blob: 18e3164de09a9ce3b5c69b9fd53b6533b39e9486 [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
Simon Arlott183ff222007-10-20 01:27:18 +020029 * slabs and you must pass objects with the same initializations to
Linus Torvalds1da177e2005-04-16 15:20:36 -070030 * 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>
Alexey Dobriyana0ec95a82008-10-06 00:59:10 +040098#include <linux/proc_fs.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070099#include <linux/seq_file.h>
100#include <linux/notifier.h>
101#include <linux/kallsyms.h>
102#include <linux/cpu.h>
103#include <linux/sysctl.h>
104#include <linux/module.h>
Zhaolei02af61b2009-04-10 14:26:18 +0800105#include <linux/kmemtrace.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700106#include <linux/rcupdate.h>
Paulo Marques543537b2005-06-23 00:09:02 -0700107#include <linux/string.h>
Andrew Morton138ae662006-12-06 20:36:41 -0800108#include <linux/uaccess.h>
Christoph Lametere498be72005-09-09 13:03:32 -0700109#include <linux/nodemask.h>
Catalin Marinasd5cff632009-06-11 13:22:40 +0100110#include <linux/kmemleak.h>
Christoph Lameterdc85da12006-01-18 17:42:36 -0800111#include <linux/mempolicy.h>
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800112#include <linux/mutex.h>
Akinobu Mita8a8b6502006-12-08 02:39:44 -0800113#include <linux/fault-inject.h>
Ingo Molnare7eebaf2006-06-27 02:54:55 -0700114#include <linux/rtmutex.h>
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800115#include <linux/reciprocal_div.h>
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -0700116#include <linux/debugobjects.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700117
Linus Torvalds1da177e2005-04-16 15:20:36 -0700118#include <asm/cacheflush.h>
119#include <asm/tlbflush.h>
120#include <asm/page.h>
121
122/*
Christoph Lameter50953fe2007-05-06 14:50:16 -0700123 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700124 * 0 for faster, smaller code (especially in the critical paths).
125 *
126 * STATS - 1 to collect stats for /proc/slabinfo.
127 * 0 for faster, smaller code (especially in the critical paths).
128 *
129 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
130 */
131
132#ifdef CONFIG_DEBUG_SLAB
133#define DEBUG 1
134#define STATS 1
135#define FORCED_DEBUG 1
136#else
137#define DEBUG 0
138#define STATS 0
139#define FORCED_DEBUG 0
140#endif
141
Linus Torvalds1da177e2005-04-16 15:20:36 -0700142/* Shouldn't this be in a header file somewhere? */
143#define BYTES_PER_WORD sizeof(void *)
David Woodhouse87a927c2007-07-04 21:26:44 -0400144#define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700145
Linus Torvalds1da177e2005-04-16 15:20:36 -0700146#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
David Woodhouseb46b8f12007-05-08 00:22:59 -0700152 * alignment larger than the alignment of a 64-bit integer.
153 * ARCH_KMALLOC_MINALIGN allows that.
154 * Note that increasing this value may disable some debug features.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700155 */
David Woodhouseb46b8f12007-05-08 00:22:59 -0700156#define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700157#endif
158
159#ifndef ARCH_SLAB_MINALIGN
160/*
161 * Enforce a minimum alignment for all caches.
162 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
163 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
164 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
165 * some debug features.
166 */
167#define ARCH_SLAB_MINALIGN 0
168#endif
169
170#ifndef ARCH_KMALLOC_FLAGS
171#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
172#endif
173
174/* Legal flag mask for kmem_cache_create(). */
175#if DEBUG
Christoph Lameter50953fe2007-05-06 14:50:16 -0700176# define CREATE_MASK (SLAB_RED_ZONE | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700177 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
Christoph Lameterac2b8982006-03-22 00:08:15 -0800178 SLAB_CACHE_DMA | \
Christoph Lameter5af60832007-05-06 14:49:56 -0700179 SLAB_STORE_USER | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700180 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -0700181 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
Catalin Marinasd5cff632009-06-11 13:22:40 +0100182 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700183#else
Christoph Lameterac2b8982006-03-22 00:08:15 -0800184# define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
Christoph Lameter5af60832007-05-06 14:49:56 -0700185 SLAB_CACHE_DMA | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700186 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -0700187 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
Catalin Marinasd5cff632009-06-11 13:22:40 +0100188 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700189#endif
190
191/*
192 * kmem_bufctl_t:
193 *
194 * Bufctl's are used for linking objs within a slab
195 * linked offsets.
196 *
197 * This implementation relies on "struct page" for locating the cache &
198 * slab an object belongs to.
199 * This allows the bufctl structure to be small (one int), but limits
200 * the number of objects a slab (not a cache) can contain when off-slab
201 * bufctls are used. The limit is the size of the largest general cache
202 * that does not use off-slab slabs.
203 * For 32bit archs with 4 kB pages, is this 56.
204 * This is not serious, as it is only for large objects, when it is unwise
205 * to have too many per slab.
206 * Note: This limit can be raised by introducing a general cache whose size
207 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
208 */
209
Kyle Moffettfa5b08d2005-09-03 15:55:03 -0700210typedef unsigned int kmem_bufctl_t;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700211#define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
212#define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
Al Viro871751e2006-03-25 03:06:39 -0800213#define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
214#define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700215
Linus Torvalds1da177e2005-04-16 15:20:36 -0700216/*
217 * struct slab
218 *
219 * Manages the objs in a slab. Placed either at the beginning of mem allocated
220 * for a slab, or allocated from an general cache.
221 * Slabs are chained into three list: fully used, partial, fully free slabs.
222 */
223struct slab {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800224 struct list_head list;
225 unsigned long colouroff;
226 void *s_mem; /* including colour offset */
227 unsigned int inuse; /* num of objs active in slab */
228 kmem_bufctl_t free;
229 unsigned short nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700230};
231
232/*
233 * struct slab_rcu
234 *
235 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
236 * arrange for kmem_freepages to be called via RCU. This is useful if
237 * we need to approach a kernel structure obliquely, from its address
238 * obtained without the usual locking. We can lock the structure to
239 * stabilize it and check it's still at the given address, only if we
240 * can be sure that the memory has not been meanwhile reused for some
241 * other kind of object (which our subsystem's lock might corrupt).
242 *
243 * rcu_read_lock before reading the address, then rcu_read_unlock after
244 * taking the spinlock within the structure expected at that address.
245 *
246 * We assume struct slab_rcu can overlay struct slab when destroying.
247 */
248struct slab_rcu {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800249 struct rcu_head head;
Pekka Enberg343e0d72006-02-01 03:05:50 -0800250 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800251 void *addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700252};
253
254/*
255 * struct array_cache
256 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700257 * Purpose:
258 * - LIFO ordering, to hand out cache-warm objects from _alloc
259 * - reduce the number of linked list operations
260 * - reduce spinlock operations
261 *
262 * The limit is stored in the per-cpu structure to reduce the data cache
263 * footprint.
264 *
265 */
266struct array_cache {
267 unsigned int avail;
268 unsigned int limit;
269 unsigned int batchcount;
270 unsigned int touched;
Christoph Lametere498be72005-09-09 13:03:32 -0700271 spinlock_t lock;
Robert P. J. Daybda5b652007-10-16 23:30:05 -0700272 void *entry[]; /*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800273 * Must have this definition in here for the proper
274 * alignment of array_cache. Also simplifies accessing
275 * the entries.
Andrew Mortona737b3e2006-03-22 00:08:11 -0800276 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700277};
278
Andrew Mortona737b3e2006-03-22 00:08:11 -0800279/*
280 * bootstrap: The caches do not work without cpuarrays anymore, but the
281 * cpuarrays are allocated from the generic caches...
Linus Torvalds1da177e2005-04-16 15:20:36 -0700282 */
283#define BOOT_CPUCACHE_ENTRIES 1
284struct arraycache_init {
285 struct array_cache cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800286 void *entries[BOOT_CPUCACHE_ENTRIES];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700287};
288
289/*
Christoph Lametere498be72005-09-09 13:03:32 -0700290 * The slab lists for all objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700291 */
292struct kmem_list3 {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800293 struct list_head slabs_partial; /* partial list first, better asm code */
294 struct list_head slabs_full;
295 struct list_head slabs_free;
296 unsigned long free_objects;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800297 unsigned int free_limit;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800298 unsigned int colour_next; /* Per-node cache coloring */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800299 spinlock_t list_lock;
300 struct array_cache *shared; /* shared per node */
301 struct array_cache **alien; /* on other nodes */
Christoph Lameter35386e32006-03-22 00:09:05 -0800302 unsigned long next_reap; /* updated without locking */
303 int free_touched; /* updated without locking */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700304};
305
Christoph Lametere498be72005-09-09 13:03:32 -0700306/*
Pekka Enberg7e85ee02009-06-12 14:03:06 +0300307 * The slab allocator is initialized with interrupts disabled. Therefore, make
308 * sure early boot allocations don't accidentally enable interrupts.
309 */
310static gfp_t slab_gfp_mask __read_mostly = SLAB_GFP_BOOT_MASK;
311
312/*
Christoph Lametere498be72005-09-09 13:03:32 -0700313 * Need this for bootstrapping a per node allocator.
314 */
Pekka Enberg556a1692008-01-25 08:20:51 +0200315#define NUM_INIT_LISTS (3 * MAX_NUMNODES)
Christoph Lametere498be72005-09-09 13:03:32 -0700316struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
317#define CACHE_CACHE 0
Pekka Enberg556a1692008-01-25 08:20:51 +0200318#define SIZE_AC MAX_NUMNODES
319#define SIZE_L3 (2 * MAX_NUMNODES)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700320
Christoph Lametered11d9e2006-06-30 01:55:45 -0700321static int drain_freelist(struct kmem_cache *cache,
322 struct kmem_list3 *l3, int tofree);
323static void free_block(struct kmem_cache *cachep, void **objpp, int len,
324 int node);
Pekka Enberg83b519e2009-06-10 19:40:04 +0300325static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
David Howells65f27f32006-11-22 14:55:48 +0000326static void cache_reap(struct work_struct *unused);
Christoph Lametered11d9e2006-06-30 01:55:45 -0700327
Christoph Lametere498be72005-09-09 13:03:32 -0700328/*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800329 * This function must be completely optimized away if a constant is passed to
330 * it. Mostly the same as what is in linux/slab.h except it returns an index.
Christoph Lametere498be72005-09-09 13:03:32 -0700331 */
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700332static __always_inline int index_of(const size_t size)
Christoph Lametere498be72005-09-09 13:03:32 -0700333{
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800334 extern void __bad_size(void);
335
Christoph Lametere498be72005-09-09 13:03:32 -0700336 if (__builtin_constant_p(size)) {
337 int i = 0;
338
339#define CACHE(x) \
340 if (size <=x) \
341 return i; \
342 else \
343 i++;
Joe Perches1c61fc42008-03-05 13:58:17 -0800344#include <linux/kmalloc_sizes.h>
Christoph Lametere498be72005-09-09 13:03:32 -0700345#undef CACHE
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800346 __bad_size();
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700347 } else
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800348 __bad_size();
Christoph Lametere498be72005-09-09 13:03:32 -0700349 return 0;
350}
351
Ingo Molnare0a42722006-06-23 02:03:46 -0700352static int slab_early_init = 1;
353
Christoph Lametere498be72005-09-09 13:03:32 -0700354#define INDEX_AC index_of(sizeof(struct arraycache_init))
355#define INDEX_L3 index_of(sizeof(struct kmem_list3))
356
Pekka Enberg5295a742006-02-01 03:05:48 -0800357static void kmem_list3_init(struct kmem_list3 *parent)
Christoph Lametere498be72005-09-09 13:03:32 -0700358{
359 INIT_LIST_HEAD(&parent->slabs_full);
360 INIT_LIST_HEAD(&parent->slabs_partial);
361 INIT_LIST_HEAD(&parent->slabs_free);
362 parent->shared = NULL;
363 parent->alien = NULL;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800364 parent->colour_next = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700365 spin_lock_init(&parent->list_lock);
366 parent->free_objects = 0;
367 parent->free_touched = 0;
368}
369
Andrew Mortona737b3e2006-03-22 00:08:11 -0800370#define MAKE_LIST(cachep, listp, slab, nodeid) \
371 do { \
372 INIT_LIST_HEAD(listp); \
373 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
Christoph Lametere498be72005-09-09 13:03:32 -0700374 } while (0)
375
Andrew Mortona737b3e2006-03-22 00:08:11 -0800376#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
377 do { \
Christoph Lametere498be72005-09-09 13:03:32 -0700378 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
379 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
380 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
381 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700382
383/*
Pekka Enberg343e0d72006-02-01 03:05:50 -0800384 * struct kmem_cache
Linus Torvalds1da177e2005-04-16 15:20:36 -0700385 *
386 * manages a cache.
387 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800388
Pekka J Enberg2109a2d2005-11-07 00:58:01 -0800389struct kmem_cache {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700390/* 1) per-cpu data, touched during every alloc/free */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800391 struct array_cache *array[NR_CPUS];
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800392/* 2) Cache tunables. Protected by cache_chain_mutex */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800393 unsigned int batchcount;
394 unsigned int limit;
395 unsigned int shared;
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800396
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800397 unsigned int buffer_size;
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800398 u32 reciprocal_buffer_size;
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800399/* 3) touched by every alloc & free from the backend */
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800400
Andrew Mortona737b3e2006-03-22 00:08:11 -0800401 unsigned int flags; /* constant flags */
402 unsigned int num; /* # of objs per slab */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700403
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800404/* 4) cache_grow/shrink */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700405 /* order of pgs per slab (2^n) */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800406 unsigned int gfporder;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700407
408 /* force GFP flags, e.g. GFP_DMA */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800409 gfp_t gfpflags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700410
Andrew Mortona737b3e2006-03-22 00:08:11 -0800411 size_t colour; /* cache colouring range */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800412 unsigned int colour_off; /* colour offset */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800413 struct kmem_cache *slabp_cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800414 unsigned int slab_size;
Andrew Mortona737b3e2006-03-22 00:08:11 -0800415 unsigned int dflags; /* dynamic flags */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700416
417 /* constructor func */
Alexey Dobriyan51cc5062008-07-25 19:45:34 -0700418 void (*ctor)(void *obj);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700419
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800420/* 5) cache creation/removal */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800421 const char *name;
422 struct list_head next;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700423
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800424/* 6) statistics */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700425#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800426 unsigned long num_active;
427 unsigned long num_allocations;
428 unsigned long high_mark;
429 unsigned long grown;
430 unsigned long reaped;
431 unsigned long errors;
432 unsigned long max_freeable;
433 unsigned long node_allocs;
434 unsigned long node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700435 unsigned long node_overflow;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800436 atomic_t allochit;
437 atomic_t allocmiss;
438 atomic_t freehit;
439 atomic_t freemiss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700440#endif
441#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800442 /*
443 * If debugging is enabled, then the allocator can add additional
444 * fields and/or padding to every object. buffer_size contains the total
445 * object size including these internal fields, the following two
446 * variables contain the offset to the user object and its size.
447 */
448 int obj_offset;
449 int obj_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700450#endif
Eric Dumazet8da34302007-05-06 14:49:29 -0700451 /*
452 * We put nodelists[] at the end of kmem_cache, because we want to size
453 * this array to nr_node_ids slots instead of MAX_NUMNODES
454 * (see kmem_cache_init())
455 * We still use [MAX_NUMNODES] and not [1] or [0] because cache_cache
456 * is statically defined, so we reserve the max number of nodes.
457 */
458 struct kmem_list3 *nodelists[MAX_NUMNODES];
459 /*
460 * Do not add fields after nodelists[]
461 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700462};
463
464#define CFLGS_OFF_SLAB (0x80000000UL)
465#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
466
467#define BATCHREFILL_LIMIT 16
Andrew Mortona737b3e2006-03-22 00:08:11 -0800468/*
469 * Optimization question: fewer reaps means less probability for unnessary
470 * cpucache drain/refill cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700471 *
Adrian Bunkdc6f3f22005-11-08 16:44:08 +0100472 * OTOH the cpuarrays can contain lots of objects,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700473 * which could lock up otherwise freeable slabs.
474 */
475#define REAPTIMEOUT_CPUC (2*HZ)
476#define REAPTIMEOUT_LIST3 (4*HZ)
477
478#if STATS
479#define STATS_INC_ACTIVE(x) ((x)->num_active++)
480#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
481#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
482#define STATS_INC_GROWN(x) ((x)->grown++)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700483#define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
Andrew Mortona737b3e2006-03-22 00:08:11 -0800484#define STATS_SET_HIGH(x) \
485 do { \
486 if ((x)->num_active > (x)->high_mark) \
487 (x)->high_mark = (x)->num_active; \
488 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700489#define STATS_INC_ERR(x) ((x)->errors++)
490#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
Christoph Lametere498be72005-09-09 13:03:32 -0700491#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700492#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800493#define STATS_SET_FREEABLE(x, i) \
494 do { \
495 if ((x)->max_freeable < i) \
496 (x)->max_freeable = i; \
497 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700498#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
499#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
500#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
501#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
502#else
503#define STATS_INC_ACTIVE(x) do { } while (0)
504#define STATS_DEC_ACTIVE(x) do { } while (0)
505#define STATS_INC_ALLOCED(x) do { } while (0)
506#define STATS_INC_GROWN(x) do { } while (0)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700507#define STATS_ADD_REAPED(x,y) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700508#define STATS_SET_HIGH(x) do { } while (0)
509#define STATS_INC_ERR(x) do { } while (0)
510#define STATS_INC_NODEALLOCS(x) do { } while (0)
Christoph Lametere498be72005-09-09 13:03:32 -0700511#define STATS_INC_NODEFREES(x) do { } while (0)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700512#define STATS_INC_ACOVERFLOW(x) do { } while (0)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800513#define STATS_SET_FREEABLE(x, i) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700514#define STATS_INC_ALLOCHIT(x) do { } while (0)
515#define STATS_INC_ALLOCMISS(x) do { } while (0)
516#define STATS_INC_FREEHIT(x) do { } while (0)
517#define STATS_INC_FREEMISS(x) do { } while (0)
518#endif
519
520#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -0700521
Andrew Mortona737b3e2006-03-22 00:08:11 -0800522/*
523 * memory layout of objects:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700524 * 0 : objp
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800525 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
Linus Torvalds1da177e2005-04-16 15:20:36 -0700526 * the end of an object is aligned with the end of the real
527 * allocation. Catches writes behind the end of the allocation.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800528 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700529 * redzone word.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800530 * cachep->obj_offset: The real object.
531 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
Andrew Mortona737b3e2006-03-22 00:08:11 -0800532 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
533 * [BYTES_PER_WORD long]
Linus Torvalds1da177e2005-04-16 15:20:36 -0700534 */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800535static int obj_offset(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700536{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800537 return cachep->obj_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700538}
539
Pekka Enberg343e0d72006-02-01 03:05:50 -0800540static int obj_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700541{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800542 return cachep->obj_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700543}
544
David Woodhouseb46b8f12007-05-08 00:22:59 -0700545static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700546{
547 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
David Woodhouseb46b8f12007-05-08 00:22:59 -0700548 return (unsigned long long*) (objp + obj_offset(cachep) -
549 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700550}
551
David Woodhouseb46b8f12007-05-08 00:22:59 -0700552static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700553{
554 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
555 if (cachep->flags & SLAB_STORE_USER)
David Woodhouseb46b8f12007-05-08 00:22:59 -0700556 return (unsigned long long *)(objp + cachep->buffer_size -
557 sizeof(unsigned long long) -
David Woodhouse87a927c2007-07-04 21:26:44 -0400558 REDZONE_ALIGN);
David Woodhouseb46b8f12007-05-08 00:22:59 -0700559 return (unsigned long long *) (objp + cachep->buffer_size -
560 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700561}
562
Pekka Enberg343e0d72006-02-01 03:05:50 -0800563static void **dbg_userword(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700564{
565 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800566 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700567}
568
569#else
570
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800571#define obj_offset(x) 0
572#define obj_size(cachep) (cachep->buffer_size)
David Woodhouseb46b8f12007-05-08 00:22:59 -0700573#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
574#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
Linus Torvalds1da177e2005-04-16 15:20:36 -0700575#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
576
577#endif
578
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +0300579#ifdef CONFIG_KMEMTRACE
580size_t slab_buffer_size(struct kmem_cache *cachep)
581{
582 return cachep->buffer_size;
583}
584EXPORT_SYMBOL(slab_buffer_size);
585#endif
586
Linus Torvalds1da177e2005-04-16 15:20:36 -0700587/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700588 * Do not go above this order unless 0 objects fit into the slab.
