blob: 909975f6e0903f4a4f065ded0f109a338193173f [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * linux/mm/slab.c
3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
5 *
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
7 *
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
10 *
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
13 *
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
21 *
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
27 *
28 * This means, that your constructor is used only for newly allocated
29 * slabs and you must pass objects with the same intializations to
30 * kmem_cache_free.
31 *
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
35 *
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
38 * partial slabs
39 * empty slabs with no allocated objects
40 *
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
43 *
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
46 *
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
52 *
Andrew Mortona737b3e2006-03-22 00:08:11 -080053 * The c_cpuarray may not be read with enabled local interrupts -
Linus Torvalds1da177e2005-04-16 15:20:36 -070054 * it's changed with a smp_call_function().
55 *
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
Pekka Enberg343e0d72006-02-01 03:05:50 -080058 * Several members in struct kmem_cache and struct slab never change, they
Linus Torvalds1da177e2005-04-16 15:20:36 -070059 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
63 *
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
66 * his patch.
67 *
68 * Further notes from the original documentation:
69 *
70 * 11 April '97. Started multi-threading - markhe
Ingo Molnarfc0abb12006-01-18 17:42:33 -080071 * The global cache-chain is protected by the mutex 'cache_chain_mutex'.
Linus Torvalds1da177e2005-04-16 15:20:36 -070072 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
75 *
76 * At present, each engine can be growing a cache. This should be blocked.
77 *
Christoph Lametere498be72005-09-09 13:03:32 -070078 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
83 *
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
Linus Torvalds1da177e2005-04-16 15:20:36 -070087 */
88
Linus Torvalds1da177e2005-04-16 15:20:36 -070089#include <linux/slab.h>
90#include <linux/mm.h>
Randy Dunlapc9cf5522006-06-27 02:53:52 -070091#include <linux/poison.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070092#include <linux/swap.h>
93#include <linux/cache.h>
94#include <linux/interrupt.h>
95#include <linux/init.h>
96#include <linux/compiler.h>
Paul Jackson101a5002006-03-24 03:16:07 -080097#include <linux/cpuset.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070098#include <linux/seq_file.h>
99#include <linux/notifier.h>
100#include <linux/kallsyms.h>
101#include <linux/cpu.h>
102#include <linux/sysctl.h>
103#include <linux/module.h>
104#include <linux/rcupdate.h>
Paulo Marques543537b2005-06-23 00:09:02 -0700105#include <linux/string.h>
Andrew Morton138ae662006-12-06 20:36:41 -0800106#include <linux/uaccess.h>
Christoph Lametere498be72005-09-09 13:03:32 -0700107#include <linux/nodemask.h>
Christoph Lameterdc85da12006-01-18 17:42:36 -0800108#include <linux/mempolicy.h>
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800109#include <linux/mutex.h>
Akinobu Mita8a8b6502006-12-08 02:39:44 -0800110#include <linux/fault-inject.h>
Ingo Molnare7eebaf2006-06-27 02:54:55 -0700111#include <linux/rtmutex.h>
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800112#include <linux/reciprocal_div.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700113
Linus Torvalds1da177e2005-04-16 15:20:36 -0700114#include <asm/cacheflush.h>
115#include <asm/tlbflush.h>
116#include <asm/page.h>
117
118/*
119 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL,
120 * SLAB_RED_ZONE & SLAB_POISON.
121 * 0 for faster, smaller code (especially in the critical paths).
122 *
123 * STATS - 1 to collect stats for /proc/slabinfo.
124 * 0 for faster, smaller code (especially in the critical paths).
125 *
126 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
127 */
128
129#ifdef CONFIG_DEBUG_SLAB
130#define DEBUG 1
131#define STATS 1
132#define FORCED_DEBUG 1
133#else
134#define DEBUG 0
135#define STATS 0
136#define FORCED_DEBUG 0
137#endif
138
Linus Torvalds1da177e2005-04-16 15:20:36 -0700139/* Shouldn't this be in a header file somewhere? */
140#define BYTES_PER_WORD sizeof(void *)
141
142#ifndef cache_line_size
143#define cache_line_size() L1_CACHE_BYTES
144#endif
145
146#ifndef ARCH_KMALLOC_MINALIGN
147/*
148 * Enforce a minimum alignment for the kmalloc caches.
149 * Usually, the kmalloc caches are cache_line_size() aligned, except when
150 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
151 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
152 * alignment larger than BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that.
153 * Note that this flag disables some debug features.
154 */
155#define ARCH_KMALLOC_MINALIGN 0
156#endif
157
158#ifndef ARCH_SLAB_MINALIGN
159/*
160 * Enforce a minimum alignment for all caches.
161 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
162 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
163 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
164 * some debug features.
165 */
166#define ARCH_SLAB_MINALIGN 0
167#endif
168
169#ifndef ARCH_KMALLOC_FLAGS
170#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
171#endif
172
173/* Legal flag mask for kmem_cache_create(). */
174#if DEBUG
175# define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \
176 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
Christoph Lameterac2b8982006-03-22 00:08:15 -0800177 SLAB_CACHE_DMA | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700178 SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \
179 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
Paul Jackson101a5002006-03-24 03:16:07 -0800180 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700181#else
Christoph Lameterac2b8982006-03-22 00:08:15 -0800182# define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700183 SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \
184 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
Paul Jackson101a5002006-03-24 03:16:07 -0800185 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700186#endif
187
188/*
189 * kmem_bufctl_t:
190 *
191 * Bufctl's are used for linking objs within a slab
192 * linked offsets.
193 *
194 * This implementation relies on "struct page" for locating the cache &
195 * slab an object belongs to.
196 * This allows the bufctl structure to be small (one int), but limits
197 * the number of objects a slab (not a cache) can contain when off-slab
198 * bufctls are used. The limit is the size of the largest general cache
199 * that does not use off-slab slabs.
200 * For 32bit archs with 4 kB pages, is this 56.
201 * This is not serious, as it is only for large objects, when it is unwise
202 * to have too many per slab.
203 * Note: This limit can be raised by introducing a general cache whose size
204 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
205 */
206
Kyle Moffettfa5b08d2005-09-03 15:55:03 -0700207typedef unsigned int kmem_bufctl_t;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700208#define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
209#define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
Al Viro871751e2006-03-25 03:06:39 -0800210#define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
211#define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700212
Linus Torvalds1da177e2005-04-16 15:20:36 -0700213/*
214 * struct slab
215 *
216 * Manages the objs in a slab. Placed either at the beginning of mem allocated
217 * for a slab, or allocated from an general cache.
218 * Slabs are chained into three list: fully used, partial, fully free slabs.
219 */
220struct slab {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800221 struct list_head list;
222 unsigned long colouroff;
223 void *s_mem; /* including colour offset */
224 unsigned int inuse; /* num of objs active in slab */
225 kmem_bufctl_t free;
226 unsigned short nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700227};
228
229/*
230 * struct slab_rcu
231 *
232 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
233 * arrange for kmem_freepages to be called via RCU. This is useful if
234 * we need to approach a kernel structure obliquely, from its address
235 * obtained without the usual locking. We can lock the structure to
236 * stabilize it and check it's still at the given address, only if we
237 * can be sure that the memory has not been meanwhile reused for some
238 * other kind of object (which our subsystem's lock might corrupt).
239 *
240 * rcu_read_lock before reading the address, then rcu_read_unlock after
241 * taking the spinlock within the structure expected at that address.
242 *
243 * We assume struct slab_rcu can overlay struct slab when destroying.
244 */
245struct slab_rcu {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800246 struct rcu_head head;
Pekka Enberg343e0d72006-02-01 03:05:50 -0800247 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800248 void *addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700249};
250
251/*
252 * struct array_cache
253 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700254 * Purpose:
255 * - LIFO ordering, to hand out cache-warm objects from _alloc
256 * - reduce the number of linked list operations
257 * - reduce spinlock operations
258 *
259 * The limit is stored in the per-cpu structure to reduce the data cache
260 * footprint.
261 *
262 */
263struct array_cache {
264 unsigned int avail;
265 unsigned int limit;
266 unsigned int batchcount;
267 unsigned int touched;
Christoph Lametere498be72005-09-09 13:03:32 -0700268 spinlock_t lock;
Andrew Mortona737b3e2006-03-22 00:08:11 -0800269 void *entry[0]; /*
270 * Must have this definition in here for the proper
271 * alignment of array_cache. Also simplifies accessing
272 * the entries.
273 * [0] is for gcc 2.95. It should really be [].
274 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700275};
276
Andrew Mortona737b3e2006-03-22 00:08:11 -0800277/*
278 * bootstrap: The caches do not work without cpuarrays anymore, but the
279 * cpuarrays are allocated from the generic caches...
Linus Torvalds1da177e2005-04-16 15:20:36 -0700280 */
281#define BOOT_CPUCACHE_ENTRIES 1
282struct arraycache_init {
283 struct array_cache cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800284 void *entries[BOOT_CPUCACHE_ENTRIES];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700285};
286
287/*
Christoph Lametere498be72005-09-09 13:03:32 -0700288 * The slab lists for all objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700289 */
290struct kmem_list3 {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800291 struct list_head slabs_partial; /* partial list first, better asm code */
292 struct list_head slabs_full;
293 struct list_head slabs_free;
294 unsigned long free_objects;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800295 unsigned int free_limit;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800296 unsigned int colour_next; /* Per-node cache coloring */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800297 spinlock_t list_lock;
298 struct array_cache *shared; /* shared per node */
299 struct array_cache **alien; /* on other nodes */
Christoph Lameter35386e32006-03-22 00:09:05 -0800300 unsigned long next_reap; /* updated without locking */
301 int free_touched; /* updated without locking */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700302};
303
Christoph Lametere498be72005-09-09 13:03:32 -0700304/*
305 * Need this for bootstrapping a per node allocator.
306 */
307#define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
308struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
309#define CACHE_CACHE 0
310#define SIZE_AC 1
311#define SIZE_L3 (1 + MAX_NUMNODES)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700312
Christoph Lametered11d9e2006-06-30 01:55:45 -0700313static int drain_freelist(struct kmem_cache *cache,
314 struct kmem_list3 *l3, int tofree);
315static void free_block(struct kmem_cache *cachep, void **objpp, int len,
316 int node);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -0700317static int enable_cpucache(struct kmem_cache *cachep);
David Howells65f27f32006-11-22 14:55:48 +0000318static void cache_reap(struct work_struct *unused);
Christoph Lametered11d9e2006-06-30 01:55:45 -0700319
Christoph Lametere498be72005-09-09 13:03:32 -0700320/*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800321 * This function must be completely optimized away if a constant is passed to
322 * it. Mostly the same as what is in linux/slab.h except it returns an index.
Christoph Lametere498be72005-09-09 13:03:32 -0700323 */
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700324static __always_inline int index_of(const size_t size)
Christoph Lametere498be72005-09-09 13:03:32 -0700325{
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800326 extern void __bad_size(void);
327
Christoph Lametere498be72005-09-09 13:03:32 -0700328 if (__builtin_constant_p(size)) {
329 int i = 0;
330
331#define CACHE(x) \
332 if (size <=x) \
333 return i; \
334 else \
335 i++;
336#include "linux/kmalloc_sizes.h"
337#undef CACHE
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800338 __bad_size();
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700339 } else
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800340 __bad_size();
Christoph Lametere498be72005-09-09 13:03:32 -0700341 return 0;
342}
343
Ingo Molnare0a42722006-06-23 02:03:46 -0700344static int slab_early_init = 1;
345
Christoph Lametere498be72005-09-09 13:03:32 -0700346#define INDEX_AC index_of(sizeof(struct arraycache_init))
347#define INDEX_L3 index_of(sizeof(struct kmem_list3))
348
Pekka Enberg5295a742006-02-01 03:05:48 -0800349static void kmem_list3_init(struct kmem_list3 *parent)
Christoph Lametere498be72005-09-09 13:03:32 -0700350{
351 INIT_LIST_HEAD(&parent->slabs_full);
352 INIT_LIST_HEAD(&parent->slabs_partial);
353 INIT_LIST_HEAD(&parent->slabs_free);
354 parent->shared = NULL;
355 parent->alien = NULL;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800356 parent->colour_next = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700357 spin_lock_init(&parent->list_lock);
358 parent->free_objects = 0;
359 parent->free_touched = 0;
360}
361
Andrew Mortona737b3e2006-03-22 00:08:11 -0800362#define MAKE_LIST(cachep, listp, slab, nodeid) \
363 do { \
364 INIT_LIST_HEAD(listp); \
365 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
Christoph Lametere498be72005-09-09 13:03:32 -0700366 } while (0)
367
Andrew Mortona737b3e2006-03-22 00:08:11 -0800368#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
369 do { \
Christoph Lametere498be72005-09-09 13:03:32 -0700370 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
371 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
372 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
373 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700374
375/*
Pekka Enberg343e0d72006-02-01 03:05:50 -0800376 * struct kmem_cache
Linus Torvalds1da177e2005-04-16 15:20:36 -0700377 *
378 * manages a cache.
379 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800380
Pekka J Enberg2109a2d2005-11-07 00:58:01 -0800381struct kmem_cache {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700382/* 1) per-cpu data, touched during every alloc/free */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800383 struct array_cache *array[NR_CPUS];
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800384/* 2) Cache tunables. Protected by cache_chain_mutex */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800385 unsigned int batchcount;
386 unsigned int limit;
387 unsigned int shared;
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800388
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800389 unsigned int buffer_size;
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800390 u32 reciprocal_buffer_size;
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800391/* 3) touched by every alloc & free from the backend */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800392 struct kmem_list3 *nodelists[MAX_NUMNODES];
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800393
Andrew Mortona737b3e2006-03-22 00:08:11 -0800394 unsigned int flags; /* constant flags */
395 unsigned int num; /* # of objs per slab */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700396
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800397/* 4) cache_grow/shrink */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700398 /* order of pgs per slab (2^n) */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800399 unsigned int gfporder;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700400
401 /* force GFP flags, e.g. GFP_DMA */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800402 gfp_t gfpflags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700403
Andrew Mortona737b3e2006-03-22 00:08:11 -0800404 size_t colour; /* cache colouring range */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800405 unsigned int colour_off; /* colour offset */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800406 struct kmem_cache *slabp_cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800407 unsigned int slab_size;
Andrew Mortona737b3e2006-03-22 00:08:11 -0800408 unsigned int dflags; /* dynamic flags */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700409
410 /* constructor func */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800411 void (*ctor) (void *, struct kmem_cache *, unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700412
413 /* de-constructor func */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800414 void (*dtor) (void *, struct kmem_cache *, unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700415
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800416/* 5) cache creation/removal */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800417 const char *name;
418 struct list_head next;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700419
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800420/* 6) statistics */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700421#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800422 unsigned long num_active;
423 unsigned long num_allocations;
424 unsigned long high_mark;
425 unsigned long grown;
426 unsigned long reaped;
427 unsigned long errors;
428 unsigned long max_freeable;
429 unsigned long node_allocs;
430 unsigned long node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700431 unsigned long node_overflow;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800432 atomic_t allochit;
433 atomic_t allocmiss;
434 atomic_t freehit;
435 atomic_t freemiss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700436#endif
437#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800438 /*
439 * If debugging is enabled, then the allocator can add additional
440 * fields and/or padding to every object. buffer_size contains the total
441 * object size including these internal fields, the following two
442 * variables contain the offset to the user object and its size.
443 */
444 int obj_offset;
445 int obj_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700446#endif
447};
448
449#define CFLGS_OFF_SLAB (0x80000000UL)
450#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
451
452#define BATCHREFILL_LIMIT 16
Andrew Mortona737b3e2006-03-22 00:08:11 -0800453/*
454 * Optimization question: fewer reaps means less probability for unnessary
455 * cpucache drain/refill cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700456 *
Adrian Bunkdc6f3f22005-11-08 16:44:08 +0100457 * OTOH the cpuarrays can contain lots of objects,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700458 * which could lock up otherwise freeable slabs.
459 */
460#define REAPTIMEOUT_CPUC (2*HZ)
461#define REAPTIMEOUT_LIST3 (4*HZ)
462
463#if STATS
464#define STATS_INC_ACTIVE(x) ((x)->num_active++)
465#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
466#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
467#define STATS_INC_GROWN(x) ((x)->grown++)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700468#define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
Andrew Mortona737b3e2006-03-22 00:08:11 -0800469#define STATS_SET_HIGH(x) \
470 do { \
471 if ((x)->num_active > (x)->high_mark) \
472 (x)->high_mark = (x)->num_active; \
473 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700474#define STATS_INC_ERR(x) ((x)->errors++)
475#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
Christoph Lametere498be72005-09-09 13:03:32 -0700476#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700477#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800478#define STATS_SET_FREEABLE(x, i) \
479 do { \
480 if ((x)->max_freeable < i) \
481 (x)->max_freeable = i; \
482 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700483#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
484#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
485#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
486#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
487#else
488#define STATS_INC_ACTIVE(x) do { } while (0)
489#define STATS_DEC_ACTIVE(x) do { } while (0)
490#define STATS_INC_ALLOCED(x) do { } while (0)
491#define STATS_INC_GROWN(x) do { } while (0)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700492#define STATS_ADD_REAPED(x,y) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700493#define STATS_SET_HIGH(x) do { } while (0)
494#define STATS_INC_ERR(x) do { } while (0)
495#define STATS_INC_NODEALLOCS(x) do { } while (0)
Christoph Lametere498be72005-09-09 13:03:32 -0700496#define STATS_INC_NODEFREES(x) do { } while (0)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700497#define STATS_INC_ACOVERFLOW(x) do { } while (0)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800498#define STATS_SET_FREEABLE(x, i) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700499#define STATS_INC_ALLOCHIT(x) do { } while (0)
500#define STATS_INC_ALLOCMISS(x) do { } while (0)
501#define STATS_INC_FREEHIT(x) do { } while (0)
502#define STATS_INC_FREEMISS(x) do { } while (0)
503#endif
504
505#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -0700506
Andrew Mortona737b3e2006-03-22 00:08:11 -0800507/*
508 * memory layout of objects:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700509 * 0 : objp
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800510 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
Linus Torvalds1da177e2005-04-16 15:20:36 -0700511 * the end of an object is aligned with the end of the real
512 * allocation. Catches writes behind the end of the allocation.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800513 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700514 * redzone word.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800515 * cachep->obj_offset: The real object.
516 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
Andrew Mortona737b3e2006-03-22 00:08:11 -0800517 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
518 * [BYTES_PER_WORD long]
Linus Torvalds1da177e2005-04-16 15:20:36 -0700519 */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800520static int obj_offset(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700521{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800522 return cachep->obj_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700523}
524
Pekka Enberg343e0d72006-02-01 03:05:50 -0800525static int obj_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700526{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800527 return cachep->obj_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700528}
529
Pekka Enberg343e0d72006-02-01 03:05:50 -0800530static unsigned long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700531{
532 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800533 return (unsigned long*) (objp+obj_offset(cachep)-BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700534}
535
Pekka Enberg343e0d72006-02-01 03:05:50 -0800536static unsigned long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700537{
538 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
539 if (cachep->flags & SLAB_STORE_USER)
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800540 return (unsigned long *)(objp + cachep->buffer_size -
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800541 2 * BYTES_PER_WORD);
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800542 return (unsigned long *)(objp + cachep->buffer_size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700543}
544
Pekka Enberg343e0d72006-02-01 03:05:50 -0800545static void **dbg_userword(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700546{
547 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800548 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700549}
550
551#else
552
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800553#define obj_offset(x) 0
554#define obj_size(cachep) (cachep->buffer_size)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700555#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;})
556#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;})
557#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
558
559#endif
560
561/*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800562 * Maximum size of an obj (in 2^order pages) and absolute limit for the gfp
563 * order.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700564 */
565#if defined(CONFIG_LARGE_ALLOCS)
566#define MAX_OBJ_ORDER 13 /* up to 32Mb */
567#define MAX_GFP_ORDER 13 /* up to 32Mb */
568#elif defined(CONFIG_MMU)
569#define MAX_OBJ_ORDER 5 /* 32 pages */
570#define MAX_GFP_ORDER 5 /* 32 pages */
571#else
572#define MAX_OBJ_ORDER 8 /* up to 1Mb */
573#define MAX_GFP_ORDER 8 /* up to 1Mb */
574#endif
575
576/*
577 * Do not go above this order unless 0 objects fit into the slab.
