002991 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002992#ifdef CONFIG_DEBUG_PAGEALLOC
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002993 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002994 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002995 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002996 else
2997 check_poison_obj(cachep, objp);
2998#else
2999 check_poison_obj(cachep, objp);
3000#endif
3001 poison_obj(cachep, objp, POISON_INUSE);
3002 }
3003 if (cachep->flags & SLAB_STORE_USER)
3004 *dbg_userword(cachep, objp) = caller;
3005
3006 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003007 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3008 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3009 slab_error(cachep, "double free, or memory outside"
3010 " object was overwritten");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003011 printk(KERN_ERR
Andrew Mortona737b3e2006-03-22 00:08:11 -08003012 "%p: redzone 1:0x%lx, redzone 2:0x%lx\n",
3013 objp, *dbg_redzone1(cachep, objp),
3014 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003015 }
3016 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3017 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3018 }
Al Viro871751e2006-03-25 03:06:39 -08003019#ifdef CONFIG_DEBUG_SLAB_LEAK
3020 {
3021 struct slab *slabp;
3022 unsigned objnr;
3023
3024 slabp = page_get_slab(virt_to_page(objp));
3025 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3026 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3027 }
3028#endif
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003029 objp += obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003030 if (cachep->ctor && cachep->flags & SLAB_POISON) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003031 unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003032
3033 if (!(flags & __GFP_WAIT))
3034 ctor_flags |= SLAB_CTOR_ATOMIC;
3035
3036 cachep->ctor(objp, cachep, ctor_flags);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003037 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003038 return objp;
3039}
3040#else
3041#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3042#endif
3043
Pekka Enberg343e0d72006-02-01 03:05:50 -08003044static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003045{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003046 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003047 struct array_cache *ac;
3048
Alok N Kataria5c382302005-09-27 21:45:46 -07003049 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003050 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003051 if (likely(ac->avail)) {
3052 STATS_INC_ALLOCHIT(cachep);
3053 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07003054 objp = ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003055 } else {
3056 STATS_INC_ALLOCMISS(cachep);
3057 objp = cache_alloc_refill(cachep, flags);
3058 }
Alok N Kataria5c382302005-09-27 21:45:46 -07003059 return objp;
3060}
3061
Andrew Mortona737b3e2006-03-22 00:08:11 -08003062static __always_inline void *__cache_alloc(struct kmem_cache *cachep,
3063 gfp_t flags, void *caller)
Alok N Kataria5c382302005-09-27 21:45:46 -07003064{
3065 unsigned long save_flags;
Christoph Lameterde3083e2006-09-27 01:50:03 -07003066 void *objp = NULL;
Alok N Kataria5c382302005-09-27 21:45:46 -07003067
3068 cache_alloc_debugcheck_before(cachep, flags);
3069
3070 local_irq_save(save_flags);
Christoph Lameterde3083e2006-09-27 01:50:03 -07003071
Christoph Lameter765c4502006-09-27 01:50:08 -07003072 if (unlikely(NUMA_BUILD &&
3073 current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY)))
Christoph Lameterde3083e2006-09-27 01:50:03 -07003074 objp = alternate_node_alloc(cachep, flags);
Christoph Lameterde3083e2006-09-27 01:50:03 -07003075
3076 if (!objp)
3077 objp = ____cache_alloc(cachep, flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003078 /*
3079 * We may just have run out of memory on the local node.
3080 * __cache_alloc_node() knows how to locate memory on other nodes
3081 */
3082 if (NUMA_BUILD && !objp)
3083 objp = __cache_alloc_node(cachep, flags, numa_node_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -07003084 local_irq_restore(save_flags);
Eric Dumazet34342e82005-09-03 15:55:06 -07003085 objp = cache_alloc_debugcheck_after(cachep, flags, objp,
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003086 caller);
Eric Dumazet34342e82005-09-03 15:55:06 -07003087 prefetchw(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003088 return objp;
3089}
3090
Christoph Lametere498be72005-09-09 13:03:32 -07003091#ifdef CONFIG_NUMA
3092/*
Paul Jacksonb2455392006-03-24 03:16:12 -08003093 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
Paul Jacksonc61afb12006-03-24 03:16:08 -08003094 *
3095 * If we are in_interrupt, then process context, including cpusets and
3096 * mempolicy, may not apply and should not be used for allocation policy.
3097 */
3098static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3099{
3100 int nid_alloc, nid_here;
3101
Christoph Lameter765c4502006-09-27 01:50:08 -07003102 if (in_interrupt() || (flags & __GFP_THISNODE))
Paul Jacksonc61afb12006-03-24 03:16:08 -08003103 return NULL;
3104 nid_alloc = nid_here = numa_node_id();
3105 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3106 nid_alloc = cpuset_mem_spread_node();
3107 else if (current->mempolicy)
3108 nid_alloc = slab_node(current->mempolicy);
3109 if (nid_alloc != nid_here)
3110 return __cache_alloc_node(cachep, flags, nid_alloc);
3111 return NULL;
3112}
3113
3114/*
Christoph Lameter765c4502006-09-27 01:50:08 -07003115 * Fallback function if there was no memory available and no objects on a
3116 * certain node and we are allowed to fall back. We mimick the behavior of
3117 * the page allocator. We fall back according to a zonelist determined by
3118 * the policy layer while obeying cpuset constraints.
