blob: 9a90b00d2f9140e2aaa67ecb949e33d89431e9c5 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * linux/mm/slab.c
3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
5 *
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
7 *
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
10 *
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
13 *
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
21 *
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
27 *
28 * This means, that your constructor is used only for newly allocated
Simon Arlott183ff222007-10-20 01:27:18 +020029 * slabs and you must pass objects with the same initializations to
Linus Torvalds1da177e2005-04-16 15:20:36 -070030 * kmem_cache_free.
31 *
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
35 *
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
38 * partial slabs
39 * empty slabs with no allocated objects
40 *
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
43 *
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
46 *
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
52 *
Andrew Mortona737b3e2006-03-22 00:08:11 -080053 * The c_cpuarray may not be read with enabled local interrupts -
Linus Torvalds1da177e2005-04-16 15:20:36 -070054 * it's changed with a smp_call_function().
55 *
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
Pekka Enberg343e0d72006-02-01 03:05:50 -080058 * Several members in struct kmem_cache and struct slab never change, they
Linus Torvalds1da177e2005-04-16 15:20:36 -070059 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
63 *
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
66 * his patch.
67 *
68 * Further notes from the original documentation:
69 *
70 * 11 April '97. Started multi-threading - markhe
Ingo Molnarfc0abb12006-01-18 17:42:33 -080071 * The global cache-chain is protected by the mutex 'cache_chain_mutex'.
Linus Torvalds1da177e2005-04-16 15:20:36 -070072 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
75 *
76 * At present, each engine can be growing a cache. This should be blocked.
77 *
Christoph Lametere498be72005-09-09 13:03:32 -070078 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
83 *
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
Linus Torvalds1da177e2005-04-16 15:20:36 -070087 */
88
Linus Torvalds1da177e2005-04-16 15:20:36 -070089#include <linux/slab.h>
90#include <linux/mm.h>
Randy Dunlapc9cf5522006-06-27 02:53:52 -070091#include <linux/poison.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070092#include <linux/swap.h>
93#include <linux/cache.h>
94#include <linux/interrupt.h>
95#include <linux/init.h>
96#include <linux/compiler.h>
Paul Jackson101a5002006-03-24 03:16:07 -080097#include <linux/cpuset.h>
Alexey Dobriyana0ec95a82008-10-06 00:59:10 +040098#include <linux/proc_fs.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070099#include <linux/seq_file.h>
100#include <linux/notifier.h>
101#include <linux/kallsyms.h>
102#include <linux/cpu.h>
103#include <linux/sysctl.h>
104#include <linux/module.h>
Ingo Molnarf09eac92008-12-31 09:43:46 +0100105#include <trace/kmemtrace.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700106#include <linux/rcupdate.h>
Paulo Marques543537b2005-06-23 00:09:02 -0700107#include <linux/string.h>
Andrew Morton138ae662006-12-06 20:36:41 -0800108#include <linux/uaccess.h>
Christoph Lametere498be72005-09-09 13:03:32 -0700109#include <linux/nodemask.h>
Christoph Lameterdc85da12006-01-18 17:42:36 -0800110#include <linux/mempolicy.h>
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800111#include <linux/mutex.h>
Akinobu Mita8a8b6502006-12-08 02:39:44 -0800112#include <linux/fault-inject.h>
Ingo Molnare7eebaf2006-06-27 02:54:55 -0700113#include <linux/rtmutex.h>
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800114#include <linux/reciprocal_div.h>
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -0700115#include <linux/debugobjects.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700116
Linus Torvalds1da177e2005-04-16 15:20:36 -0700117#include <asm/cacheflush.h>
118#include <asm/tlbflush.h>
119#include <asm/page.h>
120
121/*
Christoph Lameter50953fe2007-05-06 14:50:16 -0700122 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700123 * 0 for faster, smaller code (especially in the critical paths).
124 *
125 * STATS - 1 to collect stats for /proc/slabinfo.
126 * 0 for faster, smaller code (especially in the critical paths).
127 *
128 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
129 */
130
131#ifdef CONFIG_DEBUG_SLAB
132#define DEBUG 1
133#define STATS 1
134#define FORCED_DEBUG 1
135#else
136#define DEBUG 0
137#define STATS 0
138#define FORCED_DEBUG 0
139#endif
140
Linus Torvalds1da177e2005-04-16 15:20:36 -0700141/* Shouldn't this be in a header file somewhere? */
142#define BYTES_PER_WORD sizeof(void *)
David Woodhouse87a927c2007-07-04 21:26:44 -0400143#define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700144
Linus Torvalds1da177e2005-04-16 15:20:36 -0700145#ifndef ARCH_KMALLOC_MINALIGN
146/*
147 * Enforce a minimum alignment for the kmalloc caches.
148 * Usually, the kmalloc caches are cache_line_size() aligned, except when
149 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
150 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
David Woodhouseb46b8f12007-05-08 00:22:59 -0700151 * alignment larger than the alignment of a 64-bit integer.
152 * ARCH_KMALLOC_MINALIGN allows that.
153 * Note that increasing this value may disable some debug features.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700154 */
David Woodhouseb46b8f12007-05-08 00:22:59 -0700155#define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700156#endif
157
158#ifndef ARCH_SLAB_MINALIGN
159/*
160 * Enforce a minimum alignment for all caches.
161 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
162 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
163 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
164 * some debug features.
165 */
166#define ARCH_SLAB_MINALIGN 0
167#endif
168
169#ifndef ARCH_KMALLOC_FLAGS
170#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
171#endif
172
173/* Legal flag mask for kmem_cache_create(). */
174#if DEBUG
Christoph Lameter50953fe2007-05-06 14:50:16 -0700175# define CREATE_MASK (SLAB_RED_ZONE | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700176 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
Christoph Lameterac2b8982006-03-22 00:08:15 -0800177 SLAB_CACHE_DMA | \
Christoph Lameter5af60832007-05-06 14:49:56 -0700178 SLAB_STORE_USER | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700179 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -0700180 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
181 SLAB_DEBUG_OBJECTS)
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 | \
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -0700186 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
187 SLAB_DEBUG_OBJECTS)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700188#endif
189
190/*
191 * kmem_bufctl_t:
192 *
193 * Bufctl's are used for linking objs within a slab
194 * linked offsets.
195 *
196 * This implementation relies on "struct page" for locating the cache &
197 * slab an object belongs to.
198 * This allows the bufctl structure to be small (one int), but limits
199 * the number of objects a slab (not a cache) can contain when off-slab
200 * bufctls are used. The limit is the size of the largest general cache
201 * that does not use off-slab slabs.
202 * For 32bit archs with 4 kB pages, is this 56.
203 * This is not serious, as it is only for large objects, when it is unwise
204 * to have too many per slab.
205 * Note: This limit can be raised by introducing a general cache whose size
206 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
207 */
208
Kyle Moffettfa5b08d2005-09-03 15:55:03 -0700209typedef unsigned int kmem_bufctl_t;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700210#define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
211#define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
Al Viro871751e2006-03-25 03:06:39 -0800212#define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
213#define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700214
Linus Torvalds1da177e2005-04-16 15:20:36 -0700215/*
216 * struct slab
217 *
218 * Manages the objs in a slab. Placed either at the beginning of mem allocated
219 * for a slab, or allocated from an general cache.
220 * Slabs are chained into three list: fully used, partial, fully free slabs.
221 */
222struct slab {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800223 struct list_head list;
224 unsigned long colouroff;
225 void *s_mem; /* including colour offset */
226 unsigned int inuse; /* num of objs active in slab */
227 kmem_bufctl_t free;
228 unsigned short nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700229};
230
231/*
232 * struct slab_rcu
233 *
234 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
235 * arrange for kmem_freepages to be called via RCU. This is useful if
236 * we need to approach a kernel structure obliquely, from its address
237 * obtained without the usual locking. We can lock the structure to
238 * stabilize it and check it's still at the given address, only if we
239 * can be sure that the memory has not been meanwhile reused for some
240 * other kind of object (which our subsystem's lock might corrupt).
241 *
242 * rcu_read_lock before reading the address, then rcu_read_unlock after
243 * taking the spinlock within the structure expected at that address.
244 *
245 * We assume struct slab_rcu can overlay struct slab when destroying.