578 */
579#define BREAK_GFP_ORDER_HI 1
580#define BREAK_GFP_ORDER_LO 0
581static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
582
Andrew Mortona737b3e2006-03-22 00:08:11 -0800583/*
584 * Functions for storing/retrieving the cachep and or slab from the page
585 * allocator. These are used to find the slab an obj belongs to. With kfree(),
586 * these are used to find the cache which an obj belongs to.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700587 */
Pekka Enberg065d41c2005-11-13 16:06:46 -0800588static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
589{
590 page->lru.next = (struct list_head *)cache;
591}
592
593static inline struct kmem_cache *page_get_cache(struct page *page)
594{
Nick Piggin84097512006-03-22 00:08:34 -0800595 if (unlikely(PageCompound(page)))
596 page = (struct page *)page_private(page);
Pekka Enbergddc2e812006-06-23 02:03:40 -0700597 BUG_ON(!PageSlab(page));
Pekka Enberg065d41c2005-11-13 16:06:46 -0800598 return (struct kmem_cache *)page->lru.next;
599}
600
601static inline void page_set_slab(struct page *page, struct slab *slab)
602{
603 page->lru.prev = (struct list_head *)slab;
604}
605
606static inline struct slab *page_get_slab(struct page *page)
607{
Nick Piggin84097512006-03-22 00:08:34 -0800608 if (unlikely(PageCompound(page)))
609 page = (struct page *)page_private(page);
Pekka Enbergddc2e812006-06-23 02:03:40 -0700610 BUG_ON(!PageSlab(page));
Pekka Enberg065d41c2005-11-13 16:06:46 -0800611 return (struct slab *)page->lru.prev;
612}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700613
Pekka Enberg6ed5eb22006-02-01 03:05:49 -0800614static inline struct kmem_cache *virt_to_cache(const void *obj)
615{
616 struct page *page = virt_to_page(obj);
617 return page_get_cache(page);
618}
619
620static inline struct slab *virt_to_slab(const void *obj)
621{
622 struct page *page = virt_to_page(obj);
623 return page_get_slab(page);
624}
625
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800626static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
627 unsigned int idx)
628{
629 return slab->s_mem + cache->buffer_size * idx;
630}
631
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800632/*
633 * We want to avoid an expensive divide : (offset / cache->buffer_size)
634 * Using the fact that buffer_size is a constant for a particular cache,
635 * we can replace (offset / cache->buffer_size) by
636 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
637 */
638static inline unsigned int obj_to_index(const struct kmem_cache *cache,
639 const struct slab *slab, void *obj)
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800640{
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800641 u32 offset = (obj - slab->s_mem);
642 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800643}
644
Andrew Mortona737b3e2006-03-22 00:08:11 -0800645/*
646 * These are the default caches for kmalloc. Custom caches can have other sizes.
647 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700648struct cache_sizes malloc_sizes[] = {
649#define CACHE(x) { .cs_size = (x) },
650#include <linux/kmalloc_sizes.h>
651 CACHE(ULONG_MAX)
652#undef CACHE
653};
654EXPORT_SYMBOL(malloc_sizes);
655
656/* Must match cache_sizes above. Out of line to keep cache footprint low. */
657struct cache_names {
658 char *name;
659 char *name_dma;
660};
661
662static struct cache_names __initdata cache_names[] = {
663#define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
664#include <linux/kmalloc_sizes.h>
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800665 {NULL,}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700666#undef CACHE
667};
668
669static struct arraycache_init initarray_cache __initdata =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800670 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700671static struct arraycache_init initarray_generic =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800672 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700673
674/* internal cache of cache description objs */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800675static struct kmem_cache cache_cache = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800676 .batchcount = 1,
677 .limit = BOOT_CPUCACHE_ENTRIES,
678 .shared = 1,
Pekka Enberg343e0d72006-02-01 03:05:50 -0800679 .buffer_size = sizeof(struct kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800680 .name = "kmem_cache",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700681#if DEBUG
Pekka Enberg343e0d72006-02-01 03:05:50 -0800682 .obj_size = sizeof(struct kmem_cache),
Linus Torvalds1da177e2005-04-16 15:20:36 -0700683#endif
684};
685
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700686#define BAD_ALIEN_MAGIC 0x01020304ul
687
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200688#ifdef CONFIG_LOCKDEP
689
690/*
691 * Slab sometimes uses the kmalloc slabs to store the slab headers
692 * for other slabs "off slab".
693 * The locking for this is tricky in that it nests within the locks
694 * of all other slabs in a few places; to deal with this special
695 * locking we put on-slab caches into a separate lock-class.
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700696 *
697 * We set lock class for alien array caches which are up during init.
698 * The lock annotation will be lost if all cpus of a node goes down and
699 * then comes back up during hotplug
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200700 */
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700701static struct lock_class_key on_slab_l3_key;
702static struct lock_class_key on_slab_alc_key;
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200703
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700704static inline void init_lock_keys(void)
705
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200706{
707 int q;
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700708 struct cache_sizes *s = malloc_sizes;
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200709
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700710 while (s->cs_size != ULONG_MAX) {
711 for_each_node(q) {
712 struct array_cache **alc;
713 int r;
714 struct kmem_list3 *l3 = s->cs_cachep->nodelists[q];
715 if (!l3 || OFF_SLAB(s->cs_cachep))
716 continue;
717 lockdep_set_class(&l3->list_lock, &on_slab_l3_key);
718 alc = l3->alien;
719 /*
720 * FIXME: This check for BAD_ALIEN_MAGIC
721 * should go away when common slab code is taught to
722 * work even without alien caches.
723 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
724 * for alloc_alien_cache,
725 */
726 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
727 continue;
728 for_each_node(r) {
729 if (alc[r])
730 lockdep_set_class(&alc[r]->lock,
731 &on_slab_alc_key);
732 }
733 }
734 s++;
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200735 }
736}
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200737#else
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700738static inline void init_lock_keys(void)
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200739{
740}
741#endif
742
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -0800743/*
744 * 1. Guard access to the cache-chain.
745 * 2. Protect sanity of cpu_online_map against cpu hotplug events
746 */
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800747static DEFINE_MUTEX(cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700748static struct list_head cache_chain;
749
750/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700751 * chicken and egg problem: delay the per-cpu array allocation
752 * until the general caches are up.
753 */
754static enum {
755 NONE,
Christoph Lametere498be72005-09-09 13:03:32 -0700756 PARTIAL_AC,
757 PARTIAL_L3,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700758 FULL
759} g_cpucache_up;
760
Mike Kravetz39d24e62006-05-15 09:44:13 -0700761/*
762 * used by boot code to determine if it can use slab based allocator
763 */
764int slab_is_available(void)
765{
766 return g_cpucache_up == FULL;
767}
768
David Howells52bad642006-11-22 14:54:01 +0000769static DEFINE_PER_CPU(struct delayed_work, reap_work);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700770
Pekka Enberg343e0d72006-02-01 03:05:50 -0800771static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700772{
773 return cachep->array[smp_processor_id()];
774}
775
Andrew Mortona737b3e2006-03-22 00:08:11 -0800776static inline struct kmem_cache *__find_general_cachep(size_t size,
777 gfp_t gfpflags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700778{
779 struct cache_sizes *csizep = malloc_sizes;
780
781#if DEBUG
782 /* This happens if someone tries to call
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800783 * kmem_cache_create(), or __kmalloc(), before
784 * the generic caches are initialized.
785 */
Alok Katariac7e43c72005-09-14 12:17:53 -0700786 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700787#endif
788 while (size > csizep->cs_size)
789 csizep++;
790
791 /*
Martin Hicks0abf40c2005-09-03 15:54:54 -0700792 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
Linus Torvalds1da177e2005-04-16 15:20:36 -0700793 * has cs_{dma,}cachep==NULL. Thus no special case
794 * for large kmalloc calls required.
795 */
796 if (unlikely(gfpflags & GFP_DMA))
797 return csizep->cs_dmacachep;
798 return csizep->cs_cachep;
799}
800
Adrian Bunkb2213852006-09-25 23:31:02 -0700801static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700802{
803 return __find_general_cachep(size, gfpflags);
804}
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700805
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800806static size_t slab_mgmt_size(size_t nr_objs, size_t align)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700807{
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800808 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
809}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700810
Andrew Mortona737b3e2006-03-22 00:08:11 -0800811/*
812 * Calculate the number of objects and left-over bytes for a given buffer size.
813 */
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800814static void cache_estimate(unsigned long gfporder, size_t buffer_size,
815 size_t align, int flags, size_t *left_over,
816 unsigned int *num)
817{
818 int nr_objs;
819 size_t mgmt_size;
820 size_t slab_size = PAGE_SIZE << gfporder;
821
822 /*
823 * The slab management structure can be either off the slab or
824 * on it. For the latter case, the memory allocated for a
825 * slab is used for:
826 *
827 * - The struct slab
828 * - One kmem_bufctl_t for each object
829 * - Padding to respect alignment of @align
830 * - @buffer_size bytes for each object
831 *
832 * If the slab management structure is off the slab, then the
833 * alignment will already be calculated into the size. Because
834 * the slabs are all pages aligned, the objects will be at the
835 * correct alignment when allocated.
836 */
837 if (flags & CFLGS_OFF_SLAB) {
838 mgmt_size = 0;
839 nr_objs = slab_size / buffer_size;
840
841 if (nr_objs > SLAB_LIMIT)
842 nr_objs = SLAB_LIMIT;
843 } else {
844 /*
845 * Ignore padding for the initial guess. The padding
846 * is at most @align-1 bytes, and @buffer_size is at
847 * least @align. In the worst case, this result will
848 * be one greater than the number of objects that fit
849 * into the memory allocation when taking the padding
850 * into account.
851 */
852 nr_objs = (slab_size - sizeof(struct slab)) /
853 (buffer_size + sizeof(kmem_bufctl_t));
854
855 /*
856 * This calculated number will be either the right
857 * amount, or one greater than what we want.
858 */
859 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
860 > slab_size)
861 nr_objs--;
862
863 if (nr_objs > SLAB_LIMIT)
864 nr_objs = SLAB_LIMIT;
865
866 mgmt_size = slab_mgmt_size(nr_objs, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700867 }
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800868 *num = nr_objs;
869 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700870}
871
872#define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
873
Andrew Mortona737b3e2006-03-22 00:08:11 -0800874static void __slab_error(const char *function, struct kmem_cache *cachep,
875 char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700876{
877 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800878 function, cachep->name, msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700879 dump_stack();
880}
881
Paul Menage3395ee02006-12-06 20:32:16 -0800882/*
883 * By default on NUMA we use alien caches to stage the freeing of
884 * objects allocated from other nodes. This causes massive memory
885 * inefficiencies when using fake NUMA setup to split memory into a
886 * large number of small nodes, so it can be disabled on the command
887 * line
888 */
889
890static int use_alien_caches __read_mostly = 1;
891static int __init noaliencache_setup(char *s)
892{
893 use_alien_caches = 0;
894 return 1;
895}
896__setup("noaliencache", noaliencache_setup);
897
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800898#ifdef CONFIG_NUMA
899/*
900 * Special reaping functions for NUMA systems called from cache_reap().
901 * These take care of doing round robin flushing of alien caches (containing
902 * objects freed on different nodes from which they were allocated) and the
903 * flushing of remote pcps by calling drain_node_pages.
904 */
905static DEFINE_PER_CPU(unsigned long, reap_node);
906
907static void init_reap_node(int cpu)
908{
909 int node;
910
911 node = next_node(cpu_to_node(cpu), node_online_map);
912 if (node == MAX_NUMNODES)
Paul Jackson442295c2006-03-22 00:09:11 -0800913 node = first_node(node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800914
Daniel Yeisley7f6b8872006-11-02 22:07:14 -0800915 per_cpu(reap_node, cpu) = node;
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800916}
917
918static void next_reap_node(void)
919{
920 int node = __get_cpu_var(reap_node);
921
922 /*
923 * Also drain per cpu pages on remote zones
924 */
925 if (node != numa_node_id())
926 drain_node_pages(node);
927
928 node = next_node(node, node_online_map);
929 if (unlikely(node >= MAX_NUMNODES))
930 node = first_node(node_online_map);
931 __get_cpu_var(reap_node) = node;
932}
933
934#else
935#define init_reap_node(cpu) do { } while (0)
936#define next_reap_node(void) do { } while (0)
937#endif
938
Linus Torvalds1da177e2005-04-16 15:20:36 -0700939/*
940 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
941 * via the workqueue/eventd.
942 * Add the CPU number into the expiration time to minimize the possibility of
943 * the CPUs getting into lockstep and contending for the global cache chain
944 * lock.
945 */
946static void __devinit start_cpu_timer(int cpu)
947{
David Howells52bad642006-11-22 14:54:01 +0000948 struct delayed_work *reap_work = &per_cpu(reap_work, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700949
950 /*
951 * When this gets called from do_initcalls via cpucache_init(),
952 * init_workqueues() has already run, so keventd will be setup
953 * at that time.
954 */
David Howells52bad642006-11-22 14:54:01 +0000955 if (keventd_up() && reap_work->work.func == NULL) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800956 init_reap_node(cpu);
David Howells65f27f32006-11-22 14:55:48 +0000957 INIT_DELAYED_WORK(reap_work, cache_reap);
Arjan van de Ven2b284212006-12-10 02:21:28 -0800958 schedule_delayed_work_on(cpu, reap_work,
959 __round_jiffies_relative(HZ, cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700960 }
961}
962
Christoph Lametere498be72005-09-09 13:03:32 -0700963static struct array_cache *alloc_arraycache(int node, int entries,
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800964 int batchcount)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700965{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800966 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700967 struct array_cache *nc = NULL;
968
Christoph Lametere498be72005-09-09 13:03:32 -0700969 nc = kmalloc_node(memsize, GFP_KERNEL, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700970 if (nc) {
971 nc->avail = 0;
972 nc->limit = entries;
973 nc->batchcount = batchcount;
974 nc->touched = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700975 spin_lock_init(&nc->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700976 }
977 return nc;
978}
979
Christoph Lameter3ded1752006-03-25 03:06:44 -0800980/*
981 * Transfer objects in one arraycache to another.
982 * Locking must be handled by the caller.
983 *
984 * Return the number of entries transferred.
985 */
986static int transfer_objects(struct array_cache *to,
987 struct array_cache *from, unsigned int max)
988{
989 /* Figure out how many entries to transfer */
990 int nr = min(min(from->avail, max), to->limit - to->avail);
991
992 if (!nr)
993 return 0;
994
995 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
996 sizeof(void *) *nr);
997
998 from->avail -= nr;
999 to->avail += nr;
1000 to->touched = 1;
1001 return nr;
1002}
1003
Christoph Lameter765c4502006-09-27 01:50:08 -07001004#ifndef CONFIG_NUMA
1005
1006#define drain_alien_cache(cachep, alien) do { } while (0)
1007#define reap_alien(cachep, l3) do { } while (0)
1008
1009static inline struct array_cache **alloc_alien_cache(int node, int limit)
1010{
1011 return (struct array_cache **)BAD_ALIEN_MAGIC;
1012}
1013
1014static inline void free_alien_cache(struct array_cache **ac_ptr)
1015{
1016}
1017
1018static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1019{
1020 return 0;
1021}
1022
1023static inline void *alternate_node_alloc(struct kmem_cache *cachep,
1024 gfp_t flags)
1025{
1026 return NULL;
1027}
1028
Christoph Hellwig8b98c162006-12-06 20:32:30 -08001029static inline void *____cache_alloc_node(struct kmem_cache *cachep,
Christoph Lameter765c4502006-09-27 01:50:08 -07001030 gfp_t flags, int nodeid)
1031{
1032 return NULL;
1033}
1034
1035#else /* CONFIG_NUMA */
1036
Christoph Hellwig8b98c162006-12-06 20:32:30 -08001037static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
Paul Jacksonc61afb12006-03-24 03:16:08 -08001038static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
Christoph Lameterdc85da12006-01-18 17:42:36 -08001039
Pekka Enberg5295a742006-02-01 03:05:48 -08001040static struct array_cache **alloc_alien_cache(int node, int limit)
Christoph Lametere498be72005-09-09 13:03:32 -07001041{
1042 struct array_cache **ac_ptr;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001043 int memsize = sizeof(void *) * MAX_NUMNODES;
Christoph Lametere498be72005-09-09 13:03:32 -07001044 int i;
1045
1046 if (limit > 1)
1047 limit = 12;
1048 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
1049 if (ac_ptr) {
1050 for_each_node(i) {
1051 if (i == node || !node_online(i)) {
1052 ac_ptr[i] = NULL;
1053 continue;
1054 }
1055 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
1056 if (!ac_ptr[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001057 for (i--; i <= 0; i--)
Christoph Lametere498be72005-09-09 13:03:32 -07001058 kfree(ac_ptr[i]);
1059 kfree(ac_ptr);
1060 return NULL;
1061 }
1062 }
1063 }
1064 return ac_ptr;
1065}
1066
Pekka Enberg5295a742006-02-01 03:05:48 -08001067static void free_alien_cache(struct array_cache **ac_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -07001068{
1069 int i;
1070
1071 if (!ac_ptr)
1072 return;
Christoph Lametere498be72005-09-09 13:03:32 -07001073 for_each_node(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001074 kfree(ac_ptr[i]);
Christoph Lametere498be72005-09-09 13:03:32 -07001075 kfree(ac_ptr);
1076}
1077
Pekka Enberg343e0d72006-02-01 03:05:50 -08001078static void __drain_alien_cache(struct kmem_cache *cachep,
Pekka Enberg5295a742006-02-01 03:05:48 -08001079 struct array_cache *ac, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07001080{
1081 struct kmem_list3 *rl3 = cachep->nodelists[node];
1082
1083 if (ac->avail) {
1084 spin_lock(&rl3->list_lock);
Christoph Lametere00946f2006-03-25 03:06:45 -08001085 /*
1086 * Stuff objects into the remote nodes shared array first.
1087 * That way we could avoid the overhead of putting the objects
1088 * into the free lists and getting them back later.
1089 */
shin, jacob693f7d32006-04-28 10:54:37 -05001090 if (rl3->shared)
1091 transfer_objects(rl3->shared, ac, ac->limit);
Christoph Lametere00946f2006-03-25 03:06:45 -08001092
Christoph Lameterff694162005-09-22 21:44:02 -07001093 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001094 ac->avail = 0;
1095 spin_unlock(&rl3->list_lock);
1096 }
1097}
1098
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001099/*
1100 * Called from cache_reap() to regularly drain alien caches round robin.
1101 */
1102static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1103{
1104 int node = __get_cpu_var(reap_node);
1105
1106 if (l3->alien) {
1107 struct array_cache *ac = l3->alien[node];
Christoph Lametere00946f2006-03-25 03:06:45 -08001108
1109 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001110 __drain_alien_cache(cachep, ac, node);
1111 spin_unlock_irq(&ac->lock);
1112 }
1113 }
1114}
1115
Andrew Mortona737b3e2006-03-22 00:08:11 -08001116static void drain_alien_cache(struct kmem_cache *cachep,
1117 struct array_cache **alien)
Christoph Lametere498be72005-09-09 13:03:32 -07001118{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001119 int i = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07001120 struct array_cache *ac;
1121 unsigned long flags;
1122
1123 for_each_online_node(i) {
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001124 ac = alien[i];
Christoph Lametere498be72005-09-09 13:03:32 -07001125 if (ac) {
1126 spin_lock_irqsave(&ac->lock, flags);
1127 __drain_alien_cache(cachep, ac, i);
1128 spin_unlock_irqrestore(&ac->lock, flags);
1129 }
1130 }
1131}
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001132
Ingo Molnar873623d2006-07-13 14:44:38 +02001133static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001134{
1135 struct slab *slabp = virt_to_slab(objp);
1136 int nodeid = slabp->nodeid;
1137 struct kmem_list3 *l3;
1138 struct array_cache *alien = NULL;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001139 int node;
1140
1141 node = numa_node_id();
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001142
1143 /*
1144 * Make sure we are not freeing a object from another node to the array
1145 * cache on this cpu.