3119 */
3120void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3121{
3122 struct zonelist *zonelist = &NODE_DATA(slab_node(current->mempolicy))
3123 ->node_zonelists[gfp_zone(flags)];
3124 struct zone **z;
3125 void *obj = NULL;
3126
3127 for (z = zonelist->zones; *z && !obj; z++)
3128 if (zone_idx(*z) <= ZONE_NORMAL &&
3129 cpuset_zone_allowed(*z, flags))
3130 obj = __cache_alloc_node(cache,
3131 flags | __GFP_THISNODE,
3132 zone_to_nid(*z));
3133 return obj;
3134}
3135
3136/*
Christoph Lametere498be72005-09-09 13:03:32 -07003137 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07003138 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08003139static void *__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3140 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07003141{
3142 struct list_head *entry;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003143 struct slab *slabp;
3144 struct kmem_list3 *l3;
3145 void *obj;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003146 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003147
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003148 l3 = cachep->nodelists[nodeid];
3149 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07003150
Andrew Mortona737b3e2006-03-22 00:08:11 -08003151retry:
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003152 check_irq_off();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003153 spin_lock(&l3->list_lock);
3154 entry = l3->slabs_partial.next;
3155 if (entry == &l3->slabs_partial) {
3156 l3->free_touched = 1;
3157 entry = l3->slabs_free.next;
3158 if (entry == &l3->slabs_free)
3159 goto must_grow;
3160 }
Christoph Lametere498be72005-09-09 13:03:32 -07003161
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003162 slabp = list_entry(entry, struct slab, list);
3163 check_spinlock_acquired_node(cachep, nodeid);
3164 check_slabp(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07003165
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003166 STATS_INC_NODEALLOCS(cachep);
3167 STATS_INC_ACTIVE(cachep);
3168 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003169
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003170 BUG_ON(slabp->inuse == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07003171
Matthew Dobson78d382d2006-02-01 03:05:47 -08003172 obj = slab_get_obj(cachep, slabp, nodeid);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003173 check_slabp(cachep, slabp);
3174 l3->free_objects--;
3175 /* move slabp to correct slabp list: */
3176 list_del(&slabp->list);
Christoph Lametere498be72005-09-09 13:03:32 -07003177
Andrew Mortona737b3e2006-03-22 00:08:11 -08003178 if (slabp->free == BUFCTL_END)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003179 list_add(&slabp->list, &l3->slabs_full);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003180 else
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003181 list_add(&slabp->list, &l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07003182
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003183 spin_unlock(&l3->list_lock);
3184 goto done;
Christoph Lametere498be72005-09-09 13:03:32 -07003185
Andrew Mortona737b3e2006-03-22 00:08:11 -08003186must_grow:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003187 spin_unlock(&l3->list_lock);
3188 x = cache_grow(cachep, flags, nodeid);
Christoph Lameter765c4502006-09-27 01:50:08 -07003189 if (x)
3190 goto retry;
Christoph Lametere498be72005-09-09 13:03:32 -07003191
Christoph Lameter765c4502006-09-27 01:50:08 -07003192 if (!(flags & __GFP_THISNODE))
3193 /* Unable to grow the cache. Fall back to other nodes. */
3194 return fallback_alloc(cachep, flags);
Christoph Lametere498be72005-09-09 13:03:32 -07003195
Christoph Lameter765c4502006-09-27 01:50:08 -07003196 return NULL;
3197
Andrew Mortona737b3e2006-03-22 00:08:11 -08003198done:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003199 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07003200}
3201#endif
3202
3203/*
3204 * Caller needs to acquire correct kmem_list's list_lock
3205 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003206static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003207 int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003208{
3209 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07003210 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003211
3212 for (i = 0; i < nr_objects; i++) {
3213 void *objp = objpp[i];
3214 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003215
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08003216 slabp = virt_to_slab(objp);
Christoph Lameterff694162005-09-22 21:44:02 -07003217 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003218 list_del(&slabp->list);
Christoph Lameterff694162005-09-22 21:44:02 -07003219 check_spinlock_acquired_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003220 check_slabp(cachep, slabp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08003221 slab_put_obj(cachep, slabp, objp, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003222 STATS_DEC_ACTIVE(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003223 l3->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003224 check_slabp(cachep, slabp);
3225
3226 /* fixup slab chains */
3227 if (slabp->inuse == 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07003228 if (l3->free_objects > l3->free_limit) {
3229 l3->free_objects -= cachep->num;
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07003230 /* No need to drop any previously held
3231 * lock here, even if we have a off-slab slab
3232 * descriptor it is guaranteed to come from
3233 * a different cache, refer to comments before
3234 * alloc_slabmgmt.
3235 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003236 slab_destroy(cachep, slabp);
3237 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07003238 list_add(&slabp->list, &l3->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003239 }
3240 } else {
3241 /* Unconditionally move a slab to the end of the
3242 * partial list on free - maximum time for the
3243 * other objects to be freed, too.
3244 */
Christoph Lametere498be72005-09-09 13:03:32 -07003245 list_add_tail(&slabp->list, &l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003246 }
3247 }
3248}
3249
Pekka Enberg343e0d72006-02-01 03:05:50 -08003250static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003251{
3252 int batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07003253 struct kmem_list3 *l3;
Christoph Lameterff694162005-09-22 21:44:02 -07003254 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003255
3256 batchcount = ac->batchcount;
3257#if DEBUG
3258 BUG_ON(!batchcount || batchcount > ac->avail);
3259#endif
3260 check_irq_off();
Christoph Lameterff694162005-09-22 21:44:02 -07003261 l3 = cachep->nodelists[node];
Ingo Molnar873623d2006-07-13 14:44:38 +02003262 spin_lock(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003263 if (l3->shared) {
3264 struct array_cache *shared_array = l3->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003265 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003266 if (max) {
3267 if (batchcount > max)
3268 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07003269 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003270 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003271 shared_array->avail += batchcount;
3272 goto free_done;
3273 }
3274 }
3275
Christoph Lameterff694162005-09-22 21:44:02 -07003276 free_block(cachep, ac->entry, batchcount, node);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003277free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003278#if STATS
3279 {
3280 int i = 0;
3281 struct list_head *p;
3282
Christoph Lametere498be72005-09-09 13:03:32 -07003283 p = l3->slabs_free.next;
3284 while (p != &(l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003285 struct slab *slabp;
3286
3287 slabp = list_entry(p, struct slab, list);
3288 BUG_ON(slabp->inuse);
3289
3290 i++;
3291 p = p->next;
3292 }
3293 STATS_SET_FREEABLE(cachep, i);
3294 }
3295#endif
Christoph Lametere498be72005-09-09 13:03:32 -07003296 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003297 ac->avail -= batchcount;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003298 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003299}
3300
3301/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08003302 * Release an obj back to its cache. If the obj has a constructed state, it must
3303 * be in this state _before_ it is released. Called with disabled ints.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003304 */
Ingo Molnar873623d2006-07-13 14:44:38 +02003305static inline void __cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003306{
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003307 struct array_cache *ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003308
3309 check_irq_off();
3310 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3311
Ingo Molnar873623d2006-07-13 14:44:38 +02003312 if (cache_free_alien(cachep, objp))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07003313 return;
Christoph Lametere498be72005-09-09 13:03:32 -07003314
Linus Torvalds1da177e2005-04-16 15:20:36 -07003315 if (likely(ac->avail < ac->limit)) {
3316 STATS_INC_FREEHIT(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003317 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003318 return;
3319 } else {
3320 STATS_INC_FREEMISS(cachep);
3321 cache_flusharray(cachep, ac);
Christoph Lametere498be72005-09-09 13:03:32 -07003322 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003323 }
3324}
3325
3326/**
3327 * kmem_cache_alloc - Allocate an object
3328 * @cachep: The cache to allocate from.