589 */
590#define BREAK_GFP_ORDER_HI 1
591#define BREAK_GFP_ORDER_LO 0
592static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
593
Andrew Mortona737b3e2006-03-22 00:08:11 -0800594/*
595 * Functions for storing/retrieving the cachep and or slab from the page
596 * allocator. These are used to find the slab an obj belongs to. With kfree(),
597 * these are used to find the cache which an obj belongs to.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700598 */
Pekka Enberg065d41c2005-11-13 16:06:46 -0800599static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
600{
601 page->lru.next = (struct list_head *)cache;
602}
603
604static inline struct kmem_cache *page_get_cache(struct page *page)
605{
Christoph Lameterd85f3382007-05-06 14:49:39 -0700606 page = compound_head(page);
Pekka Enbergddc2e812006-06-23 02:03:40 -0700607 BUG_ON(!PageSlab(page));
Pekka Enberg065d41c2005-11-13 16:06:46 -0800608 return (struct kmem_cache *)page->lru.next;
609}
610
611static inline void page_set_slab(struct page *page, struct slab *slab)
612{
613 page->lru.prev = (struct list_head *)slab;
614}
615
616static inline struct slab *page_get_slab(struct page *page)
617{
Pekka Enbergddc2e812006-06-23 02:03:40 -0700618 BUG_ON(!PageSlab(page));
Pekka Enberg065d41c2005-11-13 16:06:46 -0800619 return (struct slab *)page->lru.prev;
620}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700621
Pekka Enberg6ed5eb22006-02-01 03:05:49 -0800622static inline struct kmem_cache *virt_to_cache(const void *obj)
623{
Christoph Lameterb49af682007-05-06 14:49:41 -0700624 struct page *page = virt_to_head_page(obj);
Pekka Enberg6ed5eb22006-02-01 03:05:49 -0800625 return page_get_cache(page);
626}
627
628static inline struct slab *virt_to_slab(const void *obj)
629{
Christoph Lameterb49af682007-05-06 14:49:41 -0700630 struct page *page = virt_to_head_page(obj);
Pekka Enberg6ed5eb22006-02-01 03:05:49 -0800631 return page_get_slab(page);
632}
633
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800634static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
635 unsigned int idx)
636{
637 return slab->s_mem + cache->buffer_size * idx;
638}
639
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800640/*
641 * We want to avoid an expensive divide : (offset / cache->buffer_size)
642 * Using the fact that buffer_size is a constant for a particular cache,
643 * we can replace (offset / cache->buffer_size) by
644 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
645 */
646static inline unsigned int obj_to_index(const struct kmem_cache *cache,
647 const struct slab *slab, void *obj)
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800648{
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800649 u32 offset = (obj - slab->s_mem);
650 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800651}
652
Andrew Mortona737b3e2006-03-22 00:08:11 -0800653/*
654 * These are the default caches for kmalloc. Custom caches can have other sizes.
655 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700656struct cache_sizes malloc_sizes[] = {
657#define CACHE(x) { .cs_size = (x) },
658#include <linux/kmalloc_sizes.h>
659 CACHE(ULONG_MAX)
660#undef CACHE
661};
662EXPORT_SYMBOL(malloc_sizes);
663
664/* Must match cache_sizes above. Out of line to keep cache footprint low. */
665struct cache_names {
666 char *name;
667 char *name_dma;
668};
669
670static struct cache_names __initdata cache_names[] = {
671#define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
672#include <linux/kmalloc_sizes.h>
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800673 {NULL,}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700674#undef CACHE
675};
676
677static struct arraycache_init initarray_cache __initdata =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800678 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700679static struct arraycache_init initarray_generic =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800680 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700681
682/* internal cache of cache description objs */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800683static struct kmem_cache cache_cache = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800684 .batchcount = 1,
685 .limit = BOOT_CPUCACHE_ENTRIES,
686 .shared = 1,
Pekka Enberg343e0d72006-02-01 03:05:50 -0800687 .buffer_size = sizeof(struct kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800688 .name = "kmem_cache",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700689};
690
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700691#define BAD_ALIEN_MAGIC 0x01020304ul
692
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200693#ifdef CONFIG_LOCKDEP
694
695/*
696 * Slab sometimes uses the kmalloc slabs to store the slab headers
697 * for other slabs "off slab".
698 * The locking for this is tricky in that it nests within the locks
699 * of all other slabs in a few places; to deal with this special
700 * locking we put on-slab caches into a separate lock-class.
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700701 *
702 * We set lock class for alien array caches which are up during init.
703 * The lock annotation will be lost if all cpus of a node goes down and
704 * then comes back up during hotplug
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200705 */
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700706static struct lock_class_key on_slab_l3_key;
707static struct lock_class_key on_slab_alc_key;
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200708
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700709static inline void init_lock_keys(void)
710
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200711{
712 int q;
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700713 struct cache_sizes *s = malloc_sizes;
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200714
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700715 while (s->cs_size != ULONG_MAX) {
716 for_each_node(q) {
717 struct array_cache **alc;
718 int r;
719 struct kmem_list3 *l3 = s->cs_cachep->nodelists[q];
720 if (!l3 || OFF_SLAB(s->cs_cachep))
721 continue;
722 lockdep_set_class(&l3->list_lock, &on_slab_l3_key);
723 alc = l3->alien;
724 /*
725 * FIXME: This check for BAD_ALIEN_MAGIC
726 * should go away when common slab code is taught to
727 * work even without alien caches.
728 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
729 * for alloc_alien_cache,
730 */
731 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
732 continue;
733 for_each_node(r) {
734 if (alc[r])
735 lockdep_set_class(&alc[r]->lock,
736 &on_slab_alc_key);
737 }
738 }
739 s++;
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200740 }
741}
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200742#else
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700743static inline void init_lock_keys(void)
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200744{
745}
746#endif
747
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -0800748/*
Gautham R Shenoy95402b32008-01-25 21:08:02 +0100749 * Guard access to the cache-chain.
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -0800750 */
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800751static DEFINE_MUTEX(cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700752static struct list_head cache_chain;
753
754/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700755 * chicken and egg problem: delay the per-cpu array allocation
756 * until the general caches are up.
757 */
758static enum {
759 NONE,
Christoph Lametere498be72005-09-09 13:03:32 -0700760 PARTIAL_AC,
761 PARTIAL_L3,
Pekka Enberg8429db52009-06-12 15:58:59 +0300762 EARLY,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700763 FULL
764} g_cpucache_up;
765
Mike Kravetz39d24e62006-05-15 09:44:13 -0700766/*
767 * used by boot code to determine if it can use slab based allocator
768 */
769int slab_is_available(void)
770{
Pekka Enberg8429db52009-06-12 15:58:59 +0300771 return g_cpucache_up >= EARLY;
Mike Kravetz39d24e62006-05-15 09:44:13 -0700772}
773
David Howells52bad642006-11-22 14:54:01 +0000774static DEFINE_PER_CPU(struct delayed_work, reap_work);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700775
Pekka Enberg343e0d72006-02-01 03:05:50 -0800776static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700777{
778 return cachep->array[smp_processor_id()];
779}
780
Andrew Mortona737b3e2006-03-22 00:08:11 -0800781static inline struct kmem_cache *__find_general_cachep(size_t size,
782 gfp_t gfpflags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700783{
784 struct cache_sizes *csizep = malloc_sizes;
785
786#if DEBUG
787 /* This happens if someone tries to call
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800788 * kmem_cache_create(), or __kmalloc(), before
789 * the generic caches are initialized.
790 */
Alok Katariac7e43c72005-09-14 12:17:53 -0700791 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700792#endif
Christoph Lameter6cb8f912007-07-17 04:03:22 -0700793 if (!size)
794 return ZERO_SIZE_PTR;
795
Linus Torvalds1da177e2005-04-16 15:20:36 -0700796 while (size > csizep->cs_size)
797 csizep++;
798
799 /*
Martin Hicks0abf40c2005-09-03 15:54:54 -0700800 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
Linus Torvalds1da177e2005-04-16 15:20:36 -0700801 * has cs_{dma,}cachep==NULL. Thus no special case
802 * for large kmalloc calls required.
803 */
Christoph Lameter4b51d662007-02-10 01:43:10 -0800804#ifdef CONFIG_ZONE_DMA
Linus Torvalds1da177e2005-04-16 15:20:36 -0700805 if (unlikely(gfpflags & GFP_DMA))
806 return csizep->cs_dmacachep;
Christoph Lameter4b51d662007-02-10 01:43:10 -0800807#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700808 return csizep->cs_cachep;
809}
810
Adrian Bunkb2213852006-09-25 23:31:02 -0700811static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700812{
813 return __find_general_cachep(size, gfpflags);
814}
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700815
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800816static size_t slab_mgmt_size(size_t nr_objs, size_t align)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700817{
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800818 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
819}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700820
Andrew Mortona737b3e2006-03-22 00:08:11 -0800821/*
822 * Calculate the number of objects and left-over bytes for a given buffer size.
823 */
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800824static void cache_estimate(unsigned long gfporder, size_t buffer_size,
825 size_t align, int flags, size_t *left_over,
826 unsigned int *num)
827{
828 int nr_objs;
829 size_t mgmt_size;
830 size_t slab_size = PAGE_SIZE << gfporder;
831
832 /*
833 * The slab management structure can be either off the slab or
834 * on it. For the latter case, the memory allocated for a
835 * slab is used for:
836 *
837 * - The struct slab
838 * - One kmem_bufctl_t for each object
839 * - Padding to respect alignment of @align
840 * - @buffer_size bytes for each object
841 *
842 * If the slab management structure is off the slab, then the
843 * alignment will already be calculated into the size. Because
844 * the slabs are all pages aligned, the objects will be at the
845 * correct alignment when allocated.
846 */
847 if (flags & CFLGS_OFF_SLAB) {
848 mgmt_size = 0;
849 nr_objs = slab_size / buffer_size;
850
851 if (nr_objs > SLAB_LIMIT)
852 nr_objs = SLAB_LIMIT;
853 } else {
854 /*
855 * Ignore padding for the initial guess. The padding
856 * is at most @align-1 bytes, and @buffer_size is at
857 * least @align. In the worst case, this result will
858 * be one greater than the number of objects that fit
859 * into the memory allocation when taking the padding
860 * into account.
861 */
862 nr_objs = (slab_size - sizeof(struct slab)) /
863 (buffer_size + sizeof(kmem_bufctl_t));
864
865 /*
866 * This calculated number will be either the right
867 * amount, or one greater than what we want.
868 */
869 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
870 > slab_size)
871 nr_objs--;
872
873 if (nr_objs > SLAB_LIMIT)
874 nr_objs = SLAB_LIMIT;
875
876 mgmt_size = slab_mgmt_size(nr_objs, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700877 }
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800878 *num = nr_objs;
879 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700880}
881
Harvey Harrisond40cee22008-04-30 00:55:07 -0700882#define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700883
Andrew Mortona737b3e2006-03-22 00:08:11 -0800884static void __slab_error(const char *function, struct kmem_cache *cachep,
885 char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700886{
887 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800888 function, cachep->name, msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700889 dump_stack();
890}
891
Paul Menage3395ee02006-12-06 20:32:16 -0800892/*
893 * By default on NUMA we use alien caches to stage the freeing of
894 * objects allocated from other nodes. This causes massive memory
895 * inefficiencies when using fake NUMA setup to split memory into a
896 * large number of small nodes, so it can be disabled on the command
897 * line
898 */
899
900static int use_alien_caches __read_mostly = 1;
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -0700901static int numa_platform __read_mostly = 1;
Paul Menage3395ee02006-12-06 20:32:16 -0800902static int __init noaliencache_setup(char *s)
903{
904 use_alien_caches = 0;
905 return 1;
906}
907__setup("noaliencache", noaliencache_setup);
908
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800909#ifdef CONFIG_NUMA
910/*
911 * Special reaping functions for NUMA systems called from cache_reap().
912 * These take care of doing round robin flushing of alien caches (containing
913 * objects freed on different nodes from which they were allocated) and the
914 * flushing of remote pcps by calling drain_node_pages.
915 */
916static DEFINE_PER_CPU(unsigned long, reap_node);
917
918static void init_reap_node(int cpu)
919{
920 int node;
921
922 node = next_node(cpu_to_node(cpu), node_online_map);
923 if (node == MAX_NUMNODES)
Paul Jackson442295c2006-03-22 00:09:11 -0800924 node = first_node(node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800925
Daniel Yeisley7f6b8872006-11-02 22:07:14 -0800926 per_cpu(reap_node, cpu) = node;
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800927}
928
929static void next_reap_node(void)
930{
931 int node = __get_cpu_var(reap_node);
932
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800933 node = next_node(node, node_online_map);
934 if (unlikely(node >= MAX_NUMNODES))
935 node = first_node(node_online_map);
936 __get_cpu_var(reap_node) = node;
937}
938
939#else
940#define init_reap_node(cpu) do { } while (0)
941#define next_reap_node(void) do { } while (0)
942#endif
943
Linus Torvalds1da177e2005-04-16 15:20:36 -0700944/*
945 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
946 * via the workqueue/eventd.
947 * Add the CPU number into the expiration time to minimize the possibility of
948 * the CPUs getting into lockstep and contending for the global cache chain
949 * lock.
950 */
Adrian Bunk897e6792007-07-15 23:38:20 -0700951static void __cpuinit start_cpu_timer(int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700952{
David Howells52bad642006-11-22 14:54:01 +0000953 struct delayed_work *reap_work = &per_cpu(reap_work, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700954
955 /*
956 * When this gets called from do_initcalls via cpucache_init(),
957 * init_workqueues() has already run, so keventd will be setup
958 * at that time.
959 */
David Howells52bad642006-11-22 14:54:01 +0000960 if (keventd_up() && reap_work->work.func == NULL) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800961 init_reap_node(cpu);
David Howells65f27f32006-11-22 14:55:48 +0000962 INIT_DELAYED_WORK(reap_work, cache_reap);
Arjan van de Ven2b284212006-12-10 02:21:28 -0800963 schedule_delayed_work_on(cpu, reap_work,
964 __round_jiffies_relative(HZ, cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700965 }
966}
967
Christoph Lametere498be72005-09-09 13:03:32 -0700968static struct array_cache *alloc_arraycache(int node, int entries,
Pekka Enberg83b519e2009-06-10 19:40:04 +0300969 int batchcount, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700970{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800971 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700972 struct array_cache *nc = NULL;
973
Pekka Enberg83b519e2009-06-10 19:40:04 +0300974 nc = kmalloc_node(memsize, gfp, node);
Catalin Marinasd5cff632009-06-11 13:22:40 +0100975 /*
976 * The array_cache structures contain pointers to free object.
977 * However, when such objects are allocated or transfered to another
978 * cache the pointers are not cleared and they could be counted as
979 * valid references during a kmemleak scan. Therefore, kmemleak must
980 * not scan such objects.
981 */
982 kmemleak_no_scan(nc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700983 if (nc) {
984 nc->avail = 0;
985 nc->limit = entries;
986 nc->batchcount = batchcount;
987 nc->touched = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700988 spin_lock_init(&nc->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700989 }
990 return nc;
991}
992
Christoph Lameter3ded1752006-03-25 03:06:44 -0800993/*
994 * Transfer objects in one arraycache to another.
995 * Locking must be handled by the caller.
996 *
997 * Return the number of entries transferred.
998 */
999static int transfer_objects(struct array_cache *to,
1000 struct array_cache *from, unsigned int max)
1001{
1002 /* Figure out how many entries to transfer */
1003 int nr = min(min(from->avail, max), to->limit - to->avail);
1004
1005 if (!nr)
1006 return 0;
1007
1008 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
1009 sizeof(void *) *nr);
1010
1011 from->avail -= nr;
1012 to->avail += nr;
1013 to->touched = 1;
1014 return nr;
1015}
1016
Christoph Lameter765c4502006-09-27 01:50:08 -07001017#ifndef CONFIG_NUMA
1018
1019#define drain_alien_cache(cachep, alien) do { } while (0)
1020#define reap_alien(cachep, l3) do { } while (0)
1021
Pekka Enberg83b519e2009-06-10 19:40:04 +03001022static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
Christoph Lameter765c4502006-09-27 01:50:08 -07001023{
1024 return (struct array_cache **)BAD_ALIEN_MAGIC;
1025}
1026
1027static inline void free_alien_cache(struct array_cache **ac_ptr)
1028{
1029}
1030
1031static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1032{
1033 return 0;
1034}
1035
1036static inline void *alternate_node_alloc(struct kmem_cache *cachep,
1037 gfp_t flags)
1038{
1039 return NULL;
1040}
1041
Christoph Hellwig8b98c162006-12-06 20:32:30 -08001042static inline void *____cache_alloc_node(struct kmem_cache *cachep,
Christoph Lameter765c4502006-09-27 01:50:08 -07001043 gfp_t flags, int nodeid)
1044{
1045 return NULL;
1046}
1047
1048#else /* CONFIG_NUMA */
1049
Christoph Hellwig8b98c162006-12-06 20:32:30 -08001050static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
Paul Jacksonc61afb12006-03-24 03:16:08 -08001051static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
Christoph Lameterdc85da12006-01-18 17:42:36 -08001052
Pekka Enberg83b519e2009-06-10 19:40:04 +03001053static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -07001054{
1055 struct array_cache **ac_ptr;
Christoph Lameter8ef82862007-02-20 13:57:52 -08001056 int memsize = sizeof(void *) * nr_node_ids;
Christoph Lametere498be72005-09-09 13:03:32 -07001057 int i;
1058
1059 if (limit > 1)
1060 limit = 12;
Pekka Enberg83b519e2009-06-10 19:40:04 +03001061 ac_ptr = kmalloc_node(memsize, gfp, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001062 if (ac_ptr) {
1063 for_each_node(i) {
1064 if (i == node || !node_online(i)) {
1065 ac_ptr[i] = NULL;
1066 continue;
1067 }
Pekka Enberg83b519e2009-06-10 19:40:04 +03001068 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
Christoph Lametere498be72005-09-09 13:03:32 -07001069 if (!ac_ptr[i]) {
Akinobu Mitacc550de2007-11-14 16:58:35 -08001070 for (i--; i >= 0; i--)
Christoph Lametere498be72005-09-09 13:03:32 -07001071 kfree(ac_ptr[i]);
1072 kfree(ac_ptr);
1073 return NULL;
1074 }
1075 }
1076 }
1077 return ac_ptr;
1078}
1079
Pekka Enberg5295a742006-02-01 03:05:48 -08001080static void free_alien_cache(struct array_cache **ac_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -07001081{
1082 int i;
1083
1084 if (!ac_ptr)
1085 return;
Christoph Lametere498be72005-09-09 13:03:32 -07001086 for_each_node(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001087 kfree(ac_ptr[i]);
Christoph Lametere498be72005-09-09 13:03:32 -07001088 kfree(ac_ptr);
1089}
1090
Pekka Enberg343e0d72006-02-01 03:05:50 -08001091static void __drain_alien_cache(struct kmem_cache *cachep,
Pekka Enberg5295a742006-02-01 03:05:48 -08001092 struct array_cache *ac, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07001093{
1094 struct kmem_list3 *rl3 = cachep->nodelists[node];
1095
1096 if (ac->avail) {
1097 spin_lock(&rl3->list_lock);
Christoph Lametere00946f2006-03-25 03:06:45 -08001098 /*
1099 * Stuff objects into the remote nodes shared array first.
1100 * That way we could avoid the overhead of putting the objects
1101 * into the free lists and getting them back later.
1102 */
shin, jacob693f7d32006-04-28 10:54:37 -05001103 if (rl3->shared)
1104 transfer_objects(rl3->shared, ac, ac->limit);
Christoph Lametere00946f2006-03-25 03:06:45 -08001105
Christoph Lameterff694162005-09-22 21:44:02 -07001106 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001107 ac->avail = 0;
1108 spin_unlock(&rl3->list_lock);
1109 }
1110}
1111
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001112/*
1113 * Called from cache_reap() to regularly drain alien caches round robin.
1114 */
1115static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1116{
1117 int node = __get_cpu_var(reap_node);
1118
1119 if (l3->alien) {
1120 struct array_cache *ac = l3->alien[node];
Christoph Lametere00946f2006-03-25 03:06:45 -08001121
1122 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001123 __drain_alien_cache(cachep, ac, node);
1124 spin_unlock_irq(&ac->lock);
1125 }
1126 }
1127}
1128
Andrew Mortona737b3e2006-03-22 00:08:11 -08001129static void drain_alien_cache(struct kmem_cache *cachep,
1130 struct array_cache **alien)
Christoph Lametere498be72005-09-09 13:03:32 -07001131{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001132 int i = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07001133 struct array_cache *ac;
1134 unsigned long flags;
1135
1136 for_each_online_node(i) {
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001137 ac = alien[i];
Christoph Lametere498be72005-09-09 13:03:32 -07001138 if (ac) {
1139 spin_lock_irqsave(&ac->lock, flags);
1140 __drain_alien_cache(cachep, ac, i);
1141 spin_unlock_irqrestore(&ac->lock, flags);
1142 }
1143 }
1144}
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001145
Ingo Molnar873623d2006-07-13 14:44:38 +02001146static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001147{
1148 struct slab *slabp = virt_to_slab(objp);
1149 int nodeid = slabp->nodeid;
1150 struct kmem_list3 *l3;
1151 struct array_cache *alien = NULL;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001152 int node;
1153
1154 node = numa_node_id();
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001155
1156 /*
1157 * Make sure we are not freeing a object from another node to the array
1158 * cache on this cpu.