246 */
247struct slab_rcu {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800248 struct rcu_head head;
Pekka Enberg343e0d72006-02-01 03:05:50 -0800249 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800250 void *addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700251};
252
253/*
254 * struct array_cache
255 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700256 * Purpose:
257 * - LIFO ordering, to hand out cache-warm objects from _alloc
258 * - reduce the number of linked list operations
259 * - reduce spinlock operations
260 *
261 * The limit is stored in the per-cpu structure to reduce the data cache
262 * footprint.
263 *
264 */
265struct array_cache {
266 unsigned int avail;
267 unsigned int limit;
268 unsigned int batchcount;
269 unsigned int touched;
Christoph Lametere498be72005-09-09 13:03:32 -0700270 spinlock_t lock;
Robert P. J. Daybda5b652007-10-16 23:30:05 -0700271 void *entry[]; /*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800272 * Must have this definition in here for the proper
273 * alignment of array_cache. Also simplifies accessing
274 * the entries.
Andrew Mortona737b3e2006-03-22 00:08:11 -0800275 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700276};
277
Andrew Mortona737b3e2006-03-22 00:08:11 -0800278/*
279 * bootstrap: The caches do not work without cpuarrays anymore, but the
280 * cpuarrays are allocated from the generic caches...
Linus Torvalds1da177e2005-04-16 15:20:36 -0700281 */
282#define BOOT_CPUCACHE_ENTRIES 1
283struct arraycache_init {
284 struct array_cache cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800285 void *entries[BOOT_CPUCACHE_ENTRIES];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700286};
287
288/*
Christoph Lametere498be72005-09-09 13:03:32 -0700289 * The slab lists for all objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700290 */
291struct kmem_list3 {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800292 struct list_head slabs_partial; /* partial list first, better asm code */
293 struct list_head slabs_full;
294 struct list_head slabs_free;
295 unsigned long free_objects;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800296 unsigned int free_limit;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800297 unsigned int colour_next; /* Per-node cache coloring */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800298 spinlock_t list_lock;
299 struct array_cache *shared; /* shared per node */
300 struct array_cache **alien; /* on other nodes */
Christoph Lameter35386e32006-03-22 00:09:05 -0800301 unsigned long next_reap; /* updated without locking */
302 int free_touched; /* updated without locking */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700303};
304
Christoph Lametere498be72005-09-09 13:03:32 -0700305/*
306 * Need this for bootstrapping a per node allocator.
307 */
Pekka Enberg556a1692008-01-25 08:20:51 +0200308#define NUM_INIT_LISTS (3 * MAX_NUMNODES)
Christoph Lametere498be72005-09-09 13:03:32 -0700309struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
310#define CACHE_CACHE 0
Pekka Enberg556a1692008-01-25 08:20:51 +0200311#define SIZE_AC MAX_NUMNODES
312#define SIZE_L3 (2 * MAX_NUMNODES)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700313
Christoph Lametered11d9e2006-06-30 01:55:45 -0700314static int drain_freelist(struct kmem_cache *cache,
315 struct kmem_list3 *l3, int tofree);
316static void free_block(struct kmem_cache *cachep, void **objpp, int len,
317 int node);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -0700318static int enable_cpucache(struct kmem_cache *cachep);
David Howells65f27f32006-11-22 14:55:48 +0000319static void cache_reap(struct work_struct *unused);
Christoph Lametered11d9e2006-06-30 01:55:45 -0700320
Christoph Lametere498be72005-09-09 13:03:32 -0700321/*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800322 * This function must be completely optimized away if a constant is passed to
323 * it. Mostly the same as what is in linux/slab.h except it returns an index.
Christoph Lametere498be72005-09-09 13:03:32 -0700324 */
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700325static __always_inline int index_of(const size_t size)
Christoph Lametere498be72005-09-09 13:03:32 -0700326{
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800327 extern void __bad_size(void);
328
Christoph Lametere498be72005-09-09 13:03:32 -0700329 if (__builtin_constant_p(size)) {
330 int i = 0;
331
332#define CACHE(x) \
333 if (size <=x) \
334 return i; \
335 else \
336 i++;
Joe Perches1c61fc42008-03-05 13:58:17 -0800337#include <linux/kmalloc_sizes.h>
Christoph Lametere498be72005-09-09 13:03:32 -0700338#undef CACHE
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800339 __bad_size();
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700340 } else
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800341 __bad_size();
Christoph Lametere498be72005-09-09 13:03:32 -0700342 return 0;
343}
344
Ingo Molnare0a42722006-06-23 02:03:46 -0700345static int slab_early_init = 1;
346
Christoph Lametere498be72005-09-09 13:03:32 -0700347#define INDEX_AC index_of(sizeof(struct arraycache_init))
348#define INDEX_L3 index_of(sizeof(struct kmem_list3))
349
Pekka Enberg5295a742006-02-01 03:05:48 -0800350static void kmem_list3_init(struct kmem_list3 *parent)
Christoph Lametere498be72005-09-09 13:03:32 -0700351{
352 INIT_LIST_HEAD(&parent->slabs_full);
353 INIT_LIST_HEAD(&parent->slabs_partial);
354 INIT_LIST_HEAD(&parent->slabs_free);
355 parent->shared = NULL;
356 parent->alien = NULL;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800357 parent->colour_next = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700358 spin_lock_init(&parent->list_lock);
359 parent->free_objects = 0;
360 parent->free_touched = 0;
361}
362
Andrew Mortona737b3e2006-03-22 00:08:11 -0800363#define MAKE_LIST(cachep, listp, slab, nodeid) \
364 do { \
365 INIT_LIST_HEAD(listp); \
366 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
Christoph Lametere498be72005-09-09 13:03:32 -0700367 } while (0)
368
Andrew Mortona737b3e2006-03-22 00:08:11 -0800369#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
370 do { \
Christoph Lametere498be72005-09-09 13:03:32 -0700371 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
372 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
373 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
374 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700375
376/*
Pekka Enberg343e0d72006-02-01 03:05:50 -0800377 * struct kmem_cache
Linus Torvalds1da177e2005-04-16 15:20:36 -0700378 *
379 * manages a cache.
380 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800381
Pekka J Enberg2109a2d2005-11-07 00:58:01 -0800382struct kmem_cache {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700383/* 1) per-cpu data, touched during every alloc/free */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800384 struct array_cache *array[NR_CPUS];
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800385/* 2) Cache tunables. Protected by cache_chain_mutex */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800386 unsigned int batchcount;
387 unsigned int limit;
388 unsigned int shared;
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800389
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800390 unsigned int buffer_size;
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800391 u32 reciprocal_buffer_size;
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800392/* 3) touched by every alloc & free from the backend */
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800393
Andrew Mortona737b3e2006-03-22 00:08:11 -0800394 unsigned int flags; /* constant flags */
395 unsigned int num; /* # of objs per slab */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700396
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800397/* 4) cache_grow/shrink */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700398 /* order of pgs per slab (2^n) */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800399 unsigned int gfporder;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700400
401 /* force GFP flags, e.g. GFP_DMA */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800402 gfp_t gfpflags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700403
Andrew Mortona737b3e2006-03-22 00:08:11 -0800404 size_t colour; /* cache colouring range */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800405 unsigned int colour_off; /* colour offset */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800406 struct kmem_cache *slabp_cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800407 unsigned int slab_size;
Andrew Mortona737b3e2006-03-22 00:08:11 -0800408 unsigned int dflags; /* dynamic flags */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700409
410 /* constructor func */
Alexey Dobriyan51cc5062008-07-25 19:45:34 -0700411 void (*ctor)(void *obj);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700412
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800413/* 5) cache creation/removal */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800414 const char *name;
415 struct list_head next;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700416
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800417/* 6) statistics */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700418#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800419 unsigned long num_active;
420 unsigned long num_allocations;
421 unsigned long high_mark;
422 unsigned long grown;
423 unsigned long reaped;
424 unsigned long errors;
425 unsigned long max_freeable;
426 unsigned long node_allocs;
427 unsigned long node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700428 unsigned long node_overflow;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800429 atomic_t allochit;
430 atomic_t allocmiss;
431 atomic_t freehit;
432 atomic_t freemiss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700433#endif
434#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800435 /*
436 * If debugging is enabled, then the allocator can add additional
437 * fields and/or padding to every object. buffer_size contains the total
438 * object size including these internal fields, the following two
439 * variables contain the offset to the user object and its size.