3329 * @flags: See kmalloc().
3330 *
3331 * Allocate an object from this cache. The flags are only relevant
3332 * if the cache has no available objects.
3333 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003334void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003335{
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003336 return __cache_alloc(cachep, flags, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003337}
3338EXPORT_SYMBOL(kmem_cache_alloc);
3339
3340/**
Rolf Eike Beerb8008b22006-07-30 03:04:04 -07003341 * kmem_cache_zalloc - Allocate an object. The memory is set to zero.
Pekka Enberga8c0f9a2006-03-25 03:06:42 -08003342 * @cache: The cache to allocate from.
3343 * @flags: See kmalloc().
3344 *
3345 * Allocate an object from this cache and set the allocated memory to zero.
3346 * The flags are only relevant if the cache has no available objects.
3347 */
3348void *kmem_cache_zalloc(struct kmem_cache *cache, gfp_t flags)
3349{
3350 void *ret = __cache_alloc(cache, flags, __builtin_return_address(0));
3351 if (ret)
3352 memset(ret, 0, obj_size(cache));
3353 return ret;
3354}
3355EXPORT_SYMBOL(kmem_cache_zalloc);
3356
3357/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003358 * kmem_ptr_validate - check if an untrusted pointer might
3359 * be a slab entry.
3360 * @cachep: the cache we're checking against
3361 * @ptr: pointer to validate
3362 *
3363 * This verifies that the untrusted pointer looks sane:
3364 * it is _not_ a guarantee that the pointer is actually
3365 * part of the slab cache in question, but it at least
3366 * validates that the pointer can be dereferenced and
3367 * looks half-way sane.
3368 *
3369 * Currently only used for dentry validation.
3370 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003371int fastcall kmem_ptr_validate(struct kmem_cache *cachep, void *ptr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003372{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003373 unsigned long addr = (unsigned long)ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003374 unsigned long min_addr = PAGE_OFFSET;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003375 unsigned long align_mask = BYTES_PER_WORD - 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003376 unsigned long size = cachep->buffer_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003377 struct page *page;
3378
3379 if (unlikely(addr < min_addr))
3380 goto out;
3381 if (unlikely(addr > (unsigned long)high_memory - size))
3382 goto out;
3383 if (unlikely(addr & align_mask))
3384 goto out;
3385 if (unlikely(!kern_addr_valid(addr)))
3386 goto out;
3387 if (unlikely(!kern_addr_valid(addr + size - 1)))
3388 goto out;
3389 page = virt_to_page(ptr);
3390 if (unlikely(!PageSlab(page)))
3391 goto out;
Pekka Enberg065d41c2005-11-13 16:06:46 -08003392 if (unlikely(page_get_cache(page) != cachep))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003393 goto out;
3394 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003395out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003396 return 0;
3397}
3398
3399#ifdef CONFIG_NUMA
3400/**
3401 * kmem_cache_alloc_node - Allocate an object on the specified node
3402 * @cachep: The cache to allocate from.
3403 * @flags: See kmalloc().
3404 * @nodeid: node number of the target node.
3405 *
3406 * Identical to kmem_cache_alloc, except that this function is slow
3407 * and can sleep. And it will allocate memory on the given node, which
3408 * can improve the performance for cpu bound structures.
Christoph Lametere498be72005-09-09 13:03:32 -07003409 * New and improved: it will now make sure that the object gets
3410 * put on the correct node list so that there is no false sharing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003411 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003412void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003413{
Christoph Lametere498be72005-09-09 13:03:32 -07003414 unsigned long save_flags;
3415 void *ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003416
Christoph Lametere498be72005-09-09 13:03:32 -07003417 cache_alloc_debugcheck_before(cachep, flags);
3418 local_irq_save(save_flags);
Christoph Lameter18f820f2006-02-01 03:05:43 -08003419
3420 if (nodeid == -1 || nodeid == numa_node_id() ||
Andrew Mortona737b3e2006-03-22 00:08:11 -08003421 !cachep->nodelists[nodeid])
Alok N Kataria5c382302005-09-27 21:45:46 -07003422 ptr = ____cache_alloc(cachep, flags);
3423 else
3424 ptr = __cache_alloc_node(cachep, flags, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07003425 local_irq_restore(save_flags);
Christoph Lameter18f820f2006-02-01 03:05:43 -08003426
3427 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr,
3428 __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003429
Christoph Lametere498be72005-09-09 13:03:32 -07003430 return ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003431}
3432EXPORT_SYMBOL(kmem_cache_alloc_node);
3433
Christoph Hellwigdbe5e692006-09-25 23:31:36 -07003434void *__kmalloc_node(size_t size, gfp_t flags, int node)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003435{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003436 struct kmem_cache *cachep;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003437
3438 cachep = kmem_find_general_cachep(size, flags);
3439 if (unlikely(cachep == NULL))
3440 return NULL;
3441 return kmem_cache_alloc_node(cachep, flags, node);
3442}
Christoph Hellwigdbe5e692006-09-25 23:31:36 -07003443EXPORT_SYMBOL(__kmalloc_node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003444#endif
3445
3446/**
Paul Drynoff800590f2006-06-23 02:03:48 -07003447 * __do_kmalloc - allocate memory
Linus Torvalds1da177e2005-04-16 15:20:36 -07003448 * @size: how many bytes of memory are required.
