blob: 03927cb5ec9e119ca06cf337d69de017303d82ba [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>
Linus Torvalds1da177e2005-04-16 15:20:36 -070098#include <linux/seq_file.h>
99#include <linux/notifier.h>
100#include <linux/kallsyms.h>
101#include <linux/cpu.h>
102#include <linux/sysctl.h>
103#include <linux/module.h>
104#include <linux/rcupdate.h>
Paulo Marques543537b2005-06-23 00:09:02 -0700105#include <linux/string.h>
Andrew Morton138ae662006-12-06 20:36:41 -0800106#include <linux/uaccess.h>
Christoph Lametere498be72005-09-09 13:03:32 -0700107#include <linux/nodemask.h>
Christoph Lameterdc85da12006-01-18 17:42:36 -0800108#include <linux/mempolicy.h>
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800109#include <linux/mutex.h>
Akinobu Mita8a8b6502006-12-08 02:39:44 -0800110#include <linux/fault-inject.h>
Ingo Molnare7eebaf2006-06-27 02:54:55 -0700111#include <linux/rtmutex.h>
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800112#include <linux/reciprocal_div.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700113
Linus Torvalds1da177e2005-04-16 15:20:36 -0700114#include <asm/cacheflush.h>
115#include <asm/tlbflush.h>
116#include <asm/page.h>
117
118/*
Christoph Lameter50953fe2007-05-06 14:50:16 -0700119 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700120 * 0 for faster, smaller code (especially in the critical paths).
121 *
122 * STATS - 1 to collect stats for /proc/slabinfo.
123 * 0 for faster, smaller code (especially in the critical paths).
124 *
125 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
126 */
127
128#ifdef CONFIG_DEBUG_SLAB
129#define DEBUG 1
130#define STATS 1
131#define FORCED_DEBUG 1
132#else
133#define DEBUG 0
134#define STATS 0
135#define FORCED_DEBUG 0
136#endif
137
Linus Torvalds1da177e2005-04-16 15:20:36 -0700138/* Shouldn't this be in a header file somewhere? */
139#define BYTES_PER_WORD sizeof(void *)
David Woodhouse87a927c2007-07-04 21:26:44 -0400140#define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700141
142#ifndef cache_line_size
143#define cache_line_size() L1_CACHE_BYTES
144#endif
145
146#ifndef ARCH_KMALLOC_MINALIGN
147/*
148 * Enforce a minimum alignment for the kmalloc caches.
149 * Usually, the kmalloc caches are cache_line_size() aligned, except when
150 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
151 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
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 | \
Paul Jackson101a5002006-03-24 03:16:07 -0800181 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700182#else
Christoph Lameterac2b8982006-03-22 00:08:15 -0800183# define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
Christoph Lameter5af60832007-05-06 14:49:56 -0700184 SLAB_CACHE_DMA | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700185 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
Paul Jackson101a5002006-03-24 03:16:07 -0800186 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700187#endif
188
189/*
190 * kmem_bufctl_t:
191 *
192 * Bufctl's are used for linking objs within a slab
193 * linked offsets.
194 *
195 * This implementation relies on "struct page" for locating the cache &
196 * slab an object belongs to.
197 * This allows the bufctl structure to be small (one int), but limits
198 * the number of objects a slab (not a cache) can contain when off-slab
199 * bufctls are used. The limit is the size of the largest general cache
200 * that does not use off-slab slabs.
201 * For 32bit archs with 4 kB pages, is this 56.
202 * This is not serious, as it is only for large objects, when it is unwise
203 * to have too many per slab.
204 * Note: This limit can be raised by introducing a general cache whose size
205 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
206 */
207
Kyle Moffettfa5b08d2005-09-03 15:55:03 -0700208typedef unsigned int kmem_bufctl_t;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700209#define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
210#define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
Al Viro871751e2006-03-25 03:06:39 -0800211#define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
212#define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700213
Linus Torvalds1da177e2005-04-16 15:20:36 -0700214/*
215 * struct slab
216 *
217 * Manages the objs in a slab. Placed either at the beginning of mem allocated
218 * for a slab, or allocated from an general cache.
219 * Slabs are chained into three list: fully used, partial, fully free slabs.
220 */
221struct slab {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800222 struct list_head list;
223 unsigned long colouroff;
224 void *s_mem; /* including colour offset */
225 unsigned int inuse; /* num of objs active in slab */
226 kmem_bufctl_t free;
227 unsigned short nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700228};
229
230/*
231 * struct slab_rcu
232 *
233 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
234 * arrange for kmem_freepages to be called via RCU. This is useful if
235 * we need to approach a kernel structure obliquely, from its address
236 * obtained without the usual locking. We can lock the structure to
237 * stabilize it and check it's still at the given address, only if we
238 * can be sure that the memory has not been meanwhile reused for some
239 * other kind of object (which our subsystem's lock might corrupt).
240 *
241 * rcu_read_lock before reading the address, then rcu_read_unlock after
242 * taking the spinlock within the structure expected at that address.
243 *
244 * We assume struct slab_rcu can overlay struct slab when destroying.
245 */
246struct slab_rcu {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800247 struct rcu_head head;
Pekka Enberg343e0d72006-02-01 03:05:50 -0800248 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800249 void *addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700250};
251
252/*
253 * struct array_cache
254 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700255 * Purpose:
256 * - LIFO ordering, to hand out cache-warm objects from _alloc
257 * - reduce the number of linked list operations
258 * - reduce spinlock operations
259 *
260 * The limit is stored in the per-cpu structure to reduce the data cache
261 * footprint.
262 *
263 */
264struct array_cache {
265 unsigned int avail;
266 unsigned int limit;
267 unsigned int batchcount;
268 unsigned int touched;
Christoph Lametere498be72005-09-09 13:03:32 -0700269 spinlock_t lock;
Robert P. J. Daybda5b652007-10-16 23:30:05 -0700270 void *entry[]; /*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800271 * Must have this definition in here for the proper
272 * alignment of array_cache. Also simplifies accessing
273 * the entries.
Andrew Mortona737b3e2006-03-22 00:08:11 -0800274 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700275};
276
Andrew Mortona737b3e2006-03-22 00:08:11 -0800277/*
278 * bootstrap: The caches do not work without cpuarrays anymore, but the
279 * cpuarrays are allocated from the generic caches...
Linus Torvalds1da177e2005-04-16 15:20:36 -0700280 */
281#define BOOT_CPUCACHE_ENTRIES 1
282struct arraycache_init {
283 struct array_cache cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800284 void *entries[BOOT_CPUCACHE_ENTRIES];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700285};
286
287/*
Christoph Lametere498be72005-09-09 13:03:32 -0700288 * The slab lists for all objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700289 */
290struct kmem_list3 {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800291 struct list_head slabs_partial; /* partial list first, better asm code */
292 struct list_head slabs_full;
293 struct list_head slabs_free;
294 unsigned long free_objects;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800295 unsigned int free_limit;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800296 unsigned int colour_next; /* Per-node cache coloring */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800297 spinlock_t list_lock;
298 struct array_cache *shared; /* shared per node */
299 struct array_cache **alien; /* on other nodes */
Christoph Lameter35386e32006-03-22 00:09:05 -0800300 unsigned long next_reap; /* updated without locking */
301 int free_touched; /* updated without locking */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700302};
303
Christoph Lametere498be72005-09-09 13:03:32 -0700304/*
305 * Need this for bootstrapping a per node allocator.
306 */
Pekka Enberg556a1692008-01-25 08:20:51 +0200307#define NUM_INIT_LISTS (3 * MAX_NUMNODES)
Christoph Lametere498be72005-09-09 13:03:32 -0700308struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
309#define CACHE_CACHE 0
Pekka Enberg556a1692008-01-25 08:20:51 +0200310#define SIZE_AC MAX_NUMNODES
311#define SIZE_L3 (2 * MAX_NUMNODES)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700312
Christoph Lametered11d9e2006-06-30 01:55:45 -0700313static int drain_freelist(struct kmem_cache *cache,
314 struct kmem_list3 *l3, int tofree);
315static void free_block(struct kmem_cache *cachep, void **objpp, int len,
316 int node);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -0700317static int enable_cpucache(struct kmem_cache *cachep);
David Howells65f27f32006-11-22 14:55:48 +0000318static void cache_reap(struct work_struct *unused);
Christoph Lametered11d9e2006-06-30 01:55:45 -0700319
Christoph Lametere498be72005-09-09 13:03:32 -0700320/*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800321 * This function must be completely optimized away if a constant is passed to
322 * it. Mostly the same as what is in linux/slab.h except it returns an index.
Christoph Lametere498be72005-09-09 13:03:32 -0700323 */
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700324static __always_inline int index_of(const size_t size)
Christoph Lametere498be72005-09-09 13:03:32 -0700325{
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800326 extern void __bad_size(void);
327
Christoph Lametere498be72005-09-09 13:03:32 -0700328 if (__builtin_constant_p(size)) {
329 int i = 0;
330
331#define CACHE(x) \
332 if (size <=x) \
333 return i; \
334 else \
335 i++;
Joe Perches1c61fc42008-03-05 13:58:17 -0800336#include <linux/kmalloc_sizes.h>
Christoph Lametere498be72005-09-09 13:03:32 -0700337#undef CACHE
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800338 __bad_size();
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700339 } else
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800340 __bad_size();
Christoph Lametere498be72005-09-09 13:03:32 -0700341 return 0;
342}
343
Ingo Molnare0a42722006-06-23 02:03:46 -0700344static int slab_early_init = 1;
345
Christoph Lametere498be72005-09-09 13:03:32 -0700346#define INDEX_AC index_of(sizeof(struct arraycache_init))
347#define INDEX_L3 index_of(sizeof(struct kmem_list3))
348
Pekka Enberg5295a742006-02-01 03:05:48 -0800349static void kmem_list3_init(struct kmem_list3 *parent)
Christoph Lametere498be72005-09-09 13:03:32 -0700350{
351 INIT_LIST_HEAD(&parent->slabs_full);
352 INIT_LIST_HEAD(&parent->slabs_partial);
353 INIT_LIST_HEAD(&parent->slabs_free);
354 parent->shared = NULL;
355 parent->alien = NULL;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800356 parent->colour_next = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700357 spin_lock_init(&parent->list_lock);
358 parent->free_objects = 0;
359 parent->free_touched = 0;
360}
361
Andrew Mortona737b3e2006-03-22 00:08:11 -0800362#define MAKE_LIST(cachep, listp, slab, nodeid) \
363 do { \
364 INIT_LIST_HEAD(listp); \
365 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
Christoph Lametere498be72005-09-09 13:03:32 -0700366 } while (0)
367
Andrew Mortona737b3e2006-03-22 00:08:11 -0800368#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
369 do { \
Christoph Lametere498be72005-09-09 13:03:32 -0700370 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
371 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
372 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
373 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700374
375/*
Pekka Enberg343e0d72006-02-01 03:05:50 -0800376 * struct kmem_cache
Linus Torvalds1da177e2005-04-16 15:20:36 -0700377 *
378 * manages a cache.
379 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800380
Pekka J Enberg2109a2d2005-11-07 00:58:01 -0800381struct kmem_cache {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700382/* 1) per-cpu data, touched during every alloc/free */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800383 struct array_cache *array[NR_CPUS];
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800384/* 2) Cache tunables. Protected by cache_chain_mutex */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800385 unsigned int batchcount;
386 unsigned int limit;
387 unsigned int shared;
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800388
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800389 unsigned int buffer_size;
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800390 u32 reciprocal_buffer_size;
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800391/* 3) touched by every alloc & free from the backend */
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800392
Andrew Mortona737b3e2006-03-22 00:08:11 -0800393 unsigned int flags; /* constant flags */
394 unsigned int num; /* # of objs per slab */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700395
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800396/* 4) cache_grow/shrink */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700397 /* order of pgs per slab (2^n) */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800398 unsigned int gfporder;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700399
400 /* force GFP flags, e.g. GFP_DMA */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800401 gfp_t gfpflags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700402
Andrew Mortona737b3e2006-03-22 00:08:11 -0800403 size_t colour; /* cache colouring range */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800404 unsigned int colour_off; /* colour offset */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800405 struct kmem_cache *slabp_cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800406 unsigned int slab_size;
Andrew Mortona737b3e2006-03-22 00:08:11 -0800407 unsigned int dflags; /* dynamic flags */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700408
409 /* constructor func */
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -0700410 void (*ctor)(struct kmem_cache *, void *);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700411
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800412/* 5) cache creation/removal */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800413 const char *name;
414 struct list_head next;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700415
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800416/* 6) statistics */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700417#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800418 unsigned long num_active;
419 unsigned long num_allocations;
420 unsigned long high_mark;
421 unsigned long grown;
422 unsigned long reaped;
423 unsigned long errors;
424 unsigned long max_freeable;
425 unsigned long node_allocs;
426 unsigned long node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700427 unsigned long node_overflow;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800428 atomic_t allochit;
429 atomic_t allocmiss;
430 atomic_t freehit;
431 atomic_t freemiss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700432#endif
433#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800434 /*
435 * If debugging is enabled, then the allocator can add additional
436 * fields and/or padding to every object. buffer_size contains the total
437 * object size including these internal fields, the following two
438 * variables contain the offset to the user object and its size.
439 */
440 int obj_offset;
441 int obj_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700442#endif
Eric Dumazet8da34302007-05-06 14:49:29 -0700443 /*
444 * We put nodelists[] at the end of kmem_cache, because we want to size
445 * this array to nr_node_ids slots instead of MAX_NUMNODES
446 * (see kmem_cache_init())
447 * We still use [MAX_NUMNODES] and not [1] or [0] because cache_cache
448 * is statically defined, so we reserve the max number of nodes.
449 */
450 struct kmem_list3 *nodelists[MAX_NUMNODES];
451 /*
452 * Do not add fields after nodelists[]
453 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700454};
455
456#define CFLGS_OFF_SLAB (0x80000000UL)
457#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
458
459#define BATCHREFILL_LIMIT 16
Andrew Mortona737b3e2006-03-22 00:08:11 -0800460/*
461 * Optimization question: fewer reaps means less probability for unnessary
462 * cpucache drain/refill cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700463 *
Adrian Bunkdc6f3f22005-11-08 16:44:08 +0100464 * OTOH the cpuarrays can contain lots of objects,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700465 * which could lock up otherwise freeable slabs.
466 */
467#define REAPTIMEOUT_CPUC (2*HZ)
468#define REAPTIMEOUT_LIST3 (4*HZ)
469
470#if STATS
471#define STATS_INC_ACTIVE(x) ((x)->num_active++)
472#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
473#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
474#define STATS_INC_GROWN(x) ((x)->grown++)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700475#define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
Andrew Mortona737b3e2006-03-22 00:08:11 -0800476#define STATS_SET_HIGH(x) \
477 do { \
478 if ((x)->num_active > (x)->high_mark) \
479 (x)->high_mark = (x)->num_active; \
480 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700481#define STATS_INC_ERR(x) ((x)->errors++)
482#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
Christoph Lametere498be72005-09-09 13:03:32 -0700483#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700484#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800485#define STATS_SET_FREEABLE(x, i) \
486 do { \
487 if ((x)->max_freeable < i) \
488 (x)->max_freeable = i; \
489 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700490#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
491#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
492#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
493#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
494#else
495#define STATS_INC_ACTIVE(x) do { } while (0)
496#define STATS_DEC_ACTIVE(x) do { } while (0)
497#define STATS_INC_ALLOCED(x) do { } while (0)
498#define STATS_INC_GROWN(x) do { } while (0)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700499#define STATS_ADD_REAPED(x,y) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700500#define STATS_SET_HIGH(x) do { } while (0)
501#define STATS_INC_ERR(x) do { } while (0)
502#define STATS_INC_NODEALLOCS(x) do { } while (0)
Christoph Lametere498be72005-09-09 13:03:32 -0700503#define STATS_INC_NODEFREES(x) do { } while (0)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700504#define STATS_INC_ACOVERFLOW(x) do { } while (0)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800505#define STATS_SET_FREEABLE(x, i) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700506#define STATS_INC_ALLOCHIT(x) do { } while (0)
507#define STATS_INC_ALLOCMISS(x) do { } while (0)
508#define STATS_INC_FREEHIT(x) do { } while (0)
509#define STATS_INC_FREEMISS(x) do { } while (0)
510#endif
511
512#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -0700513
Andrew Mortona737b3e2006-03-22 00:08:11 -0800514/*
515 * memory layout of objects:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700516 * 0 : objp
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800517 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
Linus Torvalds1da177e2005-04-16 15:20:36 -0700518 * the end of an object is aligned with the end of the real
519 * allocation. Catches writes behind the end of the allocation.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800520 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700521 * redzone word.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800522 * cachep->obj_offset: The real object.
523 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
Andrew Mortona737b3e2006-03-22 00:08:11 -0800524 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
525 * [BYTES_PER_WORD long]
Linus Torvalds1da177e2005-04-16 15:20:36 -0700526 */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800527static int obj_offset(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700528{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800529 return cachep->obj_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700530}
531
Pekka Enberg343e0d72006-02-01 03:05:50 -0800532static int obj_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700533{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800534 return cachep->obj_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700535}
536
David Woodhouseb46b8f12007-05-08 00:22:59 -0700537static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700538{
539 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
David Woodhouseb46b8f12007-05-08 00:22:59 -0700540 return (unsigned long long*) (objp + obj_offset(cachep) -
541 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700542}
543
David Woodhouseb46b8f12007-05-08 00:22:59 -0700544static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700545{
546 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
547 if (cachep->flags & SLAB_STORE_USER)
David Woodhouseb46b8f12007-05-08 00:22:59 -0700548 return (unsigned long long *)(objp + cachep->buffer_size -
549 sizeof(unsigned long long) -
David Woodhouse87a927c2007-07-04 21:26:44 -0400550 REDZONE_ALIGN);
David Woodhouseb46b8f12007-05-08 00:22:59 -0700551 return (unsigned long long *) (objp + cachep->buffer_size -
552 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700553}
554
Pekka Enberg343e0d72006-02-01 03:05:50 -0800555static void **dbg_userword(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700556{
557 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800558 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700559}
560
561#else
562
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800563#define obj_offset(x) 0
564#define obj_size(cachep) (cachep->buffer_size)
David Woodhouseb46b8f12007-05-08 00:22:59 -0700565#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
566#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
Linus Torvalds1da177e2005-04-16 15:20:36 -0700567#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
568
569#endif
570
571/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700572 * Do not go above this order unless 0 objects fit into the slab.
573 */
574#define BREAK_GFP_ORDER_HI 1
575#define BREAK_GFP_ORDER_LO 0
576static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
577
Andrew Mortona737b3e2006-03-22 00:08:11 -0800578/*
579 * Functions for storing/retrieving the cachep and or slab from the page
580 * allocator. These are used to find the slab an obj belongs to. With kfree(),
581 * these are used to find the cache which an obj belongs to.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700582 */
Pekka Enberg065d41c2005-11-13 16:06:46 -0800583static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
584{
585 page->lru.next = (struct list_head *)cache;
586}
587
588static inline struct kmem_cache *page_get_cache(struct page *page)
589{
Christoph Lameterd85f3382007-05-06 14:49:39 -0700590 page = compound_head(page);
Pekka Enbergddc2e812006-06-23 02:03:40 -0700591 BUG_ON(!PageSlab(page));
Pekka Enberg065d41c2005-11-13 16:06:46 -0800592 return (struct kmem_cache *)page->lru.next;
593}
594
595static inline void page_set_slab(struct page *page, struct slab *slab)
596{
597 page->lru.prev = (struct list_head *)slab;
598}
599
600static inline struct slab *page_get_slab(struct page *page)
601{
Pekka Enbergddc2e812006-06-23 02:03:40 -0700602 BUG_ON(!PageSlab(page));
Pekka Enberg065d41c2005-11-13 16:06:46 -0800603 return (struct slab *)page->lru.prev;
604}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700605
Pekka Enberg6ed5eb22006-02-01 03:05:49 -0800606static inline struct kmem_cache *virt_to_cache(const void *obj)
607{
Christoph Lameterb49af682007-05-06 14:49:41 -0700608 struct page *page = virt_to_head_page(obj);
Pekka Enberg6ed5eb22006-02-01 03:05:49 -0800609 return page_get_cache(page);
610}
611
612static inline struct slab *virt_to_slab(const void *obj)
613{
Christoph Lameterb49af682007-05-06 14:49:41 -0700614 struct page *page = virt_to_head_page(obj);
Pekka Enberg6ed5eb22006-02-01 03:05:49 -0800615 return page_get_slab(page);
616}
617
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800618static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
619 unsigned int idx)
620{
621 return slab->s_mem + cache->buffer_size * idx;
622}
623
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800624/*
625 * We want to avoid an expensive divide : (offset / cache->buffer_size)
626 * Using the fact that buffer_size is a constant for a particular cache,
627 * we can replace (offset / cache->buffer_size) by
628 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
629 */
630static inline unsigned int obj_to_index(const struct kmem_cache *cache,
631 const struct slab *slab, void *obj)
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800632{
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800633 u32 offset = (obj - slab->s_mem);
634 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800635}
636
Andrew Mortona737b3e2006-03-22 00:08:11 -0800637/*
638 * These are the default caches for kmalloc. Custom caches can have other sizes.