1159 */
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001160 if (likely(slabp->nodeid == node))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001161 return 0;
1162
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001163 l3 = cachep->nodelists[node];
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001164 STATS_INC_NODEFREES(cachep);
1165 if (l3->alien && l3->alien[nodeid]) {
1166 alien = l3->alien[nodeid];
Ingo Molnar873623d2006-07-13 14:44:38 +02001167 spin_lock(&alien->lock);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001168 if (unlikely(alien->avail == alien->limit)) {
1169 STATS_INC_ACOVERFLOW(cachep);
1170 __drain_alien_cache(cachep, alien, nodeid);
1171 }
1172 alien->entry[alien->avail++] = objp;
1173 spin_unlock(&alien->lock);
1174 } else {
1175 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1176 free_block(cachep, &objp, 1, nodeid);
1177 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1178 }
1179 return 1;
1180}
Christoph Lametere498be72005-09-09 13:03:32 -07001181#endif
1182
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001183static void __cpuinit cpuup_canceled(long cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001184{
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001185 struct kmem_cache *cachep;
1186 struct kmem_list3 *l3 = NULL;
1187 int node = cpu_to_node(cpu);
Rusty Russella70f7302009-03-13 14:49:46 +10301188 const struct cpumask *mask = cpumask_of_node(node);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001189
1190 list_for_each_entry(cachep, &cache_chain, next) {
1191 struct array_cache *nc;
1192 struct array_cache *shared;
1193 struct array_cache **alien;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001194
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001195 /* cpu is dead; no one can alloc from it. */
1196 nc = cachep->array[cpu];
1197 cachep->array[cpu] = NULL;
1198 l3 = cachep->nodelists[node];
1199
1200 if (!l3)
1201 goto free_array_cache;
1202
1203 spin_lock_irq(&l3->list_lock);
1204
1205 /* Free limit for this kmem_list3 */
1206 l3->free_limit -= cachep->batchcount;
1207 if (nc)
1208 free_block(cachep, nc->entry, nc->avail, node);
1209
Mike Travisc5f59f02008-04-04 18:11:10 -07001210 if (!cpus_empty(*mask)) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001211 spin_unlock_irq(&l3->list_lock);
1212 goto free_array_cache;
1213 }
1214
1215 shared = l3->shared;
1216 if (shared) {
1217 free_block(cachep, shared->entry,
1218 shared->avail, node);
1219 l3->shared = NULL;
1220 }
1221
1222 alien = l3->alien;
1223 l3->alien = NULL;
1224
1225 spin_unlock_irq(&l3->list_lock);
1226
1227 kfree(shared);
1228 if (alien) {
1229 drain_alien_cache(cachep, alien);
1230 free_alien_cache(alien);
1231 }
1232free_array_cache:
1233 kfree(nc);
1234 }
1235 /*
1236 * In the previous loop, all the objects were freed to
1237 * the respective cache's slabs, now we can go ahead and
1238 * shrink each nodelist to its limit.
1239 */
1240 list_for_each_entry(cachep, &cache_chain, next) {
1241 l3 = cachep->nodelists[node];
1242 if (!l3)
1243 continue;
1244 drain_freelist(cachep, l3, l3->free_objects);
1245 }
1246}
1247
1248static int __cpuinit cpuup_prepare(long cpu)
1249{
Pekka Enberg343e0d72006-02-01 03:05:50 -08001250 struct kmem_cache *cachep;
Christoph Lametere498be72005-09-09 13:03:32 -07001251 struct kmem_list3 *l3 = NULL;
1252 int node = cpu_to_node(cpu);
David Howellsea02e3d2007-07-19 01:49:09 -07001253 const int memsize = sizeof(struct kmem_list3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001254
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001255 /*
1256 * We need to do this right in the beginning since
1257 * alloc_arraycache's are going to use this list.
1258 * kmalloc_node allows us to add the slab to the right
1259 * kmem_list3 and not this cpu's kmem_list3
1260 */
1261
1262 list_for_each_entry(cachep, &cache_chain, next) {
1263 /*
1264 * Set up the size64 kmemlist for cpu before we can
1265 * begin anything. Make sure some other cpu on this
1266 * node has not already allocated this
1267 */
1268 if (!cachep->nodelists[node]) {
1269 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1270 if (!l3)
1271 goto bad;
1272 kmem_list3_init(l3);
1273 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1274 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1275
1276 /*
1277 * The l3s don't come and go as CPUs come and
1278 * go. cache_chain_mutex is sufficient
1279 * protection here.
1280 */
1281 cachep->nodelists[node] = l3;
1282 }
1283
1284 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1285 cachep->nodelists[node]->free_limit =
1286 (1 + nr_cpus_node(node)) *
1287 cachep->batchcount + cachep->num;
1288 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1289 }
1290
1291 /*
1292 * Now we can go ahead with allocating the shared arrays and
1293 * array caches
1294 */
1295 list_for_each_entry(cachep, &cache_chain, next) {
1296 struct array_cache *nc;
1297 struct array_cache *shared = NULL;
1298 struct array_cache **alien = NULL;
1299
1300 nc = alloc_arraycache(node, cachep->limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03001301 cachep->batchcount, GFP_KERNEL);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001302 if (!nc)
1303 goto bad;
1304 if (cachep->shared) {
1305 shared = alloc_arraycache(node,
1306 cachep->shared * cachep->batchcount,
Pekka Enberg83b519e2009-06-10 19:40:04 +03001307 0xbaadf00d, GFP_KERNEL);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001308 if (!shared) {
1309 kfree(nc);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001310 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001311 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001312 }
1313 if (use_alien_caches) {
Pekka Enberg83b519e2009-06-10 19:40:04 +03001314 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001315 if (!alien) {
1316 kfree(shared);
1317 kfree(nc);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001318 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001319 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001320 }
1321 cachep->array[cpu] = nc;
1322 l3 = cachep->nodelists[node];
1323 BUG_ON(!l3);
1324
1325 spin_lock_irq(&l3->list_lock);
1326 if (!l3->shared) {
1327 /*
1328 * We are serialised from CPU_DEAD or
1329 * CPU_UP_CANCELLED by the cpucontrol lock
1330 */
1331 l3->shared = shared;
1332 shared = NULL;
1333 }
1334#ifdef CONFIG_NUMA
1335 if (!l3->alien) {
1336 l3->alien = alien;
1337 alien = NULL;
1338 }
1339#endif
1340 spin_unlock_irq(&l3->list_lock);
1341 kfree(shared);
1342 free_alien_cache(alien);
1343 }
1344 return 0;
1345bad:
Akinobu Mita12d00f62007-10-18 03:05:11 -07001346 cpuup_canceled(cpu);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001347 return -ENOMEM;
1348}
1349
1350static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1351 unsigned long action, void *hcpu)
1352{
1353 long cpu = (long)hcpu;
1354 int err = 0;
1355
Linus Torvalds1da177e2005-04-16 15:20:36 -07001356 switch (action) {
Heiko Carstens38c3bd92007-05-09 02:34:05 -07001357 case CPU_UP_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001358 case CPU_UP_PREPARE_FROZEN:
Gautham R Shenoy95402b32008-01-25 21:08:02 +01001359 mutex_lock(&cache_chain_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001360 err = cpuup_prepare(cpu);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01001361 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001362 break;
1363 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001364 case CPU_ONLINE_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001365 start_cpu_timer(cpu);
1366 break;
1367#ifdef CONFIG_HOTPLUG_CPU
Christoph Lameter5830c592007-05-09 02:34:22 -07001368 case CPU_DOWN_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001369 case CPU_DOWN_PREPARE_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001370 /*
1371 * Shutdown cache reaper. Note that the cache_chain_mutex is
1372 * held so that if cache_reap() is invoked it cannot do
1373 * anything expensive but will only modify reap_work
1374 * and reschedule the timer.
1375 */
1376 cancel_rearming_delayed_work(&per_cpu(reap_work, cpu));
1377 /* Now the cache_reaper is guaranteed to be not running. */
1378 per_cpu(reap_work, cpu).work.func = NULL;
1379 break;
1380 case CPU_DOWN_FAILED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001381 case CPU_DOWN_FAILED_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001382 start_cpu_timer(cpu);
1383 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001384 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001385 case CPU_DEAD_FROZEN:
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001386 /*
1387 * Even if all the cpus of a node are down, we don't free the
1388 * kmem_list3 of any cache. This to avoid a race between
1389 * cpu_down, and a kmalloc allocation from another cpu for
1390 * memory from the node of the cpu going down. The list3
1391 * structure is usually allocated from kmem_cache_create() and
1392 * gets destroyed at kmem_cache_destroy().
1393 */
Simon Arlott183ff222007-10-20 01:27:18 +02001394 /* fall through */
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08001395#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001396 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001397 case CPU_UP_CANCELED_FROZEN:
Gautham R Shenoy95402b32008-01-25 21:08:02 +01001398 mutex_lock(&cache_chain_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001399 cpuup_canceled(cpu);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001400 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001401 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001402 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001403 return err ? NOTIFY_BAD : NOTIFY_OK;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001404}
1405
Chandra Seetharaman74b85f32006-06-27 02:54:09 -07001406static struct notifier_block __cpuinitdata cpucache_notifier = {
1407 &cpuup_callback, NULL, 0
1408};
Linus Torvalds1da177e2005-04-16 15:20:36 -07001409
Christoph Lametere498be72005-09-09 13:03:32 -07001410/*
1411 * swap the static kmem_list3 with kmalloced memory
1412 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001413static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1414 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001415{
1416 struct kmem_list3 *ptr;
1417
Pekka Enberg83b519e2009-06-10 19:40:04 +03001418 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07001419 BUG_ON(!ptr);
1420
Christoph Lametere498be72005-09-09 13:03:32 -07001421 memcpy(ptr, list, sizeof(struct kmem_list3));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001422 /*
1423 * Do not assume that spinlocks can be initialized via memcpy:
1424 */
1425 spin_lock_init(&ptr->list_lock);
1426
Christoph Lametere498be72005-09-09 13:03:32 -07001427 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1428 cachep->nodelists[nodeid] = ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001429}
1430
Andrew Mortona737b3e2006-03-22 00:08:11 -08001431/*
Pekka Enberg556a1692008-01-25 08:20:51 +02001432 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1433 * size of kmem_list3.
1434 */
1435static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1436{
1437 int node;
1438
1439 for_each_online_node(node) {
1440 cachep->nodelists[node] = &initkmem_list3[index + node];
1441 cachep->nodelists[node]->next_reap = jiffies +
1442 REAPTIMEOUT_LIST3 +
1443 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1444 }
1445}
1446
1447/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08001448 * Initialisation. Called after the page allocator have been initialised and
1449 * before smp_init().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001450 */
1451void __init kmem_cache_init(void)
1452{
1453 size_t left_over;
1454 struct cache_sizes *sizes;
1455 struct cache_names *names;
Christoph Lametere498be72005-09-09 13:03:32 -07001456 int i;
Jack Steiner07ed76b2006-03-07 21:55:46 -08001457 int order;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001458 int node;
Christoph Lametere498be72005-09-09 13:03:32 -07001459
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -07001460 if (num_possible_nodes() == 1) {
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001461 use_alien_caches = 0;
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -07001462 numa_platform = 0;
1463 }
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001464
Christoph Lametere498be72005-09-09 13:03:32 -07001465 for (i = 0; i < NUM_INIT_LISTS; i++) {
1466 kmem_list3_init(&initkmem_list3[i]);
1467 if (i < MAX_NUMNODES)
1468 cache_cache.nodelists[i] = NULL;
1469 }
Pekka Enberg556a1692008-01-25 08:20:51 +02001470 set_up_list3s(&cache_cache, CACHE_CACHE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001471
1472 /*
1473 * Fragmentation resistance on low memory - only use bigger
1474 * page orders on machines with more than 32MB of memory.
1475 */
1476 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1477 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1478
Linus Torvalds1da177e2005-04-16 15:20:36 -07001479 /* Bootstrap is tricky, because several objects are allocated
1480 * from caches that do not exist yet:
Andrew Mortona737b3e2006-03-22 00:08:11 -08001481 * 1) initialize the cache_cache cache: it contains the struct
1482 * kmem_cache structures of all caches, except cache_cache itself:
1483 * cache_cache is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001484 * Initially an __init data area is used for the head array and the
1485 * kmem_list3 structures, it's replaced with a kmalloc allocated
1486 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001487 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001488 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001489 * An __init data area is used for the head array.
1490 * 3) Create the remaining kmalloc caches, with minimally sized
1491 * head arrays.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001492 * 4) Replace the __init data head arrays for cache_cache and the first
1493 * kmalloc cache with kmalloc allocated arrays.
Christoph Lametere498be72005-09-09 13:03:32 -07001494 * 5) Replace the __init data for kmem_list3 for cache_cache and
1495 * the other cache's with kmalloc allocated memory.
1496 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001497 */
1498
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001499 node = numa_node_id();
1500
Linus Torvalds1da177e2005-04-16 15:20:36 -07001501 /* 1) create the cache_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001502 INIT_LIST_HEAD(&cache_chain);
1503 list_add(&cache_cache.next, &cache_chain);
1504 cache_cache.colour_off = cache_line_size();
1505 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
Daniel Yeisleyec1f5ee2008-03-25 23:59:08 +02001506 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001507
Eric Dumazet8da34302007-05-06 14:49:29 -07001508 /*
1509 * struct kmem_cache size depends on nr_node_ids, which
1510 * can be less than MAX_NUMNODES.
1511 */
1512 cache_cache.buffer_size = offsetof(struct kmem_cache, nodelists) +
1513 nr_node_ids * sizeof(struct kmem_list3 *);
1514#if DEBUG
1515 cache_cache.obj_size = cache_cache.buffer_size;
1516#endif
Andrew Mortona737b3e2006-03-22 00:08:11 -08001517 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1518 cache_line_size());
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08001519 cache_cache.reciprocal_buffer_size =
1520 reciprocal_value(cache_cache.buffer_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001521
Jack Steiner07ed76b2006-03-07 21:55:46 -08001522 for (order = 0; order < MAX_ORDER; order++) {
1523 cache_estimate(order, cache_cache.buffer_size,
1524 cache_line_size(), 0, &left_over, &cache_cache.num);
1525 if (cache_cache.num)
1526 break;
1527 }
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02001528 BUG_ON(!cache_cache.num);
Jack Steiner07ed76b2006-03-07 21:55:46 -08001529 cache_cache.gfporder = order;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001530 cache_cache.colour = left_over / cache_cache.colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001531 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1532 sizeof(struct slab), cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001533
1534 /* 2+3) create the kmalloc caches */
1535 sizes = malloc_sizes;
1536 names = cache_names;
1537
Andrew Mortona737b3e2006-03-22 00:08:11 -08001538 /*
1539 * Initialize the caches that provide memory for the array cache and the
1540 * kmem_list3 structures first. Without this, further allocations will
1541 * bug.
Christoph Lametere498be72005-09-09 13:03:32 -07001542 */
1543
1544 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001545 sizes[INDEX_AC].cs_size,
1546 ARCH_KMALLOC_MINALIGN,
1547 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001548 NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07001549
Andrew Mortona737b3e2006-03-22 00:08:11 -08001550 if (INDEX_AC != INDEX_L3) {
Christoph Lametere498be72005-09-09 13:03:32 -07001551 sizes[INDEX_L3].cs_cachep =
Andrew Mortona737b3e2006-03-22 00:08:11 -08001552 kmem_cache_create(names[INDEX_L3].name,
1553 sizes[INDEX_L3].cs_size,
1554 ARCH_KMALLOC_MINALIGN,
1555 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001556 NULL);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001557 }
Christoph Lametere498be72005-09-09 13:03:32 -07001558
Ingo Molnare0a42722006-06-23 02:03:46 -07001559 slab_early_init = 0;
1560
Linus Torvalds1da177e2005-04-16 15:20:36 -07001561 while (sizes->cs_size != ULONG_MAX) {
Christoph Lametere498be72005-09-09 13:03:32 -07001562 /*
1563 * For performance, all the general caches are L1 aligned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001564 * This should be particularly beneficial on SMP boxes, as it
1565 * eliminates "false sharing".
1566 * Note for systems short on memory removing the alignment will
Christoph Lametere498be72005-09-09 13:03:32 -07001567 * allow tighter packing of the smaller caches.
1568 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001569 if (!sizes->cs_cachep) {
Christoph Lametere498be72005-09-09 13:03:32 -07001570 sizes->cs_cachep = kmem_cache_create(names->name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001571 sizes->cs_size,
1572 ARCH_KMALLOC_MINALIGN,
1573 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001574 NULL);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001575 }
Christoph Lameter4b51d662007-02-10 01:43:10 -08001576#ifdef CONFIG_ZONE_DMA
1577 sizes->cs_dmacachep = kmem_cache_create(
1578 names->name_dma,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001579 sizes->cs_size,
1580 ARCH_KMALLOC_MINALIGN,
1581 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1582 SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001583 NULL);
Christoph Lameter4b51d662007-02-10 01:43:10 -08001584#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001585 sizes++;
1586 names++;
1587 }
1588 /* 4) Replace the bootstrap head arrays */
1589 {
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001590 struct array_cache *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001591
Pekka Enberg83b519e2009-06-10 19:40:04 +03001592 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
Christoph Lametere498be72005-09-09 13:03:32 -07001593
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001594 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1595 memcpy(ptr, cpu_cache_get(&cache_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001596 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001597 /*
1598 * Do not assume that spinlocks can be initialized via memcpy:
1599 */
1600 spin_lock_init(&ptr->lock);
1601
Linus Torvalds1da177e2005-04-16 15:20:36 -07001602 cache_cache.array[smp_processor_id()] = ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001603
Pekka Enberg83b519e2009-06-10 19:40:04 +03001604 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
Christoph Lametere498be72005-09-09 13:03:32 -07001605
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001606 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001607 != &initarray_generic.cache);
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001608 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001609 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001610 /*
1611 * Do not assume that spinlocks can be initialized via memcpy:
1612 */
1613 spin_lock_init(&ptr->lock);
1614
Christoph Lametere498be72005-09-09 13:03:32 -07001615 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001616 ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001617 }
Christoph Lametere498be72005-09-09 13:03:32 -07001618 /* 5) Replace the bootstrap kmem_list3's */
1619 {
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001620 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001621
Mel Gorman9c09a952008-01-24 05:49:54 -08001622 for_each_online_node(nid) {
Daniel Yeisleyec1f5ee2008-03-25 23:59:08 +02001623 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid);
Pekka Enberg556a1692008-01-25 08:20:51 +02001624
Christoph Lametere498be72005-09-09 13:03:32 -07001625 init_list(malloc_sizes[INDEX_AC].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001626 &initkmem_list3[SIZE_AC + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001627
1628 if (INDEX_AC != INDEX_L3) {
1629 init_list(malloc_sizes[INDEX_L3].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001630 &initkmem_list3[SIZE_L3 + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001631 }
1632 }
1633 }
1634
Pekka Enberg8429db52009-06-12 15:58:59 +03001635 g_cpucache_up = EARLY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001636
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001637 /* Annotate slab for lockdep -- annotate the malloc caches */
1638 init_lock_keys();
Pekka Enberg8429db52009-06-12 15:58:59 +03001639}
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001640
Pekka Enberg8429db52009-06-12 15:58:59 +03001641void __init kmem_cache_init_late(void)
1642{
1643 struct kmem_cache *cachep;
1644
1645 /*
1646 * Interrupts are enabled now so all GFP allocations are safe.
1647 */
1648 slab_gfp_mask = __GFP_BITS_MASK;
1649
1650 /* 6) resize the head arrays to their final sizes */
1651 mutex_lock(&cache_chain_mutex);
1652 list_for_each_entry(cachep, &cache_chain, next)
1653 if (enable_cpucache(cachep, GFP_NOWAIT))
1654 BUG();
1655 mutex_unlock(&cache_chain_mutex);
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001656
Linus Torvalds1da177e2005-04-16 15:20:36 -07001657 /* Done! */
1658 g_cpucache_up = FULL;
1659
Andrew Mortona737b3e2006-03-22 00:08:11 -08001660 /*
1661 * Register a cpu startup notifier callback that initializes
1662 * cpu_cache_get for all new cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07001663 */
1664 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001665
Andrew Mortona737b3e2006-03-22 00:08:11 -08001666 /*
1667 * The reap timers are started later, with a module init call: That part
1668 * of the kernel is not yet operational.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001669 */
1670}
1671
1672static int __init cpucache_init(void)
1673{
1674 int cpu;
1675
Andrew Mortona737b3e2006-03-22 00:08:11 -08001676 /*
1677 * Register the timers that return unneeded pages to the page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -07001678 */
Christoph Lametere498be72005-09-09 13:03:32 -07001679 for_each_online_cpu(cpu)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001680 start_cpu_timer(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001681 return 0;
1682}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001683__initcall(cpucache_init);
1684
1685/*
1686 * Interface to system's page allocator. No need to hold the cache-lock.
1687 *
1688 * If we requested dmaable memory, we will get it. Even if we
1689 * did not request dmaable memory, we might get it, but that
1690 * would be relatively rare and ignorable.