Paul Drynoff800590f2006-06-23 02:03:48 -07003449 * @flags: the type of memory to allocate (see kmalloc).
Randy Dunlap911851e2006-03-22 00:08:14 -08003450 * @caller: function caller for debug tracking of the caller
Linus Torvalds1da177e2005-04-16 15:20:36 -07003451 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003452static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3453 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003454{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003455 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003456
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003457 /* If you want to save a few bytes .text space: replace
3458 * __ with kmem_.
3459 * Then kmalloc uses the uninlined functions instead of the inline
3460 * functions.
3461 */
3462 cachep = __find_general_cachep(size, flags);
Andrew Mortondbdb9042005-09-23 13:24:10 -07003463 if (unlikely(cachep == NULL))
3464 return NULL;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003465 return __cache_alloc(cachep, flags, caller);
3466}
3467
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003468
3469void *__kmalloc(size_t size, gfp_t flags)
3470{
Al Viro871751e2006-03-25 03:06:39 -08003471#ifndef CONFIG_DEBUG_SLAB
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003472 return __do_kmalloc(size, flags, NULL);
Al Viro871751e2006-03-25 03:06:39 -08003473#else
3474 return __do_kmalloc(size, flags, __builtin_return_address(0));
3475#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003476}
3477EXPORT_SYMBOL(__kmalloc);
3478
Al Viro871751e2006-03-25 03:06:39 -08003479#ifdef CONFIG_DEBUG_SLAB
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003480void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller)
3481{
3482 return __do_kmalloc(size, flags, caller);
3483}
3484EXPORT_SYMBOL(__kmalloc_track_caller);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003485#endif
3486
Linus Torvalds1da177e2005-04-16 15:20:36 -07003487/**
3488 * kmem_cache_free - Deallocate an object
3489 * @cachep: The cache the allocation was from.
3490 * @objp: The previously allocated object.
3491 *
3492 * Free an object which was previously allocated from this
3493 * cache.
3494 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003495void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003496{
3497 unsigned long flags;
3498
Pekka Enbergddc2e812006-06-23 02:03:40 -07003499 BUG_ON(virt_to_cache(objp) != cachep);
3500
Linus Torvalds1da177e2005-04-16 15:20:36 -07003501 local_irq_save(flags);
Ingo Molnar873623d2006-07-13 14:44:38 +02003502 __cache_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003503 local_irq_restore(flags);
3504}
3505EXPORT_SYMBOL(kmem_cache_free);
3506
3507/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003508 * kfree - free previously allocated memory
3509 * @objp: pointer returned by kmalloc.
3510 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003511 * If @objp is NULL, no operation is performed.
3512 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003513 * Don't free memory not originally allocated by kmalloc()
3514 * or you will run into trouble.
3515 */
3516void kfree(const void *objp)
3517{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003518 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003519 unsigned long flags;
3520
3521 if (unlikely(!objp))
3522 return;
3523 local_irq_save(flags);
3524 kfree_debugcheck(objp);
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08003525 c = virt_to_cache(objp);
Ingo Molnarf9b84042006-06-27 02:54:49 -07003526 debug_check_no_locks_freed(objp, obj_size(c));
Ingo Molnar873623d2006-07-13 14:44:38 +02003527 __cache_free(c, (void *)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003528 local_irq_restore(flags);
3529}
3530EXPORT_SYMBOL(kfree);
3531
Pekka Enberg343e0d72006-02-01 03:05:50 -08003532unsigned int kmem_cache_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003533{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003534 return obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003535}
3536EXPORT_SYMBOL(kmem_cache_size);
3537
Pekka Enberg343e0d72006-02-01 03:05:50 -08003538const char *kmem_cache_name(struct kmem_cache *cachep)
Arnaldo Carvalho de Melo19449722005-06-18 22:46:19 -07003539{
3540 return cachep->name;
3541}
3542EXPORT_SYMBOL_GPL(kmem_cache_name);
3543
Christoph Lametere498be72005-09-09 13:03:32 -07003544/*
Christoph Lameter0718dc22006-03-25 03:06:47 -08003545 * This initializes kmem_list3 or resizes varioius caches for all nodes.