1146 */
Paul Menage3395ee02006-12-06 20:32:16 -08001147 if (likely(slabp->nodeid == node) || unlikely(!use_alien_caches))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001148 return 0;
1149
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001150 l3 = cachep->nodelists[node];
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001151 STATS_INC_NODEFREES(cachep);
1152 if (l3->alien && l3->alien[nodeid]) {
1153 alien = l3->alien[nodeid];
Ingo Molnar873623d2006-07-13 14:44:38 +02001154 spin_lock(&alien->lock);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001155 if (unlikely(alien->avail == alien->limit)) {
1156 STATS_INC_ACOVERFLOW(cachep);
1157 __drain_alien_cache(cachep, alien, nodeid);
1158 }
1159 alien->entry[alien->avail++] = objp;
1160 spin_unlock(&alien->lock);
1161 } else {
1162 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1163 free_block(cachep, &objp, 1, nodeid);
1164 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1165 }
1166 return 1;
1167}
Christoph Lametere498be72005-09-09 13:03:32 -07001168#endif
1169
Chandra Seetharaman8c78f302006-07-30 03:03:35 -07001170static int __cpuinit cpuup_callback(struct notifier_block *nfb,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001171 unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001172{
1173 long cpu = (long)hcpu;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001174 struct kmem_cache *cachep;
Christoph Lametere498be72005-09-09 13:03:32 -07001175 struct kmem_list3 *l3 = NULL;
1176 int node = cpu_to_node(cpu);
1177 int memsize = sizeof(struct kmem_list3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001178
1179 switch (action) {
1180 case CPU_UP_PREPARE:
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001181 mutex_lock(&cache_chain_mutex);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001182 /*
1183 * We need to do this right in the beginning since
Christoph Lametere498be72005-09-09 13:03:32 -07001184 * alloc_arraycache's are going to use this list.
1185 * kmalloc_node allows us to add the slab to the right
1186 * kmem_list3 and not this cpu's kmem_list3
1187 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001188
Christoph Lametere498be72005-09-09 13:03:32 -07001189 list_for_each_entry(cachep, &cache_chain, next) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08001190 /*
1191 * Set up the size64 kmemlist for cpu before we can
Christoph Lametere498be72005-09-09 13:03:32 -07001192 * begin anything. Make sure some other cpu on this
1193 * node has not already allocated this
1194 */
1195 if (!cachep->nodelists[node]) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08001196 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1197 if (!l3)
Christoph Lametere498be72005-09-09 13:03:32 -07001198 goto bad;
1199 kmem_list3_init(l3);
1200 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001201 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -07001202
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001203 /*
1204 * The l3s don't come and go as CPUs come and
1205 * go. cache_chain_mutex is sufficient
1206 * protection here.
1207 */
Christoph Lametere498be72005-09-09 13:03:32 -07001208 cachep->nodelists[node] = l3;
1209 }
1210
1211 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1212 cachep->nodelists[node]->free_limit =
Andrew Mortona737b3e2006-03-22 00:08:11 -08001213 (1 + nr_cpus_node(node)) *
1214 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07001215 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1216 }
1217
Andrew Mortona737b3e2006-03-22 00:08:11 -08001218 /*
1219 * Now we can go ahead with allocating the shared arrays and
1220 * array caches
1221 */
Christoph Lametere498be72005-09-09 13:03:32 -07001222 list_for_each_entry(cachep, &cache_chain, next) {
Tobias Klausercd105df2006-01-08 01:00:59 -08001223 struct array_cache *nc;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001224 struct array_cache *shared;
Paul Menage3395ee02006-12-06 20:32:16 -08001225 struct array_cache **alien = NULL;
Tobias Klausercd105df2006-01-08 01:00:59 -08001226
Christoph Lametere498be72005-09-09 13:03:32 -07001227 nc = alloc_arraycache(node, cachep->limit,
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001228 cachep->batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001229 if (!nc)
1230 goto bad;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001231 shared = alloc_arraycache(node,
1232 cachep->shared * cachep->batchcount,
1233 0xbaadf00d);
1234 if (!shared)
1235 goto bad;
Linus Torvalds7a21ef62006-02-05 11:26:38 -08001236
Paul Menage3395ee02006-12-06 20:32:16 -08001237 if (use_alien_caches) {
1238 alien = alloc_alien_cache(node, cachep->limit);
1239 if (!alien)
1240 goto bad;
1241 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001242 cachep->array[cpu] = nc;
Christoph Lametere498be72005-09-09 13:03:32 -07001243 l3 = cachep->nodelists[node];
1244 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07001245
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001246 spin_lock_irq(&l3->list_lock);
1247 if (!l3->shared) {
1248 /*
1249 * We are serialised from CPU_DEAD or
1250 * CPU_UP_CANCELLED by the cpucontrol lock
1251 */
1252 l3->shared = shared;
1253 shared = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07001254 }
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001255#ifdef CONFIG_NUMA
1256 if (!l3->alien) {
1257 l3->alien = alien;
1258 alien = NULL;
1259 }
1260#endif
1261 spin_unlock_irq(&l3->list_lock);
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001262 kfree(shared);
1263 free_alien_cache(alien);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001264 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001265 break;
1266 case CPU_ONLINE:
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08001267 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001268 start_cpu_timer(cpu);
1269 break;
1270#ifdef CONFIG_HOTPLUG_CPU
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08001271 case CPU_DOWN_PREPARE:
1272 mutex_lock(&cache_chain_mutex);
1273 break;
1274 case CPU_DOWN_FAILED:
1275 mutex_unlock(&cache_chain_mutex);
1276 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001277 case CPU_DEAD:
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001278 /*
1279 * Even if all the cpus of a node are down, we don't free the
1280 * kmem_list3 of any cache. This to avoid a race between
1281 * cpu_down, and a kmalloc allocation from another cpu for
1282 * memory from the node of the cpu going down. The list3
1283 * structure is usually allocated from kmem_cache_create() and
1284 * gets destroyed at kmem_cache_destroy().
1285 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001286 /* fall thru */
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08001287#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001288 case CPU_UP_CANCELED:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001289 list_for_each_entry(cachep, &cache_chain, next) {
1290 struct array_cache *nc;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001291 struct array_cache *shared;
1292 struct array_cache **alien;
Christoph Lametere498be72005-09-09 13:03:32 -07001293 cpumask_t mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001294
Christoph Lametere498be72005-09-09 13:03:32 -07001295 mask = node_to_cpumask(node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001296 /* cpu is dead; no one can alloc from it. */
1297 nc = cachep->array[cpu];
1298 cachep->array[cpu] = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07001299 l3 = cachep->nodelists[node];
1300
1301 if (!l3)
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001302 goto free_array_cache;
Christoph Lametere498be72005-09-09 13:03:32 -07001303
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08001304 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07001305
1306 /* Free limit for this kmem_list3 */
1307 l3->free_limit -= cachep->batchcount;
1308 if (nc)
Christoph Lameterff694162005-09-22 21:44:02 -07001309 free_block(cachep, nc->entry, nc->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001310
1311 if (!cpus_empty(mask)) {
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08001312 spin_unlock_irq(&l3->list_lock);
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001313 goto free_array_cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001314 }
Christoph Lametere498be72005-09-09 13:03:32 -07001315
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001316 shared = l3->shared;
1317 if (shared) {
Christoph Lametere498be72005-09-09 13:03:32 -07001318 free_block(cachep, l3->shared->entry,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001319 l3->shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001320 l3->shared = NULL;
1321 }
Christoph Lametere498be72005-09-09 13:03:32 -07001322
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001323 alien = l3->alien;
1324 l3->alien = NULL;
1325
1326 spin_unlock_irq(&l3->list_lock);
1327
1328 kfree(shared);
1329 if (alien) {
1330 drain_alien_cache(cachep, alien);
1331 free_alien_cache(alien);
Christoph Lametere498be72005-09-09 13:03:32 -07001332 }
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001333free_array_cache:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001334 kfree(nc);
1335 }
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001336 /*
1337 * In the previous loop, all the objects were freed to
1338 * the respective cache's slabs, now we can go ahead and
1339 * shrink each nodelist to its limit.
1340 */
1341 list_for_each_entry(cachep, &cache_chain, next) {
1342 l3 = cachep->nodelists[node];
1343 if (!l3)
1344 continue;
Christoph Lametered11d9e2006-06-30 01:55:45 -07001345 drain_freelist(cachep, l3, l3->free_objects);
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001346 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001347 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001348 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001349 }
1350 return NOTIFY_OK;
Andrew Mortona737b3e2006-03-22 00:08:11 -08001351bad:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001352 return NOTIFY_BAD;
1353}
1354
Chandra Seetharaman74b85f32006-06-27 02:54:09 -07001355static struct notifier_block __cpuinitdata cpucache_notifier = {
1356 &cpuup_callback, NULL, 0
1357};
Linus Torvalds1da177e2005-04-16 15:20:36 -07001358
Christoph Lametere498be72005-09-09 13:03:32 -07001359/*
1360 * swap the static kmem_list3 with kmalloced memory
1361 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001362static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1363 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001364{
1365 struct kmem_list3 *ptr;
1366
Christoph Lametere498be72005-09-09 13:03:32 -07001367 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1368 BUG_ON(!ptr);
1369
1370 local_irq_disable();
1371 memcpy(ptr, list, sizeof(struct kmem_list3));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001372 /*
1373 * Do not assume that spinlocks can be initialized via memcpy:
1374 */
1375 spin_lock_init(&ptr->list_lock);
1376
Christoph Lametere498be72005-09-09 13:03:32 -07001377 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1378 cachep->nodelists[nodeid] = ptr;
1379 local_irq_enable();
1380}
1381
Andrew Mortona737b3e2006-03-22 00:08:11 -08001382/*
1383 * Initialisation. Called after the page allocator have been initialised and
1384 * before smp_init().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001385 */
1386void __init kmem_cache_init(void)
1387{
1388 size_t left_over;
1389 struct cache_sizes *sizes;
1390 struct cache_names *names;
Christoph Lametere498be72005-09-09 13:03:32 -07001391 int i;
Jack Steiner07ed76b2006-03-07 21:55:46 -08001392 int order;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001393 int node;
Christoph Lametere498be72005-09-09 13:03:32 -07001394
1395 for (i = 0; i < NUM_INIT_LISTS; i++) {
1396 kmem_list3_init(&initkmem_list3[i]);
1397 if (i < MAX_NUMNODES)
1398 cache_cache.nodelists[i] = NULL;
1399 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001400
1401 /*
1402 * Fragmentation resistance on low memory - only use bigger
1403 * page orders on machines with more than 32MB of memory.
1404 */
1405 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1406 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1407
Linus Torvalds1da177e2005-04-16 15:20:36 -07001408 /* Bootstrap is tricky, because several objects are allocated
1409 * from caches that do not exist yet:
Andrew Mortona737b3e2006-03-22 00:08:11 -08001410 * 1) initialize the cache_cache cache: it contains the struct
1411 * kmem_cache structures of all caches, except cache_cache itself:
1412 * cache_cache is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001413 * Initially an __init data area is used for the head array and the
1414 * kmem_list3 structures, it's replaced with a kmalloc allocated
1415 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001416 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001417 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001418 * An __init data area is used for the head array.
1419 * 3) Create the remaining kmalloc caches, with minimally sized
1420 * head arrays.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001421 * 4) Replace the __init data head arrays for cache_cache and the first
1422 * kmalloc cache with kmalloc allocated arrays.
Christoph Lametere498be72005-09-09 13:03:32 -07001423 * 5) Replace the __init data for kmem_list3 for cache_cache and
1424 * the other cache's with kmalloc allocated memory.
1425 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001426 */
1427
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001428 node = numa_node_id();
1429
Linus Torvalds1da177e2005-04-16 15:20:36 -07001430 /* 1) create the cache_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001431 INIT_LIST_HEAD(&cache_chain);
1432 list_add(&cache_cache.next, &cache_chain);
1433 cache_cache.colour_off = cache_line_size();
1434 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001435 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001436
Andrew Mortona737b3e2006-03-22 00:08:11 -08001437 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1438 cache_line_size());
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08001439 cache_cache.reciprocal_buffer_size =
1440 reciprocal_value(cache_cache.buffer_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001441
Jack Steiner07ed76b2006-03-07 21:55:46 -08001442 for (order = 0; order < MAX_ORDER; order++) {
1443 cache_estimate(order, cache_cache.buffer_size,
1444 cache_line_size(), 0, &left_over, &cache_cache.num);
1445 if (cache_cache.num)
1446 break;
1447 }
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02001448 BUG_ON(!cache_cache.num);
Jack Steiner07ed76b2006-03-07 21:55:46 -08001449 cache_cache.gfporder = order;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001450 cache_cache.colour = left_over / cache_cache.colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001451 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1452 sizeof(struct slab), cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001453
1454 /* 2+3) create the kmalloc caches */
1455 sizes = malloc_sizes;
1456 names = cache_names;
1457
Andrew Mortona737b3e2006-03-22 00:08:11 -08001458 /*
1459 * Initialize the caches that provide memory for the array cache and the
1460 * kmem_list3 structures first. Without this, further allocations will
1461 * bug.
Christoph Lametere498be72005-09-09 13:03:32 -07001462 */
1463
1464 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001465 sizes[INDEX_AC].cs_size,
1466 ARCH_KMALLOC_MINALIGN,
1467 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1468 NULL, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07001469
Andrew Mortona737b3e2006-03-22 00:08:11 -08001470 if (INDEX_AC != INDEX_L3) {
Christoph Lametere498be72005-09-09 13:03:32 -07001471 sizes[INDEX_L3].cs_cachep =
Andrew Mortona737b3e2006-03-22 00:08:11 -08001472 kmem_cache_create(names[INDEX_L3].name,
1473 sizes[INDEX_L3].cs_size,
1474 ARCH_KMALLOC_MINALIGN,
1475 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1476 NULL, NULL);
1477 }
Christoph Lametere498be72005-09-09 13:03:32 -07001478
Ingo Molnare0a42722006-06-23 02:03:46 -07001479 slab_early_init = 0;
1480
Linus Torvalds1da177e2005-04-16 15:20:36 -07001481 while (sizes->cs_size != ULONG_MAX) {
Christoph Lametere498be72005-09-09 13:03:32 -07001482 /*
1483 * For performance, all the general caches are L1 aligned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001484 * This should be particularly beneficial on SMP boxes, as it
1485 * eliminates "false sharing".
1486 * Note for systems short on memory removing the alignment will
Christoph Lametere498be72005-09-09 13:03:32 -07001487 * allow tighter packing of the smaller caches.
1488 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001489 if (!sizes->cs_cachep) {
Christoph Lametere498be72005-09-09 13:03:32 -07001490 sizes->cs_cachep = kmem_cache_create(names->name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001491 sizes->cs_size,
1492 ARCH_KMALLOC_MINALIGN,
1493 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1494 NULL, NULL);
1495 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001496
Linus Torvalds1da177e2005-04-16 15:20:36 -07001497 sizes->cs_dmacachep = kmem_cache_create(names->name_dma,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001498 sizes->cs_size,
1499 ARCH_KMALLOC_MINALIGN,
1500 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1501 SLAB_PANIC,
1502 NULL, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001503 sizes++;
1504 names++;
1505 }
1506 /* 4) Replace the bootstrap head arrays */
1507 {
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001508 struct array_cache *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001509
Linus Torvalds1da177e2005-04-16 15:20:36 -07001510 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001511
Linus Torvalds1da177e2005-04-16 15:20:36 -07001512 local_irq_disable();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001513 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1514 memcpy(ptr, cpu_cache_get(&cache_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001515 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001516 /*
1517 * Do not assume that spinlocks can be initialized via memcpy:
1518 */
1519 spin_lock_init(&ptr->lock);
1520
Linus Torvalds1da177e2005-04-16 15:20:36 -07001521 cache_cache.array[smp_processor_id()] = ptr;
1522 local_irq_enable();
Christoph Lametere498be72005-09-09 13:03:32 -07001523
Linus Torvalds1da177e2005-04-16 15:20:36 -07001524 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001525
Linus Torvalds1da177e2005-04-16 15:20:36 -07001526 local_irq_disable();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001527 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001528 != &initarray_generic.cache);
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001529 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001530 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001531 /*
1532 * Do not assume that spinlocks can be initialized via memcpy:
1533 */
1534 spin_lock_init(&ptr->lock);
1535
Christoph Lametere498be72005-09-09 13:03:32 -07001536 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001537 ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001538 local_irq_enable();
1539 }
Christoph Lametere498be72005-09-09 13:03:32 -07001540 /* 5) Replace the bootstrap kmem_list3's */
1541 {
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001542 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001543
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001544 /* Replace the static kmem_list3 structures for the boot cpu */
1545 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE], node);
1546
1547 for_each_online_node(nid) {
Christoph Lametere498be72005-09-09 13:03:32 -07001548 init_list(malloc_sizes[INDEX_AC].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001549 &initkmem_list3[SIZE_AC + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001550
1551 if (INDEX_AC != INDEX_L3) {
1552 init_list(malloc_sizes[INDEX_L3].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001553 &initkmem_list3[SIZE_L3 + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001554 }
1555 }
1556 }
1557
1558 /* 6) resize the head arrays to their final sizes */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001559 {
Pekka Enberg343e0d72006-02-01 03:05:50 -08001560 struct kmem_cache *cachep;
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001561 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001562 list_for_each_entry(cachep, &cache_chain, next)
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07001563 if (enable_cpucache(cachep))
1564 BUG();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001565 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001566 }
1567
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001568 /* Annotate slab for lockdep -- annotate the malloc caches */
1569 init_lock_keys();
1570
1571
Linus Torvalds1da177e2005-04-16 15:20:36 -07001572 /* Done! */
1573 g_cpucache_up = FULL;
1574
Andrew Mortona737b3e2006-03-22 00:08:11 -08001575 /*
1576 * Register a cpu startup notifier callback that initializes
1577 * cpu_cache_get for all new cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07001578 */
1579 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001580
Andrew Mortona737b3e2006-03-22 00:08:11 -08001581 /*
1582 * The reap timers are started later, with a module init call: That part
1583 * of the kernel is not yet operational.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001584 */
1585}
1586
1587static int __init cpucache_init(void)
1588{
1589 int cpu;
1590
Andrew Mortona737b3e2006-03-22 00:08:11 -08001591 /*
1592 * Register the timers that return unneeded pages to the page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -07001593 */
Christoph Lametere498be72005-09-09 13:03:32 -07001594 for_each_online_cpu(cpu)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001595 start_cpu_timer(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001596 return 0;
1597}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001598__initcall(cpucache_init);
1599
1600/*
1601 * Interface to system's page allocator. No need to hold the cache-lock.
1602 *
1603 * If we requested dmaable memory, we will get it. Even if we
1604 * did not request dmaable memory, we might get it, but that
1605 * would be relatively rare and ignorable.
1606 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001607static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001608{
1609 struct page *page;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001610 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001611 int i;
1612
Luke Yangd6fef9d2006-04-10 22:52:56 -07001613#ifndef CONFIG_MMU
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001614 /*
1615 * Nommu uses slab's for process anonymous memory allocations, and thus
1616 * requires __GFP_COMP to properly refcount higher order allocations
Luke Yangd6fef9d2006-04-10 22:52:56 -07001617 */
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001618 flags |= __GFP_COMP;
Luke Yangd6fef9d2006-04-10 22:52:56 -07001619#endif
Christoph Lameter765c4502006-09-27 01:50:08 -07001620
Christoph Lameter3c517a62006-12-06 20:33:29 -08001621 flags |= cachep->gfpflags;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001622
1623 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001624 if (!page)
1625 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001626
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001627 nr_pages = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001628 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
Christoph Lameter972d1a72006-09-25 23:31:51 -07001629 add_zone_page_state(page_zone(page),
1630 NR_SLAB_RECLAIMABLE, nr_pages);
1631 else
1632 add_zone_page_state(page_zone(page),
1633 NR_SLAB_UNRECLAIMABLE, nr_pages);
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001634 for (i = 0; i < nr_pages; i++)
1635 __SetPageSlab(page + i);
1636 return page_address(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001637}
1638
1639/*
1640 * Interface to system's page release.