1691 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001692static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001693{
1694 struct page *page;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001695 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001696 int i;
1697
Luke Yangd6fef9d2006-04-10 22:52:56 -07001698#ifndef CONFIG_MMU
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001699 /*
1700 * Nommu uses slab's for process anonymous memory allocations, and thus
1701 * requires __GFP_COMP to properly refcount higher order allocations
Luke Yangd6fef9d2006-04-10 22:52:56 -07001702 */
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001703 flags |= __GFP_COMP;
Luke Yangd6fef9d2006-04-10 22:52:56 -07001704#endif
Christoph Lameter765c4502006-09-27 01:50:08 -07001705
Christoph Lameter3c517a62006-12-06 20:33:29 -08001706 flags |= cachep->gfpflags;
Mel Gormane12ba742007-10-16 01:25:52 -07001707 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1708 flags |= __GFP_RECLAIMABLE;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001709
1710 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001711 if (!page)
1712 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001713
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001714 nr_pages = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001715 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
Christoph Lameter972d1a72006-09-25 23:31:51 -07001716 add_zone_page_state(page_zone(page),
1717 NR_SLAB_RECLAIMABLE, nr_pages);
1718 else
1719 add_zone_page_state(page_zone(page),
1720 NR_SLAB_UNRECLAIMABLE, nr_pages);
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001721 for (i = 0; i < nr_pages; i++)
1722 __SetPageSlab(page + i);
1723 return page_address(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001724}
1725
1726/*
1727 * Interface to system's page release.
1728 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001729static void kmem_freepages(struct kmem_cache *cachep, void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001730{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001731 unsigned long i = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001732 struct page *page = virt_to_page(addr);
1733 const unsigned long nr_freed = i;
1734
Christoph Lameter972d1a72006-09-25 23:31:51 -07001735 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1736 sub_zone_page_state(page_zone(page),
1737 NR_SLAB_RECLAIMABLE, nr_freed);
1738 else
1739 sub_zone_page_state(page_zone(page),
1740 NR_SLAB_UNRECLAIMABLE, nr_freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001741 while (i--) {
Nick Pigginf205b2f2006-03-22 00:08:02 -08001742 BUG_ON(!PageSlab(page));
1743 __ClearPageSlab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001744 page++;
1745 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001746 if (current->reclaim_state)
1747 current->reclaim_state->reclaimed_slab += nr_freed;
1748 free_pages((unsigned long)addr, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001749}
1750
1751static void kmem_rcu_free(struct rcu_head *head)
1752{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001753 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001754 struct kmem_cache *cachep = slab_rcu->cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001755
1756 kmem_freepages(cachep, slab_rcu->addr);
1757 if (OFF_SLAB(cachep))
1758 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1759}
1760
1761#if DEBUG
1762
1763#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001764static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001765 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001766{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001767 int size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001768
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001769 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001770
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001771 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001772 return;
1773
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001774 *addr++ = 0x12345678;
1775 *addr++ = caller;
1776 *addr++ = smp_processor_id();
1777 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001778 {
1779 unsigned long *sptr = &caller;
1780 unsigned long svalue;
1781
1782 while (!kstack_end(sptr)) {
1783 svalue = *sptr++;
1784 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001785 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001786 size -= sizeof(unsigned long);
1787 if (size <= sizeof(unsigned long))
1788 break;
1789 }
1790 }
1791
1792 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001793 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001794}
1795#endif
1796
Pekka Enberg343e0d72006-02-01 03:05:50 -08001797static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001798{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001799 int size = obj_size(cachep);
1800 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001801
1802 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001803 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001804}
1805
1806static void dump_line(char *data, int offset, int limit)
1807{
1808 int i;
Dave Jonesaa83aa42006-09-29 01:59:51 -07001809 unsigned char error = 0;
1810 int bad_count = 0;
1811
Linus Torvalds1da177e2005-04-16 15:20:36 -07001812 printk(KERN_ERR "%03x:", offset);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001813 for (i = 0; i < limit; i++) {
1814 if (data[offset + i] != POISON_FREE) {
1815 error = data[offset + i];
1816 bad_count++;
1817 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001818 printk(" %02x", (unsigned char)data[offset + i]);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001819 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001820 printk("\n");
Dave Jonesaa83aa42006-09-29 01:59:51 -07001821
1822 if (bad_count == 1) {
1823 error ^= POISON_FREE;
1824 if (!(error & (error - 1))) {
1825 printk(KERN_ERR "Single bit error detected. Probably "
1826 "bad RAM.\n");
1827#ifdef CONFIG_X86
1828 printk(KERN_ERR "Run memtest86+ or a similar memory "
1829 "test tool.\n");
1830#else
1831 printk(KERN_ERR "Run a memory test tool.\n");
1832#endif
1833 }
1834 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001835}
1836#endif
1837
1838#if DEBUG
1839
Pekka Enberg343e0d72006-02-01 03:05:50 -08001840static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001841{
1842 int i, size;
1843 char *realobj;
1844
1845 if (cachep->flags & SLAB_RED_ZONE) {
David Woodhouseb46b8f12007-05-08 00:22:59 -07001846 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001847 *dbg_redzone1(cachep, objp),
1848 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001849 }
1850
1851 if (cachep->flags & SLAB_STORE_USER) {
1852 printk(KERN_ERR "Last user: [<%p>]",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001853 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001854 print_symbol("(%s)",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001855 (unsigned long)*dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001856 printk("\n");
1857 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001858 realobj = (char *)objp + obj_offset(cachep);
1859 size = obj_size(cachep);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001860 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001861 int limit;
1862 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001863 if (i + limit > size)
1864 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001865 dump_line(realobj, i, limit);
1866 }
1867}
1868
Pekka Enberg343e0d72006-02-01 03:05:50 -08001869static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001870{
1871 char *realobj;
1872 int size, i;
1873 int lines = 0;
1874
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001875 realobj = (char *)objp + obj_offset(cachep);
1876 size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001877
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001878 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001879 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001880 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001881 exp = POISON_END;
1882 if (realobj[i] != exp) {
1883 int limit;
1884 /* Mismatch ! */
1885 /* Print header */
1886 if (lines == 0) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001887 printk(KERN_ERR
David Howellse94a40c2007-04-02 23:46:28 +01001888 "Slab corruption: %s start=%p, len=%d\n",
1889 cachep->name, realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001890 print_objinfo(cachep, objp, 0);
1891 }
1892 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001893 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001894 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001895 if (i + limit > size)
1896 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001897 dump_line(realobj, i, limit);
1898 i += 16;
1899 lines++;
1900 /* Limit to 5 lines */
1901 if (lines > 5)
1902 break;
1903 }
1904 }
1905 if (lines != 0) {
1906 /* Print some data about the neighboring objects, if they
1907 * exist:
1908 */
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08001909 struct slab *slabp = virt_to_slab(objp);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001910 unsigned int objnr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001911
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001912 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001913 if (objnr) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001914 objp = index_to_obj(cachep, slabp, objnr - 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001915 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001916 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001917 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001918 print_objinfo(cachep, objp, 2);
1919 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001920 if (objnr + 1 < cachep->num) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001921 objp = index_to_obj(cachep, slabp, objnr + 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001922 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001923 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001924 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001925 print_objinfo(cachep, objp, 2);
1926 }
1927 }
1928}
1929#endif
1930
Linus Torvalds1da177e2005-04-16 15:20:36 -07001931#if DEBUG
Rabin Vincente79aec22008-07-04 00:40:32 +05301932static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001933{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001934 int i;
1935 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001936 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001937
1938 if (cachep->flags & SLAB_POISON) {
1939#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08001940 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1941 OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001942 kernel_map_pages(virt_to_page(objp),
Andrew Mortona737b3e2006-03-22 00:08:11 -08001943 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001944 else
1945 check_poison_obj(cachep, objp);
1946#else
1947 check_poison_obj(cachep, objp);
1948#endif
1949 }
1950 if (cachep->flags & SLAB_RED_ZONE) {
1951 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1952 slab_error(cachep, "start of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001953 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001954 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1955 slab_error(cachep, "end of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001956 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001957 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001958 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001959}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001960#else
Rabin Vincente79aec22008-07-04 00:40:32 +05301961static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001962{
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001963}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001964#endif
1965
Randy Dunlap911851e2006-03-22 00:08:14 -08001966/**
1967 * slab_destroy - destroy and release all objects in a slab
1968 * @cachep: cache pointer being destroyed
1969 * @slabp: slab pointer being destroyed
1970 *
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001971 * Destroy all the objs in a slab, and release the mem back to the system.
Andrew Mortona737b3e2006-03-22 00:08:11 -08001972 * Before calling the slab must have been unlinked from the cache. The
1973 * cache-lock is not held/needed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001974 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001975static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001976{
1977 void *addr = slabp->s_mem - slabp->colouroff;
1978
Rabin Vincente79aec22008-07-04 00:40:32 +05301979 slab_destroy_debugcheck(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001980 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1981 struct slab_rcu *slab_rcu;
1982
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001983 slab_rcu = (struct slab_rcu *)slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001984 slab_rcu->cachep = cachep;
1985 slab_rcu->addr = addr;
1986 call_rcu(&slab_rcu->head, kmem_rcu_free);
1987 } else {
1988 kmem_freepages(cachep, addr);
Ingo Molnar873623d2006-07-13 14:44:38 +02001989 if (OFF_SLAB(cachep))
1990 kmem_cache_free(cachep->slabp_cache, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001991 }
1992}
1993
Christoph Lameter117f6eb2006-09-25 23:31:37 -07001994static void __kmem_cache_destroy(struct kmem_cache *cachep)
1995{
1996 int i;
1997 struct kmem_list3 *l3;
1998
1999 for_each_online_cpu(i)
2000 kfree(cachep->array[i]);
2001
2002 /* NUMA: free the list3 structures */
2003 for_each_online_node(i) {
2004 l3 = cachep->nodelists[i];
2005 if (l3) {
2006 kfree(l3->shared);
2007 free_alien_cache(l3->alien);
2008 kfree(l3);
2009 }
2010 }
2011 kmem_cache_free(&cache_cache, cachep);
2012}
2013
2014
Linus Torvalds1da177e2005-04-16 15:20:36 -07002015/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08002016 * calculate_slab_order - calculate size (page order) of slabs
2017 * @cachep: pointer to the cache that is being created
2018 * @size: size of objects to be created in this cache.
2019 * @align: required alignment for the objects.
2020 * @flags: slab allocation flags
2021 *
2022 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002023 *
2024 * This could be made much more intelligent. For now, try to avoid using
2025 * high order pages for slabs. When the gfp() functions are more friendly
2026 * towards high-order requests, this should be changed.
2027 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002028static size_t calculate_slab_order(struct kmem_cache *cachep,
Randy Dunlapee13d782006-02-01 03:05:53 -08002029 size_t size, size_t align, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002030{
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002031 unsigned long offslab_limit;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002032 size_t left_over = 0;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002033 int gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002034
Christoph Lameter0aa817f2007-05-16 22:11:01 -07002035 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002036 unsigned int num;
2037 size_t remainder;
2038
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002039 cache_estimate(gfporder, size, align, flags, &remainder, &num);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002040 if (!num)
2041 continue;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002042
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002043 if (flags & CFLGS_OFF_SLAB) {
2044 /*
2045 * Max number of objs-per-slab for caches which
2046 * use off-slab slabs. Needed to avoid a possible
2047 * looping condition in cache_grow().
2048 */
2049 offslab_limit = size - sizeof(struct slab);
2050 offslab_limit /= sizeof(kmem_bufctl_t);
2051
2052 if (num > offslab_limit)
2053 break;
2054 }
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002055
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002056 /* Found something acceptable - save it away */
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002057 cachep->num = num;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002058 cachep->gfporder = gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002059 left_over = remainder;
2060
2061 /*
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002062 * A VFS-reclaimable slab tends to have most allocations
2063 * as GFP_NOFS and we really don't want to have to be allocating
2064 * higher-order pages when we are unable to shrink dcache.
2065 */
2066 if (flags & SLAB_RECLAIM_ACCOUNT)
2067 break;
2068
2069 /*
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002070 * Large number of objects is good, but very large slabs are
2071 * currently bad for the gfp()s.
2072 */
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002073 if (gfporder >= slab_break_gfp_order)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002074 break;
2075
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002076 /*
2077 * Acceptable internal fragmentation?
2078 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002079 if (left_over * 8 <= (PAGE_SIZE << gfporder))
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002080 break;
2081 }
2082 return left_over;
2083}
2084
Pekka Enberg83b519e2009-06-10 19:40:04 +03002085static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002086{
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002087 if (g_cpucache_up == FULL)
Pekka Enberg83b519e2009-06-10 19:40:04 +03002088 return enable_cpucache(cachep, gfp);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002089
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002090 if (g_cpucache_up == NONE) {
2091 /*
2092 * Note: the first kmem_cache_create must create the cache
2093 * that's used by kmalloc(24), otherwise the creation of
2094 * further caches will BUG().
2095 */
2096 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2097
2098 /*
2099 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2100 * the first cache, then we need to set up all its list3s,
2101 * otherwise the creation of further caches will BUG().
2102 */
2103 set_up_list3s(cachep, SIZE_AC);
2104 if (INDEX_AC == INDEX_L3)
2105 g_cpucache_up = PARTIAL_L3;
2106 else
2107 g_cpucache_up = PARTIAL_AC;
2108 } else {
2109 cachep->array[smp_processor_id()] =
Pekka Enberg83b519e2009-06-10 19:40:04 +03002110 kmalloc(sizeof(struct arraycache_init), gfp);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002111
2112 if (g_cpucache_up == PARTIAL_AC) {
2113 set_up_list3s(cachep, SIZE_L3);
2114 g_cpucache_up = PARTIAL_L3;
2115 } else {
2116 int node;
Pekka Enberg556a1692008-01-25 08:20:51 +02002117 for_each_online_node(node) {
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002118 cachep->nodelists[node] =
2119 kmalloc_node(sizeof(struct kmem_list3),
Pekka Enbergeb91f1d2009-06-12 14:56:09 +03002120 gfp, node);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002121 BUG_ON(!cachep->nodelists[node]);
2122 kmem_list3_init(cachep->nodelists[node]);
2123 }
2124 }
2125 }
2126 cachep->nodelists[numa_node_id()]->next_reap =
2127 jiffies + REAPTIMEOUT_LIST3 +
2128 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2129
2130 cpu_cache_get(cachep)->avail = 0;
2131 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2132 cpu_cache_get(cachep)->batchcount = 1;
2133 cpu_cache_get(cachep)->touched = 0;
2134 cachep->batchcount = 1;
2135 cachep->limit = BOOT_CPUCACHE_ENTRIES;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002136 return 0;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002137}
2138
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002139/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07002140 * kmem_cache_create - Create a cache.
2141 * @name: A string which is used in /proc/slabinfo to identify this cache.
2142 * @size: The size of objects to be created in this cache.
2143 * @align: The required alignment for the objects.
2144 * @flags: SLAB flags
2145 * @ctor: A constructor for the objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002146 *
2147 * Returns a ptr to the cache on success, NULL on failure.
2148 * Cannot be called within a int, but can be interrupted.
Paul Mundt20c2df82007-07-20 10:11:58 +09002149 * The @ctor is run when new pages are allocated by the cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002150 *
2151 * @name must be valid until the cache is destroyed. This implies that
Andrew Mortona737b3e2006-03-22 00:08:11 -08002152 * the module calling this has to destroy the cache before getting unloaded.
Catalin Marinas249da162008-11-21 12:56:22 +00002153 * Note that kmem_cache_name() is not guaranteed to return the same pointer,
2154 * therefore applications must manage it themselves.
Andrew Mortona737b3e2006-03-22 00:08:11 -08002155 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002156 * The flags are
2157 *
2158 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2159 * to catch references to uninitialised memory.
2160 *
2161 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2162 * for buffer overruns.
2163 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002164 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2165 * cacheline. This can be beneficial if you're counting cycles as closely
2166 * as davem.
2167 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002168struct kmem_cache *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002169kmem_cache_create (const char *name, size_t size, size_t align,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002170 unsigned long flags, void (*ctor)(void *))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002171{
2172 size_t left_over, slab_size, ralign;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07002173 struct kmem_cache *cachep = NULL, *pc;
Pekka Enberg83b519e2009-06-10 19:40:04 +03002174 gfp_t gfp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002175
2176 /*
2177 * Sanity checks... these are all serious usage bugs.
2178 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002179 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
Paul Mundt20c2df82007-07-20 10:11:58 +09002180 size > KMALLOC_MAX_SIZE) {
Harvey Harrisond40cee22008-04-30 00:55:07 -07002181 printk(KERN_ERR "%s: Early error in slab %s\n", __func__,
Andrew Mortona737b3e2006-03-22 00:08:11 -08002182 name);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002183 BUG();
2184 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002185
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002186 /*
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002187 * We use cache_chain_mutex to ensure a consistent view of
Rusty Russell174596a2009-01-01 10:12:29 +10302188 * cpu_online_mask as well. Please see cpuup_callback
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002189 */
Pekka Enberg83b519e2009-06-10 19:40:04 +03002190 if (slab_is_available()) {
2191 get_online_cpus();
2192 mutex_lock(&cache_chain_mutex);
2193 }
Andrew Morton4f12bb42005-11-07 00:58:00 -08002194
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07002195 list_for_each_entry(pc, &cache_chain, next) {
Andrew Morton4f12bb42005-11-07 00:58:00 -08002196 char tmp;
2197 int res;
2198
2199 /*
2200 * This happens when the module gets unloaded and doesn't
2201 * destroy its slab cache and no-one else reuses the vmalloc
2202 * area of the module. Print a warning.
2203 */
Andrew Morton138ae662006-12-06 20:36:41 -08002204 res = probe_kernel_address(pc->name, tmp);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002205 if (res) {
matzeb4169522007-05-06 14:49:52 -07002206 printk(KERN_ERR
2207 "SLAB: cache with size %d has lost its name\n",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002208 pc->buffer_size);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002209 continue;
2210 }
2211
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002212 if (!strcmp(pc->name, name)) {
matzeb4169522007-05-06 14:49:52 -07002213 printk(KERN_ERR
2214 "kmem_cache_create: duplicate cache %s\n", name);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002215 dump_stack();
2216 goto oops;
2217 }
2218 }
2219
Linus Torvalds1da177e2005-04-16 15:20:36 -07002220#if DEBUG
2221 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002222#if FORCED_DEBUG
2223 /*
2224 * Enable redzoning and last user accounting, except for caches with
2225 * large objects, if the increased size would increase the object size
2226 * above the next power of two: caches with object sizes just above a
2227 * power of two have a significant amount of internal fragmentation.
2228 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002229 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2230 2 * sizeof(unsigned long long)))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002231 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002232 if (!(flags & SLAB_DESTROY_BY_RCU))
2233 flags |= SLAB_POISON;
2234#endif
2235 if (flags & SLAB_DESTROY_BY_RCU)
2236 BUG_ON(flags & SLAB_POISON);
2237#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002238 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002239 * Always checks flags, a caller might be expecting debug support which
2240 * isn't available.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002241 */
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002242 BUG_ON(flags & ~CREATE_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002243
Andrew Mortona737b3e2006-03-22 00:08:11 -08002244 /*
2245 * Check that size is in terms of words. This is needed to avoid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002246 * unaligned accesses for some archs when redzoning is used, and makes
2247 * sure any on-slab bufctl's are also correctly aligned.
2248 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002249 if (size & (BYTES_PER_WORD - 1)) {
2250 size += (BYTES_PER_WORD - 1);
2251 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002252 }
2253
Andrew Mortona737b3e2006-03-22 00:08:11 -08002254 /* calculate the final buffer alignment: */
2255
Linus Torvalds1da177e2005-04-16 15:20:36 -07002256 /* 1) arch recommendation: can be overridden for debug */
2257 if (flags & SLAB_HWCACHE_ALIGN) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002258 /*
2259 * Default alignment: as specified by the arch code. Except if
2260 * an object is really small, then squeeze multiple objects into
2261 * one cacheline.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002262 */
2263 ralign = cache_line_size();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002264 while (size <= ralign / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002265 ralign /= 2;
2266 } else {
2267 ralign = BYTES_PER_WORD;
2268 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002269
2270 /*
David Woodhouse87a927c2007-07-04 21:26:44 -04002271 * Redzoning and user store require word alignment or possibly larger.
2272 * Note this will be overridden by architecture or caller mandated
2273 * alignment if either is greater than BYTES_PER_WORD.