Christoph Lametere498be72005-09-09 13:03:32 -07003546 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003547static int alloc_kmemlist(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07003548{
3549 int node;
3550 struct kmem_list3 *l3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003551 struct array_cache *new_shared;
3552 struct array_cache **new_alien;
Christoph Lametere498be72005-09-09 13:03:32 -07003553
3554 for_each_online_node(node) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003555
Andrew Mortona737b3e2006-03-22 00:08:11 -08003556 new_alien = alloc_alien_cache(node, cachep->limit);
3557 if (!new_alien)
Christoph Lametere498be72005-09-09 13:03:32 -07003558 goto fail;
Christoph Lametercafeb022006-03-25 03:06:46 -08003559
Christoph Lameter0718dc22006-03-25 03:06:47 -08003560 new_shared = alloc_arraycache(node,
3561 cachep->shared*cachep->batchcount,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003562 0xbaadf00d);
Christoph Lameter0718dc22006-03-25 03:06:47 -08003563 if (!new_shared) {
3564 free_alien_cache(new_alien);
Christoph Lametere498be72005-09-09 13:03:32 -07003565 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003566 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003567
Andrew Mortona737b3e2006-03-22 00:08:11 -08003568 l3 = cachep->nodelists[node];
3569 if (l3) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003570 struct array_cache *shared = l3->shared;
3571
Christoph Lametere498be72005-09-09 13:03:32 -07003572 spin_lock_irq(&l3->list_lock);
3573
Christoph Lametercafeb022006-03-25 03:06:46 -08003574 if (shared)
Christoph Lameter0718dc22006-03-25 03:06:47 -08003575 free_block(cachep, shared->entry,
3576 shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07003577
Christoph Lametercafeb022006-03-25 03:06:46 -08003578 l3->shared = new_shared;
3579 if (!l3->alien) {
Christoph Lametere498be72005-09-09 13:03:32 -07003580 l3->alien = new_alien;
3581 new_alien = NULL;
3582 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003583 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003584 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003585 spin_unlock_irq(&l3->list_lock);
Christoph Lametercafeb022006-03-25 03:06:46 -08003586 kfree(shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003587 free_alien_cache(new_alien);
3588 continue;
3589 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08003590 l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node);
Christoph Lameter0718dc22006-03-25 03:06:47 -08003591 if (!l3) {
3592 free_alien_cache(new_alien);
3593 kfree(new_shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003594 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003595 }
Christoph Lametere498be72005-09-09 13:03:32 -07003596
3597 kmem_list3_init(l3);
3598 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Andrew Mortona737b3e2006-03-22 00:08:11 -08003599 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003600 l3->shared = new_shared;
Christoph Lametere498be72005-09-09 13:03:32 -07003601 l3->alien = new_alien;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003602 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003603 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003604 cachep->nodelists[node] = l3;
3605 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003606 return 0;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003607
Andrew Mortona737b3e2006-03-22 00:08:11 -08003608fail:
Christoph Lameter0718dc22006-03-25 03:06:47 -08003609 if (!cachep->next.next) {
3610 /* Cache is not active yet. Roll back what we did */
3611 node--;
3612 while (node >= 0) {
3613 if (cachep->nodelists[node]) {
3614 l3 = cachep->nodelists[node];
3615
3616 kfree(l3->shared);
3617 free_alien_cache(l3->alien);
3618 kfree(l3);
3619 cachep->nodelists[node] = NULL;
3620 }
3621 node--;
3622 }
3623 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003624 return -ENOMEM;
Christoph Lametere498be72005-09-09 13:03:32 -07003625}
3626
Linus Torvalds1da177e2005-04-16 15:20:36 -07003627struct ccupdate_struct {
Pekka Enberg343e0d72006-02-01 03:05:50 -08003628 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003629 struct array_cache *new[NR_CPUS];
3630};
3631
3632static void do_ccupdate_local(void *info)
3633{
Andrew Mortona737b3e2006-03-22 00:08:11 -08003634 struct ccupdate_struct *new = info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003635 struct array_cache *old;
3636
3637 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003638 old = cpu_cache_get(new->cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003639
Linus Torvalds1da177e2005-04-16 15:20:36 -07003640 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3641 new->new[smp_processor_id()] = old;
3642}
3643
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003644/* Always called with the cache_chain_mutex held */
Andrew Mortona737b3e2006-03-22 00:08:11 -08003645static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3646 int batchcount, int shared)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003647{
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003648 struct ccupdate_struct *new;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003649 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003650
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003651 new = kzalloc(sizeof(*new), GFP_KERNEL);
3652 if (!new)
3653 return -ENOMEM;
3654
Christoph Lametere498be72005-09-09 13:03:32 -07003655 for_each_online_cpu(i) {
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003656 new->new[i] = alloc_arraycache(cpu_to_node(i), limit,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003657 batchcount);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003658 if (!new->new[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003659 for (i--; i >= 0; i--)
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003660 kfree(new->new[i]);
3661 kfree(new);
Christoph Lametere498be72005-09-09 13:03:32 -07003662 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003663 }
3664 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003665 new->cachep = cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003666
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003667 on_each_cpu(do_ccupdate_local, (void *)new, 1, 1);
Christoph Lametere498be72005-09-09 13:03:32 -07003668
Linus Torvalds1da177e2005-04-16 15:20:36 -07003669 check_irq_on();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003670 cachep->batchcount = batchcount;
3671 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07003672 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003673
Christoph Lametere498be72005-09-09 13:03:32 -07003674 for_each_online_cpu(i) {
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003675 struct array_cache *ccold = new->new[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003676 if (!ccold)
3677 continue;
Christoph Lametere498be72005-09-09 13:03:32 -07003678 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -07003679 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
Christoph Lametere498be72005-09-09 13:03:32 -07003680 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003681 kfree(ccold);
3682 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003683 kfree(new);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003684 return alloc_kmemlist(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003685}
3686
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003687/* Called with cache_chain_mutex held always */
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003688static int enable_cpucache(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003689{
3690 int err;
3691 int limit, shared;
3692
Andrew Mortona737b3e2006-03-22 00:08:11 -08003693 /*
3694 * The head array serves three purposes:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003695 * - create a LIFO ordering, i.e. return objects that are cache-warm
3696 * - reduce the number of spinlock operations.
Andrew Mortona737b3e2006-03-22 00:08:11 -08003697 * - reduce the number of linked list operations on the slab and
Linus Torvalds1da177e2005-04-16 15:20:36 -07003698 * bufctl chains: array operations are cheaper.
3699 * The numbers are guessed, we should auto-tune as described by
3700 * Bonwick.
3701 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003702 if (cachep->buffer_size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003703 limit = 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003704 else if (cachep->buffer_size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003705 limit = 8;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003706 else if (cachep->buffer_size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003707 limit = 24;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003708 else if (cachep->buffer_size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003709 limit = 54;
3710 else
3711 limit = 120;
3712
Andrew Mortona737b3e2006-03-22 00:08:11 -08003713 /*
3714 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
Linus Torvalds1da177e2005-04-16 15:20:36 -07003715 * allocation behaviour: Most allocs on one cpu, most free operations
3716 * on another cpu. For these cases, an efficient object passing between
3717 * cpus is necessary. This is provided by a shared array. The array
3718 * replaces Bonwick's magazine layer.