639 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700640struct cache_sizes malloc_sizes[] = {
641#define CACHE(x) { .cs_size = (x) },
642#include <linux/kmalloc_sizes.h>
643 CACHE(ULONG_MAX)
644#undef CACHE
645};
646EXPORT_SYMBOL(malloc_sizes);
647
648/* Must match cache_sizes above. Out of line to keep cache footprint low. */
649struct cache_names {
650 char *name;
651 char *name_dma;
652};
653
654static struct cache_names __initdata cache_names[] = {
655#define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
656#include <linux/kmalloc_sizes.h>
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800657 {NULL,}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700658#undef CACHE
659};
660
661static struct arraycache_init initarray_cache __initdata =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800662 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700663static struct arraycache_init initarray_generic =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800664 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700665
666/* internal cache of cache description objs */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800667static struct kmem_cache cache_cache = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800668 .batchcount = 1,
669 .limit = BOOT_CPUCACHE_ENTRIES,
670 .shared = 1,
Pekka Enberg343e0d72006-02-01 03:05:50 -0800671 .buffer_size = sizeof(struct kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800672 .name = "kmem_cache",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700673};
674
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700675#define BAD_ALIEN_MAGIC 0x01020304ul
676
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200677#ifdef CONFIG_LOCKDEP
678
679/*
680 * Slab sometimes uses the kmalloc slabs to store the slab headers
681 * for other slabs "off slab".
682 * The locking for this is tricky in that it nests within the locks
683 * of all other slabs in a few places; to deal with this special
684 * locking we put on-slab caches into a separate lock-class.
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700685 *
686 * We set lock class for alien array caches which are up during init.
687 * The lock annotation will be lost if all cpus of a node goes down and
688 * then comes back up during hotplug
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200689 */
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700690static struct lock_class_key on_slab_l3_key;
691static struct lock_class_key on_slab_alc_key;
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200692
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700693static inline void init_lock_keys(void)
694
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200695{
696 int q;
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700697 struct cache_sizes *s = malloc_sizes;
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200698
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700699 while (s->cs_size != ULONG_MAX) {
700 for_each_node(q) {
701 struct array_cache **alc;
702 int r;
703 struct kmem_list3 *l3 = s->cs_cachep->nodelists[q];
704 if (!l3 || OFF_SLAB(s->cs_cachep))
705 continue;
706 lockdep_set_class(&l3->list_lock, &on_slab_l3_key);
707 alc = l3->alien;
708 /*
709 * FIXME: This check for BAD_ALIEN_MAGIC
710 * should go away when common slab code is taught to
711 * work even without alien caches.
712 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
713 * for alloc_alien_cache,
714 */
715 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
716 continue;
717 for_each_node(r) {
718 if (alc[r])
719 lockdep_set_class(&alc[r]->lock,
720 &on_slab_alc_key);
721 }
722 }
723 s++;
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200724 }
725}
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200726#else
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700727static inline void init_lock_keys(void)
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200728{
729}
730#endif
731
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -0800732/*
Gautham R Shenoy95402b32008-01-25 21:08:02 +0100733 * Guard access to the cache-chain.
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -0800734 */
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800735static DEFINE_MUTEX(cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700736static struct list_head cache_chain;
737
738/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700739 * chicken and egg problem: delay the per-cpu array allocation
740 * until the general caches are up.
741 */
742static enum {
743 NONE,
Christoph Lametere498be72005-09-09 13:03:32 -0700744 PARTIAL_AC,
745 PARTIAL_L3,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700746 FULL
747} g_cpucache_up;
748
Mike Kravetz39d24e62006-05-15 09:44:13 -0700749/*
750 * used by boot code to determine if it can use slab based allocator
751 */
752int slab_is_available(void)
753{
754 return g_cpucache_up == FULL;
755}
756
David Howells52bad642006-11-22 14:54:01 +0000757static DEFINE_PER_CPU(struct delayed_work, reap_work);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700758
Pekka Enberg343e0d72006-02-01 03:05:50 -0800759static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700760{
761 return cachep->array[smp_processor_id()];
762}
763
Andrew Mortona737b3e2006-03-22 00:08:11 -0800764static inline struct kmem_cache *__find_general_cachep(size_t size,
765 gfp_t gfpflags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700766{
767 struct cache_sizes *csizep = malloc_sizes;
768
769#if DEBUG
770 /* This happens if someone tries to call
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800771 * kmem_cache_create(), or __kmalloc(), before
772 * the generic caches are initialized.
773 */
Alok Katariac7e43c72005-09-14 12:17:53 -0700774 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700775#endif
Christoph Lameter6cb8f912007-07-17 04:03:22 -0700776 if (!size)
777 return ZERO_SIZE_PTR;
778
Linus Torvalds1da177e2005-04-16 15:20:36 -0700779 while (size > csizep->cs_size)
780 csizep++;
781
782 /*
Martin Hicks0abf40c2005-09-03 15:54:54 -0700783 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
Linus Torvalds1da177e2005-04-16 15:20:36 -0700784 * has cs_{dma,}cachep==NULL. Thus no special case
785 * for large kmalloc calls required.
786 */
Christoph Lameter4b51d662007-02-10 01:43:10 -0800787#ifdef CONFIG_ZONE_DMA
Linus Torvalds1da177e2005-04-16 15:20:36 -0700788 if (unlikely(gfpflags & GFP_DMA))
789 return csizep->cs_dmacachep;
Christoph Lameter4b51d662007-02-10 01:43:10 -0800790#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700791 return csizep->cs_cachep;
792}
793
Adrian Bunkb2213852006-09-25 23:31:02 -0700794static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700795{
796 return __find_general_cachep(size, gfpflags);
797}
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700798
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800799static size_t slab_mgmt_size(size_t nr_objs, size_t align)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700800{
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800801 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
802}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700803
Andrew Mortona737b3e2006-03-22 00:08:11 -0800804/*
805 * Calculate the number of objects and left-over bytes for a given buffer size.
806 */
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800807static void cache_estimate(unsigned long gfporder, size_t buffer_size,
808 size_t align, int flags, size_t *left_over,
809 unsigned int *num)
810{
811 int nr_objs;
812 size_t mgmt_size;
813 size_t slab_size = PAGE_SIZE << gfporder;
814
815 /*
816 * The slab management structure can be either off the slab or
817 * on it. For the latter case, the memory allocated for a
818 * slab is used for:
819 *
820 * - The struct slab
821 * - One kmem_bufctl_t for each object
822 * - Padding to respect alignment of @align
823 * - @buffer_size bytes for each object
824 *
825 * If the slab management structure is off the slab, then the
826 * alignment will already be calculated into the size. Because
827 * the slabs are all pages aligned, the objects will be at the
828 * correct alignment when allocated.
829 */
830 if (flags & CFLGS_OFF_SLAB) {
831 mgmt_size = 0;
832 nr_objs = slab_size / buffer_size;
833
834 if (nr_objs > SLAB_LIMIT)
835 nr_objs = SLAB_LIMIT;
836 } else {
837 /*
838 * Ignore padding for the initial guess. The padding
839 * is at most @align-1 bytes, and @buffer_size is at
840 * least @align. In the worst case, this result will
841 * be one greater than the number of objects that fit
842 * into the memory allocation when taking the padding
843 * into account.
844 */
845 nr_objs = (slab_size - sizeof(struct slab)) /
846 (buffer_size + sizeof(kmem_bufctl_t));
847
848 /*
849 * This calculated number will be either the right
850 * amount, or one greater than what we want.
851 */
852 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
853 > slab_size)
854 nr_objs--;
855
856 if (nr_objs > SLAB_LIMIT)
857 nr_objs = SLAB_LIMIT;
858
859 mgmt_size = slab_mgmt_size(nr_objs, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700860 }
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800861 *num = nr_objs;
862 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700863}
864
865#define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
866
Andrew Mortona737b3e2006-03-22 00:08:11 -0800867static void __slab_error(const char *function, struct kmem_cache *cachep,
868 char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700869{
870 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800871 function, cachep->name, msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700872 dump_stack();
873}
874
Paul Menage3395ee02006-12-06 20:32:16 -0800875/*
876 * By default on NUMA we use alien caches to stage the freeing of
877 * objects allocated from other nodes. This causes massive memory
878 * inefficiencies when using fake NUMA setup to split memory into a
879 * large number of small nodes, so it can be disabled on the command
880 * line
881 */
882
883static int use_alien_caches __read_mostly = 1;
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -0700884static int numa_platform __read_mostly = 1;
Paul Menage3395ee02006-12-06 20:32:16 -0800885static int __init noaliencache_setup(char *s)
886{
887 use_alien_caches = 0;
888 return 1;
889}
890__setup("noaliencache", noaliencache_setup);
891
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800892#ifdef CONFIG_NUMA
893/*
894 * Special reaping functions for NUMA systems called from cache_reap().
895 * These take care of doing round robin flushing of alien caches (containing
896 * objects freed on different nodes from which they were allocated) and the
897 * flushing of remote pcps by calling drain_node_pages.
898 */
899static DEFINE_PER_CPU(unsigned long, reap_node);
900
901static void init_reap_node(int cpu)
902{
903 int node;
904
905 node = next_node(cpu_to_node(cpu), node_online_map);
906 if (node == MAX_NUMNODES)
Paul Jackson442295c2006-03-22 00:09:11 -0800907 node = first_node(node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800908
Daniel Yeisley7f6b8872006-11-02 22:07:14 -0800909 per_cpu(reap_node, cpu) = node;
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800910}
911
912static void next_reap_node(void)
913{
914 int node = __get_cpu_var(reap_node);
915
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800916 node = next_node(node, node_online_map);
917 if (unlikely(node >= MAX_NUMNODES))
918 node = first_node(node_online_map);
919 __get_cpu_var(reap_node) = node;
920}
921
922#else
923#define init_reap_node(cpu) do { } while (0)
924#define next_reap_node(void) do { } while (0)
925#endif
926
Linus Torvalds1da177e2005-04-16 15:20:36 -0700927/*
928 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
929 * via the workqueue/eventd.
930 * Add the CPU number into the expiration time to minimize the possibility of
931 * the CPUs getting into lockstep and contending for the global cache chain
932 * lock.
933 */
Adrian Bunk897e6792007-07-15 23:38:20 -0700934static void __cpuinit start_cpu_timer(int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700935{
David Howells52bad642006-11-22 14:54:01 +0000936 struct delayed_work *reap_work = &per_cpu(reap_work, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700937
938 /*
939 * When this gets called from do_initcalls via cpucache_init(),
940 * init_workqueues() has already run, so keventd will be setup
941 * at that time.
942 */
David Howells52bad642006-11-22 14:54:01 +0000943 if (keventd_up() && reap_work->work.func == NULL) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800944 init_reap_node(cpu);
David Howells65f27f32006-11-22 14:55:48 +0000945 INIT_DELAYED_WORK(reap_work, cache_reap);
Arjan van de Ven2b284212006-12-10 02:21:28 -0800946 schedule_delayed_work_on(cpu, reap_work,
947 __round_jiffies_relative(HZ, cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700948 }
949}
950
Christoph Lametere498be72005-09-09 13:03:32 -0700951static struct array_cache *alloc_arraycache(int node, int entries,
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800952 int batchcount)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700953{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800954 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700955 struct array_cache *nc = NULL;
956
Christoph Lametere498be72005-09-09 13:03:32 -0700957 nc = kmalloc_node(memsize, GFP_KERNEL, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700958 if (nc) {
959 nc->avail = 0;
960 nc->limit = entries;
961 nc->batchcount = batchcount;
962 nc->touched = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700963 spin_lock_init(&nc->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700964 }
965 return nc;
966}
967
Christoph Lameter3ded1752006-03-25 03:06:44 -0800968/*
969 * Transfer objects in one arraycache to another.
970 * Locking must be handled by the caller.
971 *
972 * Return the number of entries transferred.
973 */
974static int transfer_objects(struct array_cache *to,
975 struct array_cache *from, unsigned int max)
976{
977 /* Figure out how many entries to transfer */
978 int nr = min(min(from->avail, max), to->limit - to->avail);
979
980 if (!nr)
981 return 0;
982
983 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
984 sizeof(void *) *nr);
985
986 from->avail -= nr;
987 to->avail += nr;
988 to->touched = 1;
989 return nr;
990}
991
Christoph Lameter765c4502006-09-27 01:50:08 -0700992#ifndef CONFIG_NUMA
993
994#define drain_alien_cache(cachep, alien) do { } while (0)
995#define reap_alien(cachep, l3) do { } while (0)
996
997static inline struct array_cache **alloc_alien_cache(int node, int limit)
998{
999 return (struct array_cache **)BAD_ALIEN_MAGIC;
1000}
1001
1002static inline void free_alien_cache(struct array_cache **ac_ptr)
1003{
1004}
1005
1006static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1007{
1008 return 0;
1009}
1010
1011static inline void *alternate_node_alloc(struct kmem_cache *cachep,
1012 gfp_t flags)
1013{
1014 return NULL;
1015}
1016
Christoph Hellwig8b98c162006-12-06 20:32:30 -08001017static inline void *____cache_alloc_node(struct kmem_cache *cachep,
Christoph Lameter765c4502006-09-27 01:50:08 -07001018 gfp_t flags, int nodeid)
1019{
1020 return NULL;
1021}
1022
1023#else /* CONFIG_NUMA */
1024
Christoph Hellwig8b98c162006-12-06 20:32:30 -08001025static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
Paul Jacksonc61afb12006-03-24 03:16:08 -08001026static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
Christoph Lameterdc85da12006-01-18 17:42:36 -08001027
Pekka Enberg5295a742006-02-01 03:05:48 -08001028static struct array_cache **alloc_alien_cache(int node, int limit)
Christoph Lametere498be72005-09-09 13:03:32 -07001029{
1030 struct array_cache **ac_ptr;
Christoph Lameter8ef82862007-02-20 13:57:52 -08001031 int memsize = sizeof(void *) * nr_node_ids;
Christoph Lametere498be72005-09-09 13:03:32 -07001032 int i;
1033
1034 if (limit > 1)
1035 limit = 12;
1036 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
1037 if (ac_ptr) {
1038 for_each_node(i) {
1039 if (i == node || !node_online(i)) {
1040 ac_ptr[i] = NULL;
1041 continue;
1042 }
1043 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
1044 if (!ac_ptr[i]) {
Akinobu Mitacc550de2007-11-14 16:58:35 -08001045 for (i--; i >= 0; i--)
Christoph Lametere498be72005-09-09 13:03:32 -07001046 kfree(ac_ptr[i]);
1047 kfree(ac_ptr);
1048 return NULL;
1049 }
1050 }
1051 }
1052 return ac_ptr;
1053}
1054
Pekka Enberg5295a742006-02-01 03:05:48 -08001055static void free_alien_cache(struct array_cache **ac_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -07001056{
1057 int i;
1058
1059 if (!ac_ptr)
1060 return;
Christoph Lametere498be72005-09-09 13:03:32 -07001061 for_each_node(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001062 kfree(ac_ptr[i]);
Christoph Lametere498be72005-09-09 13:03:32 -07001063 kfree(ac_ptr);
1064}
1065
Pekka Enberg343e0d72006-02-01 03:05:50 -08001066static void __drain_alien_cache(struct kmem_cache *cachep,
Pekka Enberg5295a742006-02-01 03:05:48 -08001067 struct array_cache *ac, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07001068{
1069 struct kmem_list3 *rl3 = cachep->nodelists[node];
1070
1071 if (ac->avail) {
1072 spin_lock(&rl3->list_lock);
Christoph Lametere00946f2006-03-25 03:06:45 -08001073 /*
1074 * Stuff objects into the remote nodes shared array first.
1075 * That way we could avoid the overhead of putting the objects
1076 * into the free lists and getting them back later.
1077 */
shin, jacob693f7d32006-04-28 10:54:37 -05001078 if (rl3->shared)
1079 transfer_objects(rl3->shared, ac, ac->limit);
Christoph Lametere00946f2006-03-25 03:06:45 -08001080
Christoph Lameterff694162005-09-22 21:44:02 -07001081 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001082 ac->avail = 0;
1083 spin_unlock(&rl3->list_lock);
1084 }
1085}
1086
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001087/*
1088 * Called from cache_reap() to regularly drain alien caches round robin.
1089 */
1090static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1091{
1092 int node = __get_cpu_var(reap_node);
1093
1094 if (l3->alien) {
1095 struct array_cache *ac = l3->alien[node];
Christoph Lametere00946f2006-03-25 03:06:45 -08001096
1097 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001098 __drain_alien_cache(cachep, ac, node);
1099 spin_unlock_irq(&ac->lock);
1100 }
1101 }
1102}
1103
Andrew Mortona737b3e2006-03-22 00:08:11 -08001104static void drain_alien_cache(struct kmem_cache *cachep,
1105 struct array_cache **alien)
Christoph Lametere498be72005-09-09 13:03:32 -07001106{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001107 int i = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07001108 struct array_cache *ac;
1109 unsigned long flags;
1110
1111 for_each_online_node(i) {
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001112 ac = alien[i];
Christoph Lametere498be72005-09-09 13:03:32 -07001113 if (ac) {
1114 spin_lock_irqsave(&ac->lock, flags);
1115 __drain_alien_cache(cachep, ac, i);
1116 spin_unlock_irqrestore(&ac->lock, flags);
1117 }
1118 }
1119}
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001120
Ingo Molnar873623d2006-07-13 14:44:38 +02001121static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001122{
1123 struct slab *slabp = virt_to_slab(objp);
1124 int nodeid = slabp->nodeid;
1125 struct kmem_list3 *l3;
1126 struct array_cache *alien = NULL;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001127 int node;
1128
1129 node = numa_node_id();
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001130
1131 /*
1132 * Make sure we are not freeing a object from another node to the array
1133 * cache on this cpu.