440 */
441 int obj_offset;
442 int obj_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700443#endif
Eric Dumazet8da34302007-05-06 14:49:29 -0700444 /*
445 * We put nodelists[] at the end of kmem_cache, because we want to size
446 * this array to nr_node_ids slots instead of MAX_NUMNODES
447 * (see kmem_cache_init())
448 * We still use [MAX_NUMNODES] and not [1] or [0] because cache_cache
449 * is statically defined, so we reserve the max number of nodes.
450 */
451 struct kmem_list3 *nodelists[MAX_NUMNODES];
452 /*
453 * Do not add fields after nodelists[]
454 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700455};
456
457#define CFLGS_OFF_SLAB (0x80000000UL)
458#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
459
460#define BATCHREFILL_LIMIT 16
Andrew Mortona737b3e2006-03-22 00:08:11 -0800461/*
462 * Optimization question: fewer reaps means less probability for unnessary
463 * cpucache drain/refill cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700464 *
Adrian Bunkdc6f3f22005-11-08 16:44:08 +0100465 * OTOH the cpuarrays can contain lots of objects,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700466 * which could lock up otherwise freeable slabs.
467 */
468#define REAPTIMEOUT_CPUC (2*HZ)
469#define REAPTIMEOUT_LIST3 (4*HZ)
470
471#if STATS
472#define STATS_INC_ACTIVE(x) ((x)->num_active++)
473#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
474#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
475#define STATS_INC_GROWN(x) ((x)->grown++)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700476#define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
Andrew Mortona737b3e2006-03-22 00:08:11 -0800477#define STATS_SET_HIGH(x) \
478 do { \
479 if ((x)->num_active > (x)->high_mark) \
480 (x)->high_mark = (x)->num_active; \
481 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700482#define STATS_INC_ERR(x) ((x)->errors++)
483#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
Christoph Lametere498be72005-09-09 13:03:32 -0700484#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700485#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800486#define STATS_SET_FREEABLE(x, i) \
487 do { \
488 if ((x)->max_freeable < i) \
489 (x)->max_freeable = i; \
490 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700491#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
492#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
493#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
494#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
495#else
496#define STATS_INC_ACTIVE(x) do { } while (0)
497#define STATS_DEC_ACTIVE(x) do { } while (0)
498#define STATS_INC_ALLOCED(x) do { } while (0)
499#define STATS_INC_GROWN(x) do { } while (0)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700500#define STATS_ADD_REAPED(x,y) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700501#define STATS_SET_HIGH(x) do { } while (0)
502#define STATS_INC_ERR(x) do { } while (0)
503#define STATS_INC_NODEALLOCS(x) do { } while (0)
Christoph Lametere498be72005-09-09 13:03:32 -0700504#define STATS_INC_NODEFREES(x) do { } while (0)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700505#define STATS_INC_ACOVERFLOW(x) do { } while (0)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800506#define STATS_SET_FREEABLE(x, i) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700507#define STATS_INC_ALLOCHIT(x) do { } while (0)
508#define STATS_INC_ALLOCMISS(x) do { } while (0)
509#define STATS_INC_FREEHIT(x) do { } while (0)
510#define STATS_INC_FREEMISS(x) do { } while (0)
511#endif
512
513#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -0700514
Andrew Mortona737b3e2006-03-22 00:08:11 -0800515/*
516 * memory layout of objects:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700517 * 0 : objp
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800518 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
Linus Torvalds1da177e2005-04-16 15:20:36 -0700519 * the end of an object is aligned with the end of the real
520 * allocation. Catches writes behind the end of the allocation.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800521 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700522 * redzone word.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800523 * cachep->obj_offset: The real object.
524 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
Andrew Mortona737b3e2006-03-22 00:08:11 -0800525 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
526 * [BYTES_PER_WORD long]
Linus Torvalds1da177e2005-04-16 15:20:36 -0700527 */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800528static int obj_offset(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700529{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800530 return cachep->obj_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700531}
532
Pekka Enberg343e0d72006-02-01 03:05:50 -0800533static int obj_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700534{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800535 return cachep->obj_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700536}
537
David Woodhouseb46b8f12007-05-08 00:22:59 -0700538static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700539{
540 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
David Woodhouseb46b8f12007-05-08 00:22:59 -0700541 return (unsigned long long*) (objp + obj_offset(cachep) -
542 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700543}
544
David Woodhouseb46b8f12007-05-08 00:22:59 -0700545static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700546{
547 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
548 if (cachep->flags & SLAB_STORE_USER)
David Woodhouseb46b8f12007-05-08 00:22:59 -0700549 return (unsigned long long *)(objp + cachep->buffer_size -
550 sizeof(unsigned long long) -
David Woodhouse87a927c2007-07-04 21:26:44 -0400551 REDZONE_ALIGN);
David Woodhouseb46b8f12007-05-08 00:22:59 -0700552 return (unsigned long long *) (objp + cachep->buffer_size -
553 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700554}
555
Pekka Enberg343e0d72006-02-01 03:05:50 -0800556static void **dbg_userword(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700557{
558 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800559 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700560}
561
562#else
563
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800564#define obj_offset(x) 0
565#define obj_size(cachep) (cachep->buffer_size)
David Woodhouseb46b8f12007-05-08 00:22:59 -0700566#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
567#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
Linus Torvalds1da177e2005-04-16 15:20:36 -0700568#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
569
570#endif
571
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +0300572#ifdef CONFIG_KMEMTRACE
573size_t slab_buffer_size(struct kmem_cache *cachep)
574{
575 return cachep->buffer_size;
576}
577EXPORT_SYMBOL(slab_buffer_size);
578#endif
579
Linus Torvalds1da177e2005-04-16 15:20:36 -0700580/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700581 * Do not go above this order unless 0 objects fit into the slab.
582 */
583#define BREAK_GFP_ORDER_HI 1
584#define BREAK_GFP_ORDER_LO 0
585static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
586
Andrew Mortona737b3e2006-03-22 00:08:11 -0800587/*
588 * Functions for storing/retrieving the cachep and or slab from the page
589 * allocator. These are used to find the slab an obj belongs to. With kfree(),
590 * these are used to find the cache which an obj belongs to.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700591 */
Pekka Enberg065d41c2005-11-13 16:06:46 -0800592static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
593{
594 page->lru.next = (struct list_head *)cache;
595}
596
597static inline struct kmem_cache *page_get_cache(struct page *page)
598{
Christoph Lameterd85f3382007-05-06 14:49:39 -0700599 page = compound_head(page);
Pekka Enbergddc2e812006-06-23 02:03:40 -0700600 BUG_ON(!PageSlab(page));
Pekka Enberg065d41c2005-11-13 16:06:46 -0800601 return (struct kmem_cache *)page->lru.next;
602}
603
604static inline void page_set_slab(struct page *page, struct slab *slab)
605{
606 page->lru.prev = (struct list_head *)slab;
607}
608
609static inline struct slab *page_get_slab(struct page *page)
610{
Pekka Enbergddc2e812006-06-23 02:03:40 -0700611 BUG_ON(!PageSlab(page));
Pekka Enberg065d41c2005-11-13 16:06:46 -0800612 return (struct slab *)page->lru.prev;
613}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700614
Pekka Enberg6ed5eb22006-02-01 03:05:49 -0800615static inline struct kmem_cache *virt_to_cache(const void *obj)
616{
Christoph Lameterb49af682007-05-06 14:49:41 -0700617 struct page *page = virt_to_head_page(obj);
Pekka Enberg6ed5eb22006-02-01 03:05:49 -0800618 return page_get_cache(page);
619}
620
621static inline struct slab *virt_to_slab(const void *obj)
622{
Christoph Lameterb49af682007-05-06 14:49:41 -0700623 struct page *page = virt_to_head_page(obj);
Pekka Enberg6ed5eb22006-02-01 03:05:49 -0800624 return page_get_slab(page);
625}
626
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800627static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
628 unsigned int idx)
629{
630 return slab->s_mem + cache->buffer_size * idx;
631}
632
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800633/*
634 * We want to avoid an expensive divide : (offset / cache->buffer_size)
635 * Using the fact that buffer_size is a constant for a particular cache,
636 * we can replace (offset / cache->buffer_size) by
637 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
638 */
639static inline unsigned int obj_to_index(const struct kmem_cache *cache,
640 const struct slab *slab, void *obj)
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800641{
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800642 u32 offset = (obj - slab->s_mem);
643 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800644}
645
Andrew Mortona737b3e2006-03-22 00:08:11 -0800646/*
647 * These are the default caches for kmalloc. Custom caches can have other sizes.