3719 * On uniprocessor, it's functionally equivalent (but less efficient)
3720 * to a larger limit. Thus disabled by default.
3721 */
3722 shared = 0;
3723#ifdef CONFIG_SMP
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003724 if (cachep->buffer_size <= PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003725 shared = 8;
3726#endif
3727
3728#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003729 /*
3730 * With debugging enabled, large batchcount lead to excessively long
3731 * periods with disabled local interrupts. Limit the batchcount
Linus Torvalds1da177e2005-04-16 15:20:36 -07003732 */
3733 if (limit > 32)
3734 limit = 32;
3735#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003736 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003737 if (err)
3738 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003739 cachep->name, -err);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003740 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003741}
3742
Christoph Lameter1b552532006-03-22 00:09:07 -08003743/*
3744 * Drain an array if it contains any elements taking the l3 lock only if
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003745 * necessary. Note that the l3 listlock also protects the array_cache
3746 * if drain_array() is used on the shared array.
Christoph Lameter1b552532006-03-22 00:09:07 -08003747 */
3748void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
3749 struct array_cache *ac, int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003750{
3751 int tofree;
3752
Christoph Lameter1b552532006-03-22 00:09:07 -08003753 if (!ac || !ac->avail)
3754 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003755 if (ac->touched && !force) {
3756 ac->touched = 0;
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003757 } else {
Christoph Lameter1b552532006-03-22 00:09:07 -08003758 spin_lock_irq(&l3->list_lock);
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003759 if (ac->avail) {
3760 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3761 if (tofree > ac->avail)
3762 tofree = (ac->avail + 1) / 2;
3763 free_block(cachep, ac->entry, tofree, node);
3764 ac->avail -= tofree;
3765 memmove(ac->entry, &(ac->entry[tofree]),
3766 sizeof(void *) * ac->avail);
3767 }
Christoph Lameter1b552532006-03-22 00:09:07 -08003768 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003769 }
3770}
3771
3772/**
3773 * cache_reap - Reclaim memory from caches.
Randy Dunlap1e5d5332005-11-07 01:01:06 -08003774 * @unused: unused parameter
Linus Torvalds1da177e2005-04-16 15:20:36 -07003775 *
3776 * Called from workqueue/eventd every few seconds.
3777 * Purpose:
3778 * - clear the per-cpu caches for this CPU.
3779 * - return freeable pages to the main free memory pool.
3780 *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003781 * If we cannot acquire the cache chain mutex then just give up - we'll try
3782 * again on the next iteration.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003783 */
3784static void cache_reap(void *unused)
3785{
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07003786 struct kmem_cache *searchp;
Christoph Lametere498be72005-09-09 13:03:32 -07003787 struct kmem_list3 *l3;
Christoph Lameteraab22072006-03-22 00:09:06 -08003788 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003789
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003790 if (!mutex_trylock(&cache_chain_mutex)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003791 /* Give up. Setup the next iteration. */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003792 schedule_delayed_work(&__get_cpu_var(reap_work),
3793 REAPTIMEOUT_CPUC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003794 return;
3795 }
3796
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07003797 list_for_each_entry(searchp, &cache_chain, next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003798 check_irq_on();
3799
Christoph Lameter35386e32006-03-22 00:09:05 -08003800 /*
3801 * We only take the l3 lock if absolutely necessary and we
3802 * have established with reasonable certainty that
3803 * we can do some work if the lock was obtained.
3804 */
Christoph Lameteraab22072006-03-22 00:09:06 -08003805 l3 = searchp->nodelists[node];
Christoph Lameter35386e32006-03-22 00:09:05 -08003806
Christoph Lameter8fce4d82006-03-09 17:33:54 -08003807 reap_alien(searchp, l3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003808
Christoph Lameteraab22072006-03-22 00:09:06 -08003809 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003810
Christoph Lameter35386e32006-03-22 00:09:05 -08003811 /*
3812 * These are racy checks but it does not matter
3813 * if we skip one check or scan twice.
3814 */
Christoph Lametere498be72005-09-09 13:03:32 -07003815 if (time_after(l3->next_reap, jiffies))
Christoph Lameter35386e32006-03-22 00:09:05 -08003816 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003817
Christoph Lametere498be72005-09-09 13:03:32 -07003818 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003819
Christoph Lameteraab22072006-03-22 00:09:06 -08003820 drain_array(searchp, l3, l3->shared, 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003821
Christoph Lametered11d9e2006-06-30 01:55:45 -07003822 if (l3->free_touched)
Christoph Lametere498be72005-09-09 13:03:32 -07003823 l3->free_touched = 0;
Christoph Lametered11d9e2006-06-30 01:55:45 -07003824 else {
3825 int freed;
3826
3827 freed = drain_freelist(searchp, l3, (l3->free_limit +
3828 5 * searchp->num - 1) / (5 * searchp->num));
3829 STATS_ADD_REAPED(searchp, freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003830 }
Christoph Lameter35386e32006-03-22 00:09:05 -08003831next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003832 cond_resched();
3833 }
3834 check_irq_on();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003835 mutex_unlock(&cache_chain_mutex);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08003836 next_reap_node();
Christoph Lameter2244b952006-06-30 01:55:33 -07003837 refresh_cpu_vm_stats(smp_processor_id());
Andrew Mortona737b3e2006-03-22 00:08:11 -08003838 /* Set up the next iteration */
Manfred Spraulcd61ef62005-11-07 00:58:02 -08003839 schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003840}
3841
3842#ifdef CONFIG_PROC_FS
3843
Pekka Enberg85289f92006-01-08 01:00:36 -08003844static void print_slabinfo_header(struct seq_file *m)
3845{
3846 /*
3847 * Output format version, so at least we can change it
3848 * without _too_ many complaints.