1134 */
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001135 if (likely(slabp->nodeid == node))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001136 return 0;
1137
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001138 l3 = cachep->nodelists[node];
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001139 STATS_INC_NODEFREES(cachep);
1140 if (l3->alien && l3->alien[nodeid]) {
1141 alien = l3->alien[nodeid];
Ingo Molnar873623d2006-07-13 14:44:38 +02001142 spin_lock(&alien->lock);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001143 if (unlikely(alien->avail == alien->limit)) {
1144 STATS_INC_ACOVERFLOW(cachep);
1145 __drain_alien_cache(cachep, alien, nodeid);
1146 }
1147 alien->entry[alien->avail++] = objp;
1148 spin_unlock(&alien->lock);
1149 } else {
1150 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1151 free_block(cachep, &objp, 1, nodeid);
1152 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1153 }
1154 return 1;
1155}
Christoph Lametere498be72005-09-09 13:03:32 -07001156#endif
1157
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001158static void __cpuinit cpuup_canceled(long cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001159{
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001160 struct kmem_cache *cachep;
1161 struct kmem_list3 *l3 = NULL;
1162 int node = cpu_to_node(cpu);
Mike Travisc5f59f02008-04-04 18:11:10 -07001163 node_to_cpumask_ptr(mask, node);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001164
1165 list_for_each_entry(cachep, &cache_chain, next) {
1166 struct array_cache *nc;
1167 struct array_cache *shared;
1168 struct array_cache **alien;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001169
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001170 /* cpu is dead; no one can alloc from it. */
1171 nc = cachep->array[cpu];
1172 cachep->array[cpu] = NULL;
1173 l3 = cachep->nodelists[node];
1174
1175 if (!l3)
1176 goto free_array_cache;
1177
1178 spin_lock_irq(&l3->list_lock);
1179
1180 /* Free limit for this kmem_list3 */
1181 l3->free_limit -= cachep->batchcount;
1182 if (nc)
1183 free_block(cachep, nc->entry, nc->avail, node);
1184
Mike Travisc5f59f02008-04-04 18:11:10 -07001185 if (!cpus_empty(*mask)) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001186 spin_unlock_irq(&l3->list_lock);
1187 goto free_array_cache;
1188 }
1189
1190 shared = l3->shared;
1191 if (shared) {
1192 free_block(cachep, shared->entry,
1193 shared->avail, node);
1194 l3->shared = NULL;
1195 }
1196
1197 alien = l3->alien;
1198 l3->alien = NULL;
1199
1200 spin_unlock_irq(&l3->list_lock);
1201
1202 kfree(shared);
1203 if (alien) {
1204 drain_alien_cache(cachep, alien);
1205 free_alien_cache(alien);
1206 }
1207free_array_cache:
1208 kfree(nc);
1209 }
1210 /*
1211 * In the previous loop, all the objects were freed to
1212 * the respective cache's slabs, now we can go ahead and
1213 * shrink each nodelist to its limit.
1214 */
1215 list_for_each_entry(cachep, &cache_chain, next) {
1216 l3 = cachep->nodelists[node];
1217 if (!l3)
1218 continue;
1219 drain_freelist(cachep, l3, l3->free_objects);
1220 }
1221}
1222
1223static int __cpuinit cpuup_prepare(long cpu)
1224{
Pekka Enberg343e0d72006-02-01 03:05:50 -08001225 struct kmem_cache *cachep;
Christoph Lametere498be72005-09-09 13:03:32 -07001226 struct kmem_list3 *l3 = NULL;
1227 int node = cpu_to_node(cpu);
David Howellsea02e3d2007-07-19 01:49:09 -07001228 const int memsize = sizeof(struct kmem_list3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001229
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001230 /*
1231 * We need to do this right in the beginning since
1232 * alloc_arraycache's are going to use this list.
1233 * kmalloc_node allows us to add the slab to the right
1234 * kmem_list3 and not this cpu's kmem_list3
1235 */
1236
1237 list_for_each_entry(cachep, &cache_chain, next) {
1238 /*
1239 * Set up the size64 kmemlist for cpu before we can
1240 * begin anything. Make sure some other cpu on this
1241 * node has not already allocated this
1242 */
1243 if (!cachep->nodelists[node]) {
1244 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1245 if (!l3)
1246 goto bad;
1247 kmem_list3_init(l3);
1248 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1249 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1250
1251 /*
1252 * The l3s don't come and go as CPUs come and
1253 * go. cache_chain_mutex is sufficient
1254 * protection here.
1255 */
1256 cachep->nodelists[node] = l3;
1257 }
1258
1259 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1260 cachep->nodelists[node]->free_limit =
1261 (1 + nr_cpus_node(node)) *
1262 cachep->batchcount + cachep->num;
1263 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1264 }
1265
1266 /*
1267 * Now we can go ahead with allocating the shared arrays and
1268 * array caches
1269 */
1270 list_for_each_entry(cachep, &cache_chain, next) {
1271 struct array_cache *nc;
1272 struct array_cache *shared = NULL;
1273 struct array_cache **alien = NULL;
1274
1275 nc = alloc_arraycache(node, cachep->limit,
1276 cachep->batchcount);
1277 if (!nc)
1278 goto bad;
1279 if (cachep->shared) {
1280 shared = alloc_arraycache(node,
1281 cachep->shared * cachep->batchcount,
1282 0xbaadf00d);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001283 if (!shared) {
1284 kfree(nc);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001285 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001286 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001287 }
1288 if (use_alien_caches) {
1289 alien = alloc_alien_cache(node, cachep->limit);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001290 if (!alien) {
1291 kfree(shared);
1292 kfree(nc);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001293 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001294 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001295 }
1296 cachep->array[cpu] = nc;
1297 l3 = cachep->nodelists[node];
1298 BUG_ON(!l3);
1299
1300 spin_lock_irq(&l3->list_lock);
1301 if (!l3->shared) {
1302 /*
1303 * We are serialised from CPU_DEAD or
1304 * CPU_UP_CANCELLED by the cpucontrol lock
1305 */
1306 l3->shared = shared;
1307 shared = NULL;
1308 }
1309#ifdef CONFIG_NUMA
1310 if (!l3->alien) {
1311 l3->alien = alien;
1312 alien = NULL;
1313 }
1314#endif
1315 spin_unlock_irq(&l3->list_lock);
1316 kfree(shared);
1317 free_alien_cache(alien);
1318 }
1319 return 0;
1320bad:
Akinobu Mita12d00f62007-10-18 03:05:11 -07001321 cpuup_canceled(cpu);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001322 return -ENOMEM;
1323}
1324
1325static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1326 unsigned long action, void *hcpu)
1327{
1328 long cpu = (long)hcpu;
1329 int err = 0;
1330
Linus Torvalds1da177e2005-04-16 15:20:36 -07001331 switch (action) {
Heiko Carstens38c3bd92007-05-09 02:34:05 -07001332 case CPU_UP_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001333 case CPU_UP_PREPARE_FROZEN:
Gautham R Shenoy95402b32008-01-25 21:08:02 +01001334 mutex_lock(&cache_chain_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001335 err = cpuup_prepare(cpu);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01001336 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001337 break;
1338 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001339 case CPU_ONLINE_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001340 start_cpu_timer(cpu);
1341 break;
1342#ifdef CONFIG_HOTPLUG_CPU
Christoph Lameter5830c592007-05-09 02:34:22 -07001343 case CPU_DOWN_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001344 case CPU_DOWN_PREPARE_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001345 /*
1346 * Shutdown cache reaper. Note that the cache_chain_mutex is
1347 * held so that if cache_reap() is invoked it cannot do
1348 * anything expensive but will only modify reap_work
1349 * and reschedule the timer.
1350 */
1351 cancel_rearming_delayed_work(&per_cpu(reap_work, cpu));
1352 /* Now the cache_reaper is guaranteed to be not running. */
1353 per_cpu(reap_work, cpu).work.func = NULL;
1354 break;
1355 case CPU_DOWN_FAILED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001356 case CPU_DOWN_FAILED_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001357 start_cpu_timer(cpu);
1358 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001359 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001360 case CPU_DEAD_FROZEN:
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001361 /*
1362 * Even if all the cpus of a node are down, we don't free the
1363 * kmem_list3 of any cache. This to avoid a race between
1364 * cpu_down, and a kmalloc allocation from another cpu for
1365 * memory from the node of the cpu going down. The list3
1366 * structure is usually allocated from kmem_cache_create() and
1367 * gets destroyed at kmem_cache_destroy().
1368 */
Simon Arlott183ff222007-10-20 01:27:18 +02001369 /* fall through */
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08001370#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001371 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001372 case CPU_UP_CANCELED_FROZEN:
Gautham R Shenoy95402b32008-01-25 21:08:02 +01001373 mutex_lock(&cache_chain_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001374 cpuup_canceled(cpu);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001375 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001376 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001377 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001378 return err ? NOTIFY_BAD : NOTIFY_OK;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001379}
1380
Chandra Seetharaman74b85f32006-06-27 02:54:09 -07001381static struct notifier_block __cpuinitdata cpucache_notifier = {
1382 &cpuup_callback, NULL, 0
1383};
Linus Torvalds1da177e2005-04-16 15:20:36 -07001384
Christoph Lametere498be72005-09-09 13:03:32 -07001385/*
1386 * swap the static kmem_list3 with kmalloced memory
1387 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001388static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1389 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001390{
1391 struct kmem_list3 *ptr;
1392
Christoph Lametere498be72005-09-09 13:03:32 -07001393 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1394 BUG_ON(!ptr);
1395
1396 local_irq_disable();
1397 memcpy(ptr, list, sizeof(struct kmem_list3));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001398 /*
1399 * Do not assume that spinlocks can be initialized via memcpy:
1400 */
1401 spin_lock_init(&ptr->list_lock);
1402
Christoph Lametere498be72005-09-09 13:03:32 -07001403 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1404 cachep->nodelists[nodeid] = ptr;
1405 local_irq_enable();
1406}
1407
Andrew Mortona737b3e2006-03-22 00:08:11 -08001408/*
Pekka Enberg556a1692008-01-25 08:20:51 +02001409 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1410 * size of kmem_list3.
1411 */
1412static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1413{
1414 int node;
1415
1416 for_each_online_node(node) {
1417 cachep->nodelists[node] = &initkmem_list3[index + node];
1418 cachep->nodelists[node]->next_reap = jiffies +
1419 REAPTIMEOUT_LIST3 +
1420 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1421 }
1422}
1423
1424/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08001425 * Initialisation. Called after the page allocator have been initialised and
1426 * before smp_init().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001427 */
1428void __init kmem_cache_init(void)
1429{
1430 size_t left_over;
1431 struct cache_sizes *sizes;
1432 struct cache_names *names;
Christoph Lametere498be72005-09-09 13:03:32 -07001433 int i;
Jack Steiner07ed76b2006-03-07 21:55:46 -08001434 int order;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001435 int node;
Christoph Lametere498be72005-09-09 13:03:32 -07001436
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -07001437 if (num_possible_nodes() == 1) {
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001438 use_alien_caches = 0;
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -07001439 numa_platform = 0;
1440 }
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001441
Christoph Lametere498be72005-09-09 13:03:32 -07001442 for (i = 0; i < NUM_INIT_LISTS; i++) {
1443 kmem_list3_init(&initkmem_list3[i]);
1444 if (i < MAX_NUMNODES)
1445 cache_cache.nodelists[i] = NULL;
1446 }
Pekka Enberg556a1692008-01-25 08:20:51 +02001447 set_up_list3s(&cache_cache, CACHE_CACHE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001448
1449 /*
1450 * Fragmentation resistance on low memory - only use bigger
1451 * page orders on machines with more than 32MB of memory.
1452 */
1453 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1454 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1455
Linus Torvalds1da177e2005-04-16 15:20:36 -07001456 /* Bootstrap is tricky, because several objects are allocated
1457 * from caches that do not exist yet:
Andrew Mortona737b3e2006-03-22 00:08:11 -08001458 * 1) initialize the cache_cache cache: it contains the struct
1459 * kmem_cache structures of all caches, except cache_cache itself:
1460 * cache_cache is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001461 * Initially an __init data area is used for the head array and the
1462 * kmem_list3 structures, it's replaced with a kmalloc allocated
1463 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001464 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001465 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001466 * An __init data area is used for the head array.
1467 * 3) Create the remaining kmalloc caches, with minimally sized
1468 * head arrays.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001469 * 4) Replace the __init data head arrays for cache_cache and the first
1470 * kmalloc cache with kmalloc allocated arrays.
Christoph Lametere498be72005-09-09 13:03:32 -07001471 * 5) Replace the __init data for kmem_list3 for cache_cache and
1472 * the other cache's with kmalloc allocated memory.
1473 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001474 */
1475
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001476 node = numa_node_id();
1477
Linus Torvalds1da177e2005-04-16 15:20:36 -07001478 /* 1) create the cache_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001479 INIT_LIST_HEAD(&cache_chain);
1480 list_add(&cache_cache.next, &cache_chain);
1481 cache_cache.colour_off = cache_line_size();
1482 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
Daniel Yeisleyec1f5ee2008-03-25 23:59:08 +02001483 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001484
Eric Dumazet8da34302007-05-06 14:49:29 -07001485 /*
1486 * struct kmem_cache size depends on nr_node_ids, which
1487 * can be less than MAX_NUMNODES.
1488 */
1489 cache_cache.buffer_size = offsetof(struct kmem_cache, nodelists) +
1490 nr_node_ids * sizeof(struct kmem_list3 *);
1491#if DEBUG
1492 cache_cache.obj_size = cache_cache.buffer_size;
1493#endif
Andrew Mortona737b3e2006-03-22 00:08:11 -08001494 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1495 cache_line_size());
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08001496 cache_cache.reciprocal_buffer_size =
1497 reciprocal_value(cache_cache.buffer_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001498
Jack Steiner07ed76b2006-03-07 21:55:46 -08001499 for (order = 0; order < MAX_ORDER; order++) {
1500 cache_estimate(order, cache_cache.buffer_size,
1501 cache_line_size(), 0, &left_over, &cache_cache.num);
1502 if (cache_cache.num)
1503 break;
1504 }
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02001505 BUG_ON(!cache_cache.num);
Jack Steiner07ed76b2006-03-07 21:55:46 -08001506 cache_cache.gfporder = order;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001507 cache_cache.colour = left_over / cache_cache.colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001508 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1509 sizeof(struct slab), cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001510
1511 /* 2+3) create the kmalloc caches */
1512 sizes = malloc_sizes;
1513 names = cache_names;
1514
Andrew Mortona737b3e2006-03-22 00:08:11 -08001515 /*
1516 * Initialize the caches that provide memory for the array cache and the
1517 * kmem_list3 structures first. Without this, further allocations will
1518 * bug.
Christoph Lametere498be72005-09-09 13:03:32 -07001519 */
1520
1521 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001522 sizes[INDEX_AC].cs_size,
1523 ARCH_KMALLOC_MINALIGN,
1524 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001525 NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07001526
Andrew Mortona737b3e2006-03-22 00:08:11 -08001527 if (INDEX_AC != INDEX_L3) {
Christoph Lametere498be72005-09-09 13:03:32 -07001528 sizes[INDEX_L3].cs_cachep =
Andrew Mortona737b3e2006-03-22 00:08:11 -08001529 kmem_cache_create(names[INDEX_L3].name,
1530 sizes[INDEX_L3].cs_size,
1531 ARCH_KMALLOC_MINALIGN,
1532 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001533 NULL);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001534 }
Christoph Lametere498be72005-09-09 13:03:32 -07001535
Ingo Molnare0a42722006-06-23 02:03:46 -07001536 slab_early_init = 0;
1537
Linus Torvalds1da177e2005-04-16 15:20:36 -07001538 while (sizes->cs_size != ULONG_MAX) {
Christoph Lametere498be72005-09-09 13:03:32 -07001539 /*
1540 * For performance, all the general caches are L1 aligned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001541 * This should be particularly beneficial on SMP boxes, as it
1542 * eliminates "false sharing".
1543 * Note for systems short on memory removing the alignment will
Christoph Lametere498be72005-09-09 13:03:32 -07001544 * allow tighter packing of the smaller caches.
1545 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001546 if (!sizes->cs_cachep) {
Christoph Lametere498be72005-09-09 13:03:32 -07001547 sizes->cs_cachep = kmem_cache_create(names->name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001548 sizes->cs_size,
1549 ARCH_KMALLOC_MINALIGN,
1550 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001551 NULL);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001552 }
Christoph Lameter4b51d662007-02-10 01:43:10 -08001553#ifdef CONFIG_ZONE_DMA
1554 sizes->cs_dmacachep = kmem_cache_create(
1555 names->name_dma,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001556 sizes->cs_size,
1557 ARCH_KMALLOC_MINALIGN,
1558 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1559 SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001560 NULL);
Christoph Lameter4b51d662007-02-10 01:43:10 -08001561#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001562 sizes++;
1563 names++;
1564 }
1565 /* 4) Replace the bootstrap head arrays */
1566 {
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001567 struct array_cache *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001568
Linus Torvalds1da177e2005-04-16 15:20:36 -07001569 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001570
Linus Torvalds1da177e2005-04-16 15:20:36 -07001571 local_irq_disable();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001572 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1573 memcpy(ptr, cpu_cache_get(&cache_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001574 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001575 /*
1576 * Do not assume that spinlocks can be initialized via memcpy:
1577 */
1578 spin_lock_init(&ptr->lock);
1579
Linus Torvalds1da177e2005-04-16 15:20:36 -07001580 cache_cache.array[smp_processor_id()] = ptr;
1581 local_irq_enable();
Christoph Lametere498be72005-09-09 13:03:32 -07001582
Linus Torvalds1da177e2005-04-16 15:20:36 -07001583 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001584
Linus Torvalds1da177e2005-04-16 15:20:36 -07001585 local_irq_disable();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001586 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001587 != &initarray_generic.cache);
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001588 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001589 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001590 /*
1591 * Do not assume that spinlocks can be initialized via memcpy:
1592 */
1593 spin_lock_init(&ptr->lock);
1594
Christoph Lametere498be72005-09-09 13:03:32 -07001595 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001596 ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001597 local_irq_enable();
1598 }
Christoph Lametere498be72005-09-09 13:03:32 -07001599 /* 5) Replace the bootstrap kmem_list3's */
1600 {
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001601 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001602
Mel Gorman9c09a952008-01-24 05:49:54 -08001603 for_each_online_node(nid) {
Daniel Yeisleyec1f5ee2008-03-25 23:59:08 +02001604 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid);
Pekka Enberg556a1692008-01-25 08:20:51 +02001605
Christoph Lametere498be72005-09-09 13:03:32 -07001606 init_list(malloc_sizes[INDEX_AC].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001607 &initkmem_list3[SIZE_AC + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001608
1609 if (INDEX_AC != INDEX_L3) {
1610 init_list(malloc_sizes[INDEX_L3].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001611 &initkmem_list3[SIZE_L3 + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001612 }
1613 }
1614 }
1615
1616 /* 6) resize the head arrays to their final sizes */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001617 {
Pekka Enberg343e0d72006-02-01 03:05:50 -08001618 struct kmem_cache *cachep;
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001619 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001620 list_for_each_entry(cachep, &cache_chain, next)
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07001621 if (enable_cpucache(cachep))
1622 BUG();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001623 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001624 }
1625
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001626 /* Annotate slab for lockdep -- annotate the malloc caches */
1627 init_lock_keys();
1628
1629
Linus Torvalds1da177e2005-04-16 15:20:36 -07001630 /* Done! */
1631 g_cpucache_up = FULL;
1632
Andrew Mortona737b3e2006-03-22 00:08:11 -08001633 /*
1634 * Register a cpu startup notifier callback that initializes
1635 * cpu_cache_get for all new cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07001636 */
1637 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001638
Andrew Mortona737b3e2006-03-22 00:08:11 -08001639 /*
1640 * The reap timers are started later, with a module init call: That part
1641 * of the kernel is not yet operational.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001642 */
1643}
1644
1645static int __init cpucache_init(void)
1646{
1647 int cpu;
1648
Andrew Mortona737b3e2006-03-22 00:08:11 -08001649 /*
1650 * Register the timers that return unneeded pages to the page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -07001651 */
Christoph Lametere498be72005-09-09 13:03:32 -07001652 for_each_online_cpu(cpu)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001653 start_cpu_timer(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001654 return 0;
1655}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001656__initcall(cpucache_init);
1657
1658/*
1659 * Interface to system's page allocator. No need to hold the cache-lock.
1660 *
1661 * If we requested dmaable memory, we will get it. Even if we
1662 * did not request dmaable memory, we might get it, but that
1663 * would be relatively rare and ignorable.