648 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700649struct cache_sizes malloc_sizes[] = {
650#define CACHE(x) { .cs_size = (x) },
651#include <linux/kmalloc_sizes.h>
652 CACHE(ULONG_MAX)
653#undef CACHE
654};
655EXPORT_SYMBOL(malloc_sizes);
656
657/* Must match cache_sizes above. Out of line to keep cache footprint low. */
658struct cache_names {
659 char *name;
660 char *name_dma;
661};
662
663static struct cache_names __initdata cache_names[] = {
664#define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
665#include <linux/kmalloc_sizes.h>
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800666 {NULL,}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700667#undef CACHE
668};
669
670static struct arraycache_init initarray_cache __initdata =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800671 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700672static struct arraycache_init initarray_generic =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800673 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700674
675/* internal cache of cache description objs */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800676static struct kmem_cache cache_cache = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800677 .batchcount = 1,
678 .limit = BOOT_CPUCACHE_ENTRIES,
679 .shared = 1,
Pekka Enberg343e0d72006-02-01 03:05:50 -0800680 .buffer_size = sizeof(struct kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800681 .name = "kmem_cache",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700682};
683
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700684#define BAD_ALIEN_MAGIC 0x01020304ul
685
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200686#ifdef CONFIG_LOCKDEP
687
688/*
689 * Slab sometimes uses the kmalloc slabs to store the slab headers
690 * for other slabs "off slab".
691 * The locking for this is tricky in that it nests within the locks
692 * of all other slabs in a few places; to deal with this special
693 * locking we put on-slab caches into a separate lock-class.
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700694 *
695 * We set lock class for alien array caches which are up during init.
696 * The lock annotation will be lost if all cpus of a node goes down and
697 * then comes back up during hotplug
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200698 */
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700699static struct lock_class_key on_slab_l3_key;
700static struct lock_class_key on_slab_alc_key;
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200701
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700702static inline void init_lock_keys(void)
703
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200704{
705 int q;
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700706 struct cache_sizes *s = malloc_sizes;
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200707
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700708 while (s->cs_size != ULONG_MAX) {
709 for_each_node(q) {
710 struct array_cache **alc;
711 int r;
712 struct kmem_list3 *l3 = s->cs_cachep->nodelists[q];
713 if (!l3 || OFF_SLAB(s->cs_cachep))
714 continue;
715 lockdep_set_class(&l3->list_lock, &on_slab_l3_key);
716 alc = l3->alien;
717 /*
718 * FIXME: This check for BAD_ALIEN_MAGIC
719 * should go away when common slab code is taught to
720 * work even without alien caches.
721 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
722 * for alloc_alien_cache,
723 */
724 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
725 continue;
726 for_each_node(r) {
727 if (alc[r])
728 lockdep_set_class(&alc[r]->lock,
729 &on_slab_alc_key);
730 }
731 }
732 s++;
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200733 }
734}
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200735#else
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700736static inline void init_lock_keys(void)
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200737{
738}
739#endif
740
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -0800741/*
Gautham R Shenoy95402b32008-01-25 21:08:02 +0100742 * Guard access to the cache-chain.
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -0800743 */
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800744static DEFINE_MUTEX(cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700745static struct list_head cache_chain;
746
747/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700748 * chicken and egg problem: delay the per-cpu array allocation
749 * until the general caches are up.
750 */
751static enum {
752 NONE,
Christoph Lametere498be72005-09-09 13:03:32 -0700753 PARTIAL_AC,
754 PARTIAL_L3,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700755 FULL
756} g_cpucache_up;
757
Mike Kravetz39d24e62006-05-15 09:44:13 -0700758/*
759 * used by boot code to determine if it can use slab based allocator
760 */
761int slab_is_available(void)
762{
763 return g_cpucache_up == FULL;
764}
765
David Howells52bad642006-11-22 14:54:01 +0000766static DEFINE_PER_CPU(struct delayed_work, reap_work);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700767
Pekka Enberg343e0d72006-02-01 03:05:50 -0800768static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700769{
770 return cachep->array[smp_processor_id()];
771}
772
Andrew Mortona737b3e2006-03-22 00:08:11 -0800773static inline struct kmem_cache *__find_general_cachep(size_t size,
774 gfp_t gfpflags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700775{
776 struct cache_sizes *csizep = malloc_sizes;
777
778#if DEBUG
779 /* This happens if someone tries to call
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800780 * kmem_cache_create(), or __kmalloc(), before
781 * the generic caches are initialized.
782 */
Alok Katariac7e43c72005-09-14 12:17:53 -0700783 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700784#endif
Christoph Lameter6cb8f912007-07-17 04:03:22 -0700785 if (!size)
786 return ZERO_SIZE_PTR;
787
Linus Torvalds1da177e2005-04-16 15:20:36 -0700788 while (size > csizep->cs_size)
789 csizep++;
790
791 /*
Martin Hicks0abf40c2005-09-03 15:54:54 -0700792 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
Linus Torvalds1da177e2005-04-16 15:20:36 -0700793 * has cs_{dma,}cachep==NULL. Thus no special case
794 * for large kmalloc calls required.
795 */
Christoph Lameter4b51d662007-02-10 01:43:10 -0800796#ifdef CONFIG_ZONE_DMA
Linus Torvalds1da177e2005-04-16 15:20:36 -0700797 if (unlikely(gfpflags & GFP_DMA))
798 return csizep->cs_dmacachep;
Christoph Lameter4b51d662007-02-10 01:43:10 -0800799#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700800 return csizep->cs_cachep;
801}
802
Adrian Bunkb2213852006-09-25 23:31:02 -0700803static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700804{
805 return __find_general_cachep(size, gfpflags);
806}
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700807
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800808static size_t slab_mgmt_size(size_t nr_objs, size_t align)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700809{
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800810 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
811}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700812
Andrew Mortona737b3e2006-03-22 00:08:11 -0800813/*
814 * Calculate the number of objects and left-over bytes for a given buffer size.
815 */
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800816static void cache_estimate(unsigned long gfporder, size_t buffer_size,
817 size_t align, int flags, size_t *left_over,
818 unsigned int *num)
819{
820 int nr_objs;
821 size_t mgmt_size;
822 size_t slab_size = PAGE_SIZE << gfporder;
823
824 /*
825 * The slab management structure can be either off the slab or
826 * on it. For the latter case, the memory allocated for a
827 * slab is used for:
828 *
829 * - The struct slab
830 * - One kmem_bufctl_t for each object
831 * - Padding to respect alignment of @align
832 * - @buffer_size bytes for each object
833 *
834 * If the slab management structure is off the slab, then the
835 * alignment will already be calculated into the size. Because
836 * the slabs are all pages aligned, the objects will be at the
837 * correct alignment when allocated.
838 */
839 if (flags & CFLGS_OFF_SLAB) {
840 mgmt_size = 0;
841 nr_objs = slab_size / buffer_size;
842
843 if (nr_objs > SLAB_LIMIT)
844 nr_objs = SLAB_LIMIT;
845 } else {
846 /*
847 * Ignore padding for the initial guess. The padding
848 * is at most @align-1 bytes, and @buffer_size is at
849 * least @align. In the worst case, this result will
850 * be one greater than the number of objects that fit
851 * into the memory allocation when taking the padding
852 * into account.
853 */
854 nr_objs = (slab_size - sizeof(struct slab)) /
855 (buffer_size + sizeof(kmem_bufctl_t));
856
857 /*
858 * This calculated number will be either the right
859 * amount, or one greater than what we want.