1641 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001642static void kmem_freepages(struct kmem_cache *cachep, void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001643{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001644 unsigned long i = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001645 struct page *page = virt_to_page(addr);
1646 const unsigned long nr_freed = i;
1647
Christoph Lameter972d1a72006-09-25 23:31:51 -07001648 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1649 sub_zone_page_state(page_zone(page),
1650 NR_SLAB_RECLAIMABLE, nr_freed);
1651 else
1652 sub_zone_page_state(page_zone(page),
1653 NR_SLAB_UNRECLAIMABLE, nr_freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001654 while (i--) {
Nick Pigginf205b2f2006-03-22 00:08:02 -08001655 BUG_ON(!PageSlab(page));
1656 __ClearPageSlab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001657 page++;
1658 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001659 if (current->reclaim_state)
1660 current->reclaim_state->reclaimed_slab += nr_freed;
1661 free_pages((unsigned long)addr, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001662}
1663
1664static void kmem_rcu_free(struct rcu_head *head)
1665{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001666 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001667 struct kmem_cache *cachep = slab_rcu->cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001668
1669 kmem_freepages(cachep, slab_rcu->addr);
1670 if (OFF_SLAB(cachep))
1671 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1672}
1673
1674#if DEBUG
1675
1676#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001677static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001678 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001679{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001680 int size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001681
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001682 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001683
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001684 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001685 return;
1686
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001687 *addr++ = 0x12345678;
1688 *addr++ = caller;
1689 *addr++ = smp_processor_id();
1690 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001691 {
1692 unsigned long *sptr = &caller;
1693 unsigned long svalue;
1694
1695 while (!kstack_end(sptr)) {
1696 svalue = *sptr++;
1697 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001698 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001699 size -= sizeof(unsigned long);
1700 if (size <= sizeof(unsigned long))
1701 break;
1702 }
1703 }
1704
1705 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001706 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001707}
1708#endif
1709
Pekka Enberg343e0d72006-02-01 03:05:50 -08001710static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001711{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001712 int size = obj_size(cachep);
1713 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001714
1715 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001716 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001717}
1718
1719static void dump_line(char *data, int offset, int limit)
1720{
1721 int i;
Dave Jonesaa83aa42006-09-29 01:59:51 -07001722 unsigned char error = 0;
1723 int bad_count = 0;
1724
Linus Torvalds1da177e2005-04-16 15:20:36 -07001725 printk(KERN_ERR "%03x:", offset);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001726 for (i = 0; i < limit; i++) {
1727 if (data[offset + i] != POISON_FREE) {
1728 error = data[offset + i];
1729 bad_count++;
1730 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001731 printk(" %02x", (unsigned char)data[offset + i]);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001732 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001733 printk("\n");
Dave Jonesaa83aa42006-09-29 01:59:51 -07001734
1735 if (bad_count == 1) {
1736 error ^= POISON_FREE;
1737 if (!(error & (error - 1))) {
1738 printk(KERN_ERR "Single bit error detected. Probably "
1739 "bad RAM.\n");
1740#ifdef CONFIG_X86
1741 printk(KERN_ERR "Run memtest86+ or a similar memory "
1742 "test tool.\n");
1743#else
1744 printk(KERN_ERR "Run a memory test tool.\n");
1745#endif
1746 }
1747 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001748}
1749#endif
1750
1751#if DEBUG
1752
Pekka Enberg343e0d72006-02-01 03:05:50 -08001753static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001754{
1755 int i, size;
1756 char *realobj;
1757
1758 if (cachep->flags & SLAB_RED_ZONE) {
1759 printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001760 *dbg_redzone1(cachep, objp),
1761 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001762 }
1763
1764 if (cachep->flags & SLAB_STORE_USER) {
1765 printk(KERN_ERR "Last user: [<%p>]",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001766 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001767 print_symbol("(%s)",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001768 (unsigned long)*dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001769 printk("\n");
1770 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001771 realobj = (char *)objp + obj_offset(cachep);
1772 size = obj_size(cachep);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001773 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001774 int limit;
1775 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001776 if (i + limit > size)
1777 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001778 dump_line(realobj, i, limit);
1779 }
1780}
1781
Pekka Enberg343e0d72006-02-01 03:05:50 -08001782static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001783{
1784 char *realobj;
1785 int size, i;
1786 int lines = 0;
1787
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001788 realobj = (char *)objp + obj_offset(cachep);
1789 size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001790
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001791 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001792 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001793 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001794 exp = POISON_END;
1795 if (realobj[i] != exp) {
1796 int limit;
1797 /* Mismatch ! */
1798 /* Print header */
1799 if (lines == 0) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001800 printk(KERN_ERR
Andrew Mortona737b3e2006-03-22 00:08:11 -08001801 "Slab corruption: start=%p, len=%d\n",
1802 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001803 print_objinfo(cachep, objp, 0);
1804 }
1805 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001806 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001807 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001808 if (i + limit > size)
1809 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001810 dump_line(realobj, i, limit);
1811 i += 16;
1812 lines++;
1813 /* Limit to 5 lines */
1814 if (lines > 5)
1815 break;
1816 }
1817 }
1818 if (lines != 0) {
1819 /* Print some data about the neighboring objects, if they
1820 * exist:
1821 */
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08001822 struct slab *slabp = virt_to_slab(objp);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001823 unsigned int objnr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001824
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001825 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001826 if (objnr) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001827 objp = index_to_obj(cachep, slabp, objnr - 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001828 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001829 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001830 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001831 print_objinfo(cachep, objp, 2);
1832 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001833 if (objnr + 1 < cachep->num) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001834 objp = index_to_obj(cachep, slabp, objnr + 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001835 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001836 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001837 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001838 print_objinfo(cachep, objp, 2);
1839 }
1840 }
1841}
1842#endif
1843
Linus Torvalds1da177e2005-04-16 15:20:36 -07001844#if DEBUG
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001845/**
Randy Dunlap911851e2006-03-22 00:08:14 -08001846 * slab_destroy_objs - destroy a slab and its objects
1847 * @cachep: cache pointer being destroyed
1848 * @slabp: slab pointer being destroyed
1849 *
1850 * Call the registered destructor for each object in a slab that is being
1851 * destroyed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001852 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001853static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001854{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001855 int i;
1856 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001857 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001858
1859 if (cachep->flags & SLAB_POISON) {
1860#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08001861 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1862 OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001863 kernel_map_pages(virt_to_page(objp),
Andrew Mortona737b3e2006-03-22 00:08:11 -08001864 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001865 else
1866 check_poison_obj(cachep, objp);
1867#else
1868 check_poison_obj(cachep, objp);
1869#endif
1870 }
1871 if (cachep->flags & SLAB_RED_ZONE) {
1872 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1873 slab_error(cachep, "start of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001874 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001875 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1876 slab_error(cachep, "end of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001877 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001878 }
1879 if (cachep->dtor && !(cachep->flags & SLAB_POISON))
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001880 (cachep->dtor) (objp + obj_offset(cachep), cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001881 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001882}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001883#else
Pekka Enberg343e0d72006-02-01 03:05:50 -08001884static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001885{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001886 if (cachep->dtor) {
1887 int i;
1888 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001889 void *objp = index_to_obj(cachep, slabp, i);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001890 (cachep->dtor) (objp, cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001891 }
1892 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001893}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001894#endif
1895
Randy Dunlap911851e2006-03-22 00:08:14 -08001896/**
1897 * slab_destroy - destroy and release all objects in a slab
1898 * @cachep: cache pointer being destroyed
1899 * @slabp: slab pointer being destroyed
1900 *
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001901 * Destroy all the objs in a slab, and release the mem back to the system.
Andrew Mortona737b3e2006-03-22 00:08:11 -08001902 * Before calling the slab must have been unlinked from the cache. The
1903 * cache-lock is not held/needed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001904 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001905static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001906{
1907 void *addr = slabp->s_mem - slabp->colouroff;
1908
1909 slab_destroy_objs(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001910 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1911 struct slab_rcu *slab_rcu;
1912
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001913 slab_rcu = (struct slab_rcu *)slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001914 slab_rcu->cachep = cachep;
1915 slab_rcu->addr = addr;
1916 call_rcu(&slab_rcu->head, kmem_rcu_free);
1917 } else {
1918 kmem_freepages(cachep, addr);
Ingo Molnar873623d2006-07-13 14:44:38 +02001919 if (OFF_SLAB(cachep))
1920 kmem_cache_free(cachep->slabp_cache, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001921 }
1922}
1923
Andrew Mortona737b3e2006-03-22 00:08:11 -08001924/*
1925 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1926 * size of kmem_list3.
1927 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001928static void set_up_list3s(struct kmem_cache *cachep, int index)
Christoph Lametere498be72005-09-09 13:03:32 -07001929{
1930 int node;
1931
1932 for_each_online_node(node) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001933 cachep->nodelists[node] = &initkmem_list3[index + node];
Christoph Lametere498be72005-09-09 13:03:32 -07001934 cachep->nodelists[node]->next_reap = jiffies +
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001935 REAPTIMEOUT_LIST3 +
1936 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -07001937 }
1938}
1939
Christoph Lameter117f6eb2006-09-25 23:31:37 -07001940static void __kmem_cache_destroy(struct kmem_cache *cachep)
1941{
1942 int i;
1943 struct kmem_list3 *l3;
1944
1945 for_each_online_cpu(i)
1946 kfree(cachep->array[i]);
1947
1948 /* NUMA: free the list3 structures */
1949 for_each_online_node(i) {
1950 l3 = cachep->nodelists[i];
1951 if (l3) {
1952 kfree(l3->shared);
1953 free_alien_cache(l3->alien);
1954 kfree(l3);
1955 }
1956 }
1957 kmem_cache_free(&cache_cache, cachep);
1958}
1959
1960
Linus Torvalds1da177e2005-04-16 15:20:36 -07001961/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08001962 * calculate_slab_order - calculate size (page order) of slabs
1963 * @cachep: pointer to the cache that is being created
1964 * @size: size of objects to be created in this cache.
1965 * @align: required alignment for the objects.
1966 * @flags: slab allocation flags
1967 *
1968 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001969 *
1970 * This could be made much more intelligent. For now, try to avoid using
1971 * high order pages for slabs. When the gfp() functions are more friendly
1972 * towards high-order requests, this should be changed.
1973 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001974static size_t calculate_slab_order(struct kmem_cache *cachep,
Randy Dunlapee13d782006-02-01 03:05:53 -08001975 size_t size, size_t align, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001976{
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001977 unsigned long offslab_limit;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001978 size_t left_over = 0;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001979 int gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001980
Andrew Mortona737b3e2006-03-22 00:08:11 -08001981 for (gfporder = 0; gfporder <= MAX_GFP_ORDER; gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001982 unsigned int num;
1983 size_t remainder;
1984
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001985 cache_estimate(gfporder, size, align, flags, &remainder, &num);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001986 if (!num)
1987 continue;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001988
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001989 if (flags & CFLGS_OFF_SLAB) {
1990 /*
1991 * Max number of objs-per-slab for caches which
1992 * use off-slab slabs. Needed to avoid a possible
1993 * looping condition in cache_grow().
1994 */
1995 offslab_limit = size - sizeof(struct slab);
1996 offslab_limit /= sizeof(kmem_bufctl_t);
1997
1998 if (num > offslab_limit)
1999 break;
2000 }
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002001
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002002 /* Found something acceptable - save it away */
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002003 cachep->num = num;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002004 cachep->gfporder = gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002005 left_over = remainder;
2006
2007 /*
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002008 * A VFS-reclaimable slab tends to have most allocations
2009 * as GFP_NOFS and we really don't want to have to be allocating
2010 * higher-order pages when we are unable to shrink dcache.
2011 */
2012 if (flags & SLAB_RECLAIM_ACCOUNT)
2013 break;
2014
2015 /*
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002016 * Large number of objects is good, but very large slabs are
2017 * currently bad for the gfp()s.
2018 */
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002019 if (gfporder >= slab_break_gfp_order)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002020 break;
2021
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002022 /*
2023 * Acceptable internal fragmentation?
2024 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002025 if (left_over * 8 <= (PAGE_SIZE << gfporder))
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002026 break;
2027 }
2028 return left_over;
2029}
2030
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002031static int setup_cpu_cache(struct kmem_cache *cachep)
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002032{
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002033 if (g_cpucache_up == FULL)
2034 return enable_cpucache(cachep);
2035
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002036 if (g_cpucache_up == NONE) {
2037 /*
2038 * Note: the first kmem_cache_create must create the cache
2039 * that's used by kmalloc(24), otherwise the creation of
2040 * further caches will BUG().
2041 */
2042 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2043
2044 /*
2045 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2046 * the first cache, then we need to set up all its list3s,
2047 * otherwise the creation of further caches will BUG().
2048 */
2049 set_up_list3s(cachep, SIZE_AC);
2050 if (INDEX_AC == INDEX_L3)
2051 g_cpucache_up = PARTIAL_L3;
2052 else
2053 g_cpucache_up = PARTIAL_AC;
2054 } else {
2055 cachep->array[smp_processor_id()] =
2056 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
2057
2058 if (g_cpucache_up == PARTIAL_AC) {
2059 set_up_list3s(cachep, SIZE_L3);
2060 g_cpucache_up = PARTIAL_L3;
2061 } else {
2062 int node;
2063 for_each_online_node(node) {
2064 cachep->nodelists[node] =
2065 kmalloc_node(sizeof(struct kmem_list3),
2066 GFP_KERNEL, node);
2067 BUG_ON(!cachep->nodelists[node]);
2068 kmem_list3_init(cachep->nodelists[node]);
2069 }
2070 }
2071 }
2072 cachep->nodelists[numa_node_id()]->next_reap =
2073 jiffies + REAPTIMEOUT_LIST3 +
2074 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2075
2076 cpu_cache_get(cachep)->avail = 0;
2077 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2078 cpu_cache_get(cachep)->batchcount = 1;
2079 cpu_cache_get(cachep)->touched = 0;
2080 cachep->batchcount = 1;
2081 cachep->limit = BOOT_CPUCACHE_ENTRIES;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002082 return 0;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002083}
2084
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002085/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07002086 * kmem_cache_create - Create a cache.
2087 * @name: A string which is used in /proc/slabinfo to identify this cache.
2088 * @size: The size of objects to be created in this cache.
2089 * @align: The required alignment for the objects.
2090 * @flags: SLAB flags
2091 * @ctor: A constructor for the objects.
2092 * @dtor: A destructor for the objects.
2093 *
2094 * Returns a ptr to the cache on success, NULL on failure.
2095 * Cannot be called within a int, but can be interrupted.
2096 * The @ctor is run when new pages are allocated by the cache
2097 * and the @dtor is run before the pages are handed back.
2098 *
2099 * @name must be valid until the cache is destroyed. This implies that
Andrew Mortona737b3e2006-03-22 00:08:11 -08002100 * the module calling this has to destroy the cache before getting unloaded.
2101 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002102 * The flags are
2103 *
2104 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2105 * to catch references to uninitialised memory.
2106 *
2107 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2108 * for buffer overruns.
2109 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002110 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2111 * cacheline. This can be beneficial if you're counting cycles as closely
2112 * as davem.
2113 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002114struct kmem_cache *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002115kmem_cache_create (const char *name, size_t size, size_t align,
Andrew Mortona737b3e2006-03-22 00:08:11 -08002116 unsigned long flags,
2117 void (*ctor)(void*, struct kmem_cache *, unsigned long),
Pekka Enberg343e0d72006-02-01 03:05:50 -08002118 void (*dtor)(void*, struct kmem_cache *, unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002119{
2120 size_t left_over, slab_size, ralign;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07002121 struct kmem_cache *cachep = NULL, *pc;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002122
2123 /*
2124 * Sanity checks... these are all serious usage bugs.
2125 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002126 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002127 (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002128 printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__,
2129 name);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002130 BUG();
2131 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002132
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002133 /*
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002134 * We use cache_chain_mutex to ensure a consistent view of
2135 * cpu_online_map as well. Please see cpuup_callback
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002136 */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002137 mutex_lock(&cache_chain_mutex);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002138
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07002139 list_for_each_entry(pc, &cache_chain, next) {
Andrew Morton4f12bb42005-11-07 00:58:00 -08002140 char tmp;
2141 int res;
2142
2143 /*
2144 * This happens when the module gets unloaded and doesn't
2145 * destroy its slab cache and no-one else reuses the vmalloc
2146 * area of the module. Print a warning.
2147 */
Andrew Morton138ae662006-12-06 20:36:41 -08002148 res = probe_kernel_address(pc->name, tmp);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002149 if (res) {
2150 printk("SLAB: cache with size %d has lost its name\n",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002151 pc->buffer_size);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002152 continue;
2153 }
2154
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002155 if (!strcmp(pc->name, name)) {
Andrew Morton4f12bb42005-11-07 00:58:00 -08002156 printk("kmem_cache_create: duplicate cache %s\n", name);
2157 dump_stack();
2158 goto oops;
2159 }
2160 }
2161
Linus Torvalds1da177e2005-04-16 15:20:36 -07002162#if DEBUG
2163 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
2164 if ((flags & SLAB_DEBUG_INITIAL) && !ctor) {
2165 /* No constructor, but inital state check requested */
2166 printk(KERN_ERR "%s: No con, but init state check "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002167 "requested - %s\n", __FUNCTION__, name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002168 flags &= ~SLAB_DEBUG_INITIAL;
2169 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002170#if FORCED_DEBUG
2171 /*
2172 * Enable redzoning and last user accounting, except for caches with
2173 * large objects, if the increased size would increase the object size
2174 * above the next power of two: caches with object sizes just above a
2175 * power of two have a significant amount of internal fragmentation.
2176 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002177 if (size < 4096 || fls(size - 1) == fls(size-1 + 3 * BYTES_PER_WORD))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002178 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002179 if (!(flags & SLAB_DESTROY_BY_RCU))
2180 flags |= SLAB_POISON;
2181#endif
2182 if (flags & SLAB_DESTROY_BY_RCU)
2183 BUG_ON(flags & SLAB_POISON);
2184#endif
2185 if (flags & SLAB_DESTROY_BY_RCU)
2186 BUG_ON(dtor);
2187
2188 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002189 * Always checks flags, a caller might be expecting debug support which
2190 * isn't available.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002191 */
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002192 BUG_ON(flags & ~CREATE_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002193
Andrew Mortona737b3e2006-03-22 00:08:11 -08002194 /*
2195 * Check that size is in terms of words. This is needed to avoid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002196 * unaligned accesses for some archs when redzoning is used, and makes
2197 * sure any on-slab bufctl's are also correctly aligned.
2198 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002199 if (size & (BYTES_PER_WORD - 1)) {
2200 size += (BYTES_PER_WORD - 1);
2201 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002202 }
2203
Andrew Mortona737b3e2006-03-22 00:08:11 -08002204 /* calculate the final buffer alignment: */
2205
Linus Torvalds1da177e2005-04-16 15:20:36 -07002206 /* 1) arch recommendation: can be overridden for debug */
2207 if (flags & SLAB_HWCACHE_ALIGN) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002208 /*
2209 * Default alignment: as specified by the arch code. Except if
2210 * an object is really small, then squeeze multiple objects into
2211 * one cacheline.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002212 */
2213 ralign = cache_line_size();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002214 while (size <= ralign / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002215 ralign /= 2;
2216 } else {
2217 ralign = BYTES_PER_WORD;
2218 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002219
2220 /*
2221 * Redzoning and user store require word alignment. Note this will be
2222 * overridden by architecture or caller mandated alignment if either
2223 * is greater than BYTES_PER_WORD.
2224 */
2225 if (flags & SLAB_RED_ZONE || flags & SLAB_STORE_USER)
2226 ralign = BYTES_PER_WORD;
2227
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002228 /* 2) arch mandated alignment */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002229 if (ralign < ARCH_SLAB_MINALIGN) {
2230 ralign = ARCH_SLAB_MINALIGN;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002231 }
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002232 /* 3) caller mandated alignment */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002233 if (ralign < align) {
2234 ralign = align;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002235 }
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002236 /* disable debug if necessary */
2237 if (ralign > BYTES_PER_WORD)
2238 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002239 /*
Pekka Enbergca5f9702006-09-25 23:31:25 -07002240 * 4) Store it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002241 */
2242 align = ralign;
2243
2244 /* Get cache's description obj. */
Christoph Lametere94b1762006-12-06 20:33:17 -08002245 cachep = kmem_cache_zalloc(&cache_cache, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002246 if (!cachep)
Andrew Morton4f12bb42005-11-07 00:58:00 -08002247 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002248
2249#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002250 cachep->obj_size = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002251
Pekka Enbergca5f9702006-09-25 23:31:25 -07002252 /*
2253 * Both debugging options require word-alignment which is calculated
2254 * into align above.