1664 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001665static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001666{
1667 struct page *page;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001668 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001669 int i;
1670
Luke Yangd6fef9d2006-04-10 22:52:56 -07001671#ifndef CONFIG_MMU
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001672 /*
1673 * Nommu uses slab's for process anonymous memory allocations, and thus
1674 * requires __GFP_COMP to properly refcount higher order allocations
Luke Yangd6fef9d2006-04-10 22:52:56 -07001675 */
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001676 flags |= __GFP_COMP;
Luke Yangd6fef9d2006-04-10 22:52:56 -07001677#endif
Christoph Lameter765c4502006-09-27 01:50:08 -07001678
Christoph Lameter3c517a62006-12-06 20:33:29 -08001679 flags |= cachep->gfpflags;
Mel Gormane12ba742007-10-16 01:25:52 -07001680 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1681 flags |= __GFP_RECLAIMABLE;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001682
1683 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001684 if (!page)
1685 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001686
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001687 nr_pages = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001688 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
Christoph Lameter972d1a72006-09-25 23:31:51 -07001689 add_zone_page_state(page_zone(page),
1690 NR_SLAB_RECLAIMABLE, nr_pages);
1691 else
1692 add_zone_page_state(page_zone(page),
1693 NR_SLAB_UNRECLAIMABLE, nr_pages);
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001694 for (i = 0; i < nr_pages; i++)
1695 __SetPageSlab(page + i);
1696 return page_address(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001697}
1698
1699/*
1700 * Interface to system's page release.
1701 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001702static void kmem_freepages(struct kmem_cache *cachep, void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001703{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001704 unsigned long i = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001705 struct page *page = virt_to_page(addr);
1706 const unsigned long nr_freed = i;
1707
Christoph Lameter972d1a72006-09-25 23:31:51 -07001708 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1709 sub_zone_page_state(page_zone(page),
1710 NR_SLAB_RECLAIMABLE, nr_freed);
1711 else
1712 sub_zone_page_state(page_zone(page),
1713 NR_SLAB_UNRECLAIMABLE, nr_freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001714 while (i--) {
Nick Pigginf205b2f2006-03-22 00:08:02 -08001715 BUG_ON(!PageSlab(page));
1716 __ClearPageSlab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001717 page++;
1718 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001719 if (current->reclaim_state)
1720 current->reclaim_state->reclaimed_slab += nr_freed;
1721 free_pages((unsigned long)addr, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001722}
1723
1724static void kmem_rcu_free(struct rcu_head *head)
1725{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001726 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001727 struct kmem_cache *cachep = slab_rcu->cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001728
1729 kmem_freepages(cachep, slab_rcu->addr);
1730 if (OFF_SLAB(cachep))
1731 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1732}
1733
1734#if DEBUG
1735
1736#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001737static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001738 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001739{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001740 int size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001741
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001742 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001743
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001744 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001745 return;
1746
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001747 *addr++ = 0x12345678;
1748 *addr++ = caller;
1749 *addr++ = smp_processor_id();
1750 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001751 {
1752 unsigned long *sptr = &caller;
1753 unsigned long svalue;
1754
1755 while (!kstack_end(sptr)) {
1756 svalue = *sptr++;
1757 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001758 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001759 size -= sizeof(unsigned long);
1760 if (size <= sizeof(unsigned long))
1761 break;
1762 }
1763 }
1764
1765 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001766 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001767}
1768#endif
1769
Pekka Enberg343e0d72006-02-01 03:05:50 -08001770static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001771{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001772 int size = obj_size(cachep);
1773 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001774
1775 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001776 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001777}
1778
1779static void dump_line(char *data, int offset, int limit)
1780{
1781 int i;
Dave Jonesaa83aa42006-09-29 01:59:51 -07001782 unsigned char error = 0;
1783 int bad_count = 0;
1784
Linus Torvalds1da177e2005-04-16 15:20:36 -07001785 printk(KERN_ERR "%03x:", offset);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001786 for (i = 0; i < limit; i++) {
1787 if (data[offset + i] != POISON_FREE) {
1788 error = data[offset + i];
1789 bad_count++;
1790 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001791 printk(" %02x", (unsigned char)data[offset + i]);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001792 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001793 printk("\n");
Dave Jonesaa83aa42006-09-29 01:59:51 -07001794
1795 if (bad_count == 1) {
1796 error ^= POISON_FREE;
1797 if (!(error & (error - 1))) {
1798 printk(KERN_ERR "Single bit error detected. Probably "
1799 "bad RAM.\n");
1800#ifdef CONFIG_X86
1801 printk(KERN_ERR "Run memtest86+ or a similar memory "
1802 "test tool.\n");
1803#else
1804 printk(KERN_ERR "Run a memory test tool.\n");
1805#endif
1806 }
1807 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001808}
1809#endif
1810
1811#if DEBUG
1812
Pekka Enberg343e0d72006-02-01 03:05:50 -08001813static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001814{
1815 int i, size;
1816 char *realobj;
1817
1818 if (cachep->flags & SLAB_RED_ZONE) {
David Woodhouseb46b8f12007-05-08 00:22:59 -07001819 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001820 *dbg_redzone1(cachep, objp),
1821 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001822 }
1823
1824 if (cachep->flags & SLAB_STORE_USER) {
1825 printk(KERN_ERR "Last user: [<%p>]",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001826 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001827 print_symbol("(%s)",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001828 (unsigned long)*dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001829 printk("\n");
1830 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001831 realobj = (char *)objp + obj_offset(cachep);
1832 size = obj_size(cachep);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001833 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001834 int limit;
1835 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001836 if (i + limit > size)
1837 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001838 dump_line(realobj, i, limit);
1839 }
1840}
1841
Pekka Enberg343e0d72006-02-01 03:05:50 -08001842static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001843{
1844 char *realobj;
1845 int size, i;
1846 int lines = 0;
1847
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001848 realobj = (char *)objp + obj_offset(cachep);
1849 size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001850
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001851 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001852 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001853 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001854 exp = POISON_END;
1855 if (realobj[i] != exp) {
1856 int limit;
1857 /* Mismatch ! */
1858 /* Print header */
1859 if (lines == 0) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001860 printk(KERN_ERR
David Howellse94a40c2007-04-02 23:46:28 +01001861 "Slab corruption: %s start=%p, len=%d\n",
1862 cachep->name, realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001863 print_objinfo(cachep, objp, 0);
1864 }
1865 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001866 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001867 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001868 if (i + limit > size)
1869 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001870 dump_line(realobj, i, limit);
1871 i += 16;
1872 lines++;
1873 /* Limit to 5 lines */
1874 if (lines > 5)
1875 break;
1876 }
1877 }
1878 if (lines != 0) {
1879 /* Print some data about the neighboring objects, if they
1880 * exist:
1881 */
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08001882 struct slab *slabp = virt_to_slab(objp);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001883 unsigned int objnr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001884
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001885 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001886 if (objnr) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001887 objp = index_to_obj(cachep, slabp, objnr - 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001888 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001889 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001890 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001891 print_objinfo(cachep, objp, 2);
1892 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001893 if (objnr + 1 < cachep->num) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001894 objp = index_to_obj(cachep, slabp, objnr + 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001895 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001896 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001897 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001898 print_objinfo(cachep, objp, 2);
1899 }
1900 }
1901}
1902#endif
1903
Linus Torvalds1da177e2005-04-16 15:20:36 -07001904#if DEBUG
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001905/**
Randy Dunlap911851e2006-03-22 00:08:14 -08001906 * slab_destroy_objs - destroy a slab and its objects
1907 * @cachep: cache pointer being destroyed
1908 * @slabp: slab pointer being destroyed
1909 *
1910 * Call the registered destructor for each object in a slab that is being
1911 * destroyed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001912 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001913static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001914{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001915 int i;
1916 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001917 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001918
1919 if (cachep->flags & SLAB_POISON) {
1920#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08001921 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1922 OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001923 kernel_map_pages(virt_to_page(objp),
Andrew Mortona737b3e2006-03-22 00:08:11 -08001924 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001925 else
1926 check_poison_obj(cachep, objp);
1927#else
1928 check_poison_obj(cachep, objp);
1929#endif
1930 }
1931 if (cachep->flags & SLAB_RED_ZONE) {
1932 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1933 slab_error(cachep, "start of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001934 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001935 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1936 slab_error(cachep, "end of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001937 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001938 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001939 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001940}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001941#else
Pekka Enberg343e0d72006-02-01 03:05:50 -08001942static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001943{
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001944}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001945#endif
1946
Randy Dunlap911851e2006-03-22 00:08:14 -08001947/**
1948 * slab_destroy - destroy and release all objects in a slab
1949 * @cachep: cache pointer being destroyed
1950 * @slabp: slab pointer being destroyed
1951 *
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001952 * Destroy all the objs in a slab, and release the mem back to the system.
Andrew Mortona737b3e2006-03-22 00:08:11 -08001953 * Before calling the slab must have been unlinked from the cache. The
1954 * cache-lock is not held/needed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001955 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001956static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001957{
1958 void *addr = slabp->s_mem - slabp->colouroff;
1959
1960 slab_destroy_objs(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001961 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1962 struct slab_rcu *slab_rcu;
1963
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001964 slab_rcu = (struct slab_rcu *)slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001965 slab_rcu->cachep = cachep;
1966 slab_rcu->addr = addr;
1967 call_rcu(&slab_rcu->head, kmem_rcu_free);
1968 } else {
1969 kmem_freepages(cachep, addr);
Ingo Molnar873623d2006-07-13 14:44:38 +02001970 if (OFF_SLAB(cachep))
1971 kmem_cache_free(cachep->slabp_cache, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001972 }
1973}
1974
Christoph Lameter117f6eb2006-09-25 23:31:37 -07001975static void __kmem_cache_destroy(struct kmem_cache *cachep)
1976{
1977 int i;
1978 struct kmem_list3 *l3;
1979
1980 for_each_online_cpu(i)
1981 kfree(cachep->array[i]);
1982
1983 /* NUMA: free the list3 structures */
1984 for_each_online_node(i) {
1985 l3 = cachep->nodelists[i];
1986 if (l3) {
1987 kfree(l3->shared);
1988 free_alien_cache(l3->alien);
1989 kfree(l3);
1990 }
1991 }
1992 kmem_cache_free(&cache_cache, cachep);
1993}
1994
1995
Linus Torvalds1da177e2005-04-16 15:20:36 -07001996/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08001997 * calculate_slab_order - calculate size (page order) of slabs
1998 * @cachep: pointer to the cache that is being created
1999 * @size: size of objects to be created in this cache.
2000 * @align: required alignment for the objects.
2001 * @flags: slab allocation flags
2002 *
2003 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002004 *
2005 * This could be made much more intelligent. For now, try to avoid using
2006 * high order pages for slabs. When the gfp() functions are more friendly
2007 * towards high-order requests, this should be changed.
2008 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002009static size_t calculate_slab_order(struct kmem_cache *cachep,
Randy Dunlapee13d782006-02-01 03:05:53 -08002010 size_t size, size_t align, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002011{
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002012 unsigned long offslab_limit;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002013 size_t left_over = 0;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002014 int gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002015
Christoph Lameter0aa817f2007-05-16 22:11:01 -07002016 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002017 unsigned int num;
2018 size_t remainder;
2019
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002020 cache_estimate(gfporder, size, align, flags, &remainder, &num);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002021 if (!num)
2022 continue;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002023
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002024 if (flags & CFLGS_OFF_SLAB) {
2025 /*
2026 * Max number of objs-per-slab for caches which
2027 * use off-slab slabs. Needed to avoid a possible
2028 * looping condition in cache_grow().
2029 */
2030 offslab_limit = size - sizeof(struct slab);
2031 offslab_limit /= sizeof(kmem_bufctl_t);
2032
2033 if (num > offslab_limit)
2034 break;
2035 }
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002036
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002037 /* Found something acceptable - save it away */
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002038 cachep->num = num;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002039 cachep->gfporder = gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002040 left_over = remainder;
2041
2042 /*
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002043 * A VFS-reclaimable slab tends to have most allocations
2044 * as GFP_NOFS and we really don't want to have to be allocating
2045 * higher-order pages when we are unable to shrink dcache.
2046 */
2047 if (flags & SLAB_RECLAIM_ACCOUNT)
2048 break;
2049
2050 /*
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002051 * Large number of objects is good, but very large slabs are
2052 * currently bad for the gfp()s.
2053 */
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002054 if (gfporder >= slab_break_gfp_order)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002055 break;
2056
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002057 /*
2058 * Acceptable internal fragmentation?
2059 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002060 if (left_over * 8 <= (PAGE_SIZE << gfporder))
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002061 break;
2062 }
2063 return left_over;
2064}
2065
Sam Ravnborg38bdc322007-05-17 23:48:19 +02002066static int __init_refok setup_cpu_cache(struct kmem_cache *cachep)
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002067{
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002068 if (g_cpucache_up == FULL)
2069 return enable_cpucache(cachep);
2070
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002071 if (g_cpucache_up == NONE) {
2072 /*
2073 * Note: the first kmem_cache_create must create the cache
2074 * that's used by kmalloc(24), otherwise the creation of
2075 * further caches will BUG().
2076 */
2077 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2078
2079 /*
2080 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2081 * the first cache, then we need to set up all its list3s,
2082 * otherwise the creation of further caches will BUG().
2083 */
2084 set_up_list3s(cachep, SIZE_AC);
2085 if (INDEX_AC == INDEX_L3)
2086 g_cpucache_up = PARTIAL_L3;
2087 else
2088 g_cpucache_up = PARTIAL_AC;
2089 } else {
2090 cachep->array[smp_processor_id()] =
2091 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
2092
2093 if (g_cpucache_up == PARTIAL_AC) {
2094 set_up_list3s(cachep, SIZE_L3);
2095 g_cpucache_up = PARTIAL_L3;
2096 } else {
2097 int node;
Pekka Enberg556a1692008-01-25 08:20:51 +02002098 for_each_online_node(node) {
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002099 cachep->nodelists[node] =
2100 kmalloc_node(sizeof(struct kmem_list3),
2101 GFP_KERNEL, node);
2102 BUG_ON(!cachep->nodelists[node]);
2103 kmem_list3_init(cachep->nodelists[node]);
2104 }
2105 }
2106 }
2107 cachep->nodelists[numa_node_id()]->next_reap =
2108 jiffies + REAPTIMEOUT_LIST3 +
2109 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2110
2111 cpu_cache_get(cachep)->avail = 0;
2112 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2113 cpu_cache_get(cachep)->batchcount = 1;
2114 cpu_cache_get(cachep)->touched = 0;
2115 cachep->batchcount = 1;
2116 cachep->limit = BOOT_CPUCACHE_ENTRIES;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002117 return 0;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002118}
2119
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002120/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07002121 * kmem_cache_create - Create a cache.
2122 * @name: A string which is used in /proc/slabinfo to identify this cache.
2123 * @size: The size of objects to be created in this cache.
2124 * @align: The required alignment for the objects.
2125 * @flags: SLAB flags
2126 * @ctor: A constructor for the objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002127 *
2128 * Returns a ptr to the cache on success, NULL on failure.
2129 * Cannot be called within a int, but can be interrupted.
Paul Mundt20c2df82007-07-20 10:11:58 +09002130 * The @ctor is run when new pages are allocated by the cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002131 *
2132 * @name must be valid until the cache is destroyed. This implies that
Andrew Mortona737b3e2006-03-22 00:08:11 -08002133 * the module calling this has to destroy the cache before getting unloaded.
2134 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002135 * The flags are
2136 *
2137 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2138 * to catch references to uninitialised memory.
2139 *
2140 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2141 * for buffer overruns.
2142 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002143 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2144 * cacheline. This can be beneficial if you're counting cycles as closely
2145 * as davem.
2146 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002147struct kmem_cache *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002148kmem_cache_create (const char *name, size_t size, size_t align,
Andrew Mortona737b3e2006-03-22 00:08:11 -08002149 unsigned long flags,
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07002150 void (*ctor)(struct kmem_cache *, void *))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002151{
2152 size_t left_over, slab_size, ralign;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07002153 struct kmem_cache *cachep = NULL, *pc;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002154
2155 /*
2156 * Sanity checks... these are all serious usage bugs.
2157 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002158 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
Paul Mundt20c2df82007-07-20 10:11:58 +09002159 size > KMALLOC_MAX_SIZE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002160 printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__,
2161 name);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002162 BUG();
2163 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002164
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002165 /*
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002166 * We use cache_chain_mutex to ensure a consistent view of
2167 * cpu_online_map as well. Please see cpuup_callback
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002168 */
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002169 get_online_cpus();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002170 mutex_lock(&cache_chain_mutex);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002171
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07002172 list_for_each_entry(pc, &cache_chain, next) {
Andrew Morton4f12bb42005-11-07 00:58:00 -08002173 char tmp;
2174 int res;
2175
2176 /*
2177 * This happens when the module gets unloaded and doesn't
2178 * destroy its slab cache and no-one else reuses the vmalloc
2179 * area of the module. Print a warning.
2180 */
Andrew Morton138ae662006-12-06 20:36:41 -08002181 res = probe_kernel_address(pc->name, tmp);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002182 if (res) {
matzeb4169522007-05-06 14:49:52 -07002183 printk(KERN_ERR
2184 "SLAB: cache with size %d has lost its name\n",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002185 pc->buffer_size);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002186 continue;
2187 }
2188
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002189 if (!strcmp(pc->name, name)) {
matzeb4169522007-05-06 14:49:52 -07002190 printk(KERN_ERR
2191 "kmem_cache_create: duplicate cache %s\n", name);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002192 dump_stack();
2193 goto oops;
2194 }
2195 }
2196
Linus Torvalds1da177e2005-04-16 15:20:36 -07002197#if DEBUG
2198 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002199#if FORCED_DEBUG
2200 /*
2201 * Enable redzoning and last user accounting, except for caches with
2202 * large objects, if the increased size would increase the object size
2203 * above the next power of two: caches with object sizes just above a
2204 * power of two have a significant amount of internal fragmentation.
2205 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002206 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2207 2 * sizeof(unsigned long long)))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002208 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002209 if (!(flags & SLAB_DESTROY_BY_RCU))
2210 flags |= SLAB_POISON;
2211#endif
2212 if (flags & SLAB_DESTROY_BY_RCU)
2213 BUG_ON(flags & SLAB_POISON);
2214#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002215 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002216 * Always checks flags, a caller might be expecting debug support which
2217 * isn't available.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002218 */
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002219 BUG_ON(flags & ~CREATE_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002220
Andrew Mortona737b3e2006-03-22 00:08:11 -08002221 /*
2222 * Check that size is in terms of words. This is needed to avoid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002223 * unaligned accesses for some archs when redzoning is used, and makes
2224 * sure any on-slab bufctl's are also correctly aligned.
2225 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002226 if (size & (BYTES_PER_WORD - 1)) {
2227 size += (BYTES_PER_WORD - 1);
2228 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002229 }
2230
Andrew Mortona737b3e2006-03-22 00:08:11 -08002231 /* calculate the final buffer alignment: */
2232
Linus Torvalds1da177e2005-04-16 15:20:36 -07002233 /* 1) arch recommendation: can be overridden for debug */
2234 if (flags & SLAB_HWCACHE_ALIGN) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002235 /*
2236 * Default alignment: as specified by the arch code. Except if
2237 * an object is really small, then squeeze multiple objects into
2238 * one cacheline.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002239 */
2240 ralign = cache_line_size();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002241 while (size <= ralign / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002242 ralign /= 2;
2243 } else {
2244 ralign = BYTES_PER_WORD;
2245 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002246
2247 /*
David Woodhouse87a927c2007-07-04 21:26:44 -04002248 * Redzoning and user store require word alignment or possibly larger.
2249 * Note this will be overridden by architecture or caller mandated
2250 * alignment if either is greater than BYTES_PER_WORD.