3849 */
3850#if STATS
3851 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
3852#else
3853 seq_puts(m, "slabinfo - version: 2.1\n");
3854#endif
3855 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
3856 "<objperslab> <pagesperslab>");
3857 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
3858 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
3859#if STATS
3860 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07003861 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
Pekka Enberg85289f92006-01-08 01:00:36 -08003862 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
3863#endif
3864 seq_putc(m, '\n');
3865}
3866
Linus Torvalds1da177e2005-04-16 15:20:36 -07003867static void *s_start(struct seq_file *m, loff_t *pos)
3868{
3869 loff_t n = *pos;
3870 struct list_head *p;
3871
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003872 mutex_lock(&cache_chain_mutex);
Pekka Enberg85289f92006-01-08 01:00:36 -08003873 if (!n)
3874 print_slabinfo_header(m);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003875 p = cache_chain.next;
3876 while (n--) {
3877 p = p->next;
3878 if (p == &cache_chain)
3879 return NULL;
3880 }
Pekka Enberg343e0d72006-02-01 03:05:50 -08003881 return list_entry(p, struct kmem_cache, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003882}
3883
3884static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3885{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003886 struct kmem_cache *cachep = p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003887 ++*pos;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003888 return cachep->next.next == &cache_chain ?
3889 NULL : list_entry(cachep->next.next, struct kmem_cache, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003890}
3891
3892static void s_stop(struct seq_file *m, void *p)
3893{
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003894 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003895}
3896
3897static int s_show(struct seq_file *m, void *p)
3898{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003899 struct kmem_cache *cachep = p;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003900 struct slab *slabp;
3901 unsigned long active_objs;
3902 unsigned long num_objs;
3903 unsigned long active_slabs = 0;
3904 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07003905 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003906 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003907 int node;
3908 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003909
Linus Torvalds1da177e2005-04-16 15:20:36 -07003910 active_objs = 0;
3911 num_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07003912 for_each_online_node(node) {
3913 l3 = cachep->nodelists[node];
3914 if (!l3)
3915 continue;
3916
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003917 check_irq_on();
3918 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003919
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07003920 list_for_each_entry(slabp, &l3->slabs_full, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07003921 if (slabp->inuse != cachep->num && !error)
3922 error = "slabs_full accounting error";
3923 active_objs += cachep->num;
3924 active_slabs++;
3925 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07003926 list_for_each_entry(slabp, &l3->slabs_partial, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07003927 if (slabp->inuse == cachep->num && !error)
3928 error = "slabs_partial inuse accounting error";
3929 if (!slabp->inuse && !error)
3930 error = "slabs_partial/inuse accounting error";
3931 active_objs += slabp->inuse;
3932 active_slabs++;
3933 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07003934 list_for_each_entry(slabp, &l3->slabs_free, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07003935 if (slabp->inuse && !error)
3936 error = "slabs_free/inuse accounting error";
3937 num_slabs++;
3938 }
3939 free_objects += l3->free_objects;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08003940 if (l3->shared)
3941 shared_avail += l3->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07003942
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003943 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003944 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003945 num_slabs += active_slabs;
3946 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003947 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003948 error = "free_objects accounting error";
3949
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003950 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003951 if (error)
3952 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
3953
3954 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003955 name, active_objs, num_objs, cachep->buffer_size,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003956 cachep->num, (1 << cachep->gfporder));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003957 seq_printf(m, " : tunables %4u %4u %4u",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003958 cachep->limit, cachep->batchcount, cachep->shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003959 seq_printf(m, " : slabdata %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003960 active_slabs, num_slabs, shared_avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003961#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003962 { /* list3 stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003963 unsigned long high = cachep->high_mark;
3964 unsigned long allocs = cachep->num_allocations;
3965 unsigned long grown = cachep->grown;
3966 unsigned long reaped = cachep->reaped;
3967 unsigned long errors = cachep->errors;
3968 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003969 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07003970 unsigned long node_frees = cachep->node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07003971 unsigned long overflows = cachep->node_overflow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003972
Christoph Lametere498be72005-09-09 13:03:32 -07003973 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07003974 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003975 reaped, errors, max_freeable, node_allocs,
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07003976 node_frees, overflows);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003977 }
3978 /* cpu stats */
3979 {
3980 unsigned long allochit = atomic_read(&cachep->allochit);
3981 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
3982 unsigned long freehit = atomic_read(&cachep->freehit);
3983 unsigned long freemiss = atomic_read(&cachep->freemiss);
3984
3985 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003986 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003987 }
3988#endif
3989 seq_putc(m, '\n');
Linus Torvalds1da177e2005-04-16 15:20:36 -07003990 return 0;
3991}
3992
3993/*
3994 * slabinfo_op - iterator that generates /proc/slabinfo
3995 *
3996 * Output layout:
3997 * cache-name
3998 * num-active-objs
3999 * total-objs
4000 * object size
4001 * num-active-slabs
4002 * total-slabs
4003 * num-pages-per-slab
4004 * + further values on SMP and with statistics enabled
4005 */
4006
4007struct seq_operations slabinfo_op = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004008 .start = s_start,