860 */
861 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
862 > slab_size)
863 nr_objs--;
864
865 if (nr_objs > SLAB_LIMIT)
866 nr_objs = SLAB_LIMIT;
867
868 mgmt_size = slab_mgmt_size(nr_objs, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700869 }
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800870 *num = nr_objs;
871 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700872}
873
Harvey Harrisond40cee22008-04-30 00:55:07 -0700874#define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700875
Andrew Mortona737b3e2006-03-22 00:08:11 -0800876static void __slab_error(const char *function, struct kmem_cache *cachep,
877 char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700878{
879 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800880 function, cachep->name, msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700881 dump_stack();
882}
883
Paul Menage3395ee02006-12-06 20:32:16 -0800884/*
885 * By default on NUMA we use alien caches to stage the freeing of
886 * objects allocated from other nodes. This causes massive memory
887 * inefficiencies when using fake NUMA setup to split memory into a
888 * large number of small nodes, so it can be disabled on the command
889 * line
890 */
891
892static int use_alien_caches __read_mostly = 1;
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -0700893static int numa_platform __read_mostly = 1;
Paul Menage3395ee02006-12-06 20:32:16 -0800894static int __init noaliencache_setup(char *s)
895{
896 use_alien_caches = 0;
897 return 1;
898}
899__setup("noaliencache", noaliencache_setup);
900
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800901#ifdef CONFIG_NUMA
902/*
903 * Special reaping functions for NUMA systems called from cache_reap().
904 * These take care of doing round robin flushing of alien caches (containing
905 * objects freed on different nodes from which they were allocated) and the
906 * flushing of remote pcps by calling drain_node_pages.
907 */
908static DEFINE_PER_CPU(unsigned long, reap_node);
909
910static void init_reap_node(int cpu)
911{
912 int node;
913
914 node = next_node(cpu_to_node(cpu), node_online_map);
915 if (node == MAX_NUMNODES)
Paul Jackson442295c2006-03-22 00:09:11 -0800916 node = first_node(node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800917
Daniel Yeisley7f6b8872006-11-02 22:07:14 -0800918 per_cpu(reap_node, cpu) = node;
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800919}
920
921static void next_reap_node(void)
922{
923 int node = __get_cpu_var(reap_node);
924
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800925 node = next_node(node, node_online_map);
926 if (unlikely(node >= MAX_NUMNODES))
927 node = first_node(node_online_map);
928 __get_cpu_var(reap_node) = node;
929}
930
931#else
932#define init_reap_node(cpu) do { } while (0)
933#define next_reap_node(void) do { } while (0)
934#endif
935
Linus Torvalds1da177e2005-04-16 15:20:36 -0700936/*
937 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
938 * via the workqueue/eventd.
939 * Add the CPU number into the expiration time to minimize the possibility of
940 * the CPUs getting into lockstep and contending for the global cache chain
941 * lock.
942 */
Adrian Bunk897e6792007-07-15 23:38:20 -0700943static void __cpuinit start_cpu_timer(int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700944{
David Howells52bad642006-11-22 14:54:01 +0000945 struct delayed_work *reap_work = &per_cpu(reap_work, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700946
947 /*
948 * When this gets called from do_initcalls via cpucache_init(),
949 * init_workqueues() has already run, so keventd will be setup
950 * at that time.
951 */
David Howells52bad642006-11-22 14:54:01 +0000952 if (keventd_up() && reap_work->work.func == NULL) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800953 init_reap_node(cpu);
David Howells65f27f32006-11-22 14:55:48 +0000954 INIT_DELAYED_WORK(reap_work, cache_reap);
Arjan van de Ven2b284212006-12-10 02:21:28 -0800955 schedule_delayed_work_on(cpu, reap_work,
956 __round_jiffies_relative(HZ, cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700957 }
958}
959
Christoph Lametere498be72005-09-09 13:03:32 -0700960static struct array_cache *alloc_arraycache(int node, int entries,
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800961 int batchcount)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700962{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800963 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700964 struct array_cache *nc = NULL;
965
Christoph Lametere498be72005-09-09 13:03:32 -0700966 nc = kmalloc_node(memsize, GFP_KERNEL, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700967 if (nc) {
968 nc->avail = 0;
969 nc->limit = entries;
970 nc->batchcount = batchcount;
971 nc->touched = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700972 spin_lock_init(&nc->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700973 }
974 return nc;
975}
976
Christoph Lameter3ded1752006-03-25 03:06:44 -0800977/*
978 * Transfer objects in one arraycache to another.
979 * Locking must be handled by the caller.
980 *
981 * Return the number of entries transferred.
982 */
983static int transfer_objects(struct array_cache *to,
984 struct array_cache *from, unsigned int max)
985{
986 /* Figure out how many entries to transfer */
987 int nr = min(min(from->avail, max), to->limit - to->avail);
988
989 if (!nr)
990 return 0;
991
992 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
993 sizeof(void *) *nr);
994
995 from->avail -= nr;
996 to->avail += nr;
997 to->touched = 1;
998 return nr;
999}
1000
Christoph Lameter765c4502006-09-27 01:50:08 -07001001#ifndef CONFIG_NUMA
1002
1003#define drain_alien_cache(cachep, alien) do { } while (0)
1004#define reap_alien(cachep, l3) do { } while (0)
1005
1006static inline struct array_cache **alloc_alien_cache(int node, int limit)
1007{
1008 return (struct array_cache **)BAD_ALIEN_MAGIC;
1009}
1010
1011static inline void free_alien_cache(struct array_cache **ac_ptr)
1012{
1013}
1014
1015static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1016{
1017 return 0;
1018}
1019
1020static inline void *alternate_node_alloc(struct kmem_cache *cachep,
1021 gfp_t flags)
1022{
1023 return NULL;
1024}
1025
Christoph Hellwig8b98c162006-12-06 20:32:30 -08001026static inline void *____cache_alloc_node(struct kmem_cache *cachep,
Christoph Lameter765c4502006-09-27 01:50:08 -07001027 gfp_t flags, int nodeid)
1028{
1029 return NULL;
1030}
1031
1032#else /* CONFIG_NUMA */
1033
Christoph Hellwig8b98c162006-12-06 20:32:30 -08001034static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
Paul Jacksonc61afb12006-03-24 03:16:08 -08001035static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
Christoph Lameterdc85da12006-01-18 17:42:36 -08001036
Pekka Enberg5295a742006-02-01 03:05:48 -08001037static struct array_cache **alloc_alien_cache(int node, int limit)
Christoph Lametere498be72005-09-09 13:03:32 -07001038{
1039 struct array_cache **ac_ptr;
Christoph Lameter8ef82862007-02-20 13:57:52 -08001040 int memsize = sizeof(void *) * nr_node_ids;
Christoph Lametere498be72005-09-09 13:03:32 -07001041 int i;
1042
1043 if (limit > 1)
1044 limit = 12;
1045 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
1046 if (ac_ptr) {
1047 for_each_node(i) {
1048 if (i == node || !node_online(i)) {
1049 ac_ptr[i] = NULL;
1050 continue;
1051 }
1052 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
1053 if (!ac_ptr[i]) {
Akinobu Mitacc550de2007-11-14 16:58:35 -08001054 for (i--; i >= 0; i--)
Christoph Lametere498be72005-09-09 13:03:32 -07001055 kfree(ac_ptr[i]);
1056 kfree(ac_ptr);
1057 return NULL;
1058 }
1059 }
1060 }
1061 return ac_ptr;
1062}
1063
Pekka Enberg5295a742006-02-01 03:05:48 -08001064static void free_alien_cache(struct array_cache **ac_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -07001065{
1066 int i;
1067
1068 if (!ac_ptr)
1069 return;
Christoph Lametere498be72005-09-09 13:03:32 -07001070 for_each_node(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001071 kfree(ac_ptr[i]);
Christoph Lametere498be72005-09-09 13:03:32 -07001072 kfree(ac_ptr);
1073}
1074
Pekka Enberg343e0d72006-02-01 03:05:50 -08001075static void __drain_alien_cache(struct kmem_cache *cachep,
Pekka Enberg5295a742006-02-01 03:05:48 -08001076 struct array_cache *ac, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07001077{
1078 struct kmem_list3 *rl3 = cachep->nodelists[node];
1079
1080 if (ac->avail) {
1081 spin_lock(&rl3->list_lock);
Christoph Lametere00946f2006-03-25 03:06:45 -08001082 /*
1083 * Stuff objects into the remote nodes shared array first.
1084 * That way we could avoid the overhead of putting the objects
1085 * into the free lists and getting them back later.
1086 */
shin, jacob693f7d32006-04-28 10:54:37 -05001087 if (rl3->shared)
1088 transfer_objects(rl3->shared, ac, ac->limit);
Christoph Lametere00946f2006-03-25 03:06:45 -08001089
Christoph Lameterff694162005-09-22 21:44:02 -07001090 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001091 ac->avail = 0;
1092 spin_unlock(&rl3->list_lock);
1093 }
1094}
1095
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001096/*
1097 * Called from cache_reap() to regularly drain alien caches round robin.