Pekka Enbergca5f9702006-09-25 23:31:25 -07002251 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002252 if (flags & SLAB_STORE_USER)
2253 ralign = BYTES_PER_WORD;
2254
2255 if (flags & SLAB_RED_ZONE) {
2256 ralign = REDZONE_ALIGN;
2257 /* If redzoning, ensure that the second redzone is suitably
2258 * aligned, by adjusting the object size accordingly. */
2259 size += REDZONE_ALIGN - 1;
2260 size &= ~(REDZONE_ALIGN - 1);
2261 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002262
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002263 /* 2) arch mandated alignment */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002264 if (ralign < ARCH_SLAB_MINALIGN) {
2265 ralign = ARCH_SLAB_MINALIGN;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002266 }
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002267 /* 3) caller mandated alignment */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002268 if (ralign < align) {
2269 ralign = align;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002270 }
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002271 /* disable debug if necessary */
David Woodhouseb46b8f12007-05-08 00:22:59 -07002272 if (ralign > __alignof__(unsigned long long))
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002273 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002274 /*
Pekka Enbergca5f9702006-09-25 23:31:25 -07002275 * 4) Store it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002276 */
2277 align = ralign;
2278
2279 /* Get cache's description obj. */
Christoph Lametere94b1762006-12-06 20:33:17 -08002280 cachep = kmem_cache_zalloc(&cache_cache, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002281 if (!cachep)
Andrew Morton4f12bb42005-11-07 00:58:00 -08002282 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002283
2284#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002285 cachep->obj_size = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002286
Pekka Enbergca5f9702006-09-25 23:31:25 -07002287 /*
2288 * Both debugging options require word-alignment which is calculated
2289 * into align above.
2290 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002291 if (flags & SLAB_RED_ZONE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002292 /* add space for red zone words */
David Woodhouseb46b8f12007-05-08 00:22:59 -07002293 cachep->obj_offset += sizeof(unsigned long long);
2294 size += 2 * sizeof(unsigned long long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002295 }
2296 if (flags & SLAB_STORE_USER) {
Pekka Enbergca5f9702006-09-25 23:31:25 -07002297 /* user store requires one word storage behind the end of
David Woodhouse87a927c2007-07-04 21:26:44 -04002298 * the real object. But if the second red zone needs to be
2299 * aligned to 64 bits, we must allow that much space.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002300 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002301 if (flags & SLAB_RED_ZONE)
2302 size += REDZONE_ALIGN;
2303 else
2304 size += BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002305 }
2306#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002307 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002308 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2309 cachep->obj_offset += PAGE_SIZE - size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002310 size = PAGE_SIZE;
2311 }
2312#endif
2313#endif
2314
Ingo Molnare0a42722006-06-23 02:03:46 -07002315 /*
2316 * Determine if the slab management is 'on' or 'off' slab.
2317 * (bootstrapping cannot cope with offslab caches so don't do
2318 * it too early on.)
2319 */
2320 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002321 /*
2322 * Size is large, assume best to place the slab management obj
2323 * off-slab (should allow better packing of objs).
2324 */
2325 flags |= CFLGS_OFF_SLAB;
2326
2327 size = ALIGN(size, align);
2328
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002329 left_over = calculate_slab_order(cachep, size, align, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002330
2331 if (!cachep->num) {
matzeb4169522007-05-06 14:49:52 -07002332 printk(KERN_ERR
2333 "kmem_cache_create: couldn't create cache %s.\n", name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002334 kmem_cache_free(&cache_cache, cachep);
2335 cachep = NULL;
Andrew Morton4f12bb42005-11-07 00:58:00 -08002336 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002337 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002338 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2339 + sizeof(struct slab), align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002340
2341 /*
2342 * If the slab has been placed off-slab, and we have enough space then
2343 * move it on-slab. This is at the expense of any extra colouring.
2344 */
2345 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2346 flags &= ~CFLGS_OFF_SLAB;
2347 left_over -= slab_size;
2348 }
2349
2350 if (flags & CFLGS_OFF_SLAB) {
2351 /* really off slab. No need for manual alignment */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002352 slab_size =
2353 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002354 }
2355
2356 cachep->colour_off = cache_line_size();
2357 /* Offset must be a multiple of the alignment. */
2358 if (cachep->colour_off < align)
2359 cachep->colour_off = align;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002360 cachep->colour = left_over / cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002361 cachep->slab_size = slab_size;
2362 cachep->flags = flags;
2363 cachep->gfpflags = 0;
Christoph Lameter4b51d662007-02-10 01:43:10 -08002364 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002365 cachep->gfpflags |= GFP_DMA;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002366 cachep->buffer_size = size;
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08002367 cachep->reciprocal_buffer_size = reciprocal_value(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002368
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002369 if (flags & CFLGS_OFF_SLAB) {
Victor Fuscob2d55072005-09-10 00:26:36 -07002370 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002371 /*
2372 * This is a possibility for one of the malloc_sizes caches.
2373 * But since we go off slab only for object size greater than
2374 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2375 * this should not happen at all.
2376 * But leave a BUG_ON for some lucky dude.
2377 */
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002378 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002379 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002380 cachep->ctor = ctor;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002381 cachep->name = name;
2382
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002383 if (setup_cpu_cache(cachep)) {
2384 __kmem_cache_destroy(cachep);
2385 cachep = NULL;
2386 goto oops;
2387 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002388
Linus Torvalds1da177e2005-04-16 15:20:36 -07002389 /* cache setup completed, link it into the list */
2390 list_add(&cachep->next, &cache_chain);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002391oops:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002392 if (!cachep && (flags & SLAB_PANIC))
2393 panic("kmem_cache_create(): failed to create slab `%s'\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002394 name);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002395 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002396 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002397 return cachep;
2398}
2399EXPORT_SYMBOL(kmem_cache_create);
2400
2401#if DEBUG
2402static void check_irq_off(void)
2403{
2404 BUG_ON(!irqs_disabled());
2405}
2406
2407static void check_irq_on(void)
2408{
2409 BUG_ON(irqs_disabled());
2410}
2411
Pekka Enberg343e0d72006-02-01 03:05:50 -08002412static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002413{
2414#ifdef CONFIG_SMP
2415 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002416 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002417#endif
2418}
Christoph Lametere498be72005-09-09 13:03:32 -07002419
Pekka Enberg343e0d72006-02-01 03:05:50 -08002420static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002421{
2422#ifdef CONFIG_SMP
2423 check_irq_off();
2424 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2425#endif
2426}
2427
Linus Torvalds1da177e2005-04-16 15:20:36 -07002428#else
2429#define check_irq_off() do { } while(0)
2430#define check_irq_on() do { } while(0)
2431#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002432#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002433#endif
2434
Christoph Lameteraab22072006-03-22 00:09:06 -08002435static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2436 struct array_cache *ac,
2437 int force, int node);
2438
Linus Torvalds1da177e2005-04-16 15:20:36 -07002439static void do_drain(void *arg)
2440{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002441 struct kmem_cache *cachep = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002442 struct array_cache *ac;
Christoph Lameterff694162005-09-22 21:44:02 -07002443 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002444
2445 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002446 ac = cpu_cache_get(cachep);
Christoph Lameterff694162005-09-22 21:44:02 -07002447 spin_lock(&cachep->nodelists[node]->list_lock);
2448 free_block(cachep, ac->entry, ac->avail, node);
2449 spin_unlock(&cachep->nodelists[node]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002450 ac->avail = 0;
2451}
2452
Pekka Enberg343e0d72006-02-01 03:05:50 -08002453static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002454{
Christoph Lametere498be72005-09-09 13:03:32 -07002455 struct kmem_list3 *l3;
2456 int node;
2457
Andrew Mortona07fa392006-03-22 00:08:17 -08002458 on_each_cpu(do_drain, cachep, 1, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002459 check_irq_on();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002460 for_each_online_node(node) {
Christoph Lametere498be72005-09-09 13:03:32 -07002461 l3 = cachep->nodelists[node];
Roland Dreiera4523a82006-05-15 11:41:00 -07002462 if (l3 && l3->alien)
2463 drain_alien_cache(cachep, l3->alien);
2464 }
2465
2466 for_each_online_node(node) {
2467 l3 = cachep->nodelists[node];
2468 if (l3)
Christoph Lameteraab22072006-03-22 00:09:06 -08002469 drain_array(cachep, l3, l3->shared, 1, node);
Christoph Lametere498be72005-09-09 13:03:32 -07002470 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002471}
2472
Christoph Lametered11d9e2006-06-30 01:55:45 -07002473/*
2474 * Remove slabs from the list of free slabs.
2475 * Specify the number of slabs to drain in tofree.
2476 *
2477 * Returns the actual number of slabs released.
2478 */
2479static int drain_freelist(struct kmem_cache *cache,
2480 struct kmem_list3 *l3, int tofree)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002481{
Christoph Lametered11d9e2006-06-30 01:55:45 -07002482 struct list_head *p;
2483 int nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002484 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002485
Christoph Lametered11d9e2006-06-30 01:55:45 -07002486 nr_freed = 0;
2487 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002488
Christoph Lametered11d9e2006-06-30 01:55:45 -07002489 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002490 p = l3->slabs_free.prev;
Christoph Lametered11d9e2006-06-30 01:55:45 -07002491 if (p == &l3->slabs_free) {
2492 spin_unlock_irq(&l3->list_lock);
2493 goto out;
2494 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002495
Christoph Lametered11d9e2006-06-30 01:55:45 -07002496 slabp = list_entry(p, struct slab, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002497#if DEBUG
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002498 BUG_ON(slabp->inuse);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002499#endif
2500 list_del(&slabp->list);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002501 /*
2502 * Safe to drop the lock. The slab is no longer linked
2503 * to the cache.
2504 */
2505 l3->free_objects -= cache->num;
Christoph Lametere498be72005-09-09 13:03:32 -07002506 spin_unlock_irq(&l3->list_lock);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002507 slab_destroy(cache, slabp);
2508 nr_freed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002509 }
Christoph Lametered11d9e2006-06-30 01:55:45 -07002510out:
2511 return nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002512}
2513
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002514/* Called with cache_chain_mutex held to protect against cpu hotplug */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002515static int __cache_shrink(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07002516{
2517 int ret = 0, i = 0;
2518 struct kmem_list3 *l3;
2519
2520 drain_cpu_caches(cachep);
2521
2522 check_irq_on();
2523 for_each_online_node(i) {
2524 l3 = cachep->nodelists[i];
Christoph Lametered11d9e2006-06-30 01:55:45 -07002525 if (!l3)
2526 continue;
2527
2528 drain_freelist(cachep, l3, l3->free_objects);
2529
2530 ret += !list_empty(&l3->slabs_full) ||
2531 !list_empty(&l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07002532 }
2533 return (ret ? 1 : 0);
2534}
2535
Linus Torvalds1da177e2005-04-16 15:20:36 -07002536/**
2537 * kmem_cache_shrink - Shrink a cache.
2538 * @cachep: The cache to shrink.
2539 *
2540 * Releases as many slabs as possible for a cache.
2541 * To help debugging, a zero exit status indicates all slabs were released.
2542 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002543int kmem_cache_shrink(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002544{
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002545 int ret;
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002546 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002547
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002548 get_online_cpus();
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002549 mutex_lock(&cache_chain_mutex);
2550 ret = __cache_shrink(cachep);
2551 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002552 put_online_cpus();
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002553 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002554}
2555EXPORT_SYMBOL(kmem_cache_shrink);
2556
2557/**
2558 * kmem_cache_destroy - delete a cache
2559 * @cachep: the cache to destroy
2560 *
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08002561 * Remove a &struct kmem_cache object from the slab cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002562 *
2563 * It is expected this function will be called by a module when it is
2564 * unloaded. This will remove the cache completely, and avoid a duplicate
2565 * cache being allocated each time a module is loaded and unloaded, if the
2566 * module doesn't have persistent in-kernel storage across loads and unloads.
2567 *
2568 * The cache must be empty before calling this function.
2569 *
2570 * The caller must guarantee that noone will allocate memory from the cache
2571 * during the kmem_cache_destroy().
2572 */
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002573void kmem_cache_destroy(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002574{
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002575 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002576
Linus Torvalds1da177e2005-04-16 15:20:36 -07002577 /* Find the cache in the chain of caches. */
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002578 get_online_cpus();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002579 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002580 /*
2581 * the chain is never empty, cache_cache is never destroyed
2582 */
2583 list_del(&cachep->next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002584 if (__cache_shrink(cachep)) {
2585 slab_error(cachep, "Can't free all objects");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002586 list_add(&cachep->next, &cache_chain);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002587 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002588 put_online_cpus();
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002589 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002590 }
2591
2592 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
Paul E. McKenneyfbd568a3e2005-05-01 08:59:04 -07002593 synchronize_rcu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002594
Christoph Lameter117f6eb2006-09-25 23:31:37 -07002595 __kmem_cache_destroy(cachep);
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002596 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002597 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002598}
2599EXPORT_SYMBOL(kmem_cache_destroy);
2600
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002601/*
2602 * Get the memory for a slab management obj.
2603 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2604 * always come from malloc_sizes caches. The slab descriptor cannot
2605 * come from the same cache which is getting created because,
2606 * when we are searching for an appropriate cache for these
2607 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2608 * If we are creating a malloc_sizes cache here it would not be visible to
2609 * kmem_find_general_cachep till the initialization is complete.
2610 * Hence we cannot have slabp_cache same as the original cache.
2611 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002612static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002613 int colour_off, gfp_t local_flags,
2614 int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002615{
2616 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002617
Linus Torvalds1da177e2005-04-16 15:20:36 -07002618 if (OFF_SLAB(cachep)) {
2619 /* Slab management obj is off-slab. */
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002620 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
Christoph Lameter3c517a62006-12-06 20:33:29 -08002621 local_flags & ~GFP_THISNODE, nodeid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002622 if (!slabp)
2623 return NULL;
2624 } else {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002625 slabp = objp + colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002626 colour_off += cachep->slab_size;
2627 }
2628 slabp->inuse = 0;
2629 slabp->colouroff = colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002630 slabp->s_mem = objp + colour_off;
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002631 slabp->nodeid = nodeid;
Marcin Slusarze51bfd02008-02-10 11:21:54 +01002632 slabp->free = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002633 return slabp;
2634}
2635
2636static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2637{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002638 return (kmem_bufctl_t *) (slabp + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002639}
2640
Pekka Enberg343e0d72006-02-01 03:05:50 -08002641static void cache_init_objs(struct kmem_cache *cachep,
Christoph Lametera35afb82007-05-16 22:10:57 -07002642 struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002643{
2644 int i;
2645
2646 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002647 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002648#if DEBUG
2649 /* need to poison the objs? */
2650 if (cachep->flags & SLAB_POISON)
2651 poison_obj(cachep, objp, POISON_FREE);
2652 if (cachep->flags & SLAB_STORE_USER)
2653 *dbg_userword(cachep, objp) = NULL;
2654
2655 if (cachep->flags & SLAB_RED_ZONE) {
2656 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2657 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2658 }
2659 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002660 * Constructors are not allowed to allocate memory from the same
2661 * cache which they are a constructor for. Otherwise, deadlock.
2662 * They must also be threaded.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002663 */
2664 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07002665 cachep->ctor(cachep, objp + obj_offset(cachep));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002666
2667 if (cachep->flags & SLAB_RED_ZONE) {
2668 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2669 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002670 " end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002671 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2672 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002673 " start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002674 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08002675 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2676 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002677 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002678 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002679#else
2680 if (cachep->ctor)
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07002681 cachep->ctor(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002682#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002683 slab_bufctl(slabp)[i] = i + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002684 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002685 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002686}
2687
Pekka Enberg343e0d72006-02-01 03:05:50 -08002688static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002689{
Christoph Lameter4b51d662007-02-10 01:43:10 -08002690 if (CONFIG_ZONE_DMA_FLAG) {
2691 if (flags & GFP_DMA)
2692 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2693 else
2694 BUG_ON(cachep->gfpflags & GFP_DMA);
2695 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002696}
2697
Andrew Mortona737b3e2006-03-22 00:08:11 -08002698static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2699 int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002700{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002701 void *objp = index_to_obj(cachep, slabp, slabp->free);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002702 kmem_bufctl_t next;
2703
2704 slabp->inuse++;
2705 next = slab_bufctl(slabp)[slabp->free];
2706#if DEBUG
2707 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2708 WARN_ON(slabp->nodeid != nodeid);
2709#endif
2710 slabp->free = next;
2711
2712 return objp;
2713}
2714
Andrew Mortona737b3e2006-03-22 00:08:11 -08002715static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2716 void *objp, int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002717{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002718 unsigned int objnr = obj_to_index(cachep, slabp, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002719
2720#if DEBUG
2721 /* Verify that the slab belongs to the intended node */
2722 WARN_ON(slabp->nodeid != nodeid);
2723
Al Viro871751e2006-03-25 03:06:39 -08002724 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
Matthew Dobson78d382d2006-02-01 03:05:47 -08002725 printk(KERN_ERR "slab: double free detected in cache "
Andrew Mortona737b3e2006-03-22 00:08:11 -08002726 "'%s', objp %p\n", cachep->name, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002727 BUG();
2728 }
2729#endif
2730 slab_bufctl(slabp)[objnr] = slabp->free;
2731 slabp->free = objnr;
2732 slabp->inuse--;
2733}
2734
Pekka Enberg47768742006-06-23 02:03:07 -07002735/*
2736 * Map pages beginning at addr to the given cache and slab. This is required
2737 * for the slab allocator to be able to lookup the cache and slab of a
2738 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2739 */
2740static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2741 void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002742{
Pekka Enberg47768742006-06-23 02:03:07 -07002743 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002744 struct page *page;
2745
Pekka Enberg47768742006-06-23 02:03:07 -07002746 page = virt_to_page(addr);
Nick Piggin84097512006-03-22 00:08:34 -08002747
Pekka Enberg47768742006-06-23 02:03:07 -07002748 nr_pages = 1;
Nick Piggin84097512006-03-22 00:08:34 -08002749 if (likely(!PageCompound(page)))
Pekka Enberg47768742006-06-23 02:03:07 -07002750 nr_pages <<= cache->gfporder;
2751
Linus Torvalds1da177e2005-04-16 15:20:36 -07002752 do {
Pekka Enberg47768742006-06-23 02:03:07 -07002753 page_set_cache(page, cache);
2754 page_set_slab(page, slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002755 page++;
Pekka Enberg47768742006-06-23 02:03:07 -07002756 } while (--nr_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002757}
2758
2759/*
2760 * Grow (by 1) the number of slabs within a cache. This is called by
2761 * kmem_cache_alloc() when there are no active objs left in a cache.
2762 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002763static int cache_grow(struct kmem_cache *cachep,
2764 gfp_t flags, int nodeid, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002765{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002766 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002767 size_t offset;
2768 gfp_t local_flags;
Christoph Lametere498be72005-09-09 13:03:32 -07002769 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002770
Andrew Mortona737b3e2006-03-22 00:08:11 -08002771 /*
2772 * Be lazy and only check for valid flags here, keeping it out of the
2773 * critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002774 */
Christoph Lameter6cb06222007-10-16 01:25:41 -07002775 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2776 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002777
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002778 /* Take the l3 list lock to change the colour_next on this node */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002779 check_irq_off();
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002780 l3 = cachep->nodelists[nodeid];
2781 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002782
2783 /* Get colour for the slab, and cal the next value. */
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002784 offset = l3->colour_next;
2785 l3->colour_next++;
2786 if (l3->colour_next >= cachep->colour)
2787 l3->colour_next = 0;
2788 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002789
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002790 offset *= cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002791
2792 if (local_flags & __GFP_WAIT)
2793 local_irq_enable();
2794
2795 /*
2796 * The test for missing atomic flag is performed here, rather than
2797 * the more obvious place, simply to reduce the critical path length
2798 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2799 * will eventually be caught here (where it matters).
2800 */
2801 kmem_flagcheck(cachep, flags);
2802
Andrew Mortona737b3e2006-03-22 00:08:11 -08002803 /*
2804 * Get mem for the objs. Attempt to allocate a physical page from
2805 * 'nodeid'.