Pekka Enbergca5f9702006-09-25 23:31:25 -07002274 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002275 if (flags & SLAB_STORE_USER)
2276 ralign = BYTES_PER_WORD;
2277
2278 if (flags & SLAB_RED_ZONE) {
2279 ralign = REDZONE_ALIGN;
2280 /* If redzoning, ensure that the second redzone is suitably
2281 * aligned, by adjusting the object size accordingly. */
2282 size += REDZONE_ALIGN - 1;
2283 size &= ~(REDZONE_ALIGN - 1);
2284 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002285
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002286 /* 2) arch mandated alignment */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002287 if (ralign < ARCH_SLAB_MINALIGN) {
2288 ralign = ARCH_SLAB_MINALIGN;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002289 }
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002290 /* 3) caller mandated alignment */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002291 if (ralign < align) {
2292 ralign = align;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002293 }
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002294 /* disable debug if necessary */
David Woodhouseb46b8f12007-05-08 00:22:59 -07002295 if (ralign > __alignof__(unsigned long long))
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002296 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002297 /*
Pekka Enbergca5f9702006-09-25 23:31:25 -07002298 * 4) Store it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002299 */
2300 align = ralign;
2301
Pekka Enberg83b519e2009-06-10 19:40:04 +03002302 if (slab_is_available())
2303 gfp = GFP_KERNEL;
2304 else
2305 gfp = GFP_NOWAIT;
2306
Linus Torvalds1da177e2005-04-16 15:20:36 -07002307 /* Get cache's description obj. */
Pekka Enberg83b519e2009-06-10 19:40:04 +03002308 cachep = kmem_cache_zalloc(&cache_cache, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002309 if (!cachep)
Andrew Morton4f12bb42005-11-07 00:58:00 -08002310 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002311
2312#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002313 cachep->obj_size = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002314
Pekka Enbergca5f9702006-09-25 23:31:25 -07002315 /*
2316 * Both debugging options require word-alignment which is calculated
2317 * into align above.
2318 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002319 if (flags & SLAB_RED_ZONE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002320 /* add space for red zone words */
David Woodhouseb46b8f12007-05-08 00:22:59 -07002321 cachep->obj_offset += sizeof(unsigned long long);
2322 size += 2 * sizeof(unsigned long long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002323 }
2324 if (flags & SLAB_STORE_USER) {
Pekka Enbergca5f9702006-09-25 23:31:25 -07002325 /* user store requires one word storage behind the end of
David Woodhouse87a927c2007-07-04 21:26:44 -04002326 * the real object. But if the second red zone needs to be
2327 * aligned to 64 bits, we must allow that much space.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002328 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002329 if (flags & SLAB_RED_ZONE)
2330 size += REDZONE_ALIGN;
2331 else
2332 size += BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002333 }
2334#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002335 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002336 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2337 cachep->obj_offset += PAGE_SIZE - size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002338 size = PAGE_SIZE;
2339 }
2340#endif
2341#endif
2342
Ingo Molnare0a42722006-06-23 02:03:46 -07002343 /*
2344 * Determine if the slab management is 'on' or 'off' slab.
2345 * (bootstrapping cannot cope with offslab caches so don't do
2346 * it too early on.)
2347 */
2348 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002349 /*
2350 * Size is large, assume best to place the slab management obj
2351 * off-slab (should allow better packing of objs).
2352 */
2353 flags |= CFLGS_OFF_SLAB;
2354
2355 size = ALIGN(size, align);
2356
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002357 left_over = calculate_slab_order(cachep, size, align, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002358
2359 if (!cachep->num) {
matzeb4169522007-05-06 14:49:52 -07002360 printk(KERN_ERR
2361 "kmem_cache_create: couldn't create cache %s.\n", name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002362 kmem_cache_free(&cache_cache, cachep);
2363 cachep = NULL;
Andrew Morton4f12bb42005-11-07 00:58:00 -08002364 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002365 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002366 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2367 + sizeof(struct slab), align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002368
2369 /*
2370 * If the slab has been placed off-slab, and we have enough space then
2371 * move it on-slab. This is at the expense of any extra colouring.
2372 */
2373 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2374 flags &= ~CFLGS_OFF_SLAB;
2375 left_over -= slab_size;
2376 }
2377
2378 if (flags & CFLGS_OFF_SLAB) {
2379 /* really off slab. No need for manual alignment */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002380 slab_size =
2381 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002382 }
2383
2384 cachep->colour_off = cache_line_size();
2385 /* Offset must be a multiple of the alignment. */
2386 if (cachep->colour_off < align)
2387 cachep->colour_off = align;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002388 cachep->colour = left_over / cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002389 cachep->slab_size = slab_size;
2390 cachep->flags = flags;
2391 cachep->gfpflags = 0;
Christoph Lameter4b51d662007-02-10 01:43:10 -08002392 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002393 cachep->gfpflags |= GFP_DMA;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002394 cachep->buffer_size = size;
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08002395 cachep->reciprocal_buffer_size = reciprocal_value(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002396
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002397 if (flags & CFLGS_OFF_SLAB) {
Victor Fuscob2d55072005-09-10 00:26:36 -07002398 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002399 /*
2400 * This is a possibility for one of the malloc_sizes caches.
2401 * But since we go off slab only for object size greater than
2402 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2403 * this should not happen at all.
2404 * But leave a BUG_ON for some lucky dude.
2405 */
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002406 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002407 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002408 cachep->ctor = ctor;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002409 cachep->name = name;
2410
Pekka Enberg83b519e2009-06-10 19:40:04 +03002411 if (setup_cpu_cache(cachep, gfp)) {
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002412 __kmem_cache_destroy(cachep);
2413 cachep = NULL;
2414 goto oops;
2415 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002416
Linus Torvalds1da177e2005-04-16 15:20:36 -07002417 /* cache setup completed, link it into the list */
2418 list_add(&cachep->next, &cache_chain);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002419oops:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002420 if (!cachep && (flags & SLAB_PANIC))
2421 panic("kmem_cache_create(): failed to create slab `%s'\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002422 name);
Pekka Enberg83b519e2009-06-10 19:40:04 +03002423 if (slab_is_available()) {
2424 mutex_unlock(&cache_chain_mutex);
2425 put_online_cpus();
2426 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002427 return cachep;
2428}
2429EXPORT_SYMBOL(kmem_cache_create);
2430
2431#if DEBUG
2432static void check_irq_off(void)
2433{
2434 BUG_ON(!irqs_disabled());
2435}
2436
2437static void check_irq_on(void)
2438{
2439 BUG_ON(irqs_disabled());
2440}
2441
Pekka Enberg343e0d72006-02-01 03:05:50 -08002442static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002443{
2444#ifdef CONFIG_SMP
2445 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002446 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002447#endif
2448}
Christoph Lametere498be72005-09-09 13:03:32 -07002449
Pekka Enberg343e0d72006-02-01 03:05:50 -08002450static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002451{
2452#ifdef CONFIG_SMP
2453 check_irq_off();
2454 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2455#endif
2456}
2457
Linus Torvalds1da177e2005-04-16 15:20:36 -07002458#else
2459#define check_irq_off() do { } while(0)
2460#define check_irq_on() do { } while(0)
2461#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002462#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002463#endif
2464
Christoph Lameteraab22072006-03-22 00:09:06 -08002465static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2466 struct array_cache *ac,
2467 int force, int node);
2468
Linus Torvalds1da177e2005-04-16 15:20:36 -07002469static void do_drain(void *arg)
2470{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002471 struct kmem_cache *cachep = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002472 struct array_cache *ac;
Christoph Lameterff694162005-09-22 21:44:02 -07002473 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002474
2475 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002476 ac = cpu_cache_get(cachep);
Christoph Lameterff694162005-09-22 21:44:02 -07002477 spin_lock(&cachep->nodelists[node]->list_lock);
2478 free_block(cachep, ac->entry, ac->avail, node);
2479 spin_unlock(&cachep->nodelists[node]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002480 ac->avail = 0;
2481}
2482
Pekka Enberg343e0d72006-02-01 03:05:50 -08002483static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002484{
Christoph Lametere498be72005-09-09 13:03:32 -07002485 struct kmem_list3 *l3;
2486 int node;
2487
Jens Axboe15c8b6c2008-05-09 09:39:44 +02002488 on_each_cpu(do_drain, cachep, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002489 check_irq_on();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002490 for_each_online_node(node) {
Christoph Lametere498be72005-09-09 13:03:32 -07002491 l3 = cachep->nodelists[node];
Roland Dreiera4523a82006-05-15 11:41:00 -07002492 if (l3 && l3->alien)
2493 drain_alien_cache(cachep, l3->alien);
2494 }
2495
2496 for_each_online_node(node) {
2497 l3 = cachep->nodelists[node];
2498 if (l3)
Christoph Lameteraab22072006-03-22 00:09:06 -08002499 drain_array(cachep, l3, l3->shared, 1, node);
Christoph Lametere498be72005-09-09 13:03:32 -07002500 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002501}
2502
Christoph Lametered11d9e2006-06-30 01:55:45 -07002503/*
2504 * Remove slabs from the list of free slabs.
2505 * Specify the number of slabs to drain in tofree.
2506 *
2507 * Returns the actual number of slabs released.
2508 */
2509static int drain_freelist(struct kmem_cache *cache,
2510 struct kmem_list3 *l3, int tofree)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002511{
Christoph Lametered11d9e2006-06-30 01:55:45 -07002512 struct list_head *p;
2513 int nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002514 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002515
Christoph Lametered11d9e2006-06-30 01:55:45 -07002516 nr_freed = 0;
2517 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002518
Christoph Lametered11d9e2006-06-30 01:55:45 -07002519 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002520 p = l3->slabs_free.prev;
Christoph Lametered11d9e2006-06-30 01:55:45 -07002521 if (p == &l3->slabs_free) {
2522 spin_unlock_irq(&l3->list_lock);
2523 goto out;
2524 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002525
Christoph Lametered11d9e2006-06-30 01:55:45 -07002526 slabp = list_entry(p, struct slab, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002527#if DEBUG
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002528 BUG_ON(slabp->inuse);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002529#endif
2530 list_del(&slabp->list);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002531 /*
2532 * Safe to drop the lock. The slab is no longer linked
2533 * to the cache.
2534 */
2535 l3->free_objects -= cache->num;
Christoph Lametere498be72005-09-09 13:03:32 -07002536 spin_unlock_irq(&l3->list_lock);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002537 slab_destroy(cache, slabp);
2538 nr_freed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002539 }
Christoph Lametered11d9e2006-06-30 01:55:45 -07002540out:
2541 return nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002542}
2543
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002544/* Called with cache_chain_mutex held to protect against cpu hotplug */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002545static int __cache_shrink(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07002546{
2547 int ret = 0, i = 0;
2548 struct kmem_list3 *l3;
2549
2550 drain_cpu_caches(cachep);
2551
2552 check_irq_on();
2553 for_each_online_node(i) {
2554 l3 = cachep->nodelists[i];
Christoph Lametered11d9e2006-06-30 01:55:45 -07002555 if (!l3)
2556 continue;
2557
2558 drain_freelist(cachep, l3, l3->free_objects);
2559
2560 ret += !list_empty(&l3->slabs_full) ||
2561 !list_empty(&l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07002562 }
2563 return (ret ? 1 : 0);
2564}
2565
Linus Torvalds1da177e2005-04-16 15:20:36 -07002566/**
2567 * kmem_cache_shrink - Shrink a cache.
2568 * @cachep: The cache to shrink.
2569 *
2570 * Releases as many slabs as possible for a cache.
2571 * To help debugging, a zero exit status indicates all slabs were released.
2572 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002573int kmem_cache_shrink(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002574{
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002575 int ret;
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002576 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002577
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002578 get_online_cpus();
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002579 mutex_lock(&cache_chain_mutex);
2580 ret = __cache_shrink(cachep);
2581 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002582 put_online_cpus();
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002583 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002584}
2585EXPORT_SYMBOL(kmem_cache_shrink);
2586
2587/**
2588 * kmem_cache_destroy - delete a cache
2589 * @cachep: the cache to destroy
2590 *
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08002591 * Remove a &struct kmem_cache object from the slab cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002592 *
2593 * It is expected this function will be called by a module when it is
2594 * unloaded. This will remove the cache completely, and avoid a duplicate
2595 * cache being allocated each time a module is loaded and unloaded, if the
2596 * module doesn't have persistent in-kernel storage across loads and unloads.
2597 *
2598 * The cache must be empty before calling this function.
2599 *
2600 * The caller must guarantee that noone will allocate memory from the cache
2601 * during the kmem_cache_destroy().
2602 */
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002603void kmem_cache_destroy(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002604{
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002605 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002606
Linus Torvalds1da177e2005-04-16 15:20:36 -07002607 /* Find the cache in the chain of caches. */
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002608 get_online_cpus();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002609 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002610 /*
2611 * the chain is never empty, cache_cache is never destroyed
2612 */
2613 list_del(&cachep->next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002614 if (__cache_shrink(cachep)) {
2615 slab_error(cachep, "Can't free all objects");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002616 list_add(&cachep->next, &cache_chain);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002617 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002618 put_online_cpus();
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002619 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002620 }
2621
2622 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
Paul E. McKenneyfbd568a3e2005-05-01 08:59:04 -07002623 synchronize_rcu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002624
Christoph Lameter117f6eb2006-09-25 23:31:37 -07002625 __kmem_cache_destroy(cachep);
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002626 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002627 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002628}
2629EXPORT_SYMBOL(kmem_cache_destroy);
2630
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002631/*
2632 * Get the memory for a slab management obj.
2633 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2634 * always come from malloc_sizes caches. The slab descriptor cannot
2635 * come from the same cache which is getting created because,
2636 * when we are searching for an appropriate cache for these
2637 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2638 * If we are creating a malloc_sizes cache here it would not be visible to
2639 * kmem_find_general_cachep till the initialization is complete.
2640 * Hence we cannot have slabp_cache same as the original cache.
2641 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002642static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002643 int colour_off, gfp_t local_flags,
2644 int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002645{
2646 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002647
Linus Torvalds1da177e2005-04-16 15:20:36 -07002648 if (OFF_SLAB(cachep)) {
2649 /* Slab management obj is off-slab. */
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002650 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
Pekka Enberg8759ec52008-11-26 10:01:31 +02002651 local_flags, nodeid);
Catalin Marinasd5cff632009-06-11 13:22:40 +01002652 /*
2653 * If the first object in the slab is leaked (it's allocated
2654 * but no one has a reference to it), we want to make sure
2655 * kmemleak does not treat the ->s_mem pointer as a reference
2656 * to the object. Otherwise we will not report the leak.
2657 */
2658 kmemleak_scan_area(slabp, offsetof(struct slab, list),
2659 sizeof(struct list_head), local_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002660 if (!slabp)
2661 return NULL;
2662 } else {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002663 slabp = objp + colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002664 colour_off += cachep->slab_size;
2665 }
2666 slabp->inuse = 0;
2667 slabp->colouroff = colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002668 slabp->s_mem = objp + colour_off;
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002669 slabp->nodeid = nodeid;
Marcin Slusarze51bfd02008-02-10 11:21:54 +01002670 slabp->free = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002671 return slabp;
2672}
2673
2674static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2675{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002676 return (kmem_bufctl_t *) (slabp + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002677}
2678
Pekka Enberg343e0d72006-02-01 03:05:50 -08002679static void cache_init_objs(struct kmem_cache *cachep,
Christoph Lametera35afb82007-05-16 22:10:57 -07002680 struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002681{
2682 int i;
2683
2684 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002685 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002686#if DEBUG
2687 /* need to poison the objs? */
2688 if (cachep->flags & SLAB_POISON)
2689 poison_obj(cachep, objp, POISON_FREE);
2690 if (cachep->flags & SLAB_STORE_USER)
2691 *dbg_userword(cachep, objp) = NULL;
2692
2693 if (cachep->flags & SLAB_RED_ZONE) {
2694 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2695 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2696 }
2697 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002698 * Constructors are not allowed to allocate memory from the same
2699 * cache which they are a constructor for. Otherwise, deadlock.
2700 * They must also be threaded.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002701 */
2702 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002703 cachep->ctor(objp + obj_offset(cachep));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002704
2705 if (cachep->flags & SLAB_RED_ZONE) {
2706 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2707 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002708 " end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002709 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2710 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002711 " start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002712 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08002713 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2714 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002715 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002716 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002717#else
2718 if (cachep->ctor)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002719 cachep->ctor(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002720#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002721 slab_bufctl(slabp)[i] = i + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002722 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002723 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002724}
2725
Pekka Enberg343e0d72006-02-01 03:05:50 -08002726static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002727{
Christoph Lameter4b51d662007-02-10 01:43:10 -08002728 if (CONFIG_ZONE_DMA_FLAG) {
2729 if (flags & GFP_DMA)
2730 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2731 else
2732 BUG_ON(cachep->gfpflags & GFP_DMA);
2733 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002734}
2735
Andrew Mortona737b3e2006-03-22 00:08:11 -08002736static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2737 int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002738{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002739 void *objp = index_to_obj(cachep, slabp, slabp->free);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002740 kmem_bufctl_t next;
2741
2742 slabp->inuse++;
2743 next = slab_bufctl(slabp)[slabp->free];
2744#if DEBUG
2745 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2746 WARN_ON(slabp->nodeid != nodeid);
2747#endif
2748 slabp->free = next;
2749
2750 return objp;
2751}
2752
Andrew Mortona737b3e2006-03-22 00:08:11 -08002753static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2754 void *objp, int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002755{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002756 unsigned int objnr = obj_to_index(cachep, slabp, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002757
2758#if DEBUG
2759 /* Verify that the slab belongs to the intended node */
2760 WARN_ON(slabp->nodeid != nodeid);
2761
Al Viro871751e2006-03-25 03:06:39 -08002762 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
Matthew Dobson78d382d2006-02-01 03:05:47 -08002763 printk(KERN_ERR "slab: double free detected in cache "
Andrew Mortona737b3e2006-03-22 00:08:11 -08002764 "'%s', objp %p\n", cachep->name, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002765 BUG();
2766 }
2767#endif
2768 slab_bufctl(slabp)[objnr] = slabp->free;
2769 slabp->free = objnr;
2770 slabp->inuse--;
2771}
2772
Pekka Enberg47768742006-06-23 02:03:07 -07002773/*
2774 * Map pages beginning at addr to the given cache and slab. This is required
2775 * for the slab allocator to be able to lookup the cache and slab of a
2776 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2777 */
2778static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2779 void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002780{
Pekka Enberg47768742006-06-23 02:03:07 -07002781 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002782 struct page *page;
2783
Pekka Enberg47768742006-06-23 02:03:07 -07002784 page = virt_to_page(addr);
Nick Piggin84097512006-03-22 00:08:34 -08002785
Pekka Enberg47768742006-06-23 02:03:07 -07002786 nr_pages = 1;
Nick Piggin84097512006-03-22 00:08:34 -08002787 if (likely(!PageCompound(page)))
Pekka Enberg47768742006-06-23 02:03:07 -07002788 nr_pages <<= cache->gfporder;
2789
Linus Torvalds1da177e2005-04-16 15:20:36 -07002790 do {
Pekka Enberg47768742006-06-23 02:03:07 -07002791 page_set_cache(page, cache);
2792 page_set_slab(page, slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002793 page++;
Pekka Enberg47768742006-06-23 02:03:07 -07002794 } while (--nr_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002795}
2796
2797/*
2798 * Grow (by 1) the number of slabs within a cache. This is called by
2799 * kmem_cache_alloc() when there are no active objs left in a cache.
2800 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002801static int cache_grow(struct kmem_cache *cachep,
2802 gfp_t flags, int nodeid, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002803{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002804 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002805 size_t offset;
2806 gfp_t local_flags;
Christoph Lametere498be72005-09-09 13:03:32 -07002807 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002808
Andrew Mortona737b3e2006-03-22 00:08:11 -08002809 /*
2810 * Be lazy and only check for valid flags here, keeping it out of the
2811 * critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002812 */
Christoph Lameter6cb06222007-10-16 01:25:41 -07002813 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2814 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002815
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002816 /* Take the l3 list lock to change the colour_next on this node */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002817 check_irq_off();
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002818 l3 = cachep->nodelists[nodeid];
2819 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002820
2821 /* Get colour for the slab, and cal the next value. */
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002822 offset = l3->colour_next;
2823 l3->colour_next++;
2824 if (l3->colour_next >= cachep->colour)
2825 l3->colour_next = 0;
2826 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002827
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002828 offset *= cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002829
2830 if (local_flags & __GFP_WAIT)
2831 local_irq_enable();
2832
2833 /*
2834 * The test for missing atomic flag is performed here, rather than
2835 * the more obvious place, simply to reduce the critical path length
2836 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2837 * will eventually be caught here (where it matters).
2838 */
2839 kmem_flagcheck(cachep, flags);
2840
Andrew Mortona737b3e2006-03-22 00:08:11 -08002841 /*
2842 * Get mem for the objs. Attempt to allocate a physical page from
2843 * 'nodeid'.
Christoph Lametere498be72005-09-09 13:03:32 -07002844 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002845 if (!objp)
Andrew Mortonb8c1c5d2007-07-24 12:02:40 -07002846 objp = kmem_getpages(cachep, local_flags, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002847 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002848 goto failed;
2849
2850 /* Get slab management. */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002851 slabp = alloc_slabmgmt(cachep, objp, offset,
Christoph Lameter6cb06222007-10-16 01:25:41 -07002852 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002853 if (!slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002854 goto opps1;
2855
Pekka Enberg47768742006-06-23 02:03:07 -07002856 slab_map_pages(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002857
Christoph Lametera35afb82007-05-16 22:10:57 -07002858 cache_init_objs(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002859
2860 if (local_flags & __GFP_WAIT)
2861 local_irq_disable();
2862 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002863 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002864
2865 /* Make slab active. */
Christoph Lametere498be72005-09-09 13:03:32 -07002866 list_add_tail(&slabp->list, &(l3->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002867 STATS_INC_GROWN(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002868 l3->free_objects += cachep->num;
2869 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002870 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002871opps1:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002872 kmem_freepages(cachep, objp);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002873failed:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002874 if (local_flags & __GFP_WAIT)
2875 local_irq_disable();
2876 return 0;
2877}
2878
2879#if DEBUG
2880
2881/*
2882 * Perform extra freeing checks:
2883 * - detect bad pointers.