2255 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002256 if (flags & SLAB_RED_ZONE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002257 /* add space for red zone words */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002258 cachep->obj_offset += BYTES_PER_WORD;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002259 size += 2 * BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002260 }
2261 if (flags & SLAB_STORE_USER) {
Pekka Enbergca5f9702006-09-25 23:31:25 -07002262 /* user store requires one word storage behind the end of
2263 * the real object.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002264 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002265 size += BYTES_PER_WORD;
2266 }
2267#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002268 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002269 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2270 cachep->obj_offset += PAGE_SIZE - size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002271 size = PAGE_SIZE;
2272 }
2273#endif
2274#endif
2275
Ingo Molnare0a42722006-06-23 02:03:46 -07002276 /*
2277 * Determine if the slab management is 'on' or 'off' slab.
2278 * (bootstrapping cannot cope with offslab caches so don't do
2279 * it too early on.)
2280 */
2281 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002282 /*
2283 * Size is large, assume best to place the slab management obj
2284 * off-slab (should allow better packing of objs).
2285 */
2286 flags |= CFLGS_OFF_SLAB;
2287
2288 size = ALIGN(size, align);
2289
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002290 left_over = calculate_slab_order(cachep, size, align, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002291
2292 if (!cachep->num) {
2293 printk("kmem_cache_create: couldn't create cache %s.\n", name);
2294 kmem_cache_free(&cache_cache, cachep);
2295 cachep = NULL;
Andrew Morton4f12bb42005-11-07 00:58:00 -08002296 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002297 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002298 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2299 + sizeof(struct slab), align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002300
2301 /*
2302 * If the slab has been placed off-slab, and we have enough space then
2303 * move it on-slab. This is at the expense of any extra colouring.
2304 */
2305 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2306 flags &= ~CFLGS_OFF_SLAB;
2307 left_over -= slab_size;
2308 }
2309
2310 if (flags & CFLGS_OFF_SLAB) {
2311 /* really off slab. No need for manual alignment */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002312 slab_size =
2313 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002314 }
2315
2316 cachep->colour_off = cache_line_size();
2317 /* Offset must be a multiple of the alignment. */
2318 if (cachep->colour_off < align)
2319 cachep->colour_off = align;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002320 cachep->colour = left_over / cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002321 cachep->slab_size = slab_size;
2322 cachep->flags = flags;
2323 cachep->gfpflags = 0;
2324 if (flags & SLAB_CACHE_DMA)
2325 cachep->gfpflags |= GFP_DMA;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002326 cachep->buffer_size = size;
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08002327 cachep->reciprocal_buffer_size = reciprocal_value(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002328
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002329 if (flags & CFLGS_OFF_SLAB) {
Victor Fuscob2d55072005-09-10 00:26:36 -07002330 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002331 /*
2332 * This is a possibility for one of the malloc_sizes caches.
2333 * But since we go off slab only for object size greater than
2334 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2335 * this should not happen at all.
2336 * But leave a BUG_ON for some lucky dude.
2337 */
2338 BUG_ON(!cachep->slabp_cache);
2339 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002340 cachep->ctor = ctor;
2341 cachep->dtor = dtor;
2342 cachep->name = name;
2343
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002344 if (setup_cpu_cache(cachep)) {
2345 __kmem_cache_destroy(cachep);
2346 cachep = NULL;
2347 goto oops;
2348 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002349
Linus Torvalds1da177e2005-04-16 15:20:36 -07002350 /* cache setup completed, link it into the list */
2351 list_add(&cachep->next, &cache_chain);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002352oops:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002353 if (!cachep && (flags & SLAB_PANIC))
2354 panic("kmem_cache_create(): failed to create slab `%s'\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002355 name);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002356 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002357 return cachep;
2358}
2359EXPORT_SYMBOL(kmem_cache_create);
2360
2361#if DEBUG
2362static void check_irq_off(void)
2363{
2364 BUG_ON(!irqs_disabled());
2365}
2366
2367static void check_irq_on(void)
2368{
2369 BUG_ON(irqs_disabled());
2370}
2371
Pekka Enberg343e0d72006-02-01 03:05:50 -08002372static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002373{
2374#ifdef CONFIG_SMP
2375 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002376 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002377#endif
2378}
Christoph Lametere498be72005-09-09 13:03:32 -07002379
Pekka Enberg343e0d72006-02-01 03:05:50 -08002380static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002381{
2382#ifdef CONFIG_SMP
2383 check_irq_off();
2384 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2385#endif
2386}
2387
Linus Torvalds1da177e2005-04-16 15:20:36 -07002388#else
2389#define check_irq_off() do { } while(0)
2390#define check_irq_on() do { } while(0)
2391#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002392#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002393#endif
2394
Christoph Lameteraab22072006-03-22 00:09:06 -08002395static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2396 struct array_cache *ac,
2397 int force, int node);
2398
Linus Torvalds1da177e2005-04-16 15:20:36 -07002399static void do_drain(void *arg)
2400{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002401 struct kmem_cache *cachep = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002402 struct array_cache *ac;
Christoph Lameterff694162005-09-22 21:44:02 -07002403 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002404
2405 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002406 ac = cpu_cache_get(cachep);
Christoph Lameterff694162005-09-22 21:44:02 -07002407 spin_lock(&cachep->nodelists[node]->list_lock);
2408 free_block(cachep, ac->entry, ac->avail, node);
2409 spin_unlock(&cachep->nodelists[node]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002410 ac->avail = 0;
2411}
2412
Pekka Enberg343e0d72006-02-01 03:05:50 -08002413static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002414{
Christoph Lametere498be72005-09-09 13:03:32 -07002415 struct kmem_list3 *l3;
2416 int node;
2417
Andrew Mortona07fa392006-03-22 00:08:17 -08002418 on_each_cpu(do_drain, cachep, 1, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002419 check_irq_on();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002420 for_each_online_node(node) {
Christoph Lametere498be72005-09-09 13:03:32 -07002421 l3 = cachep->nodelists[node];
Roland Dreiera4523a82006-05-15 11:41:00 -07002422 if (l3 && l3->alien)
2423 drain_alien_cache(cachep, l3->alien);
2424 }
2425
2426 for_each_online_node(node) {
2427 l3 = cachep->nodelists[node];
2428 if (l3)
Christoph Lameteraab22072006-03-22 00:09:06 -08002429 drain_array(cachep, l3, l3->shared, 1, node);
Christoph Lametere498be72005-09-09 13:03:32 -07002430 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002431}
2432
Christoph Lametered11d9e2006-06-30 01:55:45 -07002433/*
2434 * Remove slabs from the list of free slabs.
2435 * Specify the number of slabs to drain in tofree.
2436 *
2437 * Returns the actual number of slabs released.
2438 */
2439static int drain_freelist(struct kmem_cache *cache,
2440 struct kmem_list3 *l3, int tofree)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002441{
Christoph Lametered11d9e2006-06-30 01:55:45 -07002442 struct list_head *p;
2443 int nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002444 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002445
Christoph Lametered11d9e2006-06-30 01:55:45 -07002446 nr_freed = 0;
2447 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002448
Christoph Lametered11d9e2006-06-30 01:55:45 -07002449 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002450 p = l3->slabs_free.prev;
Christoph Lametered11d9e2006-06-30 01:55:45 -07002451 if (p == &l3->slabs_free) {
2452 spin_unlock_irq(&l3->list_lock);
2453 goto out;
2454 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002455
Christoph Lametered11d9e2006-06-30 01:55:45 -07002456 slabp = list_entry(p, struct slab, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002457#if DEBUG
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002458 BUG_ON(slabp->inuse);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002459#endif
2460 list_del(&slabp->list);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002461 /*
2462 * Safe to drop the lock. The slab is no longer linked
2463 * to the cache.
2464 */
2465 l3->free_objects -= cache->num;
Christoph Lametere498be72005-09-09 13:03:32 -07002466 spin_unlock_irq(&l3->list_lock);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002467 slab_destroy(cache, slabp);
2468 nr_freed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002469 }
Christoph Lametered11d9e2006-06-30 01:55:45 -07002470out:
2471 return nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002472}
2473
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002474/* Called with cache_chain_mutex held to protect against cpu hotplug */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002475static int __cache_shrink(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07002476{
2477 int ret = 0, i = 0;
2478 struct kmem_list3 *l3;
2479
2480 drain_cpu_caches(cachep);
2481
2482 check_irq_on();
2483 for_each_online_node(i) {
2484 l3 = cachep->nodelists[i];
Christoph Lametered11d9e2006-06-30 01:55:45 -07002485 if (!l3)
2486 continue;
2487
2488 drain_freelist(cachep, l3, l3->free_objects);
2489
2490 ret += !list_empty(&l3->slabs_full) ||
2491 !list_empty(&l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07002492 }
2493 return (ret ? 1 : 0);
2494}
2495
Linus Torvalds1da177e2005-04-16 15:20:36 -07002496/**
2497 * kmem_cache_shrink - Shrink a cache.
2498 * @cachep: The cache to shrink.
2499 *
2500 * Releases as many slabs as possible for a cache.
2501 * To help debugging, a zero exit status indicates all slabs were released.
2502 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002503int kmem_cache_shrink(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002504{
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002505 int ret;
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002506 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002507
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002508 mutex_lock(&cache_chain_mutex);
2509 ret = __cache_shrink(cachep);
2510 mutex_unlock(&cache_chain_mutex);
2511 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002512}
2513EXPORT_SYMBOL(kmem_cache_shrink);
2514
2515/**
2516 * kmem_cache_destroy - delete a cache
2517 * @cachep: the cache to destroy
2518 *
Pekka Enberg343e0d72006-02-01 03:05:50 -08002519 * Remove a struct kmem_cache object from the slab cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002520 *
2521 * It is expected this function will be called by a module when it is
2522 * unloaded. This will remove the cache completely, and avoid a duplicate
2523 * cache being allocated each time a module is loaded and unloaded, if the
2524 * module doesn't have persistent in-kernel storage across loads and unloads.
2525 *
2526 * The cache must be empty before calling this function.
2527 *
2528 * The caller must guarantee that noone will allocate memory from the cache
2529 * during the kmem_cache_destroy().
2530 */
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002531void kmem_cache_destroy(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002532{
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002533 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002534
Linus Torvalds1da177e2005-04-16 15:20:36 -07002535 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002536 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002537 /*
2538 * the chain is never empty, cache_cache is never destroyed
2539 */
2540 list_del(&cachep->next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002541 if (__cache_shrink(cachep)) {
2542 slab_error(cachep, "Can't free all objects");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002543 list_add(&cachep->next, &cache_chain);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002544 mutex_unlock(&cache_chain_mutex);
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002545 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002546 }
2547
2548 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
Paul E. McKenneyfbd568a3e2005-05-01 08:59:04 -07002549 synchronize_rcu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002550
Christoph Lameter117f6eb2006-09-25 23:31:37 -07002551 __kmem_cache_destroy(cachep);
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002552 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002553}
2554EXPORT_SYMBOL(kmem_cache_destroy);
2555
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002556/*
2557 * Get the memory for a slab management obj.
2558 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2559 * always come from malloc_sizes caches. The slab descriptor cannot
2560 * come from the same cache which is getting created because,
2561 * when we are searching for an appropriate cache for these
2562 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2563 * If we are creating a malloc_sizes cache here it would not be visible to
2564 * kmem_find_general_cachep till the initialization is complete.
2565 * Hence we cannot have slabp_cache same as the original cache.
2566 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002567static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002568 int colour_off, gfp_t local_flags,
2569 int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002570{
2571 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002572
Linus Torvalds1da177e2005-04-16 15:20:36 -07002573 if (OFF_SLAB(cachep)) {
2574 /* Slab management obj is off-slab. */
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002575 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
Christoph Lameter3c517a62006-12-06 20:33:29 -08002576 local_flags & ~GFP_THISNODE, nodeid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002577 if (!slabp)
2578 return NULL;
2579 } else {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002580 slabp = objp + colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002581 colour_off += cachep->slab_size;
2582 }
2583 slabp->inuse = 0;
2584 slabp->colouroff = colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002585 slabp->s_mem = objp + colour_off;
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002586 slabp->nodeid = nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002587 return slabp;
2588}
2589
2590static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2591{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002592 return (kmem_bufctl_t *) (slabp + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002593}
2594
Pekka Enberg343e0d72006-02-01 03:05:50 -08002595static void cache_init_objs(struct kmem_cache *cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002596 struct slab *slabp, unsigned long ctor_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002597{
2598 int i;
2599
2600 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002601 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002602#if DEBUG
2603 /* need to poison the objs? */
2604 if (cachep->flags & SLAB_POISON)
2605 poison_obj(cachep, objp, POISON_FREE);
2606 if (cachep->flags & SLAB_STORE_USER)
2607 *dbg_userword(cachep, objp) = NULL;
2608
2609 if (cachep->flags & SLAB_RED_ZONE) {
2610 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2611 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2612 }
2613 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002614 * Constructors are not allowed to allocate memory from the same
2615 * cache which they are a constructor for. Otherwise, deadlock.
2616 * They must also be threaded.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002617 */
2618 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002619 cachep->ctor(objp + obj_offset(cachep), cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002620 ctor_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002621
2622 if (cachep->flags & SLAB_RED_ZONE) {
2623 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2624 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002625 " end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002626 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2627 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002628 " start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002629 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08002630 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2631 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002632 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002633 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002634#else
2635 if (cachep->ctor)
2636 cachep->ctor(objp, cachep, ctor_flags);
2637#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002638 slab_bufctl(slabp)[i] = i + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002639 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002640 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002641 slabp->free = 0;
2642}
2643
Pekka Enberg343e0d72006-02-01 03:05:50 -08002644static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002645{
Christoph Lameter441e1432006-12-06 20:33:19 -08002646 if (flags & GFP_DMA)
Andrew Mortona737b3e2006-03-22 00:08:11 -08002647 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2648 else
2649 BUG_ON(cachep->gfpflags & GFP_DMA);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002650}
2651
Andrew Mortona737b3e2006-03-22 00:08:11 -08002652static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2653 int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002654{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002655 void *objp = index_to_obj(cachep, slabp, slabp->free);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002656 kmem_bufctl_t next;
2657
2658 slabp->inuse++;
2659 next = slab_bufctl(slabp)[slabp->free];
2660#if DEBUG
2661 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2662 WARN_ON(slabp->nodeid != nodeid);
2663#endif
2664 slabp->free = next;
2665
2666 return objp;
2667}
2668
Andrew Mortona737b3e2006-03-22 00:08:11 -08002669static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2670 void *objp, int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002671{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002672 unsigned int objnr = obj_to_index(cachep, slabp, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002673
2674#if DEBUG
2675 /* Verify that the slab belongs to the intended node */
2676 WARN_ON(slabp->nodeid != nodeid);
2677
Al Viro871751e2006-03-25 03:06:39 -08002678 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
Matthew Dobson78d382d2006-02-01 03:05:47 -08002679 printk(KERN_ERR "slab: double free detected in cache "
Andrew Mortona737b3e2006-03-22 00:08:11 -08002680 "'%s', objp %p\n", cachep->name, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002681 BUG();
2682 }
2683#endif
2684 slab_bufctl(slabp)[objnr] = slabp->free;
2685 slabp->free = objnr;
2686 slabp->inuse--;
2687}
2688
Pekka Enberg47768742006-06-23 02:03:07 -07002689/*
2690 * Map pages beginning at addr to the given cache and slab. This is required
2691 * for the slab allocator to be able to lookup the cache and slab of a
2692 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2693 */
2694static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2695 void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002696{
Pekka Enberg47768742006-06-23 02:03:07 -07002697 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002698 struct page *page;
2699
Pekka Enberg47768742006-06-23 02:03:07 -07002700 page = virt_to_page(addr);
Nick Piggin84097512006-03-22 00:08:34 -08002701
Pekka Enberg47768742006-06-23 02:03:07 -07002702 nr_pages = 1;
Nick Piggin84097512006-03-22 00:08:34 -08002703 if (likely(!PageCompound(page)))
Pekka Enberg47768742006-06-23 02:03:07 -07002704 nr_pages <<= cache->gfporder;
2705
Linus Torvalds1da177e2005-04-16 15:20:36 -07002706 do {
Pekka Enberg47768742006-06-23 02:03:07 -07002707 page_set_cache(page, cache);
2708 page_set_slab(page, slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002709 page++;
Pekka Enberg47768742006-06-23 02:03:07 -07002710 } while (--nr_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002711}
2712
2713/*
2714 * Grow (by 1) the number of slabs within a cache. This is called by
2715 * kmem_cache_alloc() when there are no active objs left in a cache.
2716 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002717static int cache_grow(struct kmem_cache *cachep,
2718 gfp_t flags, int nodeid, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002719{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002720 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002721 size_t offset;
2722 gfp_t local_flags;
2723 unsigned long ctor_flags;
Christoph Lametere498be72005-09-09 13:03:32 -07002724 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002725
Andrew Mortona737b3e2006-03-22 00:08:11 -08002726 /*
2727 * Be lazy and only check for valid flags here, keeping it out of the
2728 * critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002729 */
Christoph Lameter441e1432006-12-06 20:33:19 -08002730 BUG_ON(flags & ~(GFP_DMA | GFP_LEVEL_MASK | __GFP_NO_GROW));
Christoph Lameter6e0eaa42006-12-06 20:33:10 -08002731 if (flags & __GFP_NO_GROW)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002732 return 0;
2733
2734 ctor_flags = SLAB_CTOR_CONSTRUCTOR;
Christoph Lametera06d72c2006-12-06 20:33:12 -08002735 local_flags = (flags & GFP_LEVEL_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002736 if (!(local_flags & __GFP_WAIT))
2737 /*
2738 * Not allowed to sleep. Need to tell a constructor about
2739 * this - it might need to know...
2740 */
2741 ctor_flags |= SLAB_CTOR_ATOMIC;
2742
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002743 /* Take the l3 list lock to change the colour_next on this node */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002744 check_irq_off();
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002745 l3 = cachep->nodelists[nodeid];
2746 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002747
2748 /* Get colour for the slab, and cal the next value. */
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002749 offset = l3->colour_next;
2750 l3->colour_next++;
2751 if (l3->colour_next >= cachep->colour)
2752 l3->colour_next = 0;
2753 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002754
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002755 offset *= cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002756
2757 if (local_flags & __GFP_WAIT)
2758 local_irq_enable();
2759
2760 /*
2761 * The test for missing atomic flag is performed here, rather than
2762 * the more obvious place, simply to reduce the critical path length
2763 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2764 * will eventually be caught here (where it matters).
2765 */
2766 kmem_flagcheck(cachep, flags);
2767
Andrew Mortona737b3e2006-03-22 00:08:11 -08002768 /*
2769 * Get mem for the objs. Attempt to allocate a physical page from
2770 * 'nodeid'.
Christoph Lametere498be72005-09-09 13:03:32 -07002771 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002772 if (!objp)
2773 objp = kmem_getpages(cachep, flags, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002774 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002775 goto failed;
2776
2777 /* Get slab management. */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002778 slabp = alloc_slabmgmt(cachep, objp, offset,
2779 local_flags & ~GFP_THISNODE, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002780 if (!slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002781 goto opps1;
2782
Christoph Lametere498be72005-09-09 13:03:32 -07002783 slabp->nodeid = nodeid;
Pekka Enberg47768742006-06-23 02:03:07 -07002784 slab_map_pages(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002785
2786 cache_init_objs(cachep, slabp, ctor_flags);
2787
2788 if (local_flags & __GFP_WAIT)
2789 local_irq_disable();
2790 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002791 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002792
2793 /* Make slab active. */
Christoph Lametere498be72005-09-09 13:03:32 -07002794 list_add_tail(&slabp->list, &(l3->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002795 STATS_INC_GROWN(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002796 l3->free_objects += cachep->num;
2797 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002798 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002799opps1:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002800 kmem_freepages(cachep, objp);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002801failed:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002802 if (local_flags & __GFP_WAIT)
2803 local_irq_disable();
2804 return 0;
2805}
2806
2807#if DEBUG
2808
2809/*
2810 * Perform extra freeing checks:
2811 * - detect bad pointers.