1098 */
1099static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1100{
1101 int node = __get_cpu_var(reap_node);
1102
1103 if (l3->alien) {
1104 struct array_cache *ac = l3->alien[node];
Christoph Lametere00946f2006-03-25 03:06:45 -08001105
1106 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001107 __drain_alien_cache(cachep, ac, node);
1108 spin_unlock_irq(&ac->lock);
1109 }
1110 }
1111}
1112
Andrew Mortona737b3e2006-03-22 00:08:11 -08001113static void drain_alien_cache(struct kmem_cache *cachep,
1114 struct array_cache **alien)
Christoph Lametere498be72005-09-09 13:03:32 -07001115{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001116 int i = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07001117 struct array_cache *ac;
1118 unsigned long flags;
1119
1120 for_each_online_node(i) {
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001121 ac = alien[i];
Christoph Lametere498be72005-09-09 13:03:32 -07001122 if (ac) {
1123 spin_lock_irqsave(&ac->lock, flags);
1124 __drain_alien_cache(cachep, ac, i);
1125 spin_unlock_irqrestore(&ac->lock, flags);
1126 }
1127 }
1128}
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001129
Ingo Molnar873623d2006-07-13 14:44:38 +02001130static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001131{
1132 struct slab *slabp = virt_to_slab(objp);
1133 int nodeid = slabp->nodeid;
1134 struct kmem_list3 *l3;
1135 struct array_cache *alien = NULL;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001136 int node;
1137
1138 node = numa_node_id();
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001139
1140 /*
1141 * Make sure we are not freeing a object from another node to the array
1142 * cache on this cpu.
1143 */
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001144 if (likely(slabp->nodeid == node))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001145 return 0;
1146
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001147 l3 = cachep->nodelists[node];
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001148 STATS_INC_NODEFREES(cachep);
1149 if (l3->alien && l3->alien[nodeid]) {
1150 alien = l3->alien[nodeid];
Ingo Molnar873623d2006-07-13 14:44:38 +02001151 spin_lock(&alien->lock);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001152 if (unlikely(alien->avail == alien->limit)) {
1153 STATS_INC_ACOVERFLOW(cachep);
1154 __drain_alien_cache(cachep, alien, nodeid);
1155 }
1156 alien->entry[alien->avail++] = objp;
1157 spin_unlock(&alien->lock);
1158 } else {
1159 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1160 free_block(cachep, &objp, 1, nodeid);
1161 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1162 }
1163 return 1;
1164}
Christoph Lametere498be72005-09-09 13:03:32 -07001165#endif
1166
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001167static void __cpuinit cpuup_canceled(long cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001168{
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001169 struct kmem_cache *cachep;
1170 struct kmem_list3 *l3 = NULL;
1171 int node = cpu_to_node(cpu);
Rusty Russella70f7302009-03-13 14:49:46 +10301172 const struct cpumask *mask = cpumask_of_node(node);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001173
1174 list_for_each_entry(cachep, &cache_chain, next) {
1175 struct array_cache *nc;
1176 struct array_cache *shared;
1177 struct array_cache **alien;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001178
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001179 /* cpu is dead; no one can alloc from it. */
1180 nc = cachep->array[cpu];
1181 cachep->array[cpu] = NULL;
1182 l3 = cachep->nodelists[node];
1183
1184 if (!l3)
1185 goto free_array_cache;
1186
1187 spin_lock_irq(&l3->list_lock);
1188
1189 /* Free limit for this kmem_list3 */
1190 l3->free_limit -= cachep->batchcount;
1191 if (nc)
1192 free_block(cachep, nc->entry, nc->avail, node);
1193
Mike Travisc5f59f02008-04-04 18:11:10 -07001194 if (!cpus_empty(*mask)) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001195 spin_unlock_irq(&l3->list_lock);
1196 goto free_array_cache;
1197 }
1198
1199 shared = l3->shared;
1200 if (shared) {
1201 free_block(cachep, shared->entry,
1202 shared->avail, node);
1203 l3->shared = NULL;
1204 }
1205
1206 alien = l3->alien;
1207 l3->alien = NULL;
1208
1209 spin_unlock_irq(&l3->list_lock);
1210
1211 kfree(shared);
1212 if (alien) {
1213 drain_alien_cache(cachep, alien);
1214 free_alien_cache(alien);
1215 }
1216free_array_cache:
1217 kfree(nc);
1218 }
1219 /*
1220 * In the previous loop, all the objects were freed to
1221 * the respective cache's slabs, now we can go ahead and
1222 * shrink each nodelist to its limit.
1223 */
1224 list_for_each_entry(cachep, &cache_chain, next) {
1225 l3 = cachep->nodelists[node];
1226 if (!l3)
1227 continue;
1228 drain_freelist(cachep, l3, l3->free_objects);
1229 }
1230}
1231
1232static int __cpuinit cpuup_prepare(long cpu)
1233{
Pekka Enberg343e0d72006-02-01 03:05:50 -08001234 struct kmem_cache *cachep;
Christoph Lametere498be72005-09-09 13:03:32 -07001235 struct kmem_list3 *l3 = NULL;
1236 int node = cpu_to_node(cpu);
David Howellsea02e3d2007-07-19 01:49:09 -07001237 const int memsize = sizeof(struct kmem_list3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001238
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001239 /*
1240 * We need to do this right in the beginning since
1241 * alloc_arraycache's are going to use this list.
1242 * kmalloc_node allows us to add the slab to the right
1243 * kmem_list3 and not this cpu's kmem_list3
1244 */
1245
1246 list_for_each_entry(cachep, &cache_chain, next) {
1247 /*
1248 * Set up the size64 kmemlist for cpu before we can
1249 * begin anything. Make sure some other cpu on this
1250 * node has not already allocated this
1251 */
1252 if (!cachep->nodelists[node]) {
1253 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1254 if (!l3)
1255 goto bad;
1256 kmem_list3_init(l3);
1257 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1258 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1259
1260 /*
1261 * The l3s don't come and go as CPUs come and
1262 * go. cache_chain_mutex is sufficient
1263 * protection here.
1264 */
1265 cachep->nodelists[node] = l3;
1266 }
1267
1268 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1269 cachep->nodelists[node]->free_limit =
1270 (1 + nr_cpus_node(node)) *
1271 cachep->batchcount + cachep->num;
1272 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1273 }
1274
1275 /*
1276 * Now we can go ahead with allocating the shared arrays and
1277 * array caches
1278 */
1279 list_for_each_entry(cachep, &cache_chain, next) {
1280 struct array_cache *nc;
1281 struct array_cache *shared = NULL;
1282 struct array_cache **alien = NULL;
1283
1284 nc = alloc_arraycache(node, cachep->limit,
1285 cachep->batchcount);
1286 if (!nc)
1287 goto bad;
1288 if (cachep->shared) {
1289 shared = alloc_arraycache(node,
1290 cachep->shared * cachep->batchcount,
1291 0xbaadf00d);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001292 if (!shared) {
1293 kfree(nc);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001294 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001295 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001296 }
1297 if (use_alien_caches) {
1298 alien = alloc_alien_cache(node, cachep->limit);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001299 if (!alien) {
1300 kfree(shared);
1301 kfree(nc);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001302 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001303 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001304 }
1305 cachep->array[cpu] = nc;
1306 l3 = cachep->nodelists[node];
1307 BUG_ON(!l3);
1308
1309 spin_lock_irq(&l3->list_lock);
1310 if (!l3->shared) {
1311 /*
1312 * We are serialised from CPU_DEAD or
1313 * CPU_UP_CANCELLED by the cpucontrol lock
1314 */
1315 l3->shared = shared;
1316 shared = NULL;
1317 }
1318#ifdef CONFIG_NUMA
1319 if (!l3->alien) {
1320 l3->alien = alien;
1321 alien = NULL;
1322 }
1323#endif
1324 spin_unlock_irq(&l3->list_lock);
1325 kfree(shared);
1326 free_alien_cache(alien);
1327 }
1328 return 0;
1329bad:
Akinobu Mita12d00f62007-10-18 03:05:11 -07001330 cpuup_canceled(cpu);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001331 return -ENOMEM;
1332}
1333
1334static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1335 unsigned long action, void *hcpu)
1336{
1337 long cpu = (long)hcpu;
1338 int err = 0;
1339
Linus Torvalds1da177e2005-04-16 15:20:36 -07001340 switch (action) {
Heiko Carstens38c3bd92007-05-09 02:34:05 -07001341 case CPU_UP_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001342 case CPU_UP_PREPARE_FROZEN:
Gautham R Shenoy95402b32008-01-25 21:08:02 +01001343 mutex_lock(&cache_chain_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001344 err = cpuup_prepare(cpu);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01001345 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001346 break;
1347 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001348 case CPU_ONLINE_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001349 start_cpu_timer(cpu);
1350 break;
1351#ifdef CONFIG_HOTPLUG_CPU
Christoph Lameter5830c592007-05-09 02:34:22 -07001352 case CPU_DOWN_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001353 case CPU_DOWN_PREPARE_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001354 /*
1355 * Shutdown cache reaper. Note that the cache_chain_mutex is
1356 * held so that if cache_reap() is invoked it cannot do
1357 * anything expensive but will only modify reap_work
1358 * and reschedule the timer.