2884 * - POISON/RED_ZONE checking
Linus Torvalds1da177e2005-04-16 15:20:36 -07002885 */
2886static void kfree_debugcheck(const void *objp)
2887{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002888 if (!virt_addr_valid(objp)) {
2889 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002890 (unsigned long)objp);
2891 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002892 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002893}
2894
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002895static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2896{
David Woodhouseb46b8f12007-05-08 00:22:59 -07002897 unsigned long long redzone1, redzone2;
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002898
2899 redzone1 = *dbg_redzone1(cache, obj);
2900 redzone2 = *dbg_redzone2(cache, obj);
2901
2902 /*
2903 * Redzone is ok.
2904 */
2905 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2906 return;
2907
2908 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2909 slab_error(cache, "double free detected");
2910 else
2911 slab_error(cache, "memory outside object was overwritten");
2912
David Woodhouseb46b8f12007-05-08 00:22:59 -07002913 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002914 obj, redzone1, redzone2);
2915}
2916
Pekka Enberg343e0d72006-02-01 03:05:50 -08002917static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002918 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002919{
2920 struct page *page;
2921 unsigned int objnr;
2922 struct slab *slabp;
2923
Matthew Wilcox80cbd912007-11-29 12:05:13 -07002924 BUG_ON(virt_to_cache(objp) != cachep);
2925
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002926 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002927 kfree_debugcheck(objp);
Christoph Lameterb49af682007-05-06 14:49:41 -07002928 page = virt_to_head_page(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002929
Pekka Enberg065d41c2005-11-13 16:06:46 -08002930 slabp = page_get_slab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002931
2932 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002933 verify_redzone_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002934 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2935 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2936 }
2937 if (cachep->flags & SLAB_STORE_USER)
2938 *dbg_userword(cachep, objp) = caller;
2939
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002940 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002941
2942 BUG_ON(objnr >= cachep->num);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002943 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002944
Al Viro871751e2006-03-25 03:06:39 -08002945#ifdef CONFIG_DEBUG_SLAB_LEAK
2946 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2947#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002948 if (cachep->flags & SLAB_POISON) {
2949#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08002950 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002951 store_stackinfo(cachep, objp, (unsigned long)caller);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002952 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002953 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002954 } else {
2955 poison_obj(cachep, objp, POISON_FREE);
2956 }
2957#else
2958 poison_obj(cachep, objp, POISON_FREE);
2959#endif
2960 }
2961 return objp;
2962}
2963
Pekka Enberg343e0d72006-02-01 03:05:50 -08002964static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002965{
2966 kmem_bufctl_t i;
2967 int entries = 0;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002968
Linus Torvalds1da177e2005-04-16 15:20:36 -07002969 /* Check slab's freelist to see if this obj is there. */
2970 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2971 entries++;
2972 if (entries > cachep->num || i >= cachep->num)
2973 goto bad;
2974 }
2975 if (entries != cachep->num - slabp->inuse) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002976bad:
2977 printk(KERN_ERR "slab: Internal list corruption detected in "
2978 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2979 cachep->name, cachep->num, slabp, slabp->inuse);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002980 for (i = 0;
Linus Torvalds264132b2006-03-06 12:10:07 -08002981 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002982 i++) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002983 if (i % 16 == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002984 printk("\n%03x:", i);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002985 printk(" %02x", ((unsigned char *)slabp)[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002986 }
2987 printk("\n");
2988 BUG();
2989 }
2990}
2991#else
2992#define kfree_debugcheck(x) do { } while(0)
2993#define cache_free_debugcheck(x,objp,z) (objp)
2994#define check_slabp(x,y) do { } while(0)
2995#endif
2996
Pekka Enberg343e0d72006-02-01 03:05:50 -08002997static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002998{
2999 int batchcount;
3000 struct kmem_list3 *l3;
3001 struct array_cache *ac;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07003002 int node;
3003
Andrew Mortona737b3e2006-03-22 00:08:11 -08003004retry:
Joe Korty6d2144d2008-03-05 15:04:59 -08003005 check_irq_off();
3006 node = numa_node_id();
3007 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003008 batchcount = ac->batchcount;
3009 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003010 /*
3011 * If there was little recent activity on this cache, then
3012 * perform only a partial refill. Otherwise we could generate
3013 * refill bouncing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003014 */
3015 batchcount = BATCHREFILL_LIMIT;
3016 }
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07003017 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003018
Christoph Lametere498be72005-09-09 13:03:32 -07003019 BUG_ON(ac->avail > 0 || !l3);
3020 spin_lock(&l3->list_lock);
3021
Christoph Lameter3ded1752006-03-25 03:06:44 -08003022 /* See if we can refill from the shared array */
3023 if (l3->shared && transfer_objects(ac, l3->shared, batchcount))
3024 goto alloc_done;
3025
Linus Torvalds1da177e2005-04-16 15:20:36 -07003026 while (batchcount > 0) {
3027 struct list_head *entry;
3028 struct slab *slabp;
3029 /* Get slab alloc is to come from. */
3030 entry = l3->slabs_partial.next;
3031 if (entry == &l3->slabs_partial) {
3032 l3->free_touched = 1;
3033 entry = l3->slabs_free.next;
3034 if (entry == &l3->slabs_free)
3035 goto must_grow;
3036 }
3037
3038 slabp = list_entry(entry, struct slab, list);
3039 check_slabp(cachep, slabp);
3040 check_spinlock_acquired(cachep);
Pekka Enberg714b8172007-05-06 14:49:03 -07003041
3042 /*
3043 * The slab was either on partial or free list so
3044 * there must be at least one object available for
3045 * allocation.
3046 */
roel kluin249b9f32008-10-29 17:18:07 -04003047 BUG_ON(slabp->inuse >= cachep->num);
Pekka Enberg714b8172007-05-06 14:49:03 -07003048
Linus Torvalds1da177e2005-04-16 15:20:36 -07003049 while (slabp->inuse < cachep->num && batchcount--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003050 STATS_INC_ALLOCED(cachep);
3051 STATS_INC_ACTIVE(cachep);
3052 STATS_SET_HIGH(cachep);
3053
Matthew Dobson78d382d2006-02-01 03:05:47 -08003054 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07003055 node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003056 }
3057 check_slabp(cachep, slabp);
3058
3059 /* move slabp to correct slabp list: */
3060 list_del(&slabp->list);
3061 if (slabp->free == BUFCTL_END)
3062 list_add(&slabp->list, &l3->slabs_full);
3063 else
3064 list_add(&slabp->list, &l3->slabs_partial);
3065 }
3066
Andrew Mortona737b3e2006-03-22 00:08:11 -08003067must_grow:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003068 l3->free_objects -= ac->avail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003069alloc_done:
Christoph Lametere498be72005-09-09 13:03:32 -07003070 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003071
3072 if (unlikely(!ac->avail)) {
3073 int x;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003074 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07003075
Andrew Mortona737b3e2006-03-22 00:08:11 -08003076 /* cache_grow can reenable interrupts, then ac could change. */
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003077 ac = cpu_cache_get(cachep);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003078 if (!x && ac->avail == 0) /* no objects in sight? abort */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003079 return NULL;
3080
Andrew Mortona737b3e2006-03-22 00:08:11 -08003081 if (!ac->avail) /* objects refilled by interrupt? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003082 goto retry;
3083 }
3084 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07003085 return ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003086}
3087
Andrew Mortona737b3e2006-03-22 00:08:11 -08003088static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3089 gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003090{
3091 might_sleep_if(flags & __GFP_WAIT);
3092#if DEBUG
3093 kmem_flagcheck(cachep, flags);
3094#endif
3095}
3096
3097#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003098static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3099 gfp_t flags, void *objp, void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003100{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003101 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003102 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003103 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003104#ifdef CONFIG_DEBUG_PAGEALLOC
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003105 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003106 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003107 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003108 else
3109 check_poison_obj(cachep, objp);
3110#else
3111 check_poison_obj(cachep, objp);
3112#endif
3113 poison_obj(cachep, objp, POISON_INUSE);
3114 }
3115 if (cachep->flags & SLAB_STORE_USER)
3116 *dbg_userword(cachep, objp) = caller;
3117
3118 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003119 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3120 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3121 slab_error(cachep, "double free, or memory outside"
3122 " object was overwritten");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003123 printk(KERN_ERR
David Woodhouseb46b8f12007-05-08 00:22:59 -07003124 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08003125 objp, *dbg_redzone1(cachep, objp),
3126 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003127 }
3128 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3129 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3130 }
Al Viro871751e2006-03-25 03:06:39 -08003131#ifdef CONFIG_DEBUG_SLAB_LEAK
3132 {
3133 struct slab *slabp;
3134 unsigned objnr;
3135
Christoph Lameterb49af682007-05-06 14:49:41 -07003136 slabp = page_get_slab(virt_to_head_page(objp));
Al Viro871751e2006-03-25 03:06:39 -08003137 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3138 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3139 }
3140#endif
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003141 objp += obj_offset(cachep);
Christoph Lameter4f104932007-05-06 14:50:17 -07003142 if (cachep->ctor && cachep->flags & SLAB_POISON)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07003143 cachep->ctor(objp);
Kevin Hilmana44b56d2006-12-06 20:32:11 -08003144#if ARCH_SLAB_MINALIGN
3145 if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) {
3146 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3147 objp, ARCH_SLAB_MINALIGN);
3148 }
3149#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003150 return objp;
3151}
3152#else
3153#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3154#endif
3155
Akinobu Mita773ff602008-12-23 19:37:01 +09003156static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003157{
3158 if (cachep == &cache_cache)
Akinobu Mita773ff602008-12-23 19:37:01 +09003159 return false;
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003160
Akinobu Mita773ff602008-12-23 19:37:01 +09003161 return should_failslab(obj_size(cachep), flags);
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003162}
3163
Pekka Enberg343e0d72006-02-01 03:05:50 -08003164static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003165{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003166 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003167 struct array_cache *ac;
3168
Alok N Kataria5c382302005-09-27 21:45:46 -07003169 check_irq_off();
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003170
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003171 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003172 if (likely(ac->avail)) {
3173 STATS_INC_ALLOCHIT(cachep);
3174 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07003175 objp = ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003176 } else {
3177 STATS_INC_ALLOCMISS(cachep);
3178 objp = cache_alloc_refill(cachep, flags);
3179 }
Catalin Marinasd5cff632009-06-11 13:22:40 +01003180 /*
3181 * To avoid a false negative, if an object that is in one of the
3182 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3183 * treat the array pointers as a reference to the object.
3184 */
3185 kmemleak_erase(&ac->entry[ac->avail]);
Alok N Kataria5c382302005-09-27 21:45:46 -07003186 return objp;
3187}
3188
Christoph Lametere498be72005-09-09 13:03:32 -07003189#ifdef CONFIG_NUMA
3190/*
Paul Jacksonb2455392006-03-24 03:16:12 -08003191 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
Paul Jacksonc61afb12006-03-24 03:16:08 -08003192 *
3193 * If we are in_interrupt, then process context, including cpusets and
3194 * mempolicy, may not apply and should not be used for allocation policy.
3195 */
3196static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3197{
3198 int nid_alloc, nid_here;
3199
Christoph Lameter765c4502006-09-27 01:50:08 -07003200 if (in_interrupt() || (flags & __GFP_THISNODE))
Paul Jacksonc61afb12006-03-24 03:16:08 -08003201 return NULL;
3202 nid_alloc = nid_here = numa_node_id();
3203 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3204 nid_alloc = cpuset_mem_spread_node();
3205 else if (current->mempolicy)
3206 nid_alloc = slab_node(current->mempolicy);
3207 if (nid_alloc != nid_here)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003208 return ____cache_alloc_node(cachep, flags, nid_alloc);
Paul Jacksonc61afb12006-03-24 03:16:08 -08003209 return NULL;
3210}
3211
3212/*
Christoph Lameter765c4502006-09-27 01:50:08 -07003213 * Fallback function if there was no memory available and no objects on a
Christoph Lameter3c517a62006-12-06 20:33:29 -08003214 * certain node and fall back is permitted. First we scan all the
3215 * available nodelists for available objects. If that fails then we
3216 * perform an allocation without specifying a node. This allows the page
3217 * allocator to do its reclaim / fallback magic. We then insert the
3218 * slab into the proper nodelist and then allocate from it.
Christoph Lameter765c4502006-09-27 01:50:08 -07003219 */
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003220static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
Christoph Lameter765c4502006-09-27 01:50:08 -07003221{
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003222 struct zonelist *zonelist;
3223 gfp_t local_flags;
Mel Gormandd1a2392008-04-28 02:12:17 -07003224 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07003225 struct zone *zone;
3226 enum zone_type high_zoneidx = gfp_zone(flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003227 void *obj = NULL;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003228 int nid;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003229
3230 if (flags & __GFP_THISNODE)
3231 return NULL;
3232
Mel Gorman0e884602008-04-28 02:12:14 -07003233 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
Christoph Lameter6cb06222007-10-16 01:25:41 -07003234 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Christoph Lameter765c4502006-09-27 01:50:08 -07003235
Christoph Lameter3c517a62006-12-06 20:33:29 -08003236retry:
3237 /*
3238 * Look through allowed nodes for objects available
3239 * from existing per node queues.
3240 */
Mel Gorman54a6eb52008-04-28 02:12:16 -07003241 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3242 nid = zone_to_nid(zone);
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003243
Mel Gorman54a6eb52008-04-28 02:12:16 -07003244 if (cpuset_zone_allowed_hardwall(zone, flags) &&
Christoph Lameter3c517a62006-12-06 20:33:29 -08003245 cache->nodelists[nid] &&
Christoph Lameter481c5342008-06-21 16:46:35 -07003246 cache->nodelists[nid]->free_objects) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003247 obj = ____cache_alloc_node(cache,
3248 flags | GFP_THISNODE, nid);
Christoph Lameter481c5342008-06-21 16:46:35 -07003249 if (obj)
3250 break;
3251 }
Christoph Lameter3c517a62006-12-06 20:33:29 -08003252 }
3253
Christoph Lametercfce6602007-05-06 14:50:17 -07003254 if (!obj) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003255 /*
3256 * This allocation will be performed within the constraints
3257 * of the current cpuset / memory policy requirements.
3258 * We may trigger various forms of reclaim on the allowed
3259 * set and go into memory reserves if necessary.
3260 */
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003261 if (local_flags & __GFP_WAIT)
3262 local_irq_enable();
3263 kmem_flagcheck(cache, flags);
Christoph Lameter9ac33b22008-03-04 12:24:22 -08003264 obj = kmem_getpages(cache, local_flags, -1);
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003265 if (local_flags & __GFP_WAIT)
3266 local_irq_disable();
Christoph Lameter3c517a62006-12-06 20:33:29 -08003267 if (obj) {
3268 /*
3269 * Insert into the appropriate per node queues
3270 */
3271 nid = page_to_nid(virt_to_page(obj));
3272 if (cache_grow(cache, flags, nid, obj)) {
3273 obj = ____cache_alloc_node(cache,
3274 flags | GFP_THISNODE, nid);
3275 if (!obj)
3276 /*
3277 * Another processor may allocate the
3278 * objects in the slab since we are
3279 * not holding any locks.
3280 */
3281 goto retry;
3282 } else {
Hugh Dickinsb6a60452007-01-05 16:36:36 -08003283 /* cache_grow already freed obj */
Christoph Lameter3c517a62006-12-06 20:33:29 -08003284 obj = NULL;
3285 }
3286 }
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003287 }
Christoph Lameter765c4502006-09-27 01:50:08 -07003288 return obj;
3289}
3290
3291/*
Christoph Lametere498be72005-09-09 13:03:32 -07003292 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07003293 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003294static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003295 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07003296{
3297 struct list_head *entry;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003298 struct slab *slabp;
3299 struct kmem_list3 *l3;
3300 void *obj;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003301 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003302
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003303 l3 = cachep->nodelists[nodeid];
3304 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07003305
Andrew Mortona737b3e2006-03-22 00:08:11 -08003306retry:
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003307 check_irq_off();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003308 spin_lock(&l3->list_lock);
3309 entry = l3->slabs_partial.next;
3310 if (entry == &l3->slabs_partial) {
3311 l3->free_touched = 1;
3312 entry = l3->slabs_free.next;
3313 if (entry == &l3->slabs_free)
3314 goto must_grow;
3315 }
Christoph Lametere498be72005-09-09 13:03:32 -07003316
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003317 slabp = list_entry(entry, struct slab, list);
3318 check_spinlock_acquired_node(cachep, nodeid);
3319 check_slabp(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07003320
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003321 STATS_INC_NODEALLOCS(cachep);
3322 STATS_INC_ACTIVE(cachep);
3323 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003324
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003325 BUG_ON(slabp->inuse == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07003326
Matthew Dobson78d382d2006-02-01 03:05:47 -08003327 obj = slab_get_obj(cachep, slabp, nodeid);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003328 check_slabp(cachep, slabp);
3329 l3->free_objects--;
3330 /* move slabp to correct slabp list: */
3331 list_del(&slabp->list);
Christoph Lametere498be72005-09-09 13:03:32 -07003332
Andrew Mortona737b3e2006-03-22 00:08:11 -08003333 if (slabp->free == BUFCTL_END)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003334 list_add(&slabp->list, &l3->slabs_full);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003335 else
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003336 list_add(&slabp->list, &l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07003337
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003338 spin_unlock(&l3->list_lock);
3339 goto done;
Christoph Lametere498be72005-09-09 13:03:32 -07003340
Andrew Mortona737b3e2006-03-22 00:08:11 -08003341must_grow:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003342 spin_unlock(&l3->list_lock);
Christoph Lameter3c517a62006-12-06 20:33:29 -08003343 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
Christoph Lameter765c4502006-09-27 01:50:08 -07003344 if (x)
3345 goto retry;
Christoph Lametere498be72005-09-09 13:03:32 -07003346
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003347 return fallback_alloc(cachep, flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003348
Andrew Mortona737b3e2006-03-22 00:08:11 -08003349done:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003350 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07003351}
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003352
3353/**
3354 * kmem_cache_alloc_node - Allocate an object on the specified node
3355 * @cachep: The cache to allocate from.
3356 * @flags: See kmalloc().
3357 * @nodeid: node number of the target node.
3358 * @caller: return address of caller, used for debug information
3359 *
3360 * Identical to kmem_cache_alloc but it will allocate memory on the given
3361 * node, which can improve the performance for cpu bound structures.
3362 *
3363 * Fallback to other node is possible if __GFP_THISNODE is not set.
3364 */
3365static __always_inline void *
3366__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3367 void *caller)
3368{
3369 unsigned long save_flags;
3370 void *ptr;
3371
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003372 flags &= slab_gfp_mask;
3373
Nick Piggincf40bd12009-01-21 08:12:39 +01003374 lockdep_trace_alloc(flags);
3375
Akinobu Mita773ff602008-12-23 19:37:01 +09003376 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003377 return NULL;
3378
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003379 cache_alloc_debugcheck_before(cachep, flags);
3380 local_irq_save(save_flags);
3381
3382 if (unlikely(nodeid == -1))
3383 nodeid = numa_node_id();
3384
3385 if (unlikely(!cachep->nodelists[nodeid])) {
3386 /* Node not bootstrapped yet */
3387 ptr = fallback_alloc(cachep, flags);
3388 goto out;
3389 }
3390
3391 if (nodeid == numa_node_id()) {
3392 /*
3393 * Use the locally cached objects if possible.
3394 * However ____cache_alloc does not allow fallback
3395 * to other nodes. It may fail while we still have
3396 * objects on other nodes available.
3397 */
3398 ptr = ____cache_alloc(cachep, flags);
3399 if (ptr)
3400 goto out;
3401 }
3402 /* ___cache_alloc_node can fall back to other nodes */
3403 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3404 out:
3405 local_irq_restore(save_flags);
3406 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
Catalin Marinasd5cff632009-06-11 13:22:40 +01003407 kmemleak_alloc_recursive(ptr, obj_size(cachep), 1, cachep->flags,
3408 flags);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003409
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003410 if (unlikely((flags & __GFP_ZERO) && ptr))
3411 memset(ptr, 0, obj_size(cachep));
3412
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003413 return ptr;
3414}
3415
3416static __always_inline void *
3417__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3418{
3419 void *objp;
3420
3421 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3422 objp = alternate_node_alloc(cache, flags);
3423 if (objp)
3424 goto out;
3425 }
3426 objp = ____cache_alloc(cache, flags);
3427
3428 /*
3429 * We may just have run out of memory on the local node.