Christoph Lametere498be72005-09-09 13:03:32 -07002806 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002807 if (!objp)
Andrew Mortonb8c1c5d2007-07-24 12:02:40 -07002808 objp = kmem_getpages(cachep, local_flags, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002809 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002810 goto failed;
2811
2812 /* Get slab management. */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002813 slabp = alloc_slabmgmt(cachep, objp, offset,
Christoph Lameter6cb06222007-10-16 01:25:41 -07002814 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002815 if (!slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002816 goto opps1;
2817
Pekka Enberg47768742006-06-23 02:03:07 -07002818 slab_map_pages(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002819
Christoph Lametera35afb82007-05-16 22:10:57 -07002820 cache_init_objs(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002821
2822 if (local_flags & __GFP_WAIT)
2823 local_irq_disable();
2824 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002825 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002826
2827 /* Make slab active. */
Christoph Lametere498be72005-09-09 13:03:32 -07002828 list_add_tail(&slabp->list, &(l3->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002829 STATS_INC_GROWN(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002830 l3->free_objects += cachep->num;
2831 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002832 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002833opps1:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002834 kmem_freepages(cachep, objp);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002835failed:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002836 if (local_flags & __GFP_WAIT)
2837 local_irq_disable();
2838 return 0;
2839}
2840
2841#if DEBUG
2842
2843/*
2844 * Perform extra freeing checks:
2845 * - detect bad pointers.
2846 * - POISON/RED_ZONE checking
Linus Torvalds1da177e2005-04-16 15:20:36 -07002847 */
2848static void kfree_debugcheck(const void *objp)
2849{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002850 if (!virt_addr_valid(objp)) {
2851 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002852 (unsigned long)objp);
2853 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002854 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002855}
2856
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002857static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2858{
David Woodhouseb46b8f12007-05-08 00:22:59 -07002859 unsigned long long redzone1, redzone2;
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002860
2861 redzone1 = *dbg_redzone1(cache, obj);
2862 redzone2 = *dbg_redzone2(cache, obj);
2863
2864 /*
2865 * Redzone is ok.
2866 */
2867 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2868 return;
2869
2870 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2871 slab_error(cache, "double free detected");
2872 else
2873 slab_error(cache, "memory outside object was overwritten");
2874
David Woodhouseb46b8f12007-05-08 00:22:59 -07002875 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002876 obj, redzone1, redzone2);
2877}
2878
Pekka Enberg343e0d72006-02-01 03:05:50 -08002879static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002880 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002881{
2882 struct page *page;
2883 unsigned int objnr;
2884 struct slab *slabp;
2885
Matthew Wilcox80cbd912007-11-29 12:05:13 -07002886 BUG_ON(virt_to_cache(objp) != cachep);
2887
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002888 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002889 kfree_debugcheck(objp);
Christoph Lameterb49af682007-05-06 14:49:41 -07002890 page = virt_to_head_page(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002891
Pekka Enberg065d41c2005-11-13 16:06:46 -08002892 slabp = page_get_slab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002893
2894 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002895 verify_redzone_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002896 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2897 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2898 }
2899 if (cachep->flags & SLAB_STORE_USER)
2900 *dbg_userword(cachep, objp) = caller;
2901
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002902 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002903
2904 BUG_ON(objnr >= cachep->num);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002905 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002906
Al Viro871751e2006-03-25 03:06:39 -08002907#ifdef CONFIG_DEBUG_SLAB_LEAK
2908 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2909#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002910 if (cachep->flags & SLAB_POISON) {
2911#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08002912 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002913 store_stackinfo(cachep, objp, (unsigned long)caller);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002914 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002915 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002916 } else {
2917 poison_obj(cachep, objp, POISON_FREE);
2918 }
2919#else
2920 poison_obj(cachep, objp, POISON_FREE);
2921#endif
2922 }
2923 return objp;
2924}
2925
Pekka Enberg343e0d72006-02-01 03:05:50 -08002926static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002927{
2928 kmem_bufctl_t i;
2929 int entries = 0;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002930
Linus Torvalds1da177e2005-04-16 15:20:36 -07002931 /* Check slab's freelist to see if this obj is there. */
2932 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2933 entries++;
2934 if (entries > cachep->num || i >= cachep->num)
2935 goto bad;
2936 }
2937 if (entries != cachep->num - slabp->inuse) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002938bad:
2939 printk(KERN_ERR "slab: Internal list corruption detected in "
2940 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2941 cachep->name, cachep->num, slabp, slabp->inuse);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002942 for (i = 0;
Linus Torvalds264132b2006-03-06 12:10:07 -08002943 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002944 i++) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002945 if (i % 16 == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002946 printk("\n%03x:", i);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002947 printk(" %02x", ((unsigned char *)slabp)[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002948 }
2949 printk("\n");
2950 BUG();
2951 }
2952}
2953#else
2954#define kfree_debugcheck(x) do { } while(0)
2955#define cache_free_debugcheck(x,objp,z) (objp)
2956#define check_slabp(x,y) do { } while(0)
2957#endif
2958
Pekka Enberg343e0d72006-02-01 03:05:50 -08002959static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002960{
2961 int batchcount;
2962 struct kmem_list3 *l3;
2963 struct array_cache *ac;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002964 int node;
2965
Andrew Mortona737b3e2006-03-22 00:08:11 -08002966retry:
Joe Korty6d2144d2008-03-05 15:04:59 -08002967 check_irq_off();
2968 node = numa_node_id();
2969 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002970 batchcount = ac->batchcount;
2971 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002972 /*
2973 * If there was little recent activity on this cache, then
2974 * perform only a partial refill. Otherwise we could generate
2975 * refill bouncing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002976 */
2977 batchcount = BATCHREFILL_LIMIT;
2978 }
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002979 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002980
Christoph Lametere498be72005-09-09 13:03:32 -07002981 BUG_ON(ac->avail > 0 || !l3);
2982 spin_lock(&l3->list_lock);
2983
Christoph Lameter3ded1752006-03-25 03:06:44 -08002984 /* See if we can refill from the shared array */
2985 if (l3->shared && transfer_objects(ac, l3->shared, batchcount))
2986 goto alloc_done;
2987
Linus Torvalds1da177e2005-04-16 15:20:36 -07002988 while (batchcount > 0) {
2989 struct list_head *entry;
2990 struct slab *slabp;
2991 /* Get slab alloc is to come from. */
2992 entry = l3->slabs_partial.next;
2993 if (entry == &l3->slabs_partial) {
2994 l3->free_touched = 1;
2995 entry = l3->slabs_free.next;
2996 if (entry == &l3->slabs_free)
2997 goto must_grow;
2998 }
2999
3000 slabp = list_entry(entry, struct slab, list);
3001 check_slabp(cachep, slabp);
3002 check_spinlock_acquired(cachep);
Pekka Enberg714b8172007-05-06 14:49:03 -07003003
3004 /*
3005 * The slab was either on partial or free list so
3006 * there must be at least one object available for
3007 * allocation.
3008 */
3009 BUG_ON(slabp->inuse < 0 || slabp->inuse >= cachep->num);
3010
Linus Torvalds1da177e2005-04-16 15:20:36 -07003011 while (slabp->inuse < cachep->num && batchcount--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003012 STATS_INC_ALLOCED(cachep);
3013 STATS_INC_ACTIVE(cachep);
3014 STATS_SET_HIGH(cachep);
3015
Matthew Dobson78d382d2006-02-01 03:05:47 -08003016 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07003017 node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003018 }
3019 check_slabp(cachep, slabp);
3020
3021 /* move slabp to correct slabp list: */
3022 list_del(&slabp->list);
3023 if (slabp->free == BUFCTL_END)
3024 list_add(&slabp->list, &l3->slabs_full);
3025 else
3026 list_add(&slabp->list, &l3->slabs_partial);
3027 }
3028
Andrew Mortona737b3e2006-03-22 00:08:11 -08003029must_grow:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003030 l3->free_objects -= ac->avail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003031alloc_done:
Christoph Lametere498be72005-09-09 13:03:32 -07003032 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003033
3034 if (unlikely(!ac->avail)) {
3035 int x;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003036 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07003037
Andrew Mortona737b3e2006-03-22 00:08:11 -08003038 /* cache_grow can reenable interrupts, then ac could change. */
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003039 ac = cpu_cache_get(cachep);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003040 if (!x && ac->avail == 0) /* no objects in sight? abort */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003041 return NULL;
3042
Andrew Mortona737b3e2006-03-22 00:08:11 -08003043 if (!ac->avail) /* objects refilled by interrupt? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003044 goto retry;
3045 }
3046 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07003047 return ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003048}
3049
Andrew Mortona737b3e2006-03-22 00:08:11 -08003050static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3051 gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003052{
3053 might_sleep_if(flags & __GFP_WAIT);
3054#if DEBUG
3055 kmem_flagcheck(cachep, flags);
3056#endif
3057}
3058
3059#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003060static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3061 gfp_t flags, void *objp, void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003062{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003063 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003064 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003065 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003066#ifdef CONFIG_DEBUG_PAGEALLOC
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003067 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003068 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003069 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003070 else
3071 check_poison_obj(cachep, objp);
3072#else
3073 check_poison_obj(cachep, objp);
3074#endif
3075 poison_obj(cachep, objp, POISON_INUSE);
3076 }
3077 if (cachep->flags & SLAB_STORE_USER)
3078 *dbg_userword(cachep, objp) = caller;
3079
3080 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003081 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3082 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3083 slab_error(cachep, "double free, or memory outside"
3084 " object was overwritten");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003085 printk(KERN_ERR
David Woodhouseb46b8f12007-05-08 00:22:59 -07003086 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08003087 objp, *dbg_redzone1(cachep, objp),
3088 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003089 }
3090 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3091 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3092 }
Al Viro871751e2006-03-25 03:06:39 -08003093#ifdef CONFIG_DEBUG_SLAB_LEAK
3094 {
3095 struct slab *slabp;
3096 unsigned objnr;
3097
Christoph Lameterb49af682007-05-06 14:49:41 -07003098 slabp = page_get_slab(virt_to_head_page(objp));
Al Viro871751e2006-03-25 03:06:39 -08003099 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3100 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3101 }
3102#endif
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003103 objp += obj_offset(cachep);
Christoph Lameter4f104932007-05-06 14:50:17 -07003104 if (cachep->ctor && cachep->flags & SLAB_POISON)
Christoph Lameter4ba9b9d2007-10-16 23:25:51 -07003105 cachep->ctor(cachep, objp);
Kevin Hilmana44b56d2006-12-06 20:32:11 -08003106#if ARCH_SLAB_MINALIGN
3107 if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) {
3108 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3109 objp, ARCH_SLAB_MINALIGN);
3110 }
3111#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003112 return objp;
3113}
3114#else
3115#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3116#endif
3117
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003118#ifdef CONFIG_FAILSLAB
3119
3120static struct failslab_attr {
3121
3122 struct fault_attr attr;
3123
3124 u32 ignore_gfp_wait;
3125#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3126 struct dentry *ignore_gfp_wait_file;
3127#endif
3128
3129} failslab = {
3130 .attr = FAULT_ATTR_INITIALIZER,
Don Mullis6b1b60f2006-12-08 02:39:53 -08003131 .ignore_gfp_wait = 1,
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003132};
3133
3134static int __init setup_failslab(char *str)
3135{
3136 return setup_fault_attr(&failslab.attr, str);
3137}
3138__setup("failslab=", setup_failslab);
3139
3140static int should_failslab(struct kmem_cache *cachep, gfp_t flags)
3141{
3142 if (cachep == &cache_cache)
3143 return 0;
3144 if (flags & __GFP_NOFAIL)
3145 return 0;
3146 if (failslab.ignore_gfp_wait && (flags & __GFP_WAIT))
3147 return 0;
3148
3149 return should_fail(&failslab.attr, obj_size(cachep));
3150}
3151
3152#ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3153
3154static int __init failslab_debugfs(void)
3155{
3156 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
3157 struct dentry *dir;
3158 int err;
3159
Akinobu Mita824ebef2007-05-06 14:49:58 -07003160 err = init_fault_attr_dentries(&failslab.attr, "failslab");
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003161 if (err)
3162 return err;
3163 dir = failslab.attr.dentries.dir;
3164
3165 failslab.ignore_gfp_wait_file =
3166 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3167 &failslab.ignore_gfp_wait);
3168
3169 if (!failslab.ignore_gfp_wait_file) {
3170 err = -ENOMEM;
3171 debugfs_remove(failslab.ignore_gfp_wait_file);
3172 cleanup_fault_attr_dentries(&failslab.attr);
3173 }
3174
3175 return err;
3176}
3177
3178late_initcall(failslab_debugfs);
3179
3180#endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3181
3182#else /* CONFIG_FAILSLAB */
3183
3184static inline int should_failslab(struct kmem_cache *cachep, gfp_t flags)
3185{
3186 return 0;
3187}
3188
3189#endif /* CONFIG_FAILSLAB */
3190
Pekka Enberg343e0d72006-02-01 03:05:50 -08003191static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003192{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003193 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003194 struct array_cache *ac;
3195
Alok N Kataria5c382302005-09-27 21:45:46 -07003196 check_irq_off();
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003197
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003198 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003199 if (likely(ac->avail)) {
3200 STATS_INC_ALLOCHIT(cachep);
3201 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07003202 objp = ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003203 } else {
3204 STATS_INC_ALLOCMISS(cachep);
3205 objp = cache_alloc_refill(cachep, flags);
3206 }
Alok N Kataria5c382302005-09-27 21:45:46 -07003207 return objp;
3208}
3209
Christoph Lametere498be72005-09-09 13:03:32 -07003210#ifdef CONFIG_NUMA
3211/*
Paul Jacksonb2455392006-03-24 03:16:12 -08003212 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
Paul Jacksonc61afb12006-03-24 03:16:08 -08003213 *
3214 * If we are in_interrupt, then process context, including cpusets and
3215 * mempolicy, may not apply and should not be used for allocation policy.
3216 */
3217static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3218{
3219 int nid_alloc, nid_here;
3220
Christoph Lameter765c4502006-09-27 01:50:08 -07003221 if (in_interrupt() || (flags & __GFP_THISNODE))
Paul Jacksonc61afb12006-03-24 03:16:08 -08003222 return NULL;
3223 nid_alloc = nid_here = numa_node_id();
3224 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3225 nid_alloc = cpuset_mem_spread_node();
3226 else if (current->mempolicy)
3227 nid_alloc = slab_node(current->mempolicy);
3228 if (nid_alloc != nid_here)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003229 return ____cache_alloc_node(cachep, flags, nid_alloc);
Paul Jacksonc61afb12006-03-24 03:16:08 -08003230 return NULL;
3231}
3232
3233/*
Christoph Lameter765c4502006-09-27 01:50:08 -07003234 * Fallback function if there was no memory available and no objects on a
Christoph Lameter3c517a62006-12-06 20:33:29 -08003235 * certain node and fall back is permitted. First we scan all the
3236 * available nodelists for available objects. If that fails then we
3237 * perform an allocation without specifying a node. This allows the page
3238 * allocator to do its reclaim / fallback magic. We then insert the
3239 * slab into the proper nodelist and then allocate from it.
Christoph Lameter765c4502006-09-27 01:50:08 -07003240 */
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003241static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
Christoph Lameter765c4502006-09-27 01:50:08 -07003242{
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003243 struct zonelist *zonelist;
3244 gfp_t local_flags;
Christoph Lameter765c4502006-09-27 01:50:08 -07003245 struct zone **z;
3246 void *obj = NULL;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003247 int nid;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003248
3249 if (flags & __GFP_THISNODE)
3250 return NULL;
3251
3252 zonelist = &NODE_DATA(slab_node(current->mempolicy))
3253 ->node_zonelists[gfp_zone(flags)];
Christoph Lameter6cb06222007-10-16 01:25:41 -07003254 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Christoph Lameter765c4502006-09-27 01:50:08 -07003255
Christoph Lameter3c517a62006-12-06 20:33:29 -08003256retry:
3257 /*
3258 * Look through allowed nodes for objects available
3259 * from existing per node queues.
3260 */
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003261 for (z = zonelist->zones; *z && !obj; z++) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003262 nid = zone_to_nid(*z);
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003263
Paul Jackson02a0e532006-12-13 00:34:25 -08003264 if (cpuset_zone_allowed_hardwall(*z, flags) &&
Christoph Lameter3c517a62006-12-06 20:33:29 -08003265 cache->nodelists[nid] &&
3266 cache->nodelists[nid]->free_objects)
3267 obj = ____cache_alloc_node(cache,
3268 flags | GFP_THISNODE, nid);
3269 }
3270
Christoph Lametercfce6602007-05-06 14:50:17 -07003271 if (!obj) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003272 /*
3273 * This allocation will be performed within the constraints
3274 * of the current cpuset / memory policy requirements.
3275 * We may trigger various forms of reclaim on the allowed
3276 * set and go into memory reserves if necessary.
3277 */
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003278 if (local_flags & __GFP_WAIT)
3279 local_irq_enable();
3280 kmem_flagcheck(cache, flags);
Christoph Lameter9ac33b22008-03-04 12:24:22 -08003281 obj = kmem_getpages(cache, local_flags, -1);
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003282 if (local_flags & __GFP_WAIT)
3283 local_irq_disable();
Christoph Lameter3c517a62006-12-06 20:33:29 -08003284 if (obj) {
3285 /*
3286 * Insert into the appropriate per node queues
3287 */
3288 nid = page_to_nid(virt_to_page(obj));
3289 if (cache_grow(cache, flags, nid, obj)) {
3290 obj = ____cache_alloc_node(cache,
3291 flags | GFP_THISNODE, nid);
3292 if (!obj)
3293 /*
3294 * Another processor may allocate the
3295 * objects in the slab since we are
3296 * not holding any locks.
3297 */
3298 goto retry;
3299 } else {
Hugh Dickinsb6a60452007-01-05 16:36:36 -08003300 /* cache_grow already freed obj */
Christoph Lameter3c517a62006-12-06 20:33:29 -08003301 obj = NULL;
3302 }
3303 }
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003304 }
Christoph Lameter765c4502006-09-27 01:50:08 -07003305 return obj;
3306}
3307
3308/*
Christoph Lametere498be72005-09-09 13:03:32 -07003309 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07003310 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003311static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003312 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07003313{
3314 struct list_head *entry;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003315 struct slab *slabp;
3316 struct kmem_list3 *l3;
3317 void *obj;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003318 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003319
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003320 l3 = cachep->nodelists[nodeid];
3321 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07003322
Andrew Mortona737b3e2006-03-22 00:08:11 -08003323retry:
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003324 check_irq_off();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003325 spin_lock(&l3->list_lock);
3326 entry = l3->slabs_partial.next;
3327 if (entry == &l3->slabs_partial) {
3328 l3->free_touched = 1;
3329 entry = l3->slabs_free.next;
3330 if (entry == &l3->slabs_free)
3331 goto must_grow;
3332 }
Christoph Lametere498be72005-09-09 13:03:32 -07003333
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003334 slabp = list_entry(entry, struct slab, list);
3335 check_spinlock_acquired_node(cachep, nodeid);
3336 check_slabp(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07003337
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003338 STATS_INC_NODEALLOCS(cachep);
3339 STATS_INC_ACTIVE(cachep);
3340 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003341
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003342 BUG_ON(slabp->inuse == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07003343
Matthew Dobson78d382d2006-02-01 03:05:47 -08003344 obj = slab_get_obj(cachep, slabp, nodeid);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003345 check_slabp(cachep, slabp);
3346 l3->free_objects--;
3347 /* move slabp to correct slabp list: */
3348 list_del(&slabp->list);
Christoph Lametere498be72005-09-09 13:03:32 -07003349
Andrew Mortona737b3e2006-03-22 00:08:11 -08003350 if (slabp->free == BUFCTL_END)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003351 list_add(&slabp->list, &l3->slabs_full);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003352 else
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003353 list_add(&slabp->list, &l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07003354
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003355 spin_unlock(&l3->list_lock);
3356 goto done;
Christoph Lametere498be72005-09-09 13:03:32 -07003357
Andrew Mortona737b3e2006-03-22 00:08:11 -08003358must_grow:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003359 spin_unlock(&l3->list_lock);
Christoph Lameter3c517a62006-12-06 20:33:29 -08003360 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
Christoph Lameter765c4502006-09-27 01:50:08 -07003361 if (x)
3362 goto retry;
Christoph Lametere498be72005-09-09 13:03:32 -07003363
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003364 return fallback_alloc(cachep, flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003365
Andrew Mortona737b3e2006-03-22 00:08:11 -08003366done:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003367 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07003368}
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003369
3370/**
3371 * kmem_cache_alloc_node - Allocate an object on the specified node
3372 * @cachep: The cache to allocate from.