2812 * - POISON/RED_ZONE checking
2813 * - destructor calls, for caches with POISON+dtor
2814 */
2815static void kfree_debugcheck(const void *objp)
2816{
2817 struct page *page;
2818
2819 if (!virt_addr_valid(objp)) {
2820 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002821 (unsigned long)objp);
2822 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002823 }
2824 page = virt_to_page(objp);
2825 if (!PageSlab(page)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002826 printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n",
2827 (unsigned long)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002828 BUG();
2829 }
2830}
2831
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002832static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2833{
2834 unsigned long redzone1, redzone2;
2835
2836 redzone1 = *dbg_redzone1(cache, obj);
2837 redzone2 = *dbg_redzone2(cache, obj);
2838
2839 /*
2840 * Redzone is ok.
2841 */
2842 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2843 return;
2844
2845 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2846 slab_error(cache, "double free detected");
2847 else
2848 slab_error(cache, "memory outside object was overwritten");
2849
2850 printk(KERN_ERR "%p: redzone 1:0x%lx, redzone 2:0x%lx.\n",
2851 obj, redzone1, redzone2);
2852}
2853
Pekka Enberg343e0d72006-02-01 03:05:50 -08002854static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002855 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002856{
2857 struct page *page;
2858 unsigned int objnr;
2859 struct slab *slabp;
2860
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002861 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002862 kfree_debugcheck(objp);
2863 page = virt_to_page(objp);
2864
Pekka Enberg065d41c2005-11-13 16:06:46 -08002865 slabp = page_get_slab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002866
2867 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002868 verify_redzone_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002869 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2870 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2871 }
2872 if (cachep->flags & SLAB_STORE_USER)
2873 *dbg_userword(cachep, objp) = caller;
2874
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002875 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002876
2877 BUG_ON(objnr >= cachep->num);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002878 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002879
2880 if (cachep->flags & SLAB_DEBUG_INITIAL) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002881 /*
2882 * Need to call the slab's constructor so the caller can
2883 * perform a verify of its state (debugging). Called without
2884 * the cache-lock held.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002885 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002886 cachep->ctor(objp + obj_offset(cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002887 cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002888 }
2889 if (cachep->flags & SLAB_POISON && cachep->dtor) {
2890 /* we want to cache poison the object,
2891 * call the destruction callback
2892 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002893 cachep->dtor(objp + obj_offset(cachep), cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002894 }
Al Viro871751e2006-03-25 03:06:39 -08002895#ifdef CONFIG_DEBUG_SLAB_LEAK
2896 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2897#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002898 if (cachep->flags & SLAB_POISON) {
2899#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08002900 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002901 store_stackinfo(cachep, objp, (unsigned long)caller);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002902 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002903 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002904 } else {
2905 poison_obj(cachep, objp, POISON_FREE);
2906 }
2907#else
2908 poison_obj(cachep, objp, POISON_FREE);
2909#endif
2910 }
2911 return objp;
2912}
2913
Pekka Enberg343e0d72006-02-01 03:05:50 -08002914static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002915{
2916 kmem_bufctl_t i;
2917 int entries = 0;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002918
Linus Torvalds1da177e2005-04-16 15:20:36 -07002919 /* Check slab's freelist to see if this obj is there. */
2920 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2921 entries++;
2922 if (entries > cachep->num || i >= cachep->num)
2923 goto bad;
2924 }
2925 if (entries != cachep->num - slabp->inuse) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002926bad:
2927 printk(KERN_ERR "slab: Internal list corruption detected in "
2928 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2929 cachep->name, cachep->num, slabp, slabp->inuse);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002930 for (i = 0;
Linus Torvalds264132b2006-03-06 12:10:07 -08002931 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002932 i++) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002933 if (i % 16 == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002934 printk("\n%03x:", i);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002935 printk(" %02x", ((unsigned char *)slabp)[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002936 }
2937 printk("\n");
2938 BUG();
2939 }
2940}
2941#else
2942#define kfree_debugcheck(x) do { } while(0)
2943#define cache_free_debugcheck(x,objp,z) (objp)
2944#define check_slabp(x,y) do { } while(0)
2945#endif
2946
Pekka Enberg343e0d72006-02-01 03:05:50 -08002947static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002948{
2949 int batchcount;
2950 struct kmem_list3 *l3;
2951 struct array_cache *ac;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002952 int node;
2953
2954 node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002955
2956 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002957 ac = cpu_cache_get(cachep);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002958retry:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002959 batchcount = ac->batchcount;
2960 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002961 /*
2962 * If there was little recent activity on this cache, then
2963 * perform only a partial refill. Otherwise we could generate
2964 * refill bouncing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002965 */
2966 batchcount = BATCHREFILL_LIMIT;
2967 }
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002968 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002969
Christoph Lametere498be72005-09-09 13:03:32 -07002970 BUG_ON(ac->avail > 0 || !l3);
2971 spin_lock(&l3->list_lock);
2972
Christoph Lameter3ded1752006-03-25 03:06:44 -08002973 /* See if we can refill from the shared array */
2974 if (l3->shared && transfer_objects(ac, l3->shared, batchcount))
2975 goto alloc_done;
2976
Linus Torvalds1da177e2005-04-16 15:20:36 -07002977 while (batchcount > 0) {
2978 struct list_head *entry;
2979 struct slab *slabp;
2980 /* Get slab alloc is to come from. */
2981 entry = l3->slabs_partial.next;
2982 if (entry == &l3->slabs_partial) {
2983 l3->free_touched = 1;
2984 entry = l3->slabs_free.next;
2985 if (entry == &l3->slabs_free)
2986 goto must_grow;
2987 }
2988
2989 slabp = list_entry(entry, struct slab, list);
2990 check_slabp(cachep, slabp);
2991 check_spinlock_acquired(cachep);
2992 while (slabp->inuse < cachep->num && batchcount--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002993 STATS_INC_ALLOCED(cachep);
2994 STATS_INC_ACTIVE(cachep);
2995 STATS_SET_HIGH(cachep);
2996
Matthew Dobson78d382d2006-02-01 03:05:47 -08002997 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002998 node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002999 }
3000 check_slabp(cachep, slabp);
3001
3002 /* move slabp to correct slabp list: */
3003 list_del(&slabp->list);
3004 if (slabp->free == BUFCTL_END)
3005 list_add(&slabp->list, &l3->slabs_full);
3006 else
3007 list_add(&slabp->list, &l3->slabs_partial);
3008 }
3009
Andrew Mortona737b3e2006-03-22 00:08:11 -08003010must_grow:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003011 l3->free_objects -= ac->avail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003012alloc_done:
Christoph Lametere498be72005-09-09 13:03:32 -07003013 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003014
3015 if (unlikely(!ac->avail)) {
3016 int x;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003017 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07003018
Andrew Mortona737b3e2006-03-22 00:08:11 -08003019 /* cache_grow can reenable interrupts, then ac could change. */
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003020 ac = cpu_cache_get(cachep);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003021 if (!x && ac->avail == 0) /* no objects in sight? abort */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003022 return NULL;
3023
Andrew Mortona737b3e2006-03-22 00:08:11 -08003024 if (!ac->avail) /* objects refilled by interrupt? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003025 goto retry;
3026 }
3027 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07003028 return ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003029}
3030
Andrew Mortona737b3e2006-03-22 00:08:11 -08003031static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3032 gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003033{
3034 might_sleep_if(flags & __GFP_WAIT);
3035#if DEBUG
3036 kmem_flagcheck(cachep, flags);
3037#endif
3038}
3039
3040#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003041static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3042 gfp_t flags, void *objp, void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003043{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003044 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003045 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003046 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003047#ifdef CONFIG_DEBUG_PAGEALLOC
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003048 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003049 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003050 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003051 else
3052 check_poison_obj(cachep, objp);
3053#else
3054 check_poison_obj(cachep, objp);
3055#endif
3056 poison_obj(cachep, objp, POISON_INUSE);
3057 }
3058 if (cachep->flags & SLAB_STORE_USER)
3059 *dbg_userword(cachep, objp) = caller;
3060
3061 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003062 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3063 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3064 slab_error(cachep, "double free, or memory outside"
3065 " object was overwritten");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003066 printk(KERN_ERR
Andrew Mortona737b3e2006-03-22 00:08:11 -08003067 "%p: redzone 1:0x%lx, redzone 2:0x%lx\n",
3068 objp, *dbg_redzone1(cachep, objp),
3069 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003070 }
3071 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3072 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3073 }
Al Viro871751e2006-03-25 03:06:39 -08003074#ifdef CONFIG_DEBUG_SLAB_LEAK
3075 {
3076 struct slab *slabp;
3077 unsigned objnr;
3078
3079 slabp = page_get_slab(virt_to_page(objp));
3080 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3081 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3082 }
3083#endif
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003084 objp += obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003085 if (cachep->ctor && cachep->flags & SLAB_POISON) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003086 unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003087
3088 if (!(flags & __GFP_WAIT))
3089 ctor_flags |= SLAB_CTOR_ATOMIC;
3090
3091 cachep->ctor(objp, cachep, ctor_flags);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003092 }
Kevin Hilmana44b56d2006-12-06 20:32:11 -08003093#if ARCH_SLAB_MINALIGN
3094 if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) {
3095 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3096 objp, ARCH_SLAB_MINALIGN);
3097 }
3098#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003099 return objp;
3100}
3101#else
3102#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3103#endif
3104
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003105#ifdef CONFIG_FAILSLAB
3106
3107static struct failslab_attr {
3108
3109 struct fault_attr attr;
3110
3111 u32 ignore_gfp_wait;
3112#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3113 struct dentry *ignore_gfp_wait_file;
3114#endif
3115
3116} failslab = {
3117 .attr = FAULT_ATTR_INITIALIZER,
Don Mullis6b1b60f2006-12-08 02:39:53 -08003118 .ignore_gfp_wait = 1,
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003119};
3120
3121static int __init setup_failslab(char *str)
3122{
3123 return setup_fault_attr(&failslab.attr, str);
3124}
3125__setup("failslab=", setup_failslab);
3126
3127static int should_failslab(struct kmem_cache *cachep, gfp_t flags)
3128{
3129 if (cachep == &cache_cache)
3130 return 0;
3131 if (flags & __GFP_NOFAIL)
3132 return 0;
3133 if (failslab.ignore_gfp_wait && (flags & __GFP_WAIT))
3134 return 0;
3135
3136 return should_fail(&failslab.attr, obj_size(cachep));
3137}
3138
3139#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3140
3141static int __init failslab_debugfs(void)
3142{
3143 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
3144 struct dentry *dir;
3145 int err;
3146
3147 err = init_fault_attr_dentries(&failslab.attr, "failslab");
3148 if (err)
3149 return err;
3150 dir = failslab.attr.dentries.dir;
3151
3152 failslab.ignore_gfp_wait_file =
3153 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3154 &failslab.ignore_gfp_wait);
3155
3156 if (!failslab.ignore_gfp_wait_file) {
3157 err = -ENOMEM;
3158 debugfs_remove(failslab.ignore_gfp_wait_file);
3159 cleanup_fault_attr_dentries(&failslab.attr);
3160 }
3161
3162 return err;
3163}
3164
3165late_initcall(failslab_debugfs);
3166
3167#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3168
3169#else /* CONFIG_FAILSLAB */
3170
3171static inline int should_failslab(struct kmem_cache *cachep, gfp_t flags)
3172{
3173 return 0;
3174}
3175
3176#endif /* CONFIG_FAILSLAB */
3177
Pekka Enberg343e0d72006-02-01 03:05:50 -08003178static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003179{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003180 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003181 struct array_cache *ac;
3182
Alok N Kataria5c382302005-09-27 21:45:46 -07003183 check_irq_off();
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003184
3185 if (should_failslab(cachep, flags))
3186 return NULL;
3187
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003188 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003189 if (likely(ac->avail)) {
3190 STATS_INC_ALLOCHIT(cachep);
3191 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07003192 objp = ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003193 } else {
3194 STATS_INC_ALLOCMISS(cachep);
3195 objp = cache_alloc_refill(cachep, flags);
3196 }
Alok N Kataria5c382302005-09-27 21:45:46 -07003197 return objp;
3198}
3199
Andrew Mortona737b3e2006-03-22 00:08:11 -08003200static __always_inline void *__cache_alloc(struct kmem_cache *cachep,
3201 gfp_t flags, void *caller)
Alok N Kataria5c382302005-09-27 21:45:46 -07003202{
3203 unsigned long save_flags;
Christoph Lameterde3083e2006-09-27 01:50:03 -07003204 void *objp = NULL;
Alok N Kataria5c382302005-09-27 21:45:46 -07003205
3206 cache_alloc_debugcheck_before(cachep, flags);
3207
3208 local_irq_save(save_flags);
Christoph Lameterde3083e2006-09-27 01:50:03 -07003209
Christoph Lameter765c4502006-09-27 01:50:08 -07003210 if (unlikely(NUMA_BUILD &&
3211 current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY)))
Christoph Lameterde3083e2006-09-27 01:50:03 -07003212 objp = alternate_node_alloc(cachep, flags);
Christoph Lameterde3083e2006-09-27 01:50:03 -07003213
3214 if (!objp)
3215 objp = ____cache_alloc(cachep, flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003216 /*
3217 * We may just have run out of memory on the local node.
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003218 * ____cache_alloc_node() knows how to locate memory on other nodes
Christoph Lameter765c4502006-09-27 01:50:08 -07003219 */
3220 if (NUMA_BUILD && !objp)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003221 objp = ____cache_alloc_node(cachep, flags, numa_node_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -07003222 local_irq_restore(save_flags);
Eric Dumazet34342e82005-09-03 15:55:06 -07003223 objp = cache_alloc_debugcheck_after(cachep, flags, objp,
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003224 caller);
Eric Dumazet34342e82005-09-03 15:55:06 -07003225 prefetchw(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003226 return objp;
3227}
3228
Christoph Lametere498be72005-09-09 13:03:32 -07003229#ifdef CONFIG_NUMA
3230/*
Paul Jacksonb2455392006-03-24 03:16:12 -08003231 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
Paul Jacksonc61afb12006-03-24 03:16:08 -08003232 *
3233 * If we are in_interrupt, then process context, including cpusets and
3234 * mempolicy, may not apply and should not be used for allocation policy.
3235 */
3236static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3237{
3238 int nid_alloc, nid_here;
3239
Christoph Lameter765c4502006-09-27 01:50:08 -07003240 if (in_interrupt() || (flags & __GFP_THISNODE))
Paul Jacksonc61afb12006-03-24 03:16:08 -08003241 return NULL;
3242 nid_alloc = nid_here = numa_node_id();
3243 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3244 nid_alloc = cpuset_mem_spread_node();
3245 else if (current->mempolicy)
3246 nid_alloc = slab_node(current->mempolicy);
3247 if (nid_alloc != nid_here)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003248 return ____cache_alloc_node(cachep, flags, nid_alloc);
Paul Jacksonc61afb12006-03-24 03:16:08 -08003249 return NULL;
3250}
3251
3252/*
Christoph Lameter765c4502006-09-27 01:50:08 -07003253 * Fallback function if there was no memory available and no objects on a
Christoph Lameter3c517a62006-12-06 20:33:29 -08003254 * certain node and fall back is permitted. First we scan all the
3255 * available nodelists for available objects. If that fails then we
3256 * perform an allocation without specifying a node. This allows the page
3257 * allocator to do its reclaim / fallback magic. We then insert the
3258 * slab into the proper nodelist and then allocate from it.
Christoph Lameter765c4502006-09-27 01:50:08 -07003259 */
3260void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3261{
3262 struct zonelist *zonelist = &NODE_DATA(slab_node(current->mempolicy))
3263 ->node_zonelists[gfp_zone(flags)];
3264 struct zone **z;
3265 void *obj = NULL;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003266 int nid;
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003267 gfp_t local_flags = (flags & GFP_LEVEL_MASK);
Christoph Lameter765c4502006-09-27 01:50:08 -07003268
Christoph Lameter3c517a62006-12-06 20:33:29 -08003269retry:
3270 /*
3271 * Look through allowed nodes for objects available
3272 * from existing per node queues.
3273 */
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003274 for (z = zonelist->zones; *z && !obj; z++) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003275 nid = zone_to_nid(*z);
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003276
Paul Jackson02a0e532006-12-13 00:34:25 -08003277 if (cpuset_zone_allowed_hardwall(*z, flags) &&
Christoph Lameter3c517a62006-12-06 20:33:29 -08003278 cache->nodelists[nid] &&
3279 cache->nodelists[nid]->free_objects)
3280 obj = ____cache_alloc_node(cache,
3281 flags | GFP_THISNODE, nid);
3282 }
3283
3284 if (!obj) {
3285 /*
3286 * This allocation will be performed within the constraints
3287 * of the current cpuset / memory policy requirements.
3288 * We may trigger various forms of reclaim on the allowed
3289 * set and go into memory reserves if necessary.
3290 */
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003291 if (local_flags & __GFP_WAIT)
3292 local_irq_enable();
3293 kmem_flagcheck(cache, flags);
Christoph Lameter3c517a62006-12-06 20:33:29 -08003294 obj = kmem_getpages(cache, flags, -1);
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003295 if (local_flags & __GFP_WAIT)
3296 local_irq_disable();
Christoph Lameter3c517a62006-12-06 20:33:29 -08003297 if (obj) {
3298 /*
3299 * Insert into the appropriate per node queues
3300 */
3301 nid = page_to_nid(virt_to_page(obj));
3302 if (cache_grow(cache, flags, nid, obj)) {
3303 obj = ____cache_alloc_node(cache,
3304 flags | GFP_THISNODE, nid);
3305 if (!obj)
3306 /*
3307 * Another processor may allocate the
3308 * objects in the slab since we are
3309 * not holding any locks.
3310 */
3311 goto retry;
3312 } else {
3313 kmem_freepages(cache, obj);
3314 obj = NULL;
3315 }
3316 }
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003317 }
Christoph Lameter765c4502006-09-27 01:50:08 -07003318 return obj;
3319}
3320
3321/*
Christoph Lametere498be72005-09-09 13:03:32 -07003322 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07003323 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003324static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003325 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07003326{
3327 struct list_head *entry;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003328 struct slab *slabp;
3329 struct kmem_list3 *l3;
3330 void *obj;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003331 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003332
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003333 l3 = cachep->nodelists[nodeid];
3334 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07003335
Andrew Mortona737b3e2006-03-22 00:08:11 -08003336retry:
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003337 check_irq_off();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003338 spin_lock(&l3->list_lock);
3339 entry = l3->slabs_partial.next;
3340 if (entry == &l3->slabs_partial) {
3341 l3->free_touched = 1;
3342 entry = l3->slabs_free.next;
3343 if (entry == &l3->slabs_free)
3344 goto must_grow;
3345 }
Christoph Lametere498be72005-09-09 13:03:32 -07003346
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003347 slabp = list_entry(entry, struct slab, list);
3348 check_spinlock_acquired_node(cachep, nodeid);
3349 check_slabp(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07003350
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003351 STATS_INC_NODEALLOCS(cachep);
3352 STATS_INC_ACTIVE(cachep);
3353 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003354
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003355 BUG_ON(slabp->inuse == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07003356
Matthew Dobson78d382d2006-02-01 03:05:47 -08003357 obj = slab_get_obj(cachep, slabp, nodeid);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003358 check_slabp(cachep, slabp);
3359 l3->free_objects--;
3360 /* move slabp to correct slabp list: */
3361 list_del(&slabp->list);
Christoph Lametere498be72005-09-09 13:03:32 -07003362
Andrew Mortona737b3e2006-03-22 00:08:11 -08003363 if (slabp->free == BUFCTL_END)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003364 list_add(&slabp->list, &l3->slabs_full);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003365 else
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003366 list_add(&slabp->list, &l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07003367
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003368 spin_unlock(&l3->list_lock);
3369 goto done;
Christoph Lametere498be72005-09-09 13:03:32 -07003370
Andrew Mortona737b3e2006-03-22 00:08:11 -08003371must_grow:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003372 spin_unlock(&l3->list_lock);
Christoph Lameter3c517a62006-12-06 20:33:29 -08003373 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
Christoph Lameter765c4502006-09-27 01:50:08 -07003374 if (x)
3375 goto retry;
Christoph Lametere498be72005-09-09 13:03:32 -07003376
Christoph Lameter765c4502006-09-27 01:50:08 -07003377 if (!(flags & __GFP_THISNODE))
3378 /* Unable to grow the cache. Fall back to other nodes. */
3379 return fallback_alloc(cachep, flags);
Christoph Lametere498be72005-09-09 13:03:32 -07003380
Christoph Lameter765c4502006-09-27 01:50:08 -07003381 return NULL;
3382
Andrew Mortona737b3e2006-03-22 00:08:11 -08003383done:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003384 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07003385}
3386#endif
3387
3388/*
3389 * Caller needs to acquire correct kmem_list's list_lock
3390 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003391static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003392 int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003393{
3394 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07003395 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003396
3397 for (i = 0; i < nr_objects; i++) {
3398 void *objp = objpp[i];
3399 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003400
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08003401 slabp = virt_to_slab(objp);
Christoph Lameterff694162005-09-22 21:44:02 -07003402 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003403 list_del(&slabp->list);
Christoph Lameterff694162005-09-22 21:44:02 -07003404 check_spinlock_acquired_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003405 check_slabp(cachep, slabp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08003406 slab_put_obj(cachep, slabp, objp, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003407 STATS_DEC_ACTIVE(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003408 l3->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003409 check_slabp(cachep, slabp);
3410
3411 /* fixup slab chains */
3412 if (slabp->inuse == 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07003413 if (l3->free_objects > l3->free_limit) {
3414 l3->free_objects -= cachep->num;
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07003415 /* No need to drop any previously held
3416 * lock here, even if we have a off-slab slab
3417 * descriptor it is guaranteed to come from
3418 * a different cache, refer to comments before
3419 * alloc_slabmgmt.