1359 */
1360 cancel_rearming_delayed_work(&per_cpu(reap_work, cpu));
1361 /* Now the cache_reaper is guaranteed to be not running. */
1362 per_cpu(reap_work, cpu).work.func = NULL;
1363 break;
1364 case CPU_DOWN_FAILED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001365 case CPU_DOWN_FAILED_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001366 start_cpu_timer(cpu);
1367 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001368 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001369 case CPU_DEAD_FROZEN:
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001370 /*
1371 * Even if all the cpus of a node are down, we don't free the
1372 * kmem_list3 of any cache. This to avoid a race between
1373 * cpu_down, and a kmalloc allocation from another cpu for
1374 * memory from the node of the cpu going down. The list3
1375 * structure is usually allocated from kmem_cache_create() and
1376 * gets destroyed at kmem_cache_destroy().
1377 */
Simon Arlott183ff222007-10-20 01:27:18 +02001378 /* fall through */
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08001379#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001380 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001381 case CPU_UP_CANCELED_FROZEN:
Gautham R Shenoy95402b32008-01-25 21:08:02 +01001382 mutex_lock(&cache_chain_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001383 cpuup_canceled(cpu);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001384 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001385 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001386 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001387 return err ? NOTIFY_BAD : NOTIFY_OK;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001388}
1389
Chandra Seetharaman74b85f32006-06-27 02:54:09 -07001390static struct notifier_block __cpuinitdata cpucache_notifier = {
1391 &cpuup_callback, NULL, 0
1392};
Linus Torvalds1da177e2005-04-16 15:20:36 -07001393
Christoph Lametere498be72005-09-09 13:03:32 -07001394/*
1395 * swap the static kmem_list3 with kmalloced memory
1396 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001397static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1398 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001399{
1400 struct kmem_list3 *ptr;
1401
Christoph Lametere498be72005-09-09 13:03:32 -07001402 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1403 BUG_ON(!ptr);
1404
1405 local_irq_disable();
1406 memcpy(ptr, list, sizeof(struct kmem_list3));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001407 /*
1408 * Do not assume that spinlocks can be initialized via memcpy:
1409 */
1410 spin_lock_init(&ptr->list_lock);
1411
Christoph Lametere498be72005-09-09 13:03:32 -07001412 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1413 cachep->nodelists[nodeid] = ptr;
1414 local_irq_enable();
1415}
1416
Andrew Mortona737b3e2006-03-22 00:08:11 -08001417/*
Pekka Enberg556a1692008-01-25 08:20:51 +02001418 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1419 * size of kmem_list3.
1420 */
1421static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1422{
1423 int node;
1424
1425 for_each_online_node(node) {
1426 cachep->nodelists[node] = &initkmem_list3[index + node];
1427 cachep->nodelists[node]->next_reap = jiffies +
1428 REAPTIMEOUT_LIST3 +
1429 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1430 }
1431}
1432
1433/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08001434 * Initialisation. Called after the page allocator have been initialised and
1435 * before smp_init().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001436 */
1437void __init kmem_cache_init(void)
1438{
1439 size_t left_over;
1440 struct cache_sizes *sizes;
1441 struct cache_names *names;
Christoph Lametere498be72005-09-09 13:03:32 -07001442 int i;
Jack Steiner07ed76b2006-03-07 21:55:46 -08001443 int order;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001444 int node;
Christoph Lametere498be72005-09-09 13:03:32 -07001445
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -07001446 if (num_possible_nodes() == 1) {
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001447 use_alien_caches = 0;
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -07001448 numa_platform = 0;
1449 }
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001450
Christoph Lametere498be72005-09-09 13:03:32 -07001451 for (i = 0; i < NUM_INIT_LISTS; i++) {
1452 kmem_list3_init(&initkmem_list3[i]);
1453 if (i < MAX_NUMNODES)
1454 cache_cache.nodelists[i] = NULL;
1455 }
Pekka Enberg556a1692008-01-25 08:20:51 +02001456 set_up_list3s(&cache_cache, CACHE_CACHE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001457
1458 /*
1459 * Fragmentation resistance on low memory - only use bigger
1460 * page orders on machines with more than 32MB of memory.
1461 */
1462 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1463 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1464
Linus Torvalds1da177e2005-04-16 15:20:36 -07001465 /* Bootstrap is tricky, because several objects are allocated
1466 * from caches that do not exist yet:
Andrew Mortona737b3e2006-03-22 00:08:11 -08001467 * 1) initialize the cache_cache cache: it contains the struct
1468 * kmem_cache structures of all caches, except cache_cache itself:
1469 * cache_cache is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001470 * Initially an __init data area is used for the head array and the
1471 * kmem_list3 structures, it's replaced with a kmalloc allocated
1472 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001473 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001474 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001475 * An __init data area is used for the head array.
1476 * 3) Create the remaining kmalloc caches, with minimally sized
1477 * head arrays.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001478 * 4) Replace the __init data head arrays for cache_cache and the first
1479 * kmalloc cache with kmalloc allocated arrays.
Christoph Lametere498be72005-09-09 13:03:32 -07001480 * 5) Replace the __init data for kmem_list3 for cache_cache and
1481 * the other cache's with kmalloc allocated memory.
1482 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001483 */
1484
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001485 node = numa_node_id();
1486
Linus Torvalds1da177e2005-04-16 15:20:36 -07001487 /* 1) create the cache_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001488 INIT_LIST_HEAD(&cache_chain);
1489 list_add(&cache_cache.next, &cache_chain);
1490 cache_cache.colour_off = cache_line_size();
1491 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
Daniel Yeisleyec1f5ee2008-03-25 23:59:08 +02001492 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001493
Eric Dumazet8da34302007-05-06 14:49:29 -07001494 /*
1495 * struct kmem_cache size depends on nr_node_ids, which
1496 * can be less than MAX_NUMNODES.
1497 */
1498 cache_cache.buffer_size = offsetof(struct kmem_cache, nodelists) +
1499 nr_node_ids * sizeof(struct kmem_list3 *);
1500#if DEBUG
1501 cache_cache.obj_size = cache_cache.buffer_size;
1502#endif
Andrew Mortona737b3e2006-03-22 00:08:11 -08001503 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1504 cache_line_size());
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08001505 cache_cache.reciprocal_buffer_size =
1506 reciprocal_value(cache_cache.buffer_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001507
Jack Steiner07ed76b2006-03-07 21:55:46 -08001508 for (order = 0; order < MAX_ORDER; order++) {
1509 cache_estimate(order, cache_cache.buffer_size,
1510 cache_line_size(), 0, &left_over, &cache_cache.num);
1511 if (cache_cache.num)
1512 break;
1513 }
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02001514 BUG_ON(!cache_cache.num);
Jack Steiner07ed76b2006-03-07 21:55:46 -08001515 cache_cache.gfporder = order;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001516 cache_cache.colour = left_over / cache_cache.colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001517 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1518 sizeof(struct slab), cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001519
1520 /* 2+3) create the kmalloc caches */
1521 sizes = malloc_sizes;
1522 names = cache_names;
1523
Andrew Mortona737b3e2006-03-22 00:08:11 -08001524 /*
1525 * Initialize the caches that provide memory for the array cache and the
1526 * kmem_list3 structures first. Without this, further allocations will
1527 * bug.