3430 * ____cache_alloc_node() knows how to locate memory on other nodes
3431 */
3432 if (!objp)
3433 objp = ____cache_alloc_node(cache, flags, numa_node_id());
3434
3435 out:
3436 return objp;
3437}
3438#else
3439
3440static __always_inline void *
3441__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3442{
3443 return ____cache_alloc(cachep, flags);
3444}
3445
3446#endif /* CONFIG_NUMA */
3447
3448static __always_inline void *
3449__cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3450{
3451 unsigned long save_flags;
3452 void *objp;
3453
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003454 flags &= slab_gfp_mask;
3455
Nick Piggincf40bd12009-01-21 08:12:39 +01003456 lockdep_trace_alloc(flags);
3457
Akinobu Mita773ff602008-12-23 19:37:01 +09003458 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003459 return NULL;
3460
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003461 cache_alloc_debugcheck_before(cachep, flags);
3462 local_irq_save(save_flags);
3463 objp = __do_cache_alloc(cachep, flags);
3464 local_irq_restore(save_flags);
3465 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
Catalin Marinasd5cff632009-06-11 13:22:40 +01003466 kmemleak_alloc_recursive(objp, obj_size(cachep), 1, cachep->flags,
3467 flags);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003468 prefetchw(objp);
3469
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003470 if (unlikely((flags & __GFP_ZERO) && objp))
3471 memset(objp, 0, obj_size(cachep));
3472
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003473 return objp;
3474}
Christoph Lametere498be72005-09-09 13:03:32 -07003475
3476/*
3477 * Caller needs to acquire correct kmem_list's list_lock
3478 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003479static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003480 int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003481{
3482 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07003483 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003484
3485 for (i = 0; i < nr_objects; i++) {
3486 void *objp = objpp[i];
3487 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003488
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08003489 slabp = virt_to_slab(objp);
Christoph Lameterff694162005-09-22 21:44:02 -07003490 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003491 list_del(&slabp->list);
Christoph Lameterff694162005-09-22 21:44:02 -07003492 check_spinlock_acquired_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003493 check_slabp(cachep, slabp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08003494 slab_put_obj(cachep, slabp, objp, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003495 STATS_DEC_ACTIVE(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003496 l3->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003497 check_slabp(cachep, slabp);
3498
3499 /* fixup slab chains */
3500 if (slabp->inuse == 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07003501 if (l3->free_objects > l3->free_limit) {
3502 l3->free_objects -= cachep->num;
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07003503 /* No need to drop any previously held
3504 * lock here, even if we have a off-slab slab
3505 * descriptor it is guaranteed to come from
3506 * a different cache, refer to comments before
3507 * alloc_slabmgmt.
3508 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003509 slab_destroy(cachep, slabp);
3510 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07003511 list_add(&slabp->list, &l3->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003512 }
3513 } else {
3514 /* Unconditionally move a slab to the end of the
3515 * partial list on free - maximum time for the
3516 * other objects to be freed, too.
3517 */
Christoph Lametere498be72005-09-09 13:03:32 -07003518 list_add_tail(&slabp->list, &l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003519 }
3520 }
3521}
3522
Pekka Enberg343e0d72006-02-01 03:05:50 -08003523static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003524{
3525 int batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07003526 struct kmem_list3 *l3;
Christoph Lameterff694162005-09-22 21:44:02 -07003527 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003528
3529 batchcount = ac->batchcount;
3530#if DEBUG
3531 BUG_ON(!batchcount || batchcount > ac->avail);
3532#endif
3533 check_irq_off();
Christoph Lameterff694162005-09-22 21:44:02 -07003534 l3 = cachep->nodelists[node];
Ingo Molnar873623d2006-07-13 14:44:38 +02003535 spin_lock(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003536 if (l3->shared) {
3537 struct array_cache *shared_array = l3->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003538 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003539 if (max) {
3540 if (batchcount > max)
3541 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07003542 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003543 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003544 shared_array->avail += batchcount;
3545 goto free_done;
3546 }
3547 }
3548
Christoph Lameterff694162005-09-22 21:44:02 -07003549 free_block(cachep, ac->entry, batchcount, node);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003550free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003551#if STATS
3552 {
3553 int i = 0;
3554 struct list_head *p;
3555
Christoph Lametere498be72005-09-09 13:03:32 -07003556 p = l3->slabs_free.next;
3557 while (p != &(l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003558 struct slab *slabp;
3559
3560 slabp = list_entry(p, struct slab, list);
3561 BUG_ON(slabp->inuse);
3562
3563 i++;
3564 p = p->next;
3565 }
3566 STATS_SET_FREEABLE(cachep, i);
3567 }
3568#endif
Christoph Lametere498be72005-09-09 13:03:32 -07003569 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003570 ac->avail -= batchcount;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003571 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003572}
3573
3574/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08003575 * Release an obj back to its cache. If the obj has a constructed state, it must
3576 * be in this state _before_ it is released. Called with disabled ints.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003577 */
Ingo Molnar873623d2006-07-13 14:44:38 +02003578static inline void __cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003579{
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003580 struct array_cache *ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003581
3582 check_irq_off();
Catalin Marinasd5cff632009-06-11 13:22:40 +01003583 kmemleak_free_recursive(objp, cachep->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003584 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3585
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -07003586 /*
3587 * Skip calling cache_free_alien() when the platform is not numa.
3588 * This will avoid cache misses that happen while accessing slabp (which
3589 * is per page memory reference) to get nodeid. Instead use a global
3590 * variable to skip the call, which is mostly likely to be present in
3591 * the cache.
3592 */
3593 if (numa_platform && cache_free_alien(cachep, objp))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07003594 return;
Christoph Lametere498be72005-09-09 13:03:32 -07003595
Linus Torvalds1da177e2005-04-16 15:20:36 -07003596 if (likely(ac->avail < ac->limit)) {
3597 STATS_INC_FREEHIT(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003598 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003599 return;
3600 } else {
3601 STATS_INC_FREEMISS(cachep);
3602 cache_flusharray(cachep, ac);
Christoph Lametere498be72005-09-09 13:03:32 -07003603 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003604 }
3605}
3606
3607/**
3608 * kmem_cache_alloc - Allocate an object
3609 * @cachep: The cache to allocate from.
3610 * @flags: See kmalloc().
3611 *
3612 * Allocate an object from this cache. The flags are only relevant
3613 * if the cache has no available objects.
3614 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003615void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003616{
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003617 void *ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3618
Eduard - Gabriel Munteanuca2b84c2009-03-23 15:12:24 +02003619 trace_kmem_cache_alloc(_RET_IP_, ret,
3620 obj_size(cachep), cachep->buffer_size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003621
3622 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003623}
3624EXPORT_SYMBOL(kmem_cache_alloc);
3625
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003626#ifdef CONFIG_KMEMTRACE
3627void *kmem_cache_alloc_notrace(struct kmem_cache *cachep, gfp_t flags)
3628{
3629 return __cache_alloc(cachep, flags, __builtin_return_address(0));
3630}
3631EXPORT_SYMBOL(kmem_cache_alloc_notrace);
3632#endif
3633
Linus Torvalds1da177e2005-04-16 15:20:36 -07003634/**
Randy Dunlap76824862008-03-19 17:00:40 -07003635 * kmem_ptr_validate - check if an untrusted pointer might be a slab entry.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003636 * @cachep: the cache we're checking against
3637 * @ptr: pointer to validate
3638 *
Randy Dunlap76824862008-03-19 17:00:40 -07003639 * This verifies that the untrusted pointer looks sane;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003640 * it is _not_ a guarantee that the pointer is actually
3641 * part of the slab cache in question, but it at least
3642 * validates that the pointer can be dereferenced and
3643 * looks half-way sane.
3644 *
3645 * Currently only used for dentry validation.
3646 */
Christoph Lameterb7f869a2006-12-22 01:06:44 -08003647int kmem_ptr_validate(struct kmem_cache *cachep, const void *ptr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003648{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003649 unsigned long addr = (unsigned long)ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003650 unsigned long min_addr = PAGE_OFFSET;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003651 unsigned long align_mask = BYTES_PER_WORD - 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003652 unsigned long size = cachep->buffer_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003653 struct page *page;
3654
3655 if (unlikely(addr < min_addr))
3656 goto out;
3657 if (unlikely(addr > (unsigned long)high_memory - size))
3658 goto out;
3659 if (unlikely(addr & align_mask))
3660 goto out;
3661 if (unlikely(!kern_addr_valid(addr)))
3662 goto out;
3663 if (unlikely(!kern_addr_valid(addr + size - 1)))
3664 goto out;
3665 page = virt_to_page(ptr);
3666 if (unlikely(!PageSlab(page)))
3667 goto out;
Pekka Enberg065d41c2005-11-13 16:06:46 -08003668 if (unlikely(page_get_cache(page) != cachep))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003669 goto out;
3670 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003671out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003672 return 0;
3673}
3674
3675#ifdef CONFIG_NUMA
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003676void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3677{
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003678 void *ret = __cache_alloc_node(cachep, flags, nodeid,
3679 __builtin_return_address(0));
3680
Eduard - Gabriel Munteanuca2b84c2009-03-23 15:12:24 +02003681 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3682 obj_size(cachep), cachep->buffer_size,
3683 flags, nodeid);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003684
3685 return ret;
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003686}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003687EXPORT_SYMBOL(kmem_cache_alloc_node);
3688
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003689#ifdef CONFIG_KMEMTRACE
3690void *kmem_cache_alloc_node_notrace(struct kmem_cache *cachep,
3691 gfp_t flags,
3692 int nodeid)
3693{
3694 return __cache_alloc_node(cachep, flags, nodeid,
3695 __builtin_return_address(0));
3696}
3697EXPORT_SYMBOL(kmem_cache_alloc_node_notrace);
3698#endif
3699
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003700static __always_inline void *
3701__do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003702{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003703 struct kmem_cache *cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003704 void *ret;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003705
3706 cachep = kmem_find_general_cachep(size, flags);
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003707 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3708 return cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003709 ret = kmem_cache_alloc_node_notrace(cachep, flags, node);
3710
Eduard - Gabriel Munteanuca2b84c2009-03-23 15:12:24 +02003711 trace_kmalloc_node((unsigned long) caller, ret,
3712 size, cachep->buffer_size, flags, node);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003713
3714 return ret;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003715}
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003716
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003717#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_KMEMTRACE)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003718void *__kmalloc_node(size_t size, gfp_t flags, int node)
3719{
3720 return __do_kmalloc_node(size, flags, node,
3721 __builtin_return_address(0));
3722}
Christoph Hellwigdbe5e692006-09-25 23:31:36 -07003723EXPORT_SYMBOL(__kmalloc_node);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003724
3725void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003726 int node, unsigned long caller)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003727{
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003728 return __do_kmalloc_node(size, flags, node, (void *)caller);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003729}
3730EXPORT_SYMBOL(__kmalloc_node_track_caller);
3731#else
3732void *__kmalloc_node(size_t size, gfp_t flags, int node)
3733{
3734 return __do_kmalloc_node(size, flags, node, NULL);
3735}
3736EXPORT_SYMBOL(__kmalloc_node);
3737#endif /* CONFIG_DEBUG_SLAB */
3738#endif /* CONFIG_NUMA */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003739
3740/**
Paul Drynoff800590f2006-06-23 02:03:48 -07003741 * __do_kmalloc - allocate memory
Linus Torvalds1da177e2005-04-16 15:20:36 -07003742 * @size: how many bytes of memory are required.
Paul Drynoff800590f2006-06-23 02:03:48 -07003743 * @flags: the type of memory to allocate (see kmalloc).
Randy Dunlap911851e2006-03-22 00:08:14 -08003744 * @caller: function caller for debug tracking of the caller
Linus Torvalds1da177e2005-04-16 15:20:36 -07003745 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003746static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3747 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003748{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003749 struct kmem_cache *cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003750 void *ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003751
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003752 /* If you want to save a few bytes .text space: replace
3753 * __ with kmem_.
3754 * Then kmalloc uses the uninlined functions instead of the inline
3755 * functions.
3756 */
3757 cachep = __find_general_cachep(size, flags);
Linus Torvaldsa5c96d82007-07-19 13:17:15 -07003758 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3759 return cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003760 ret = __cache_alloc(cachep, flags, caller);
3761
Eduard - Gabriel Munteanuca2b84c2009-03-23 15:12:24 +02003762 trace_kmalloc((unsigned long) caller, ret,
3763 size, cachep->buffer_size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003764
3765 return ret;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003766}
3767
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003768
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003769#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_KMEMTRACE)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003770void *__kmalloc(size_t size, gfp_t flags)
3771{
Al Viro871751e2006-03-25 03:06:39 -08003772 return __do_kmalloc(size, flags, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003773}
3774EXPORT_SYMBOL(__kmalloc);
3775
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003776void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003777{
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003778 return __do_kmalloc(size, flags, (void *)caller);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003779}
3780EXPORT_SYMBOL(__kmalloc_track_caller);
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003781
3782#else
3783void *__kmalloc(size_t size, gfp_t flags)
3784{
3785 return __do_kmalloc(size, flags, NULL);
3786}
3787EXPORT_SYMBOL(__kmalloc);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003788#endif
3789
Linus Torvalds1da177e2005-04-16 15:20:36 -07003790/**
3791 * kmem_cache_free - Deallocate an object
3792 * @cachep: The cache the allocation was from.
3793 * @objp: The previously allocated object.
3794 *
3795 * Free an object which was previously allocated from this
3796 * cache.
3797 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003798void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003799{
3800 unsigned long flags;
3801
3802 local_irq_save(flags);
Ingo Molnar898552c2007-02-10 01:44:57 -08003803 debug_check_no_locks_freed(objp, obj_size(cachep));
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07003804 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3805 debug_check_no_obj_freed(objp, obj_size(cachep));
Ingo Molnar873623d2006-07-13 14:44:38 +02003806 __cache_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003807 local_irq_restore(flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003808
Eduard - Gabriel Munteanuca2b84c2009-03-23 15:12:24 +02003809 trace_kmem_cache_free(_RET_IP_, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003810}
3811EXPORT_SYMBOL(kmem_cache_free);
3812
3813/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003814 * kfree - free previously allocated memory
3815 * @objp: pointer returned by kmalloc.
3816 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003817 * If @objp is NULL, no operation is performed.
3818 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003819 * Don't free memory not originally allocated by kmalloc()
3820 * or you will run into trouble.
3821 */
3822void kfree(const void *objp)
3823{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003824 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003825 unsigned long flags;
3826
Pekka Enberg2121db72009-03-25 11:05:57 +02003827 trace_kfree(_RET_IP_, objp);
3828
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003829 if (unlikely(ZERO_OR_NULL_PTR(objp)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003830 return;
3831 local_irq_save(flags);
3832 kfree_debugcheck(objp);
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08003833 c = virt_to_cache(objp);
Ingo Molnarf9b84042006-06-27 02:54:49 -07003834 debug_check_no_locks_freed(objp, obj_size(c));
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07003835 debug_check_no_obj_freed(objp, obj_size(c));
Ingo Molnar873623d2006-07-13 14:44:38 +02003836 __cache_free(c, (void *)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003837 local_irq_restore(flags);
3838}
3839EXPORT_SYMBOL(kfree);
3840
Pekka Enberg343e0d72006-02-01 03:05:50 -08003841unsigned int kmem_cache_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003842{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003843 return obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003844}
3845EXPORT_SYMBOL(kmem_cache_size);
3846
Pekka Enberg343e0d72006-02-01 03:05:50 -08003847const char *kmem_cache_name(struct kmem_cache *cachep)
Arnaldo Carvalho de Melo19449722005-06-18 22:46:19 -07003848{
3849 return cachep->name;
3850}
3851EXPORT_SYMBOL_GPL(kmem_cache_name);
3852
Christoph Lametere498be72005-09-09 13:03:32 -07003853/*
Simon Arlott183ff222007-10-20 01:27:18 +02003854 * This initializes kmem_list3 or resizes various caches for all nodes.
Christoph Lametere498be72005-09-09 13:03:32 -07003855 */
Pekka Enberg83b519e2009-06-10 19:40:04 +03003856static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -07003857{
3858 int node;
3859 struct kmem_list3 *l3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003860 struct array_cache *new_shared;
Paul Menage3395ee02006-12-06 20:32:16 -08003861 struct array_cache **new_alien = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003862
Mel Gorman9c09a952008-01-24 05:49:54 -08003863 for_each_online_node(node) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003864
Paul Menage3395ee02006-12-06 20:32:16 -08003865 if (use_alien_caches) {
Pekka Enberg83b519e2009-06-10 19:40:04 +03003866 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
Paul Menage3395ee02006-12-06 20:32:16 -08003867 if (!new_alien)
3868 goto fail;
3869 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003870
Eric Dumazet63109842007-05-06 14:49:28 -07003871 new_shared = NULL;
3872 if (cachep->shared) {
3873 new_shared = alloc_arraycache(node,
Christoph Lameter0718dc22006-03-25 03:06:47 -08003874 cachep->shared*cachep->batchcount,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003875 0xbaadf00d, gfp);
Eric Dumazet63109842007-05-06 14:49:28 -07003876 if (!new_shared) {
3877 free_alien_cache(new_alien);
3878 goto fail;
3879 }
Christoph Lameter0718dc22006-03-25 03:06:47 -08003880 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003881
Andrew Mortona737b3e2006-03-22 00:08:11 -08003882 l3 = cachep->nodelists[node];
3883 if (l3) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003884 struct array_cache *shared = l3->shared;
3885
Christoph Lametere498be72005-09-09 13:03:32 -07003886 spin_lock_irq(&l3->list_lock);
3887
Christoph Lametercafeb022006-03-25 03:06:46 -08003888 if (shared)
Christoph Lameter0718dc22006-03-25 03:06:47 -08003889 free_block(cachep, shared->entry,
3890 shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07003891
Christoph Lametercafeb022006-03-25 03:06:46 -08003892 l3->shared = new_shared;
3893 if (!l3->alien) {
Christoph Lametere498be72005-09-09 13:03:32 -07003894 l3->alien = new_alien;
3895 new_alien = NULL;
3896 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003897 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003898 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003899 spin_unlock_irq(&l3->list_lock);
Christoph Lametercafeb022006-03-25 03:06:46 -08003900 kfree(shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003901 free_alien_cache(new_alien);
3902 continue;
3903 }
Pekka Enberg83b519e2009-06-10 19:40:04 +03003904 l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node);
Christoph Lameter0718dc22006-03-25 03:06:47 -08003905 if (!l3) {
3906 free_alien_cache(new_alien);
3907 kfree(new_shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003908 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003909 }
Christoph Lametere498be72005-09-09 13:03:32 -07003910
3911 kmem_list3_init(l3);
3912 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Andrew Mortona737b3e2006-03-22 00:08:11 -08003913 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003914 l3->shared = new_shared;
Christoph Lametere498be72005-09-09 13:03:32 -07003915 l3->alien = new_alien;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003916 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003917 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003918 cachep->nodelists[node] = l3;
3919 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003920 return 0;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003921
Andrew Mortona737b3e2006-03-22 00:08:11 -08003922fail:
Christoph Lameter0718dc22006-03-25 03:06:47 -08003923 if (!cachep->next.next) {
3924 /* Cache is not active yet. Roll back what we did */
3925 node--;
3926 while (node >= 0) {
3927 if (cachep->nodelists[node]) {
3928 l3 = cachep->nodelists[node];
3929
3930 kfree(l3->shared);
3931 free_alien_cache(l3->alien);
3932 kfree(l3);
3933 cachep->nodelists[node] = NULL;
3934 }
3935 node--;
3936 }
3937 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003938 return -ENOMEM;
Christoph Lametere498be72005-09-09 13:03:32 -07003939}
3940
Linus Torvalds1da177e2005-04-16 15:20:36 -07003941struct ccupdate_struct {
Pekka Enberg343e0d72006-02-01 03:05:50 -08003942 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003943 struct array_cache *new[NR_CPUS];
3944};
3945
3946static void do_ccupdate_local(void *info)
3947{
Andrew Mortona737b3e2006-03-22 00:08:11 -08003948 struct ccupdate_struct *new = info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003949 struct array_cache *old;
3950
3951 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003952 old = cpu_cache_get(new->cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003953
Linus Torvalds1da177e2005-04-16 15:20:36 -07003954 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3955 new->new[smp_processor_id()] = old;
3956}
3957
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003958/* Always called with the cache_chain_mutex held */
Andrew Mortona737b3e2006-03-22 00:08:11 -08003959static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003960 int batchcount, int shared, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003961{
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003962 struct ccupdate_struct *new;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003963 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003964
Pekka Enberg83b519e2009-06-10 19:40:04 +03003965 new = kzalloc(sizeof(*new), gfp);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003966 if (!new)
3967 return -ENOMEM;
3968
Christoph Lametere498be72005-09-09 13:03:32 -07003969 for_each_online_cpu(i) {
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003970 new->new[i] = alloc_arraycache(cpu_to_node(i), limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003971 batchcount, gfp);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003972 if (!new->new[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003973 for (i--; i >= 0; i--)
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003974 kfree(new->new[i]);
3975 kfree(new);
Christoph Lametere498be72005-09-09 13:03:32 -07003976 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003977 }
3978 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003979 new->cachep = cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003980
Jens Axboe15c8b6c2008-05-09 09:39:44 +02003981 on_each_cpu(do_ccupdate_local, (void *)new, 1);
Christoph Lametere498be72005-09-09 13:03:32 -07003982
Linus Torvalds1da177e2005-04-16 15:20:36 -07003983 check_irq_on();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003984 cachep->batchcount = batchcount;
3985 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07003986 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003987
Christoph Lametere498be72005-09-09 13:03:32 -07003988 for_each_online_cpu(i) {
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003989 struct array_cache *ccold = new->new[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003990 if (!ccold)
3991 continue;
Christoph Lametere498be72005-09-09 13:03:32 -07003992 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -07003993 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
Christoph Lametere498be72005-09-09 13:03:32 -07003994 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003995 kfree(ccold);
3996 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003997 kfree(new);
Pekka Enberg83b519e2009-06-10 19:40:04 +03003998 return alloc_kmemlist(cachep, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003999}
4000
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08004001/* Called with cache_chain_mutex held always */
Pekka Enberg83b519e2009-06-10 19:40:04 +03004002static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004003{
4004 int err;
4005 int limit, shared;
4006
Andrew Mortona737b3e2006-03-22 00:08:11 -08004007 /*
4008 * The head array serves three purposes:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004009 * - create a LIFO ordering, i.e. return objects that are cache-warm
4010 * - reduce the number of spinlock operations.