3420 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003421 slab_destroy(cachep, slabp);
3422 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07003423 list_add(&slabp->list, &l3->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003424 }
3425 } else {
3426 /* Unconditionally move a slab to the end of the
3427 * partial list on free - maximum time for the
3428 * other objects to be freed, too.
3429 */
Christoph Lametere498be72005-09-09 13:03:32 -07003430 list_add_tail(&slabp->list, &l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003431 }
3432 }
3433}
3434
Pekka Enberg343e0d72006-02-01 03:05:50 -08003435static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003436{
3437 int batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07003438 struct kmem_list3 *l3;
Christoph Lameterff694162005-09-22 21:44:02 -07003439 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003440
3441 batchcount = ac->batchcount;
3442#if DEBUG
3443 BUG_ON(!batchcount || batchcount > ac->avail);
3444#endif
3445 check_irq_off();
Christoph Lameterff694162005-09-22 21:44:02 -07003446 l3 = cachep->nodelists[node];
Ingo Molnar873623d2006-07-13 14:44:38 +02003447 spin_lock(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003448 if (l3->shared) {
3449 struct array_cache *shared_array = l3->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003450 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003451 if (max) {
3452 if (batchcount > max)
3453 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07003454 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003455 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003456 shared_array->avail += batchcount;
3457 goto free_done;
3458 }
3459 }
3460
Christoph Lameterff694162005-09-22 21:44:02 -07003461 free_block(cachep, ac->entry, batchcount, node);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003462free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003463#if STATS
3464 {
3465 int i = 0;
3466 struct list_head *p;
3467
Christoph Lametere498be72005-09-09 13:03:32 -07003468 p = l3->slabs_free.next;
3469 while (p != &(l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003470 struct slab *slabp;
3471
3472 slabp = list_entry(p, struct slab, list);
3473 BUG_ON(slabp->inuse);
3474
3475 i++;
3476 p = p->next;
3477 }
3478 STATS_SET_FREEABLE(cachep, i);
3479 }
3480#endif
Christoph Lametere498be72005-09-09 13:03:32 -07003481 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003482 ac->avail -= batchcount;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003483 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003484}
3485
3486/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08003487 * Release an obj back to its cache. If the obj has a constructed state, it must
3488 * be in this state _before_ it is released. Called with disabled ints.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003489 */
Ingo Molnar873623d2006-07-13 14:44:38 +02003490static inline void __cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003491{
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003492 struct array_cache *ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003493
3494 check_irq_off();
3495 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3496
Ingo Molnar873623d2006-07-13 14:44:38 +02003497 if (cache_free_alien(cachep, objp))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07003498 return;
Christoph Lametere498be72005-09-09 13:03:32 -07003499
Linus Torvalds1da177e2005-04-16 15:20:36 -07003500 if (likely(ac->avail < ac->limit)) {
3501 STATS_INC_FREEHIT(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003502 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003503 return;
3504 } else {
3505 STATS_INC_FREEMISS(cachep);
3506 cache_flusharray(cachep, ac);
Christoph Lametere498be72005-09-09 13:03:32 -07003507 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003508 }
3509}
3510
3511/**
3512 * kmem_cache_alloc - Allocate an object
3513 * @cachep: The cache to allocate from.
3514 * @flags: See kmalloc().
3515 *
3516 * Allocate an object from this cache. The flags are only relevant
3517 * if the cache has no available objects.
3518 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003519void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003520{
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003521 return __cache_alloc(cachep, flags, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003522}
3523EXPORT_SYMBOL(kmem_cache_alloc);
3524
3525/**
Rolf Eike Beerb8008b22006-07-30 03:04:04 -07003526 * kmem_cache_zalloc - Allocate an object. The memory is set to zero.
Pekka Enberga8c0f9a2006-03-25 03:06:42 -08003527 * @cache: The cache to allocate from.
3528 * @flags: See kmalloc().
3529 *
3530 * Allocate an object from this cache and set the allocated memory to zero.
3531 * The flags are only relevant if the cache has no available objects.
3532 */
3533void *kmem_cache_zalloc(struct kmem_cache *cache, gfp_t flags)
3534{
3535 void *ret = __cache_alloc(cache, flags, __builtin_return_address(0));
3536 if (ret)
3537 memset(ret, 0, obj_size(cache));
3538 return ret;
3539}
3540EXPORT_SYMBOL(kmem_cache_zalloc);
3541
3542/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003543 * kmem_ptr_validate - check if an untrusted pointer might
3544 * be a slab entry.
3545 * @cachep: the cache we're checking against
3546 * @ptr: pointer to validate
3547 *
3548 * This verifies that the untrusted pointer looks sane:
3549 * it is _not_ a guarantee that the pointer is actually
3550 * part of the slab cache in question, but it at least
3551 * validates that the pointer can be dereferenced and
3552 * looks half-way sane.
3553 *
3554 * Currently only used for dentry validation.
3555 */
Christoph Lameter55935a32006-12-13 00:34:24 -08003556int fastcall kmem_ptr_validate(struct kmem_cache *cachep, const void *ptr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003557{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003558 unsigned long addr = (unsigned long)ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003559 unsigned long min_addr = PAGE_OFFSET;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003560 unsigned long align_mask = BYTES_PER_WORD - 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003561 unsigned long size = cachep->buffer_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003562 struct page *page;
3563
3564 if (unlikely(addr < min_addr))
3565 goto out;
3566 if (unlikely(addr > (unsigned long)high_memory - size))
3567 goto out;
3568 if (unlikely(addr & align_mask))
3569 goto out;
3570 if (unlikely(!kern_addr_valid(addr)))
3571 goto out;
3572 if (unlikely(!kern_addr_valid(addr + size - 1)))
3573 goto out;
3574 page = virt_to_page(ptr);
3575 if (unlikely(!PageSlab(page)))
3576 goto out;
Pekka Enberg065d41c2005-11-13 16:06:46 -08003577 if (unlikely(page_get_cache(page) != cachep))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003578 goto out;
3579 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003580out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003581 return 0;
3582}
3583
3584#ifdef CONFIG_NUMA
3585/**
3586 * kmem_cache_alloc_node - Allocate an object on the specified node
3587 * @cachep: The cache to allocate from.
3588 * @flags: See kmalloc().
3589 * @nodeid: node number of the target node.
3590 *
Christoph Lameter5bcd2342006-12-06 20:33:24 -08003591 * Identical to kmem_cache_alloc but it will allocate memory on the given
3592 * node, which can improve the performance for cpu bound structures.
3593 *
3594 * Fallback to other node is possible if __GFP_THISNODE is not set.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003595 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003596static __always_inline void *
3597__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3598 int nodeid, void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003599{
Christoph Lametere498be72005-09-09 13:03:32 -07003600 unsigned long save_flags;
Christoph Lameter5bcd2342006-12-06 20:33:24 -08003601 void *ptr = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003602
Christoph Lametere498be72005-09-09 13:03:32 -07003603 cache_alloc_debugcheck_before(cachep, flags);
3604 local_irq_save(save_flags);
Christoph Lameter18f820f2006-02-01 03:05:43 -08003605
Christoph Lameter5bcd2342006-12-06 20:33:24 -08003606 if (unlikely(nodeid == -1))
3607 nodeid = numa_node_id();
Christoph Lameter18f820f2006-02-01 03:05:43 -08003608
Christoph Lameter5bcd2342006-12-06 20:33:24 -08003609 if (likely(cachep->nodelists[nodeid])) {
3610 if (nodeid == numa_node_id()) {
3611 /*
3612 * Use the locally cached objects if possible.
3613 * However ____cache_alloc does not allow fallback
3614 * to other nodes. It may fail while we still have
3615 * objects on other nodes available.
3616 */
3617 ptr = ____cache_alloc(cachep, flags);
3618 }
3619 if (!ptr) {
3620 /* ___cache_alloc_node can fall back to other nodes */
3621 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3622 }
3623 } else {
3624 /* Node not bootstrapped yet */
3625 if (!(flags & __GFP_THISNODE))
3626 ptr = fallback_alloc(cachep, flags);
3627 }
3628
3629 local_irq_restore(save_flags);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003630 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003631
Christoph Lametere498be72005-09-09 13:03:32 -07003632 return ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003633}
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003634
3635void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3636{
3637 return __cache_alloc_node(cachep, flags, nodeid,
3638 __builtin_return_address(0));
3639}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003640EXPORT_SYMBOL(kmem_cache_alloc_node);
3641
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003642static __always_inline void *
3643__do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003644{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003645 struct kmem_cache *cachep;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003646
3647 cachep = kmem_find_general_cachep(size, flags);
3648 if (unlikely(cachep == NULL))
3649 return NULL;
3650 return kmem_cache_alloc_node(cachep, flags, node);
3651}
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003652
3653#ifdef CONFIG_DEBUG_SLAB
3654void *__kmalloc_node(size_t size, gfp_t flags, int node)
3655{
3656 return __do_kmalloc_node(size, flags, node,
3657 __builtin_return_address(0));
3658}
Christoph Hellwigdbe5e692006-09-25 23:31:36 -07003659EXPORT_SYMBOL(__kmalloc_node);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003660
3661void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3662 int node, void *caller)
3663{
3664 return __do_kmalloc_node(size, flags, node, caller);
3665}
3666EXPORT_SYMBOL(__kmalloc_node_track_caller);
3667#else
3668void *__kmalloc_node(size_t size, gfp_t flags, int node)
3669{
3670 return __do_kmalloc_node(size, flags, node, NULL);
3671}
3672EXPORT_SYMBOL(__kmalloc_node);
3673#endif /* CONFIG_DEBUG_SLAB */
3674#endif /* CONFIG_NUMA */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003675
3676/**
Paul Drynoff800590f2006-06-23 02:03:48 -07003677 * __do_kmalloc - allocate memory
Linus Torvalds1da177e2005-04-16 15:20:36 -07003678 * @size: how many bytes of memory are required.
Paul Drynoff800590f2006-06-23 02:03:48 -07003679 * @flags: the type of memory to allocate (see kmalloc).
Randy Dunlap911851e2006-03-22 00:08:14 -08003680 * @caller: function caller for debug tracking of the caller
Linus Torvalds1da177e2005-04-16 15:20:36 -07003681 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003682static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3683 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003684{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003685 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003686
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003687 /* If you want to save a few bytes .text space: replace
3688 * __ with kmem_.
3689 * Then kmalloc uses the uninlined functions instead of the inline
3690 * functions.
3691 */
3692 cachep = __find_general_cachep(size, flags);
Andrew Mortondbdb9042005-09-23 13:24:10 -07003693 if (unlikely(cachep == NULL))
3694 return NULL;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003695 return __cache_alloc(cachep, flags, caller);
3696}
3697
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003698
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003699#ifdef CONFIG_DEBUG_SLAB
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003700void *__kmalloc(size_t size, gfp_t flags)
3701{
Al Viro871751e2006-03-25 03:06:39 -08003702 return __do_kmalloc(size, flags, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003703}
3704EXPORT_SYMBOL(__kmalloc);
3705
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003706void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller)
3707{
3708 return __do_kmalloc(size, flags, caller);
3709}
3710EXPORT_SYMBOL(__kmalloc_track_caller);
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003711
3712#else
3713void *__kmalloc(size_t size, gfp_t flags)
3714{
3715 return __do_kmalloc(size, flags, NULL);
3716}
3717EXPORT_SYMBOL(__kmalloc);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003718#endif
3719
Linus Torvalds1da177e2005-04-16 15:20:36 -07003720/**
3721 * kmem_cache_free - Deallocate an object
3722 * @cachep: The cache the allocation was from.
3723 * @objp: The previously allocated object.
3724 *
3725 * Free an object which was previously allocated from this
3726 * cache.
3727 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003728void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003729{
3730 unsigned long flags;
3731
Pekka Enbergddc2e812006-06-23 02:03:40 -07003732 BUG_ON(virt_to_cache(objp) != cachep);
3733
Linus Torvalds1da177e2005-04-16 15:20:36 -07003734 local_irq_save(flags);
Ingo Molnar873623d2006-07-13 14:44:38 +02003735 __cache_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003736 local_irq_restore(flags);
3737}
3738EXPORT_SYMBOL(kmem_cache_free);
3739
3740/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003741 * kfree - free previously allocated memory
3742 * @objp: pointer returned by kmalloc.
3743 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003744 * If @objp is NULL, no operation is performed.
3745 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003746 * Don't free memory not originally allocated by kmalloc()
3747 * or you will run into trouble.
3748 */
3749void kfree(const void *objp)
3750{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003751 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003752 unsigned long flags;
3753
3754 if (unlikely(!objp))
3755 return;
3756 local_irq_save(flags);
3757 kfree_debugcheck(objp);
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08003758 c = virt_to_cache(objp);
Ingo Molnarf9b84042006-06-27 02:54:49 -07003759 debug_check_no_locks_freed(objp, obj_size(c));
Ingo Molnar873623d2006-07-13 14:44:38 +02003760 __cache_free(c, (void *)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003761 local_irq_restore(flags);
3762}
3763EXPORT_SYMBOL(kfree);
3764
Pekka Enberg343e0d72006-02-01 03:05:50 -08003765unsigned int kmem_cache_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003766{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003767 return obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003768}
3769EXPORT_SYMBOL(kmem_cache_size);
3770
Pekka Enberg343e0d72006-02-01 03:05:50 -08003771const char *kmem_cache_name(struct kmem_cache *cachep)
Arnaldo Carvalho de Melo19449722005-06-18 22:46:19 -07003772{
3773 return cachep->name;
3774}
3775EXPORT_SYMBOL_GPL(kmem_cache_name);
3776
Christoph Lametere498be72005-09-09 13:03:32 -07003777/*
Christoph Lameter0718dc22006-03-25 03:06:47 -08003778 * This initializes kmem_list3 or resizes varioius caches for all nodes.
Christoph Lametere498be72005-09-09 13:03:32 -07003779 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003780static int alloc_kmemlist(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07003781{
3782 int node;
3783 struct kmem_list3 *l3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003784 struct array_cache *new_shared;
Paul Menage3395ee02006-12-06 20:32:16 -08003785 struct array_cache **new_alien = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003786
3787 for_each_online_node(node) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003788
Paul Menage3395ee02006-12-06 20:32:16 -08003789 if (use_alien_caches) {
3790 new_alien = alloc_alien_cache(node, cachep->limit);
3791 if (!new_alien)
3792 goto fail;
3793 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003794
Christoph Lameter0718dc22006-03-25 03:06:47 -08003795 new_shared = alloc_arraycache(node,
3796 cachep->shared*cachep->batchcount,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003797 0xbaadf00d);
Christoph Lameter0718dc22006-03-25 03:06:47 -08003798 if (!new_shared) {
3799 free_alien_cache(new_alien);
Christoph Lametere498be72005-09-09 13:03:32 -07003800 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003801 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003802
Andrew Mortona737b3e2006-03-22 00:08:11 -08003803 l3 = cachep->nodelists[node];
3804 if (l3) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003805 struct array_cache *shared = l3->shared;
3806
Christoph Lametere498be72005-09-09 13:03:32 -07003807 spin_lock_irq(&l3->list_lock);
3808
Christoph Lametercafeb022006-03-25 03:06:46 -08003809 if (shared)
Christoph Lameter0718dc22006-03-25 03:06:47 -08003810 free_block(cachep, shared->entry,
3811 shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07003812
Christoph Lametercafeb022006-03-25 03:06:46 -08003813 l3->shared = new_shared;
3814 if (!l3->alien) {
Christoph Lametere498be72005-09-09 13:03:32 -07003815 l3->alien = new_alien;
3816 new_alien = NULL;
3817 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003818 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003819 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003820 spin_unlock_irq(&l3->list_lock);
Christoph Lametercafeb022006-03-25 03:06:46 -08003821 kfree(shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003822 free_alien_cache(new_alien);
3823 continue;
3824 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08003825 l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node);
Christoph Lameter0718dc22006-03-25 03:06:47 -08003826 if (!l3) {
3827 free_alien_cache(new_alien);
3828 kfree(new_shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003829 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003830 }
Christoph Lametere498be72005-09-09 13:03:32 -07003831
3832 kmem_list3_init(l3);
3833 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Andrew Mortona737b3e2006-03-22 00:08:11 -08003834 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003835 l3->shared = new_shared;
Christoph Lametere498be72005-09-09 13:03:32 -07003836 l3->alien = new_alien;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003837 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003838 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003839 cachep->nodelists[node] = l3;
3840 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003841 return 0;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003842
Andrew Mortona737b3e2006-03-22 00:08:11 -08003843fail:
Christoph Lameter0718dc22006-03-25 03:06:47 -08003844 if (!cachep->next.next) {
3845 /* Cache is not active yet. Roll back what we did */
3846 node--;
3847 while (node >= 0) {
3848 if (cachep->nodelists[node]) {
3849 l3 = cachep->nodelists[node];
3850
3851 kfree(l3->shared);
3852 free_alien_cache(l3->alien);
3853 kfree(l3);
3854 cachep->nodelists[node] = NULL;
3855 }
3856 node--;
3857 }
3858 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003859 return -ENOMEM;
Christoph Lametere498be72005-09-09 13:03:32 -07003860}
3861
Linus Torvalds1da177e2005-04-16 15:20:36 -07003862struct ccupdate_struct {
Pekka Enberg343e0d72006-02-01 03:05:50 -08003863 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003864 struct array_cache *new[NR_CPUS];
3865};
3866
3867static void do_ccupdate_local(void *info)
3868{
Andrew Mortona737b3e2006-03-22 00:08:11 -08003869 struct ccupdate_struct *new = info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003870 struct array_cache *old;
3871
3872 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003873 old = cpu_cache_get(new->cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003874
Linus Torvalds1da177e2005-04-16 15:20:36 -07003875 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3876 new->new[smp_processor_id()] = old;
3877}
3878
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003879/* Always called with the cache_chain_mutex held */
Andrew Mortona737b3e2006-03-22 00:08:11 -08003880static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3881 int batchcount, int shared)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003882{
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003883 struct ccupdate_struct *new;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003884 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003885
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003886 new = kzalloc(sizeof(*new), GFP_KERNEL);
3887 if (!new)
3888 return -ENOMEM;
3889
Christoph Lametere498be72005-09-09 13:03:32 -07003890 for_each_online_cpu(i) {
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003891 new->new[i] = alloc_arraycache(cpu_to_node(i), limit,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003892 batchcount);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003893 if (!new->new[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003894 for (i--; i >= 0; i--)
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003895 kfree(new->new[i]);
3896 kfree(new);
Christoph Lametere498be72005-09-09 13:03:32 -07003897 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003898 }
3899 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003900 new->cachep = cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003901
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003902 on_each_cpu(do_ccupdate_local, (void *)new, 1, 1);
Christoph Lametere498be72005-09-09 13:03:32 -07003903
Linus Torvalds1da177e2005-04-16 15:20:36 -07003904 check_irq_on();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003905 cachep->batchcount = batchcount;
3906 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07003907 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003908
Christoph Lametere498be72005-09-09 13:03:32 -07003909 for_each_online_cpu(i) {
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003910 struct array_cache *ccold = new->new[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003911 if (!ccold)
3912 continue;
Christoph Lametere498be72005-09-09 13:03:32 -07003913 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -07003914 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
Christoph Lametere498be72005-09-09 13:03:32 -07003915 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003916 kfree(ccold);
3917 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003918 kfree(new);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003919 return alloc_kmemlist(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003920}
3921
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003922/* Called with cache_chain_mutex held always */
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003923static int enable_cpucache(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003924{
3925 int err;
3926 int limit, shared;
3927
Andrew Mortona737b3e2006-03-22 00:08:11 -08003928 /*
3929 * The head array serves three purposes:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003930 * - create a LIFO ordering, i.e. return objects that are cache-warm
3931 * - reduce the number of spinlock operations.