4009 .next = s_next,
4010 .stop = s_stop,
4011 .show = s_show,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004012};
4013
4014#define MAX_SLABINFO_WRITE 128
4015/**
4016 * slabinfo_write - Tuning for the slab allocator
4017 * @file: unused
4018 * @buffer: user buffer
4019 * @count: data length
4020 * @ppos: unused
4021 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004022ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4023 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004024{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004025 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004026 int limit, batchcount, shared, res;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004027 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004028
Linus Torvalds1da177e2005-04-16 15:20:36 -07004029 if (count > MAX_SLABINFO_WRITE)
4030 return -EINVAL;
4031 if (copy_from_user(&kbuf, buffer, count))
4032 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004033 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07004034
4035 tmp = strchr(kbuf, ' ');
4036 if (!tmp)
4037 return -EINVAL;
4038 *tmp = '\0';
4039 tmp++;
4040 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4041 return -EINVAL;
4042
4043 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004044 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004045 res = -EINVAL;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004046 list_for_each_entry(cachep, &cache_chain, next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004047 if (!strcmp(cachep->name, kbuf)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08004048 if (limit < 1 || batchcount < 1 ||
4049 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07004050 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004051 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07004052 res = do_tune_cpucache(cachep, limit,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004053 batchcount, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004054 }
4055 break;
4056 }
4057 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004058 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004059 if (res >= 0)
4060 res = count;
4061 return res;
4062}
Al Viro871751e2006-03-25 03:06:39 -08004063
4064#ifdef CONFIG_DEBUG_SLAB_LEAK
4065
4066static void *leaks_start(struct seq_file *m, loff_t *pos)
4067{
4068 loff_t n = *pos;
4069 struct list_head *p;
4070
4071 mutex_lock(&cache_chain_mutex);
4072 p = cache_chain.next;
4073 while (n--) {
4074 p = p->next;
4075 if (p == &cache_chain)
4076 return NULL;
4077 }
4078 return list_entry(p, struct kmem_cache, next);
4079}
4080
4081static inline int add_caller(unsigned long *n, unsigned long v)
4082{
4083 unsigned long *p;
4084 int l;
4085 if (!v)
4086 return 1;
4087 l = n[1];
4088 p = n + 2;
4089 while (l) {
4090 int i = l/2;
4091 unsigned long *q = p + 2 * i;
4092 if (*q == v) {
4093 q[1]++;
4094 return 1;
4095 }
4096 if (*q > v) {
4097 l = i;
4098 } else {
4099 p = q + 2;
4100 l -= i + 1;
4101 }
4102 }
4103 if (++n[1] == n[0])
4104 return 0;
4105 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4106 p[0] = v;
4107 p[1] = 1;
4108 return 1;
4109}
4110
4111static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4112{
4113 void *p;
4114 int i;
4115 if (n[0] == n[1])
4116 return;
4117 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4118 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4119 continue;
4120 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4121 return;
4122 }
4123}
4124
4125static void show_symbol(struct seq_file *m, unsigned long address)
4126{
4127#ifdef CONFIG_KALLSYMS
4128 char *modname;
4129 const char *name;
4130 unsigned long offset, size;
4131 char namebuf[KSYM_NAME_LEN+1];
4132
4133 name = kallsyms_lookup(address, &size, &offset, &modname, namebuf);
4134
4135 if (name) {
4136 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4137 if (modname)
4138 seq_printf(m, " [%s]", modname);
4139 return;
4140 }
4141#endif
4142 seq_printf(m, "%p", (void *)address);
4143}
4144
4145static int leaks_show(struct seq_file *m, void *p)
4146{
4147 struct kmem_cache *cachep = p;
Al Viro871751e2006-03-25 03:06:39 -08004148 struct slab *slabp;
4149 struct kmem_list3 *l3;
4150 const char *name;
4151 unsigned long *n = m->private;
4152 int node;
4153 int i;
4154
4155 if (!(cachep->flags & SLAB_STORE_USER))
4156 return 0;
4157 if (!(cachep->flags & SLAB_RED_ZONE))
4158 return 0;
4159
4160 /* OK, we can do it */
4161
4162 n[1] = 0;
4163
4164 for_each_online_node(node) {
4165 l3 = cachep->nodelists[node];
4166 if (!l3)
4167 continue;
4168
4169 check_irq_on();
4170 spin_lock_irq(&l3->list_lock);
4171
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004172 list_for_each_entry(slabp, &l3->slabs_full, list)
Al Viro871751e2006-03-25 03:06:39 -08004173 handle_slab(n, cachep, slabp);
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004174 list_for_each_entry(slabp, &l3->slabs_partial, list)
Al Viro871751e2006-03-25 03:06:39 -08004175 handle_slab(n, cachep, slabp);
Al Viro871751e2006-03-25 03:06:39 -08004176 spin_unlock_irq(&l3->list_lock);
4177 }
4178 name = cachep->name;
4179 if (n[0] == n[1]) {
4180 /* Increase the buffer size */
4181 mutex_unlock(&cache_chain_mutex);
4182 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4183 if (!m->private) {
4184 /* Too bad, we are really out */
4185 m->private = n;
4186 mutex_lock(&cache_chain_mutex);
4187 return -ENOMEM;
4188 }
4189 *(unsigned long *)m->private = n[0] * 2;
4190 kfree(n);
4191 mutex_lock(&cache_chain_mutex);
4192 /* Now make sure this entry will be retried */
4193 m->count = m->size;
4194 return 0;
4195 }
4196 for (i = 0; i < n[1]; i++) {
4197 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4198 show_symbol(m, n[2*i+2]);
4199 seq_putc(m, '\n');
4200 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004201
Al Viro871751e2006-03-25 03:06:39 -08004202 return 0;
4203}
4204
4205struct seq_operations slabstats_op = {
4206 .start = leaks_start,
4207 .next = s_next,
4208 .stop = s_stop,
4209 .show = leaks_show,
4210};
4211#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004212#endif
4213
Manfred Spraul00e145b2005-09-03 15:55:07 -07004214/**
4215 * ksize - get the actual amount of memory allocated for a given object
4216 * @objp: Pointer to the object
4217 *
4218 * kmalloc may internally round up allocations and return more memory
4219 * than requested. ksize() can be used to determine the actual amount of
4220 * memory allocated. The caller may use this additional memory, even though
4221 * a smaller amount of memory was initially specified with the kmalloc call.
4222 * The caller must guarantee that objp points to a valid object previously
4223 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4224 * must not be freed during the duration of the call.
4225 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004226unsigned int ksize(const void *objp)
4227{
Manfred Spraul00e145b2005-09-03 15:55:07 -07004228 if (unlikely(objp == NULL))
4229 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004230
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08004231 return obj_size(virt_to_cache(objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004232}