3373 * @flags: See kmalloc().
3374 * @nodeid: node number of the target node.
3375 * @caller: return address of caller, used for debug information
3376 *
3377 * Identical to kmem_cache_alloc but it will allocate memory on the given
3378 * node, which can improve the performance for cpu bound structures.
3379 *
3380 * Fallback to other node is possible if __GFP_THISNODE is not set.
3381 */
3382static __always_inline void *
3383__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3384 void *caller)
3385{
3386 unsigned long save_flags;
3387 void *ptr;
3388
Akinobu Mita824ebef2007-05-06 14:49:58 -07003389 if (should_failslab(cachep, flags))
3390 return NULL;
3391
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003392 cache_alloc_debugcheck_before(cachep, flags);
3393 local_irq_save(save_flags);
3394
3395 if (unlikely(nodeid == -1))
3396 nodeid = numa_node_id();
3397
3398 if (unlikely(!cachep->nodelists[nodeid])) {
3399 /* Node not bootstrapped yet */
3400 ptr = fallback_alloc(cachep, flags);
3401 goto out;
3402 }
3403
3404 if (nodeid == numa_node_id()) {
3405 /*
3406 * Use the locally cached objects if possible.
3407 * However ____cache_alloc does not allow fallback
3408 * to other nodes. It may fail while we still have
3409 * objects on other nodes available.
3410 */
3411 ptr = ____cache_alloc(cachep, flags);
3412 if (ptr)
3413 goto out;
3414 }
3415 /* ___cache_alloc_node can fall back to other nodes */
3416 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3417 out:
3418 local_irq_restore(save_flags);
3419 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3420
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003421 if (unlikely((flags & __GFP_ZERO) && ptr))
3422 memset(ptr, 0, obj_size(cachep));
3423
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003424 return ptr;
3425}
3426
3427static __always_inline void *
3428__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3429{
3430 void *objp;
3431
3432 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3433 objp = alternate_node_alloc(cache, flags);
3434 if (objp)
3435 goto out;
3436 }
3437 objp = ____cache_alloc(cache, flags);
3438
3439 /*
3440 * We may just have run out of memory on the local node.
3441 * ____cache_alloc_node() knows how to locate memory on other nodes
3442 */
3443 if (!objp)
3444 objp = ____cache_alloc_node(cache, flags, numa_node_id());
3445
3446 out:
3447 return objp;
3448}
3449#else
3450
3451static __always_inline void *
3452__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3453{
3454 return ____cache_alloc(cachep, flags);
3455}
3456
3457#endif /* CONFIG_NUMA */
3458
3459static __always_inline void *
3460__cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3461{
3462 unsigned long save_flags;
3463 void *objp;
3464
Akinobu Mita824ebef2007-05-06 14:49:58 -07003465 if (should_failslab(cachep, flags))
3466 return NULL;
3467
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003468 cache_alloc_debugcheck_before(cachep, flags);
3469 local_irq_save(save_flags);
3470 objp = __do_cache_alloc(cachep, flags);
3471 local_irq_restore(save_flags);
3472 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3473 prefetchw(objp);
3474
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003475 if (unlikely((flags & __GFP_ZERO) && objp))
3476 memset(objp, 0, obj_size(cachep));
3477
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003478 return objp;
3479}
Christoph Lametere498be72005-09-09 13:03:32 -07003480
3481/*
3482 * Caller needs to acquire correct kmem_list's list_lock
3483 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003484static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003485 int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003486{
3487 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07003488 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003489
3490 for (i = 0; i < nr_objects; i++) {
3491 void *objp = objpp[i];
3492 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003493
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08003494 slabp = virt_to_slab(objp);
Christoph Lameterff694162005-09-22 21:44:02 -07003495 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003496 list_del(&slabp->list);
Christoph Lameterff694162005-09-22 21:44:02 -07003497 check_spinlock_acquired_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003498 check_slabp(cachep, slabp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08003499 slab_put_obj(cachep, slabp, objp, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003500 STATS_DEC_ACTIVE(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003501 l3->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003502 check_slabp(cachep, slabp);
3503
3504 /* fixup slab chains */
3505 if (slabp->inuse == 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07003506 if (l3->free_objects > l3->free_limit) {
3507 l3->free_objects -= cachep->num;
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07003508 /* No need to drop any previously held
3509 * lock here, even if we have a off-slab slab
3510 * descriptor it is guaranteed to come from
3511 * a different cache, refer to comments before
3512 * alloc_slabmgmt.
3513 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003514 slab_destroy(cachep, slabp);
3515 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07003516 list_add(&slabp->list, &l3->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003517 }
3518 } else {
3519 /* Unconditionally move a slab to the end of the
3520 * partial list on free - maximum time for the
3521 * other objects to be freed, too.
3522 */
Christoph Lametere498be72005-09-09 13:03:32 -07003523 list_add_tail(&slabp->list, &l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003524 }
3525 }
3526}
3527
Pekka Enberg343e0d72006-02-01 03:05:50 -08003528static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003529{
3530 int batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07003531 struct kmem_list3 *l3;
Christoph Lameterff694162005-09-22 21:44:02 -07003532 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003533
3534 batchcount = ac->batchcount;
3535#if DEBUG
3536 BUG_ON(!batchcount || batchcount > ac->avail);
3537#endif
3538 check_irq_off();
Christoph Lameterff694162005-09-22 21:44:02 -07003539 l3 = cachep->nodelists[node];
Ingo Molnar873623d2006-07-13 14:44:38 +02003540 spin_lock(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003541 if (l3->shared) {
3542 struct array_cache *shared_array = l3->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003543 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003544 if (max) {
3545 if (batchcount > max)
3546 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07003547 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003548 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003549 shared_array->avail += batchcount;
3550 goto free_done;
3551 }
3552 }
3553
Christoph Lameterff694162005-09-22 21:44:02 -07003554 free_block(cachep, ac->entry, batchcount, node);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003555free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003556#if STATS
3557 {
3558 int i = 0;
3559 struct list_head *p;
3560
Christoph Lametere498be72005-09-09 13:03:32 -07003561 p = l3->slabs_free.next;
3562 while (p != &(l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003563 struct slab *slabp;
3564
3565 slabp = list_entry(p, struct slab, list);
3566 BUG_ON(slabp->inuse);
3567
3568 i++;
3569 p = p->next;
3570 }
3571 STATS_SET_FREEABLE(cachep, i);
3572 }
3573#endif
Christoph Lametere498be72005-09-09 13:03:32 -07003574 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003575 ac->avail -= batchcount;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003576 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003577}
3578
3579/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08003580 * Release an obj back to its cache. If the obj has a constructed state, it must
3581 * be in this state _before_ it is released. Called with disabled ints.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003582 */
Ingo Molnar873623d2006-07-13 14:44:38 +02003583static inline void __cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003584{
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003585 struct array_cache *ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003586
3587 check_irq_off();
3588 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3589
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -07003590 /*
3591 * Skip calling cache_free_alien() when the platform is not numa.
3592 * This will avoid cache misses that happen while accessing slabp (which
3593 * is per page memory reference) to get nodeid. Instead use a global
3594 * variable to skip the call, which is mostly likely to be present in
3595 * the cache.
3596 */
3597 if (numa_platform && cache_free_alien(cachep, objp))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07003598 return;
Christoph Lametere498be72005-09-09 13:03:32 -07003599
Linus Torvalds1da177e2005-04-16 15:20:36 -07003600 if (likely(ac->avail < ac->limit)) {
3601 STATS_INC_FREEHIT(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003602 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003603 return;
3604 } else {
3605 STATS_INC_FREEMISS(cachep);
3606 cache_flusharray(cachep, ac);
Christoph Lametere498be72005-09-09 13:03:32 -07003607 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003608 }
3609}
3610
3611/**
3612 * kmem_cache_alloc - Allocate an object
3613 * @cachep: The cache to allocate from.
3614 * @flags: See kmalloc().
3615 *
3616 * Allocate an object from this cache. The flags are only relevant
3617 * if the cache has no available objects.
3618 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003619void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003620{
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003621 return __cache_alloc(cachep, flags, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003622}
3623EXPORT_SYMBOL(kmem_cache_alloc);
3624
3625/**
Randy Dunlap76824862008-03-19 17:00:40 -07003626 * kmem_ptr_validate - check if an untrusted pointer might be a slab entry.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003627 * @cachep: the cache we're checking against
3628 * @ptr: pointer to validate
3629 *
Randy Dunlap76824862008-03-19 17:00:40 -07003630 * This verifies that the untrusted pointer looks sane;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003631 * it is _not_ a guarantee that the pointer is actually
3632 * part of the slab cache in question, but it at least
3633 * validates that the pointer can be dereferenced and
3634 * looks half-way sane.
3635 *
3636 * Currently only used for dentry validation.
3637 */
Christoph Lameterb7f869a2006-12-22 01:06:44 -08003638int kmem_ptr_validate(struct kmem_cache *cachep, const void *ptr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003639{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003640 unsigned long addr = (unsigned long)ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003641 unsigned long min_addr = PAGE_OFFSET;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003642 unsigned long align_mask = BYTES_PER_WORD - 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003643 unsigned long size = cachep->buffer_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003644 struct page *page;
3645
3646 if (unlikely(addr < min_addr))
3647 goto out;
3648 if (unlikely(addr > (unsigned long)high_memory - size))
3649 goto out;
3650 if (unlikely(addr & align_mask))
3651 goto out;
3652 if (unlikely(!kern_addr_valid(addr)))
3653 goto out;
3654 if (unlikely(!kern_addr_valid(addr + size - 1)))
3655 goto out;
3656 page = virt_to_page(ptr);
3657 if (unlikely(!PageSlab(page)))
3658 goto out;
Pekka Enberg065d41c2005-11-13 16:06:46 -08003659 if (unlikely(page_get_cache(page) != cachep))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003660 goto out;
3661 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003662out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003663 return 0;
3664}
3665
3666#ifdef CONFIG_NUMA
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003667void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3668{
3669 return __cache_alloc_node(cachep, flags, nodeid,
3670 __builtin_return_address(0));
3671}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003672EXPORT_SYMBOL(kmem_cache_alloc_node);
3673
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003674static __always_inline void *
3675__do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003676{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003677 struct kmem_cache *cachep;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003678
3679 cachep = kmem_find_general_cachep(size, flags);
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003680 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3681 return cachep;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003682 return kmem_cache_alloc_node(cachep, flags, node);
3683}
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003684
3685#ifdef CONFIG_DEBUG_SLAB
3686void *__kmalloc_node(size_t size, gfp_t flags, int node)
3687{
3688 return __do_kmalloc_node(size, flags, node,
3689 __builtin_return_address(0));
3690}
Christoph Hellwigdbe5e692006-09-25 23:31:36 -07003691EXPORT_SYMBOL(__kmalloc_node);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003692
3693void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3694 int node, void *caller)
3695{
3696 return __do_kmalloc_node(size, flags, node, caller);
3697}
3698EXPORT_SYMBOL(__kmalloc_node_track_caller);
3699#else
3700void *__kmalloc_node(size_t size, gfp_t flags, int node)
3701{
3702 return __do_kmalloc_node(size, flags, node, NULL);
3703}
3704EXPORT_SYMBOL(__kmalloc_node);
3705#endif /* CONFIG_DEBUG_SLAB */
3706#endif /* CONFIG_NUMA */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003707
3708/**
Paul Drynoff800590f2006-06-23 02:03:48 -07003709 * __do_kmalloc - allocate memory
Linus Torvalds1da177e2005-04-16 15:20:36 -07003710 * @size: how many bytes of memory are required.
Paul Drynoff800590f2006-06-23 02:03:48 -07003711 * @flags: the type of memory to allocate (see kmalloc).
Randy Dunlap911851e2006-03-22 00:08:14 -08003712 * @caller: function caller for debug tracking of the caller
Linus Torvalds1da177e2005-04-16 15:20:36 -07003713 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003714static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3715 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003716{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003717 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003718
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003719 /* If you want to save a few bytes .text space: replace
3720 * __ with kmem_.
3721 * Then kmalloc uses the uninlined functions instead of the inline
3722 * functions.
3723 */
3724 cachep = __find_general_cachep(size, flags);
Linus Torvaldsa5c96d82007-07-19 13:17:15 -07003725 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3726 return cachep;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003727 return __cache_alloc(cachep, flags, caller);
3728}
3729
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003730
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003731#ifdef CONFIG_DEBUG_SLAB
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003732void *__kmalloc(size_t size, gfp_t flags)
3733{
Al Viro871751e2006-03-25 03:06:39 -08003734 return __do_kmalloc(size, flags, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003735}
3736EXPORT_SYMBOL(__kmalloc);
3737
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003738void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller)
3739{
3740 return __do_kmalloc(size, flags, caller);
3741}
3742EXPORT_SYMBOL(__kmalloc_track_caller);
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003743
3744#else
3745void *__kmalloc(size_t size, gfp_t flags)
3746{
3747 return __do_kmalloc(size, flags, NULL);
3748}
3749EXPORT_SYMBOL(__kmalloc);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003750#endif
3751
Linus Torvalds1da177e2005-04-16 15:20:36 -07003752/**
3753 * kmem_cache_free - Deallocate an object
3754 * @cachep: The cache the allocation was from.
3755 * @objp: The previously allocated object.
3756 *
3757 * Free an object which was previously allocated from this
3758 * cache.
3759 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003760void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003761{
3762 unsigned long flags;
3763
3764 local_irq_save(flags);
Ingo Molnar898552c2007-02-10 01:44:57 -08003765 debug_check_no_locks_freed(objp, obj_size(cachep));
Ingo Molnar873623d2006-07-13 14:44:38 +02003766 __cache_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003767 local_irq_restore(flags);
3768}
3769EXPORT_SYMBOL(kmem_cache_free);
3770
3771/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003772 * kfree - free previously allocated memory
3773 * @objp: pointer returned by kmalloc.
3774 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003775 * If @objp is NULL, no operation is performed.
3776 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003777 * Don't free memory not originally allocated by kmalloc()
3778 * or you will run into trouble.
3779 */
3780void kfree(const void *objp)
3781{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003782 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003783 unsigned long flags;
3784
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003785 if (unlikely(ZERO_OR_NULL_PTR(objp)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003786 return;
3787 local_irq_save(flags);
3788 kfree_debugcheck(objp);
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08003789 c = virt_to_cache(objp);
Ingo Molnarf9b84042006-06-27 02:54:49 -07003790 debug_check_no_locks_freed(objp, obj_size(c));
Ingo Molnar873623d2006-07-13 14:44:38 +02003791 __cache_free(c, (void *)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003792 local_irq_restore(flags);
3793}
3794EXPORT_SYMBOL(kfree);
3795
Pekka Enberg343e0d72006-02-01 03:05:50 -08003796unsigned int kmem_cache_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003797{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003798 return obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003799}
3800EXPORT_SYMBOL(kmem_cache_size);
3801
Pekka Enberg343e0d72006-02-01 03:05:50 -08003802const char *kmem_cache_name(struct kmem_cache *cachep)
Arnaldo Carvalho de Melo19449722005-06-18 22:46:19 -07003803{
3804 return cachep->name;
3805}
3806EXPORT_SYMBOL_GPL(kmem_cache_name);
3807
Christoph Lametere498be72005-09-09 13:03:32 -07003808/*
Simon Arlott183ff222007-10-20 01:27:18 +02003809 * This initializes kmem_list3 or resizes various caches for all nodes.
Christoph Lametere498be72005-09-09 13:03:32 -07003810 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003811static int alloc_kmemlist(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07003812{
3813 int node;
3814 struct kmem_list3 *l3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003815 struct array_cache *new_shared;
Paul Menage3395ee02006-12-06 20:32:16 -08003816 struct array_cache **new_alien = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003817
Mel Gorman9c09a952008-01-24 05:49:54 -08003818 for_each_online_node(node) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003819
Paul Menage3395ee02006-12-06 20:32:16 -08003820 if (use_alien_caches) {
3821 new_alien = alloc_alien_cache(node, cachep->limit);
3822 if (!new_alien)
3823 goto fail;
3824 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003825
Eric Dumazet63109842007-05-06 14:49:28 -07003826 new_shared = NULL;
3827 if (cachep->shared) {
3828 new_shared = alloc_arraycache(node,
Christoph Lameter0718dc22006-03-25 03:06:47 -08003829 cachep->shared*cachep->batchcount,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003830 0xbaadf00d);
Eric Dumazet63109842007-05-06 14:49:28 -07003831 if (!new_shared) {
3832 free_alien_cache(new_alien);
3833 goto fail;
3834 }
Christoph Lameter0718dc22006-03-25 03:06:47 -08003835 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003836
Andrew Mortona737b3e2006-03-22 00:08:11 -08003837 l3 = cachep->nodelists[node];
3838 if (l3) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003839 struct array_cache *shared = l3->shared;
3840
Christoph Lametere498be72005-09-09 13:03:32 -07003841 spin_lock_irq(&l3->list_lock);
3842
Christoph Lametercafeb022006-03-25 03:06:46 -08003843 if (shared)
Christoph Lameter0718dc22006-03-25 03:06:47 -08003844 free_block(cachep, shared->entry,
3845 shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07003846
Christoph Lametercafeb022006-03-25 03:06:46 -08003847 l3->shared = new_shared;
3848 if (!l3->alien) {
Christoph Lametere498be72005-09-09 13:03:32 -07003849 l3->alien = new_alien;
3850 new_alien = NULL;
3851 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003852 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003853 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003854 spin_unlock_irq(&l3->list_lock);
Christoph Lametercafeb022006-03-25 03:06:46 -08003855 kfree(shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003856 free_alien_cache(new_alien);
3857 continue;
3858 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08003859 l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node);
Christoph Lameter0718dc22006-03-25 03:06:47 -08003860 if (!l3) {
3861 free_alien_cache(new_alien);
3862 kfree(new_shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003863 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003864 }
Christoph Lametere498be72005-09-09 13:03:32 -07003865
3866 kmem_list3_init(l3);
3867 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Andrew Mortona737b3e2006-03-22 00:08:11 -08003868 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003869 l3->shared = new_shared;
Christoph Lametere498be72005-09-09 13:03:32 -07003870 l3->alien = new_alien;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003871 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003872 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003873 cachep->nodelists[node] = l3;
3874 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003875 return 0;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003876
Andrew Mortona737b3e2006-03-22 00:08:11 -08003877fail:
Christoph Lameter0718dc22006-03-25 03:06:47 -08003878 if (!cachep->next.next) {
3879 /* Cache is not active yet. Roll back what we did */
3880 node--;
3881 while (node >= 0) {
3882 if (cachep->nodelists[node]) {
3883 l3 = cachep->nodelists[node];
3884
3885 kfree(l3->shared);
3886 free_alien_cache(l3->alien);
3887 kfree(l3);
3888 cachep->nodelists[node] = NULL;
3889 }
3890 node--;
3891 }
3892 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003893 return -ENOMEM;
Christoph Lametere498be72005-09-09 13:03:32 -07003894}
3895
Linus Torvalds1da177e2005-04-16 15:20:36 -07003896struct ccupdate_struct {
Pekka Enberg343e0d72006-02-01 03:05:50 -08003897 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003898 struct array_cache *new[NR_CPUS];
3899};
3900
3901static void do_ccupdate_local(void *info)
3902{
Andrew Mortona737b3e2006-03-22 00:08:11 -08003903 struct ccupdate_struct *new = info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003904 struct array_cache *old;
3905
3906 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003907 old = cpu_cache_get(new->cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003908
Linus Torvalds1da177e2005-04-16 15:20:36 -07003909 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3910 new->new[smp_processor_id()] = old;
3911}
3912
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003913/* Always called with the cache_chain_mutex held */
Andrew Mortona737b3e2006-03-22 00:08:11 -08003914static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3915 int batchcount, int shared)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003916{
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003917 struct ccupdate_struct *new;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003918 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003919
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003920 new = kzalloc(sizeof(*new), GFP_KERNEL);
3921 if (!new)
3922 return -ENOMEM;
3923
Christoph Lametere498be72005-09-09 13:03:32 -07003924 for_each_online_cpu(i) {
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003925 new->new[i] = alloc_arraycache(cpu_to_node(i), limit,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003926 batchcount);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003927 if (!new->new[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003928 for (i--; i >= 0; i--)
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003929 kfree(new->new[i]);
3930 kfree(new);
Christoph Lametere498be72005-09-09 13:03:32 -07003931 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003932 }
3933 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003934 new->cachep = cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003935
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003936 on_each_cpu(do_ccupdate_local, (void *)new, 1, 1);
Christoph Lametere498be72005-09-09 13:03:32 -07003937
Linus Torvalds1da177e2005-04-16 15:20:36 -07003938 check_irq_on();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003939 cachep->batchcount = batchcount;
3940 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07003941 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003942
Christoph Lametere498be72005-09-09 13:03:32 -07003943 for_each_online_cpu(i) {
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003944 struct array_cache *ccold = new->new[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003945 if (!ccold)
3946 continue;
Christoph Lametere498be72005-09-09 13:03:32 -07003947 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -07003948 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
Christoph Lametere498be72005-09-09 13:03:32 -07003949 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003950 kfree(ccold);
3951 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003952 kfree(new);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003953 return alloc_kmemlist(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003954}
3955
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003956/* Called with cache_chain_mutex held always */
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003957static int enable_cpucache(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003958{
3959 int err;
3960 int limit, shared;
3961
Andrew Mortona737b3e2006-03-22 00:08:11 -08003962 /*
3963 * The head array serves three purposes:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003964 * - create a LIFO ordering, i.e. return objects that are cache-warm
3965 * - reduce the number of spinlock operations.