Andrew Mortona737b3e2006-03-22 00:08:11 -08003932 * - reduce the number of linked list operations on the slab and
Linus Torvalds1da177e2005-04-16 15:20:36 -07003933 * bufctl chains: array operations are cheaper.
3934 * The numbers are guessed, we should auto-tune as described by
3935 * Bonwick.
3936 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003937 if (cachep->buffer_size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003938 limit = 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003939 else if (cachep->buffer_size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003940 limit = 8;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003941 else if (cachep->buffer_size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003942 limit = 24;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003943 else if (cachep->buffer_size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003944 limit = 54;
3945 else
3946 limit = 120;
3947
Andrew Mortona737b3e2006-03-22 00:08:11 -08003948 /*
3949 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
Linus Torvalds1da177e2005-04-16 15:20:36 -07003950 * allocation behaviour: Most allocs on one cpu, most free operations
3951 * on another cpu. For these cases, an efficient object passing between
3952 * cpus is necessary. This is provided by a shared array. The array
3953 * replaces Bonwick's magazine layer.
3954 * On uniprocessor, it's functionally equivalent (but less efficient)
3955 * to a larger limit. Thus disabled by default.
3956 */
3957 shared = 0;
3958#ifdef CONFIG_SMP
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003959 if (cachep->buffer_size <= PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003960 shared = 8;
3961#endif
3962
3963#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003964 /*
3965 * With debugging enabled, large batchcount lead to excessively long
3966 * periods with disabled local interrupts. Limit the batchcount
Linus Torvalds1da177e2005-04-16 15:20:36 -07003967 */
3968 if (limit > 32)
3969 limit = 32;
3970#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003971 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003972 if (err)
3973 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003974 cachep->name, -err);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003975 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003976}
3977
Christoph Lameter1b552532006-03-22 00:09:07 -08003978/*
3979 * Drain an array if it contains any elements taking the l3 lock only if
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003980 * necessary. Note that the l3 listlock also protects the array_cache
3981 * if drain_array() is used on the shared array.
Christoph Lameter1b552532006-03-22 00:09:07 -08003982 */
3983void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
3984 struct array_cache *ac, int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003985{
3986 int tofree;
3987
Christoph Lameter1b552532006-03-22 00:09:07 -08003988 if (!ac || !ac->avail)
3989 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003990 if (ac->touched && !force) {
3991 ac->touched = 0;
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003992 } else {
Christoph Lameter1b552532006-03-22 00:09:07 -08003993 spin_lock_irq(&l3->list_lock);
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003994 if (ac->avail) {
3995 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3996 if (tofree > ac->avail)
3997 tofree = (ac->avail + 1) / 2;
3998 free_block(cachep, ac->entry, tofree, node);
3999 ac->avail -= tofree;
4000 memmove(ac->entry, &(ac->entry[tofree]),
4001 sizeof(void *) * ac->avail);
4002 }
Christoph Lameter1b552532006-03-22 00:09:07 -08004003 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004004 }
4005}
4006
4007/**
4008 * cache_reap - Reclaim memory from caches.
Randy Dunlap1e5d5332005-11-07 01:01:06 -08004009 * @unused: unused parameter
Linus Torvalds1da177e2005-04-16 15:20:36 -07004010 *
4011 * Called from workqueue/eventd every few seconds.
4012 * Purpose:
4013 * - clear the per-cpu caches for this CPU.
4014 * - return freeable pages to the main free memory pool.
4015 *
Andrew Mortona737b3e2006-03-22 00:08:11 -08004016 * If we cannot acquire the cache chain mutex then just give up - we'll try
4017 * again on the next iteration.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004018 */
David Howells65f27f32006-11-22 14:55:48 +00004019static void cache_reap(struct work_struct *unused)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004020{
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004021 struct kmem_cache *searchp;
Christoph Lametere498be72005-09-09 13:03:32 -07004022 struct kmem_list3 *l3;
Christoph Lameteraab22072006-03-22 00:09:06 -08004023 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004024
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004025 if (!mutex_trylock(&cache_chain_mutex)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004026 /* Give up. Setup the next iteration. */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004027 schedule_delayed_work(&__get_cpu_var(reap_work),
Arjan van de Ven2b284212006-12-10 02:21:28 -08004028 round_jiffies_relative(REAPTIMEOUT_CPUC));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004029 return;
4030 }
4031
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004032 list_for_each_entry(searchp, &cache_chain, next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004033 check_irq_on();
4034
Christoph Lameter35386e32006-03-22 00:09:05 -08004035 /*
4036 * We only take the l3 lock if absolutely necessary and we
4037 * have established with reasonable certainty that
4038 * we can do some work if the lock was obtained.
4039 */
Christoph Lameteraab22072006-03-22 00:09:06 -08004040 l3 = searchp->nodelists[node];
Christoph Lameter35386e32006-03-22 00:09:05 -08004041
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004042 reap_alien(searchp, l3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004043
Christoph Lameteraab22072006-03-22 00:09:06 -08004044 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004045
Christoph Lameter35386e32006-03-22 00:09:05 -08004046 /*
4047 * These are racy checks but it does not matter
4048 * if we skip one check or scan twice.
4049 */
Christoph Lametere498be72005-09-09 13:03:32 -07004050 if (time_after(l3->next_reap, jiffies))
Christoph Lameter35386e32006-03-22 00:09:05 -08004051 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004052
Christoph Lametere498be72005-09-09 13:03:32 -07004053 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004054
Christoph Lameteraab22072006-03-22 00:09:06 -08004055 drain_array(searchp, l3, l3->shared, 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004056
Christoph Lametered11d9e2006-06-30 01:55:45 -07004057 if (l3->free_touched)
Christoph Lametere498be72005-09-09 13:03:32 -07004058 l3->free_touched = 0;
Christoph Lametered11d9e2006-06-30 01:55:45 -07004059 else {
4060 int freed;
4061
4062 freed = drain_freelist(searchp, l3, (l3->free_limit +
4063 5 * searchp->num - 1) / (5 * searchp->num));
4064 STATS_ADD_REAPED(searchp, freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004065 }
Christoph Lameter35386e32006-03-22 00:09:05 -08004066next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004067 cond_resched();
4068 }
4069 check_irq_on();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004070 mutex_unlock(&cache_chain_mutex);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004071 next_reap_node();
Christoph Lameter2244b952006-06-30 01:55:33 -07004072 refresh_cpu_vm_stats(smp_processor_id());
Andrew Mortona737b3e2006-03-22 00:08:11 -08004073 /* Set up the next iteration */
Arjan van de Ven2b284212006-12-10 02:21:28 -08004074 schedule_delayed_work(&__get_cpu_var(reap_work),
4075 round_jiffies_relative(REAPTIMEOUT_CPUC));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004076}
4077
4078#ifdef CONFIG_PROC_FS
4079
Pekka Enberg85289f92006-01-08 01:00:36 -08004080static void print_slabinfo_header(struct seq_file *m)
4081{
4082 /*
4083 * Output format version, so at least we can change it
4084 * without _too_ many complaints.
4085 */
4086#if STATS
4087 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4088#else
4089 seq_puts(m, "slabinfo - version: 2.1\n");
4090#endif
4091 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4092 "<objperslab> <pagesperslab>");
4093 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4094 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4095#if STATS
4096 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004097 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
Pekka Enberg85289f92006-01-08 01:00:36 -08004098 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4099#endif
4100 seq_putc(m, '\n');
4101}
4102
Linus Torvalds1da177e2005-04-16 15:20:36 -07004103static void *s_start(struct seq_file *m, loff_t *pos)
4104{
4105 loff_t n = *pos;
4106 struct list_head *p;
4107
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004108 mutex_lock(&cache_chain_mutex);
Pekka Enberg85289f92006-01-08 01:00:36 -08004109 if (!n)
4110 print_slabinfo_header(m);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004111 p = cache_chain.next;
4112 while (n--) {
4113 p = p->next;
4114 if (p == &cache_chain)
4115 return NULL;
4116 }
Pekka Enberg343e0d72006-02-01 03:05:50 -08004117 return list_entry(p, struct kmem_cache, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004118}
4119
4120static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4121{
Pekka Enberg343e0d72006-02-01 03:05:50 -08004122 struct kmem_cache *cachep = p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004123 ++*pos;
Andrew Mortona737b3e2006-03-22 00:08:11 -08004124 return cachep->next.next == &cache_chain ?
4125 NULL : list_entry(cachep->next.next, struct kmem_cache, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004126}
4127
4128static void s_stop(struct seq_file *m, void *p)
4129{
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004130 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004131}
4132
4133static int s_show(struct seq_file *m, void *p)
4134{
Pekka Enberg343e0d72006-02-01 03:05:50 -08004135 struct kmem_cache *cachep = p;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004136 struct slab *slabp;
4137 unsigned long active_objs;
4138 unsigned long num_objs;
4139 unsigned long active_slabs = 0;
4140 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004141 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004142 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07004143 int node;
4144 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004145
Linus Torvalds1da177e2005-04-16 15:20:36 -07004146 active_objs = 0;
4147 num_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004148 for_each_online_node(node) {
4149 l3 = cachep->nodelists[node];
4150 if (!l3)
4151 continue;
4152
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004153 check_irq_on();
4154 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07004155
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004156 list_for_each_entry(slabp, &l3->slabs_full, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004157 if (slabp->inuse != cachep->num && !error)
4158 error = "slabs_full accounting error";
4159 active_objs += cachep->num;
4160 active_slabs++;
4161 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004162 list_for_each_entry(slabp, &l3->slabs_partial, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004163 if (slabp->inuse == cachep->num && !error)
4164 error = "slabs_partial inuse accounting error";
4165 if (!slabp->inuse && !error)
4166 error = "slabs_partial/inuse accounting error";
4167 active_objs += slabp->inuse;
4168 active_slabs++;
4169 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004170 list_for_each_entry(slabp, &l3->slabs_free, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004171 if (slabp->inuse && !error)
4172 error = "slabs_free/inuse accounting error";
4173 num_slabs++;
4174 }
4175 free_objects += l3->free_objects;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08004176 if (l3->shared)
4177 shared_avail += l3->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07004178
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004179 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004180 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004181 num_slabs += active_slabs;
4182 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07004183 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004184 error = "free_objects accounting error";
4185
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004186 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004187 if (error)
4188 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4189
4190 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004191 name, active_objs, num_objs, cachep->buffer_size,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004192 cachep->num, (1 << cachep->gfporder));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004193 seq_printf(m, " : tunables %4u %4u %4u",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004194 cachep->limit, cachep->batchcount, cachep->shared);
Christoph Lametere498be72005-09-09 13:03:32 -07004195 seq_printf(m, " : slabdata %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004196 active_slabs, num_slabs, shared_avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004197#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004198 { /* list3 stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004199 unsigned long high = cachep->high_mark;
4200 unsigned long allocs = cachep->num_allocations;
4201 unsigned long grown = cachep->grown;
4202 unsigned long reaped = cachep->reaped;
4203 unsigned long errors = cachep->errors;
4204 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004205 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07004206 unsigned long node_frees = cachep->node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004207 unsigned long overflows = cachep->node_overflow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004208
Christoph Lametere498be72005-09-09 13:03:32 -07004209 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004210 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown,
Andrew Mortona737b3e2006-03-22 00:08:11 -08004211 reaped, errors, max_freeable, node_allocs,
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004212 node_frees, overflows);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004213 }
4214 /* cpu stats */
4215 {
4216 unsigned long allochit = atomic_read(&cachep->allochit);
4217 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4218 unsigned long freehit = atomic_read(&cachep->freehit);
4219 unsigned long freemiss = atomic_read(&cachep->freemiss);
4220
4221 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004222 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004223 }
4224#endif
4225 seq_putc(m, '\n');
Linus Torvalds1da177e2005-04-16 15:20:36 -07004226 return 0;
4227}
4228
4229/*
4230 * slabinfo_op - iterator that generates /proc/slabinfo
4231 *
4232 * Output layout:
4233 * cache-name
4234 * num-active-objs
4235 * total-objs
4236 * object size
4237 * num-active-slabs
4238 * total-slabs
4239 * num-pages-per-slab
4240 * + further values on SMP and with statistics enabled
4241 */
4242
Helge Deller15ad7cd2006-12-06 20:40:36 -08004243const struct seq_operations slabinfo_op = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004244 .start = s_start,
4245 .next = s_next,
4246 .stop = s_stop,
4247 .show = s_show,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004248};
4249
4250#define MAX_SLABINFO_WRITE 128
4251/**
4252 * slabinfo_write - Tuning for the slab allocator
4253 * @file: unused
4254 * @buffer: user buffer
4255 * @count: data length
4256 * @ppos: unused
4257 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004258ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4259 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004260{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004261 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004262 int limit, batchcount, shared, res;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004263 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004264
Linus Torvalds1da177e2005-04-16 15:20:36 -07004265 if (count > MAX_SLABINFO_WRITE)
4266 return -EINVAL;
4267 if (copy_from_user(&kbuf, buffer, count))
4268 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004269 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07004270
4271 tmp = strchr(kbuf, ' ');
4272 if (!tmp)
4273 return -EINVAL;
4274 *tmp = '\0';
4275 tmp++;
4276 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4277 return -EINVAL;
4278
4279 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004280 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004281 res = -EINVAL;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004282 list_for_each_entry(cachep, &cache_chain, next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004283 if (!strcmp(cachep->name, kbuf)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08004284 if (limit < 1 || batchcount < 1 ||
4285 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07004286 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004287 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07004288 res = do_tune_cpucache(cachep, limit,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004289 batchcount, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004290 }
4291 break;
4292 }
4293 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004294 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004295 if (res >= 0)
4296 res = count;
4297 return res;
4298}
Al Viro871751e2006-03-25 03:06:39 -08004299
4300#ifdef CONFIG_DEBUG_SLAB_LEAK
4301
4302static void *leaks_start(struct seq_file *m, loff_t *pos)
4303{
4304 loff_t n = *pos;
4305 struct list_head *p;
4306
4307 mutex_lock(&cache_chain_mutex);
4308 p = cache_chain.next;
4309 while (n--) {
4310 p = p->next;
4311 if (p == &cache_chain)
4312 return NULL;
4313 }
4314 return list_entry(p, struct kmem_cache, next);
4315}
4316
4317static inline int add_caller(unsigned long *n, unsigned long v)
4318{
4319 unsigned long *p;
4320 int l;
4321 if (!v)
4322 return 1;
4323 l = n[1];
4324 p = n + 2;
4325 while (l) {
4326 int i = l/2;
4327 unsigned long *q = p + 2 * i;
4328 if (*q == v) {
4329 q[1]++;
4330 return 1;
4331 }
4332 if (*q > v) {
4333 l = i;
4334 } else {
4335 p = q + 2;
4336 l -= i + 1;
4337 }
4338 }
4339 if (++n[1] == n[0])
4340 return 0;
4341 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4342 p[0] = v;
4343 p[1] = 1;
4344 return 1;
4345}
4346
4347static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4348{
4349 void *p;
4350 int i;
4351 if (n[0] == n[1])
4352 return;
4353 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4354 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4355 continue;
4356 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4357 return;
4358 }
4359}
4360
4361static void show_symbol(struct seq_file *m, unsigned long address)
4362{
4363#ifdef CONFIG_KALLSYMS
4364 char *modname;
4365 const char *name;
4366 unsigned long offset, size;
4367 char namebuf[KSYM_NAME_LEN+1];
4368
4369 name = kallsyms_lookup(address, &size, &offset, &modname, namebuf);
4370
4371 if (name) {
4372 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4373 if (modname)
4374 seq_printf(m, " [%s]", modname);
4375 return;
4376 }
4377#endif
4378 seq_printf(m, "%p", (void *)address);
4379}
4380
4381static int leaks_show(struct seq_file *m, void *p)
4382{
4383 struct kmem_cache *cachep = p;
Al Viro871751e2006-03-25 03:06:39 -08004384 struct slab *slabp;
4385 struct kmem_list3 *l3;
4386 const char *name;
4387 unsigned long *n = m->private;
4388 int node;
4389 int i;
4390
4391 if (!(cachep->flags & SLAB_STORE_USER))
4392 return 0;
4393 if (!(cachep->flags & SLAB_RED_ZONE))
4394 return 0;
4395
4396 /* OK, we can do it */
4397
4398 n[1] = 0;
4399
4400 for_each_online_node(node) {
4401 l3 = cachep->nodelists[node];
4402 if (!l3)
4403 continue;
4404
4405 check_irq_on();
4406 spin_lock_irq(&l3->list_lock);
4407
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004408 list_for_each_entry(slabp, &l3->slabs_full, list)
Al Viro871751e2006-03-25 03:06:39 -08004409 handle_slab(n, cachep, slabp);
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004410 list_for_each_entry(slabp, &l3->slabs_partial, list)
Al Viro871751e2006-03-25 03:06:39 -08004411 handle_slab(n, cachep, slabp);
Al Viro871751e2006-03-25 03:06:39 -08004412 spin_unlock_irq(&l3->list_lock);
4413 }
4414 name = cachep->name;
4415 if (n[0] == n[1]) {
4416 /* Increase the buffer size */
4417 mutex_unlock(&cache_chain_mutex);
4418 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4419 if (!m->private) {
4420 /* Too bad, we are really out */
4421 m->private = n;
4422 mutex_lock(&cache_chain_mutex);
4423 return -ENOMEM;
4424 }
4425 *(unsigned long *)m->private = n[0] * 2;
4426 kfree(n);
4427 mutex_lock(&cache_chain_mutex);
4428 /* Now make sure this entry will be retried */
4429 m->count = m->size;
4430 return 0;
4431 }
4432 for (i = 0; i < n[1]; i++) {
4433 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4434 show_symbol(m, n[2*i+2]);
4435 seq_putc(m, '\n');
4436 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004437
Al Viro871751e2006-03-25 03:06:39 -08004438 return 0;
4439}
4440
Helge Deller15ad7cd2006-12-06 20:40:36 -08004441const struct seq_operations slabstats_op = {
Al Viro871751e2006-03-25 03:06:39 -08004442 .start = leaks_start,
4443 .next = s_next,
4444 .stop = s_stop,
4445 .show = leaks_show,
4446};
4447#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004448#endif
4449
Manfred Spraul00e145b2005-09-03 15:55:07 -07004450/**
4451 * ksize - get the actual amount of memory allocated for a given object
4452 * @objp: Pointer to the object
4453 *
4454 * kmalloc may internally round up allocations and return more memory
4455 * than requested. ksize() can be used to determine the actual amount of
4456 * memory allocated. The caller may use this additional memory, even though
4457 * a smaller amount of memory was initially specified with the kmalloc call.
4458 * The caller must guarantee that objp points to a valid object previously
4459 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4460 * must not be freed during the duration of the call.
4461 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004462unsigned int ksize(const void *objp)
4463{
Manfred Spraul00e145b2005-09-03 15:55:07 -07004464 if (unlikely(objp == NULL))
4465 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004466
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08004467 return obj_size(virt_to_cache(objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004468}