Andrew Mortona737b3e2006-03-22 00:08:11 -08004011 * - reduce the number of linked list operations on the slab and
Linus Torvalds1da177e2005-04-16 15:20:36 -07004012 * bufctl chains: array operations are cheaper.
4013 * The numbers are guessed, we should auto-tune as described by
4014 * Bonwick.
4015 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004016 if (cachep->buffer_size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004017 limit = 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004018 else if (cachep->buffer_size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004019 limit = 8;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004020 else if (cachep->buffer_size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004021 limit = 24;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004022 else if (cachep->buffer_size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004023 limit = 54;
4024 else
4025 limit = 120;
4026
Andrew Mortona737b3e2006-03-22 00:08:11 -08004027 /*
4028 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
Linus Torvalds1da177e2005-04-16 15:20:36 -07004029 * allocation behaviour: Most allocs on one cpu, most free operations
4030 * on another cpu. For these cases, an efficient object passing between
4031 * cpus is necessary. This is provided by a shared array. The array
4032 * replaces Bonwick's magazine layer.
4033 * On uniprocessor, it's functionally equivalent (but less efficient)
4034 * to a larger limit. Thus disabled by default.
4035 */
4036 shared = 0;
Eric Dumazet364fbb22007-05-06 14:49:27 -07004037 if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004038 shared = 8;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004039
4040#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08004041 /*
4042 * With debugging enabled, large batchcount lead to excessively long
4043 * periods with disabled local interrupts. Limit the batchcount
Linus Torvalds1da177e2005-04-16 15:20:36 -07004044 */
4045 if (limit > 32)
4046 limit = 32;
4047#endif
Pekka Enberg83b519e2009-06-10 19:40:04 +03004048 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004049 if (err)
4050 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004051 cachep->name, -err);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07004052 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004053}
4054
Christoph Lameter1b552532006-03-22 00:09:07 -08004055/*
4056 * Drain an array if it contains any elements taking the l3 lock only if
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004057 * necessary. Note that the l3 listlock also protects the array_cache
4058 * if drain_array() is used on the shared array.
Christoph Lameter1b552532006-03-22 00:09:07 -08004059 */
4060void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
4061 struct array_cache *ac, int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004062{
4063 int tofree;
4064
Christoph Lameter1b552532006-03-22 00:09:07 -08004065 if (!ac || !ac->avail)
4066 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004067 if (ac->touched && !force) {
4068 ac->touched = 0;
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004069 } else {
Christoph Lameter1b552532006-03-22 00:09:07 -08004070 spin_lock_irq(&l3->list_lock);
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004071 if (ac->avail) {
4072 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4073 if (tofree > ac->avail)
4074 tofree = (ac->avail + 1) / 2;
4075 free_block(cachep, ac->entry, tofree, node);
4076 ac->avail -= tofree;
4077 memmove(ac->entry, &(ac->entry[tofree]),
4078 sizeof(void *) * ac->avail);
4079 }
Christoph Lameter1b552532006-03-22 00:09:07 -08004080 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004081 }
4082}
4083
4084/**
4085 * cache_reap - Reclaim memory from caches.
Randy Dunlap05fb6bf2007-02-28 20:12:13 -08004086 * @w: work descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07004087 *
4088 * Called from workqueue/eventd every few seconds.
4089 * Purpose:
4090 * - clear the per-cpu caches for this CPU.
4091 * - return freeable pages to the main free memory pool.
4092 *
Andrew Mortona737b3e2006-03-22 00:08:11 -08004093 * If we cannot acquire the cache chain mutex then just give up - we'll try
4094 * again on the next iteration.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004095 */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004096static void cache_reap(struct work_struct *w)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004097{
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004098 struct kmem_cache *searchp;
Christoph Lametere498be72005-09-09 13:03:32 -07004099 struct kmem_list3 *l3;
Christoph Lameteraab22072006-03-22 00:09:06 -08004100 int node = numa_node_id();
Jean Delvarebf6aede2009-04-02 16:56:54 -07004101 struct delayed_work *work = to_delayed_work(w);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004102
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004103 if (!mutex_trylock(&cache_chain_mutex))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004104 /* Give up. Setup the next iteration. */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004105 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004106
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004107 list_for_each_entry(searchp, &cache_chain, next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004108 check_irq_on();
4109
Christoph Lameter35386e32006-03-22 00:09:05 -08004110 /*
4111 * We only take the l3 lock if absolutely necessary and we
4112 * have established with reasonable certainty that
4113 * we can do some work if the lock was obtained.
4114 */
Christoph Lameteraab22072006-03-22 00:09:06 -08004115 l3 = searchp->nodelists[node];
Christoph Lameter35386e32006-03-22 00:09:05 -08004116
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004117 reap_alien(searchp, l3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004118
Christoph Lameteraab22072006-03-22 00:09:06 -08004119 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004120
Christoph Lameter35386e32006-03-22 00:09:05 -08004121 /*
4122 * These are racy checks but it does not matter
4123 * if we skip one check or scan twice.
4124 */
Christoph Lametere498be72005-09-09 13:03:32 -07004125 if (time_after(l3->next_reap, jiffies))
Christoph Lameter35386e32006-03-22 00:09:05 -08004126 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004127
Christoph Lametere498be72005-09-09 13:03:32 -07004128 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004129
Christoph Lameteraab22072006-03-22 00:09:06 -08004130 drain_array(searchp, l3, l3->shared, 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004131
Christoph Lametered11d9e2006-06-30 01:55:45 -07004132 if (l3->free_touched)
Christoph Lametere498be72005-09-09 13:03:32 -07004133 l3->free_touched = 0;
Christoph Lametered11d9e2006-06-30 01:55:45 -07004134 else {
4135 int freed;
4136
4137 freed = drain_freelist(searchp, l3, (l3->free_limit +
4138 5 * searchp->num - 1) / (5 * searchp->num));
4139 STATS_ADD_REAPED(searchp, freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004140 }
Christoph Lameter35386e32006-03-22 00:09:05 -08004141next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004142 cond_resched();
4143 }
4144 check_irq_on();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004145 mutex_unlock(&cache_chain_mutex);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004146 next_reap_node();
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004147out:
Andrew Mortona737b3e2006-03-22 00:08:11 -08004148 /* Set up the next iteration */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004149 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004150}
4151
Linus Torvalds158a9622008-01-02 13:04:48 -08004152#ifdef CONFIG_SLABINFO
Linus Torvalds1da177e2005-04-16 15:20:36 -07004153
Pekka Enberg85289f92006-01-08 01:00:36 -08004154static void print_slabinfo_header(struct seq_file *m)
4155{
4156 /*
4157 * Output format version, so at least we can change it
4158 * without _too_ many complaints.
4159 */
4160#if STATS
4161 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4162#else
4163 seq_puts(m, "slabinfo - version: 2.1\n");
4164#endif
4165 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4166 "<objperslab> <pagesperslab>");
4167 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4168 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4169#if STATS
4170 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004171 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
Pekka Enberg85289f92006-01-08 01:00:36 -08004172 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4173#endif
4174 seq_putc(m, '\n');
4175}
4176
Linus Torvalds1da177e2005-04-16 15:20:36 -07004177static void *s_start(struct seq_file *m, loff_t *pos)
4178{
4179 loff_t n = *pos;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004180
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004181 mutex_lock(&cache_chain_mutex);
Pekka Enberg85289f92006-01-08 01:00:36 -08004182 if (!n)
4183 print_slabinfo_header(m);
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004184
4185 return seq_list_start(&cache_chain, *pos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004186}
4187
4188static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4189{
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004190 return seq_list_next(p, &cache_chain, pos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004191}
4192
4193static void s_stop(struct seq_file *m, void *p)
4194{
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004195 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004196}
4197
4198static int s_show(struct seq_file *m, void *p)
4199{
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004200 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004201 struct slab *slabp;
4202 unsigned long active_objs;
4203 unsigned long num_objs;
4204 unsigned long active_slabs = 0;
4205 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004206 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004207 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07004208 int node;
4209 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004210
Linus Torvalds1da177e2005-04-16 15:20:36 -07004211 active_objs = 0;
4212 num_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004213 for_each_online_node(node) {
4214 l3 = cachep->nodelists[node];
4215 if (!l3)
4216 continue;
4217
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004218 check_irq_on();
4219 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07004220
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004221 list_for_each_entry(slabp, &l3->slabs_full, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004222 if (slabp->inuse != cachep->num && !error)
4223 error = "slabs_full accounting error";
4224 active_objs += cachep->num;
4225 active_slabs++;
4226 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004227 list_for_each_entry(slabp, &l3->slabs_partial, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004228 if (slabp->inuse == cachep->num && !error)
4229 error = "slabs_partial inuse accounting error";
4230 if (!slabp->inuse && !error)
4231 error = "slabs_partial/inuse accounting error";
4232 active_objs += slabp->inuse;
4233 active_slabs++;
4234 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004235 list_for_each_entry(slabp, &l3->slabs_free, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004236 if (slabp->inuse && !error)
4237 error = "slabs_free/inuse accounting error";
4238 num_slabs++;
4239 }
4240 free_objects += l3->free_objects;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08004241 if (l3->shared)
4242 shared_avail += l3->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07004243
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004244 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004245 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004246 num_slabs += active_slabs;
4247 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07004248 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004249 error = "free_objects accounting error";
4250
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004251 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004252 if (error)
4253 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4254
4255 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004256 name, active_objs, num_objs, cachep->buffer_size,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004257 cachep->num, (1 << cachep->gfporder));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004258 seq_printf(m, " : tunables %4u %4u %4u",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004259 cachep->limit, cachep->batchcount, cachep->shared);
Christoph Lametere498be72005-09-09 13:03:32 -07004260 seq_printf(m, " : slabdata %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004261 active_slabs, num_slabs, shared_avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004262#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004263 { /* list3 stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004264 unsigned long high = cachep->high_mark;
4265 unsigned long allocs = cachep->num_allocations;
4266 unsigned long grown = cachep->grown;
4267 unsigned long reaped = cachep->reaped;
4268 unsigned long errors = cachep->errors;
4269 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004270 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07004271 unsigned long node_frees = cachep->node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004272 unsigned long overflows = cachep->node_overflow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004273
Christoph Lametere498be72005-09-09 13:03:32 -07004274 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004275 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown,
Andrew Mortona737b3e2006-03-22 00:08:11 -08004276 reaped, errors, max_freeable, node_allocs,
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004277 node_frees, overflows);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004278 }
4279 /* cpu stats */
4280 {
4281 unsigned long allochit = atomic_read(&cachep->allochit);
4282 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4283 unsigned long freehit = atomic_read(&cachep->freehit);
4284 unsigned long freemiss = atomic_read(&cachep->freemiss);
4285
4286 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004287 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004288 }
4289#endif
4290 seq_putc(m, '\n');
Linus Torvalds1da177e2005-04-16 15:20:36 -07004291 return 0;
4292}
4293
4294/*
4295 * slabinfo_op - iterator that generates /proc/slabinfo
4296 *
4297 * Output layout:
4298 * cache-name
4299 * num-active-objs
4300 * total-objs
4301 * object size
4302 * num-active-slabs
4303 * total-slabs
4304 * num-pages-per-slab
4305 * + further values on SMP and with statistics enabled
4306 */
4307
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04004308static const struct seq_operations slabinfo_op = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004309 .start = s_start,
4310 .next = s_next,
4311 .stop = s_stop,
4312 .show = s_show,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004313};
4314
4315#define MAX_SLABINFO_WRITE 128
4316/**
4317 * slabinfo_write - Tuning for the slab allocator
4318 * @file: unused
4319 * @buffer: user buffer
4320 * @count: data length
4321 * @ppos: unused
4322 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004323ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4324 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004325{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004326 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004327 int limit, batchcount, shared, res;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004328 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004329
Linus Torvalds1da177e2005-04-16 15:20:36 -07004330 if (count > MAX_SLABINFO_WRITE)
4331 return -EINVAL;
4332 if (copy_from_user(&kbuf, buffer, count))
4333 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004334 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07004335
4336 tmp = strchr(kbuf, ' ');
4337 if (!tmp)
4338 return -EINVAL;
4339 *tmp = '\0';
4340 tmp++;
4341 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4342 return -EINVAL;
4343
4344 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004345 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004346 res = -EINVAL;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004347 list_for_each_entry(cachep, &cache_chain, next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004348 if (!strcmp(cachep->name, kbuf)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08004349 if (limit < 1 || batchcount < 1 ||
4350 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07004351 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004352 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07004353 res = do_tune_cpucache(cachep, limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03004354 batchcount, shared,
4355 GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004356 }
4357 break;
4358 }
4359 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004360 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004361 if (res >= 0)
4362 res = count;
4363 return res;
4364}
Al Viro871751e2006-03-25 03:06:39 -08004365
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04004366static int slabinfo_open(struct inode *inode, struct file *file)
4367{
4368 return seq_open(file, &slabinfo_op);
4369}
4370
4371static const struct file_operations proc_slabinfo_operations = {
4372 .open = slabinfo_open,
4373 .read = seq_read,
4374 .write = slabinfo_write,
4375 .llseek = seq_lseek,
4376 .release = seq_release,
4377};
4378
Al Viro871751e2006-03-25 03:06:39 -08004379#ifdef CONFIG_DEBUG_SLAB_LEAK
4380
4381static void *leaks_start(struct seq_file *m, loff_t *pos)
4382{
Al Viro871751e2006-03-25 03:06:39 -08004383 mutex_lock(&cache_chain_mutex);
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004384 return seq_list_start(&cache_chain, *pos);
Al Viro871751e2006-03-25 03:06:39 -08004385}
4386
4387static inline int add_caller(unsigned long *n, unsigned long v)
4388{
4389 unsigned long *p;
4390 int l;
4391 if (!v)
4392 return 1;
4393 l = n[1];
4394 p = n + 2;
4395 while (l) {
4396 int i = l/2;
4397 unsigned long *q = p + 2 * i;
4398 if (*q == v) {
4399 q[1]++;
4400 return 1;
4401 }
4402 if (*q > v) {
4403 l = i;
4404 } else {
4405 p = q + 2;
4406 l -= i + 1;
4407 }
4408 }
4409 if (++n[1] == n[0])
4410 return 0;
4411 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4412 p[0] = v;
4413 p[1] = 1;
4414 return 1;
4415}
4416
4417static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4418{
4419 void *p;
4420 int i;
4421 if (n[0] == n[1])
4422 return;
4423 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4424 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4425 continue;
4426 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4427 return;
4428 }
4429}
4430
4431static void show_symbol(struct seq_file *m, unsigned long address)
4432{
4433#ifdef CONFIG_KALLSYMS
Al Viro871751e2006-03-25 03:06:39 -08004434 unsigned long offset, size;
Tejun Heo9281ace2007-07-17 04:03:51 -07004435 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
Al Viro871751e2006-03-25 03:06:39 -08004436
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004437 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
Al Viro871751e2006-03-25 03:06:39 -08004438 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004439 if (modname[0])
Al Viro871751e2006-03-25 03:06:39 -08004440 seq_printf(m, " [%s]", modname);
4441 return;
4442 }
4443#endif
4444 seq_printf(m, "%p", (void *)address);
4445}
4446
4447static int leaks_show(struct seq_file *m, void *p)
4448{
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004449 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
Al Viro871751e2006-03-25 03:06:39 -08004450 struct slab *slabp;
4451 struct kmem_list3 *l3;
4452 const char *name;
4453 unsigned long *n = m->private;
4454 int node;
4455 int i;
4456
4457 if (!(cachep->flags & SLAB_STORE_USER))
4458 return 0;
4459 if (!(cachep->flags & SLAB_RED_ZONE))
4460 return 0;
4461
4462 /* OK, we can do it */
4463
4464 n[1] = 0;
4465
4466 for_each_online_node(node) {
4467 l3 = cachep->nodelists[node];
4468 if (!l3)
4469 continue;
4470
4471 check_irq_on();
4472 spin_lock_irq(&l3->list_lock);
4473
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004474 list_for_each_entry(slabp, &l3->slabs_full, list)
Al Viro871751e2006-03-25 03:06:39 -08004475 handle_slab(n, cachep, slabp);
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004476 list_for_each_entry(slabp, &l3->slabs_partial, list)
Al Viro871751e2006-03-25 03:06:39 -08004477 handle_slab(n, cachep, slabp);
Al Viro871751e2006-03-25 03:06:39 -08004478 spin_unlock_irq(&l3->list_lock);
4479 }
4480 name = cachep->name;
4481 if (n[0] == n[1]) {
4482 /* Increase the buffer size */
4483 mutex_unlock(&cache_chain_mutex);
4484 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4485 if (!m->private) {
4486 /* Too bad, we are really out */
4487 m->private = n;
4488 mutex_lock(&cache_chain_mutex);
4489 return -ENOMEM;
4490 }
4491 *(unsigned long *)m->private = n[0] * 2;
4492 kfree(n);
4493 mutex_lock(&cache_chain_mutex);
4494 /* Now make sure this entry will be retried */
4495 m->count = m->size;
4496 return 0;
4497 }
4498 for (i = 0; i < n[1]; i++) {
4499 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4500 show_symbol(m, n[2*i+2]);
4501 seq_putc(m, '\n');
4502 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004503
Al Viro871751e2006-03-25 03:06:39 -08004504 return 0;
4505}
4506
Alexey Dobriyana0ec95a82008-10-06 00:59:10 +04004507static const struct seq_operations slabstats_op = {
Al Viro871751e2006-03-25 03:06:39 -08004508 .start = leaks_start,
4509 .next = s_next,
4510 .stop = s_stop,
4511 .show = leaks_show,
4512};
Alexey Dobriyana0ec95a82008-10-06 00:59:10 +04004513
4514static int slabstats_open(struct inode *inode, struct file *file)
4515{
4516 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4517 int ret = -ENOMEM;
4518 if (n) {
4519 ret = seq_open(file, &slabstats_op);
4520 if (!ret) {
4521 struct seq_file *m = file->private_data;
4522 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4523 m->private = n;
4524 n = NULL;
4525 }
4526 kfree(n);
4527 }
4528 return ret;
4529}
4530
4531static const struct file_operations proc_slabstats_operations = {
4532 .open = slabstats_open,
4533 .read = seq_read,
4534 .llseek = seq_lseek,
4535 .release = seq_release_private,
4536};
Al Viro871751e2006-03-25 03:06:39 -08004537#endif
Alexey Dobriyana0ec95a82008-10-06 00:59:10 +04004538
4539static int __init slab_proc_init(void)
4540{
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04004541 proc_create("slabinfo",S_IWUSR|S_IRUGO,NULL,&proc_slabinfo_operations);
Alexey Dobriyana0ec95a82008-10-06 00:59:10 +04004542#ifdef CONFIG_DEBUG_SLAB_LEAK
4543 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4544#endif
4545 return 0;
4546}
4547module_init(slab_proc_init);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004548#endif
4549
Manfred Spraul00e145b2005-09-03 15:55:07 -07004550/**
4551 * ksize - get the actual amount of memory allocated for a given object
4552 * @objp: Pointer to the object
4553 *
4554 * kmalloc may internally round up allocations and return more memory
4555 * than requested. ksize() can be used to determine the actual amount of
4556 * memory allocated. The caller may use this additional memory, even though
4557 * a smaller amount of memory was initially specified with the kmalloc call.
4558 * The caller must guarantee that objp points to a valid object previously
4559 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4560 * must not be freed during the duration of the call.
4561 */
Pekka Enbergfd76bab2007-05-06 14:48:40 -07004562size_t ksize(const void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004563{
Christoph Lameteref8b4522007-10-16 01:24:46 -07004564 BUG_ON(!objp);
4565 if (unlikely(objp == ZERO_SIZE_PTR))
Manfred Spraul00e145b2005-09-03 15:55:07 -07004566 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004567
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08004568 return obj_size(virt_to_cache(objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004569}
Kirill A. Shutemovb1aabec2009-02-10 15:21:44 +02004570EXPORT_SYMBOL(ksize);