Andrew Mortona737b3e2006-03-22 00:08:11 -08003966 * - reduce the number of linked list operations on the slab and
Linus Torvalds1da177e2005-04-16 15:20:36 -07003967 * bufctl chains: array operations are cheaper.
3968 * The numbers are guessed, we should auto-tune as described by
3969 * Bonwick.
3970 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003971 if (cachep->buffer_size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003972 limit = 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003973 else if (cachep->buffer_size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003974 limit = 8;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003975 else if (cachep->buffer_size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003976 limit = 24;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003977 else if (cachep->buffer_size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003978 limit = 54;
3979 else
3980 limit = 120;
3981
Andrew Mortona737b3e2006-03-22 00:08:11 -08003982 /*
3983 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
Linus Torvalds1da177e2005-04-16 15:20:36 -07003984 * allocation behaviour: Most allocs on one cpu, most free operations
3985 * on another cpu. For these cases, an efficient object passing between
3986 * cpus is necessary. This is provided by a shared array. The array
3987 * replaces Bonwick's magazine layer.
3988 * On uniprocessor, it's functionally equivalent (but less efficient)
3989 * to a larger limit. Thus disabled by default.
3990 */
3991 shared = 0;
Eric Dumazet364fbb22007-05-06 14:49:27 -07003992 if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003993 shared = 8;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003994
3995#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003996 /*
3997 * With debugging enabled, large batchcount lead to excessively long
3998 * periods with disabled local interrupts. Limit the batchcount
Linus Torvalds1da177e2005-04-16 15:20:36 -07003999 */
4000 if (limit > 32)
4001 limit = 32;
4002#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004003 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004004 if (err)
4005 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004006 cachep->name, -err);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07004007 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004008}
4009
Christoph Lameter1b552532006-03-22 00:09:07 -08004010/*
4011 * Drain an array if it contains any elements taking the l3 lock only if
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004012 * necessary. Note that the l3 listlock also protects the array_cache
4013 * if drain_array() is used on the shared array.
Christoph Lameter1b552532006-03-22 00:09:07 -08004014 */
4015void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
4016 struct array_cache *ac, int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004017{
4018 int tofree;
4019
Christoph Lameter1b552532006-03-22 00:09:07 -08004020 if (!ac || !ac->avail)
4021 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004022 if (ac->touched && !force) {
4023 ac->touched = 0;
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004024 } else {
Christoph Lameter1b552532006-03-22 00:09:07 -08004025 spin_lock_irq(&l3->list_lock);
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004026 if (ac->avail) {
4027 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4028 if (tofree > ac->avail)
4029 tofree = (ac->avail + 1) / 2;
4030 free_block(cachep, ac->entry, tofree, node);
4031 ac->avail -= tofree;
4032 memmove(ac->entry, &(ac->entry[tofree]),
4033 sizeof(void *) * ac->avail);
4034 }
Christoph Lameter1b552532006-03-22 00:09:07 -08004035 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004036 }
4037}
4038
4039/**
4040 * cache_reap - Reclaim memory from caches.
Randy Dunlap05fb6bf2007-02-28 20:12:13 -08004041 * @w: work descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07004042 *
4043 * Called from workqueue/eventd every few seconds.
4044 * Purpose:
4045 * - clear the per-cpu caches for this CPU.
4046 * - return freeable pages to the main free memory pool.
4047 *
Andrew Mortona737b3e2006-03-22 00:08:11 -08004048 * If we cannot acquire the cache chain mutex then just give up - we'll try
4049 * again on the next iteration.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004050 */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004051static void cache_reap(struct work_struct *w)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004052{
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004053 struct kmem_cache *searchp;
Christoph Lametere498be72005-09-09 13:03:32 -07004054 struct kmem_list3 *l3;
Christoph Lameteraab22072006-03-22 00:09:06 -08004055 int node = numa_node_id();
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004056 struct delayed_work *work =
4057 container_of(w, struct delayed_work, work);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004058
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004059 if (!mutex_trylock(&cache_chain_mutex))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004060 /* Give up. Setup the next iteration. */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004061 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004062
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004063 list_for_each_entry(searchp, &cache_chain, next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004064 check_irq_on();
4065
Christoph Lameter35386e32006-03-22 00:09:05 -08004066 /*
4067 * We only take the l3 lock if absolutely necessary and we
4068 * have established with reasonable certainty that
4069 * we can do some work if the lock was obtained.
4070 */
Christoph Lameteraab22072006-03-22 00:09:06 -08004071 l3 = searchp->nodelists[node];
Christoph Lameter35386e32006-03-22 00:09:05 -08004072
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004073 reap_alien(searchp, l3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004074
Christoph Lameteraab22072006-03-22 00:09:06 -08004075 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004076
Christoph Lameter35386e32006-03-22 00:09:05 -08004077 /*
4078 * These are racy checks but it does not matter
4079 * if we skip one check or scan twice.
4080 */
Christoph Lametere498be72005-09-09 13:03:32 -07004081 if (time_after(l3->next_reap, jiffies))
Christoph Lameter35386e32006-03-22 00:09:05 -08004082 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004083
Christoph Lametere498be72005-09-09 13:03:32 -07004084 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004085
Christoph Lameteraab22072006-03-22 00:09:06 -08004086 drain_array(searchp, l3, l3->shared, 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004087
Christoph Lametered11d9e2006-06-30 01:55:45 -07004088 if (l3->free_touched)
Christoph Lametere498be72005-09-09 13:03:32 -07004089 l3->free_touched = 0;
Christoph Lametered11d9e2006-06-30 01:55:45 -07004090 else {
4091 int freed;
4092
4093 freed = drain_freelist(searchp, l3, (l3->free_limit +
4094 5 * searchp->num - 1) / (5 * searchp->num));
4095 STATS_ADD_REAPED(searchp, freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004096 }
Christoph Lameter35386e32006-03-22 00:09:05 -08004097next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004098 cond_resched();
4099 }
4100 check_irq_on();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004101 mutex_unlock(&cache_chain_mutex);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004102 next_reap_node();
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004103out:
Andrew Mortona737b3e2006-03-22 00:08:11 -08004104 /* Set up the next iteration */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004105 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004106}
4107
Linus Torvalds158a9622008-01-02 13:04:48 -08004108#ifdef CONFIG_SLABINFO
Linus Torvalds1da177e2005-04-16 15:20:36 -07004109
Pekka Enberg85289f92006-01-08 01:00:36 -08004110static void print_slabinfo_header(struct seq_file *m)
4111{
4112 /*
4113 * Output format version, so at least we can change it
4114 * without _too_ many complaints.
4115 */
4116#if STATS
4117 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4118#else
4119 seq_puts(m, "slabinfo - version: 2.1\n");
4120#endif
4121 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4122 "<objperslab> <pagesperslab>");
4123 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4124 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4125#if STATS
4126 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004127 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
Pekka Enberg85289f92006-01-08 01:00:36 -08004128 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4129#endif
4130 seq_putc(m, '\n');
4131}
4132
Linus Torvalds1da177e2005-04-16 15:20:36 -07004133static void *s_start(struct seq_file *m, loff_t *pos)
4134{
4135 loff_t n = *pos;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004136
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004137 mutex_lock(&cache_chain_mutex);
Pekka Enberg85289f92006-01-08 01:00:36 -08004138 if (!n)
4139 print_slabinfo_header(m);
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004140
4141 return seq_list_start(&cache_chain, *pos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004142}
4143
4144static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4145{
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004146 return seq_list_next(p, &cache_chain, pos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004147}
4148
4149static void s_stop(struct seq_file *m, void *p)
4150{
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004151 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004152}
4153
4154static int s_show(struct seq_file *m, void *p)
4155{
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004156 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004157 struct slab *slabp;
4158 unsigned long active_objs;
4159 unsigned long num_objs;
4160 unsigned long active_slabs = 0;
4161 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004162 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004163 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07004164 int node;
4165 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004166
Linus Torvalds1da177e2005-04-16 15:20:36 -07004167 active_objs = 0;
4168 num_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004169 for_each_online_node(node) {
4170 l3 = cachep->nodelists[node];
4171 if (!l3)
4172 continue;
4173
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004174 check_irq_on();
4175 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07004176
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004177 list_for_each_entry(slabp, &l3->slabs_full, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004178 if (slabp->inuse != cachep->num && !error)
4179 error = "slabs_full accounting error";
4180 active_objs += cachep->num;
4181 active_slabs++;
4182 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004183 list_for_each_entry(slabp, &l3->slabs_partial, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004184 if (slabp->inuse == cachep->num && !error)
4185 error = "slabs_partial inuse accounting error";
4186 if (!slabp->inuse && !error)
4187 error = "slabs_partial/inuse accounting error";
4188 active_objs += slabp->inuse;
4189 active_slabs++;
4190 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004191 list_for_each_entry(slabp, &l3->slabs_free, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004192 if (slabp->inuse && !error)
4193 error = "slabs_free/inuse accounting error";
4194 num_slabs++;
4195 }
4196 free_objects += l3->free_objects;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08004197 if (l3->shared)
4198 shared_avail += l3->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07004199
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004200 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004201 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004202 num_slabs += active_slabs;
4203 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07004204 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004205 error = "free_objects accounting error";
4206
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004207 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004208 if (error)
4209 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4210
4211 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004212 name, active_objs, num_objs, cachep->buffer_size,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004213 cachep->num, (1 << cachep->gfporder));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004214 seq_printf(m, " : tunables %4u %4u %4u",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004215 cachep->limit, cachep->batchcount, cachep->shared);
Christoph Lametere498be72005-09-09 13:03:32 -07004216 seq_printf(m, " : slabdata %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004217 active_slabs, num_slabs, shared_avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004218#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004219 { /* list3 stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004220 unsigned long high = cachep->high_mark;
4221 unsigned long allocs = cachep->num_allocations;
4222 unsigned long grown = cachep->grown;
4223 unsigned long reaped = cachep->reaped;
4224 unsigned long errors = cachep->errors;
4225 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004226 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07004227 unsigned long node_frees = cachep->node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004228 unsigned long overflows = cachep->node_overflow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004229
Christoph Lametere498be72005-09-09 13:03:32 -07004230 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004231 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown,
Andrew Mortona737b3e2006-03-22 00:08:11 -08004232 reaped, errors, max_freeable, node_allocs,
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004233 node_frees, overflows);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004234 }
4235 /* cpu stats */
4236 {
4237 unsigned long allochit = atomic_read(&cachep->allochit);
4238 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4239 unsigned long freehit = atomic_read(&cachep->freehit);
4240 unsigned long freemiss = atomic_read(&cachep->freemiss);
4241
4242 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004243 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004244 }
4245#endif
4246 seq_putc(m, '\n');
Linus Torvalds1da177e2005-04-16 15:20:36 -07004247 return 0;
4248}
4249
4250/*
4251 * slabinfo_op - iterator that generates /proc/slabinfo
4252 *
4253 * Output layout:
4254 * cache-name
4255 * num-active-objs
4256 * total-objs
4257 * object size
4258 * num-active-slabs
4259 * total-slabs
4260 * num-pages-per-slab
4261 * + further values on SMP and with statistics enabled
4262 */
4263
Helge Deller15ad7cd2006-12-06 20:40:36 -08004264const struct seq_operations slabinfo_op = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004265 .start = s_start,
4266 .next = s_next,
4267 .stop = s_stop,
4268 .show = s_show,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004269};
4270
4271#define MAX_SLABINFO_WRITE 128
4272/**
4273 * slabinfo_write - Tuning for the slab allocator
4274 * @file: unused
4275 * @buffer: user buffer
4276 * @count: data length
4277 * @ppos: unused
4278 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004279ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4280 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004281{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004282 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004283 int limit, batchcount, shared, res;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004284 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004285
Linus Torvalds1da177e2005-04-16 15:20:36 -07004286 if (count > MAX_SLABINFO_WRITE)
4287 return -EINVAL;
4288 if (copy_from_user(&kbuf, buffer, count))
4289 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004290 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07004291
4292 tmp = strchr(kbuf, ' ');
4293 if (!tmp)
4294 return -EINVAL;
4295 *tmp = '\0';
4296 tmp++;
4297 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4298 return -EINVAL;
4299
4300 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004301 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004302 res = -EINVAL;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004303 list_for_each_entry(cachep, &cache_chain, next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004304 if (!strcmp(cachep->name, kbuf)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08004305 if (limit < 1 || batchcount < 1 ||
4306 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07004307 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004308 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07004309 res = do_tune_cpucache(cachep, limit,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004310 batchcount, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004311 }
4312 break;
4313 }
4314 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004315 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004316 if (res >= 0)
4317 res = count;
4318 return res;
4319}
Al Viro871751e2006-03-25 03:06:39 -08004320
4321#ifdef CONFIG_DEBUG_SLAB_LEAK
4322
4323static void *leaks_start(struct seq_file *m, loff_t *pos)
4324{
Al Viro871751e2006-03-25 03:06:39 -08004325 mutex_lock(&cache_chain_mutex);
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004326 return seq_list_start(&cache_chain, *pos);
Al Viro871751e2006-03-25 03:06:39 -08004327}
4328
4329static inline int add_caller(unsigned long *n, unsigned long v)
4330{
4331 unsigned long *p;
4332 int l;
4333 if (!v)
4334 return 1;
4335 l = n[1];
4336 p = n + 2;
4337 while (l) {
4338 int i = l/2;
4339 unsigned long *q = p + 2 * i;
4340 if (*q == v) {
4341 q[1]++;
4342 return 1;
4343 }
4344 if (*q > v) {
4345 l = i;
4346 } else {
4347 p = q + 2;
4348 l -= i + 1;
4349 }
4350 }
4351 if (++n[1] == n[0])
4352 return 0;
4353 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4354 p[0] = v;
4355 p[1] = 1;
4356 return 1;
4357}
4358
4359static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4360{
4361 void *p;
4362 int i;
4363 if (n[0] == n[1])
4364 return;
4365 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4366 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4367 continue;
4368 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4369 return;
4370 }
4371}
4372
4373static void show_symbol(struct seq_file *m, unsigned long address)
4374{
4375#ifdef CONFIG_KALLSYMS
Al Viro871751e2006-03-25 03:06:39 -08004376 unsigned long offset, size;
Tejun Heo9281ace2007-07-17 04:03:51 -07004377 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
Al Viro871751e2006-03-25 03:06:39 -08004378
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004379 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
Al Viro871751e2006-03-25 03:06:39 -08004380 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004381 if (modname[0])
Al Viro871751e2006-03-25 03:06:39 -08004382 seq_printf(m, " [%s]", modname);
4383 return;
4384 }
4385#endif
4386 seq_printf(m, "%p", (void *)address);
4387}
4388
4389static int leaks_show(struct seq_file *m, void *p)
4390{
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004391 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
Al Viro871751e2006-03-25 03:06:39 -08004392 struct slab *slabp;
4393 struct kmem_list3 *l3;
4394 const char *name;
4395 unsigned long *n = m->private;
4396 int node;
4397 int i;
4398
4399 if (!(cachep->flags & SLAB_STORE_USER))
4400 return 0;
4401 if (!(cachep->flags & SLAB_RED_ZONE))
4402 return 0;
4403
4404 /* OK, we can do it */
4405
4406 n[1] = 0;
4407
4408 for_each_online_node(node) {
4409 l3 = cachep->nodelists[node];
4410 if (!l3)
4411 continue;
4412
4413 check_irq_on();
4414 spin_lock_irq(&l3->list_lock);
4415
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004416 list_for_each_entry(slabp, &l3->slabs_full, list)
Al Viro871751e2006-03-25 03:06:39 -08004417 handle_slab(n, cachep, slabp);
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004418 list_for_each_entry(slabp, &l3->slabs_partial, list)
Al Viro871751e2006-03-25 03:06:39 -08004419 handle_slab(n, cachep, slabp);
Al Viro871751e2006-03-25 03:06:39 -08004420 spin_unlock_irq(&l3->list_lock);
4421 }
4422 name = cachep->name;
4423 if (n[0] == n[1]) {
4424 /* Increase the buffer size */
4425 mutex_unlock(&cache_chain_mutex);
4426 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4427 if (!m->private) {
4428 /* Too bad, we are really out */
4429 m->private = n;
4430 mutex_lock(&cache_chain_mutex);
4431 return -ENOMEM;
4432 }
4433 *(unsigned long *)m->private = n[0] * 2;
4434 kfree(n);
4435 mutex_lock(&cache_chain_mutex);
4436 /* Now make sure this entry will be retried */
4437 m->count = m->size;
4438 return 0;
4439 }
4440 for (i = 0; i < n[1]; i++) {
4441 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4442 show_symbol(m, n[2*i+2]);
4443 seq_putc(m, '\n');
4444 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004445
Al Viro871751e2006-03-25 03:06:39 -08004446 return 0;
4447}
4448
Helge Deller15ad7cd2006-12-06 20:40:36 -08004449const struct seq_operations slabstats_op = {
Al Viro871751e2006-03-25 03:06:39 -08004450 .start = leaks_start,
4451 .next = s_next,
4452 .stop = s_stop,
4453 .show = leaks_show,
4454};
4455#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004456#endif
4457
Manfred Spraul00e145b2005-09-03 15:55:07 -07004458/**
4459 * ksize - get the actual amount of memory allocated for a given object
4460 * @objp: Pointer to the object
4461 *
4462 * kmalloc may internally round up allocations and return more memory
4463 * than requested. ksize() can be used to determine the actual amount of
4464 * memory allocated. The caller may use this additional memory, even though
4465 * a smaller amount of memory was initially specified with the kmalloc call.
4466 * The caller must guarantee that objp points to a valid object previously
4467 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4468 * must not be freed during the duration of the call.
4469 */
Pekka Enbergfd76bab2007-05-06 14:48:40 -07004470size_t ksize(const void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004471{
Christoph Lameteref8b4522007-10-16 01:24:46 -07004472 BUG_ON(!objp);
4473 if (unlikely(objp == ZERO_SIZE_PTR))
Manfred Spraul00e145b2005-09-03 15:55:07 -07004474 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004475
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08004476 return obj_size(virt_to_cache(objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004477}
Tetsuo Handaf8fcc932007-12-04 23:45:08 -08004478EXPORT_SYMBOL(ksize);