Christoph Lametere498be72005-09-09 13:03:32 -07001528 */
1529
1530 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001531 sizes[INDEX_AC].cs_size,
1532 ARCH_KMALLOC_MINALIGN,
1533 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001534 NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07001535
Andrew Mortona737b3e2006-03-22 00:08:11 -08001536 if (INDEX_AC != INDEX_L3) {
Christoph Lametere498be72005-09-09 13:03:32 -07001537 sizes[INDEX_L3].cs_cachep =
Andrew Mortona737b3e2006-03-22 00:08:11 -08001538 kmem_cache_create(names[INDEX_L3].name,
1539 sizes[INDEX_L3].cs_size,
1540 ARCH_KMALLOC_MINALIGN,
1541 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001542 NULL);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001543 }
Christoph Lametere498be72005-09-09 13:03:32 -07001544
Ingo Molnare0a42722006-06-23 02:03:46 -07001545 slab_early_init = 0;
1546
Linus Torvalds1da177e2005-04-16 15:20:36 -07001547 while (sizes->cs_size != ULONG_MAX) {
Christoph Lametere498be72005-09-09 13:03:32 -07001548 /*
1549 * For performance, all the general caches are L1 aligned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001550 * This should be particularly beneficial on SMP boxes, as it
1551 * eliminates "false sharing".
1552 * Note for systems short on memory removing the alignment will
Christoph Lametere498be72005-09-09 13:03:32 -07001553 * allow tighter packing of the smaller caches.
1554 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001555 if (!sizes->cs_cachep) {
Christoph Lametere498be72005-09-09 13:03:32 -07001556 sizes->cs_cachep = kmem_cache_create(names->name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001557 sizes->cs_size,
1558 ARCH_KMALLOC_MINALIGN,
1559 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001560 NULL);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001561 }
Christoph Lameter4b51d662007-02-10 01:43:10 -08001562#ifdef CONFIG_ZONE_DMA
1563 sizes->cs_dmacachep = kmem_cache_create(
1564 names->name_dma,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001565 sizes->cs_size,
1566 ARCH_KMALLOC_MINALIGN,
1567 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1568 SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001569 NULL);
Christoph Lameter4b51d662007-02-10 01:43:10 -08001570#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001571 sizes++;
1572 names++;
1573 }
1574 /* 4) Replace the bootstrap head arrays */
1575 {
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001576 struct array_cache *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001577
Linus Torvalds1da177e2005-04-16 15:20:36 -07001578 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001579
Linus Torvalds1da177e2005-04-16 15:20:36 -07001580 local_irq_disable();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001581 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1582 memcpy(ptr, cpu_cache_get(&cache_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001583 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001584 /*
1585 * Do not assume that spinlocks can be initialized via memcpy:
1586 */
1587 spin_lock_init(&ptr->lock);
1588
Linus Torvalds1da177e2005-04-16 15:20:36 -07001589 cache_cache.array[smp_processor_id()] = ptr;
1590 local_irq_enable();
Christoph Lametere498be72005-09-09 13:03:32 -07001591
Linus Torvalds1da177e2005-04-16 15:20:36 -07001592 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001593
Linus Torvalds1da177e2005-04-16 15:20:36 -07001594 local_irq_disable();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001595 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001596 != &initarray_generic.cache);
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001597 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001598 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001599 /*
1600 * Do not assume that spinlocks can be initialized via memcpy:
1601 */
1602 spin_lock_init(&ptr->lock);
1603
Christoph Lametere498be72005-09-09 13:03:32 -07001604 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001605 ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001606 local_irq_enable();
1607 }
Christoph Lametere498be72005-09-09 13:03:32 -07001608 /* 5) Replace the bootstrap kmem_list3's */
1609 {
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001610 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001611
Mel Gorman9c09a952008-01-24 05:49:54 -08001612 for_each_online_node(nid) {
Daniel Yeisleyec1f5ee2008-03-25 23:59:08 +02001613 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid);
Pekka Enberg556a1692008-01-25 08:20:51 +02001614
Christoph Lametere498be72005-09-09 13:03:32 -07001615 init_list(malloc_sizes[INDEX_AC].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001616 &initkmem_list3[SIZE_AC + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001617
1618 if (INDEX_AC != INDEX_L3) {
1619 init_list(malloc_sizes[INDEX_L3].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001620 &initkmem_list3[SIZE_L3 + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001621 }
1622 }
1623 }
1624
1625 /* 6) resize the head arrays to their final sizes */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001626 {
Pekka Enberg343e0d72006-02-01 03:05:50 -08001627 struct kmem_cache *cachep;
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001628 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001629 list_for_each_entry(cachep, &cache_chain, next)
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07001630 if (enable_cpucache(cachep))
1631 BUG();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001632 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001633 }
1634
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001635 /* Annotate slab for lockdep -- annotate the malloc caches */
1636 init_lock_keys();
1637
1638
Linus Torvalds1da177e2005-04-16 15:20:36 -07001639 /* Done! */
1640 g_cpucache_up = FULL;
1641
Andrew Mortona737b3e2006-03-22 00:08:11 -08001642 /*
1643 * Register a cpu startup notifier callback that initializes
1644 * cpu_cache_get for all new cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07001645 */
1646 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001647
Andrew Mortona737b3e2006-03-22 00:08:11 -08001648 /*
1649 * The reap timers are started later, with a module init call: That part
1650 * of the kernel is not yet operational.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001651 */
1652}
1653
1654static int __init cpucache_init(void)
1655{
1656 int cpu;
1657
Andrew Mortona737b3e2006-03-22 00:08:11 -08001658 /*
1659 * Register the timers that return unneeded pages to the page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -07001660 */
Christoph Lametere498be72005-09-09 13:03:32 -07001661 for_each_online_cpu(cpu)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001662 start_cpu_timer(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001663 return 0;
1664}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001665__initcall(cpucache_init);
1666
1667/*
1668 * Interface to system's page allocator. No need to hold the cache-lock.
1669 *
1670 * If we requested dmaable memory, we will get it. Even if we
1671 * did not request dmaable memory, we might get it, but that
1672 * would be relatively rare and ignorable.
1673 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001674static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001675{
1676 struct page *page;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001677 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001678 int i;
1679
Luke Yangd6fef9d2006-04-10 22:52:56 -07001680#ifndef CONFIG_MMU
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001681 /*
1682 * Nommu uses slab's for process anonymous memory allocations, and thus
1683 * requires __GFP_COMP to properly refcount higher order allocations
Luke Yangd6fef9d2006-04-10 22:52:56 -07001684 */
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001685 flags |= __GFP_COMP;
Luke Yangd6fef9d2006-04-10 22:52:56 -07001686#endif
Christoph Lameter765c4502006-09-27 01:50:08 -07001687
Christoph Lameter3c517a62006-12-06 20:33:29 -08001688 flags |= cachep->gfpflags;
Mel Gormane12ba742007-10-16 01:25:52 -07001689 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1690 flags |= __GFP_RECLAIMABLE;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001691
1692 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001693 if (!page)
1694 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001695
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001696 nr_pages = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001697 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
Christoph Lameter972d1a72006-09-25 23:31:51 -07001698 add_zone_page_state(page_zone(page),
1699 NR_SLAB_RECLAIMABLE, nr_pages);
1700 else
1701 add_zone_page_state(page_zone(page),
1702 NR_SLAB_UNRECLAIMABLE, nr_pages);
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001703 for (i = 0; i < nr_pages; i++)
1704 __SetPageSlab(page + i);
1705 return page_address(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001706}
1707
1708/*
1709 * Interface to system's page release.
1710 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001711static void kmem_freepages(struct kmem_cache *cachep, void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001712{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001713 unsigned long i = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001714 struct page *page = virt_to_page(addr);
1715 const unsigned long nr_freed = i;
1716
Christoph Lameter972d1a72006-09-25 23:31:51 -07001717 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1718 sub_zone_page_state(page_zone(page),
1719 NR_SLAB_RECLAIMABLE, nr_freed);
1720 else
1721 sub_zone_page_state(page_zone(page),
1722 NR_SLAB_UNRECLAIMABLE, nr_freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001723 while (i--) {
Nick Pigginf205b2f2006-03-22 00:08:02 -08001724 BUG_ON(!PageSlab(page));
1725 __ClearPageSlab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001726 page++;
1727 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001728 if (current->reclaim_state)
1729 current->reclaim_state->reclaimed_slab += nr_freed;
1730 free_pages((unsigned long)addr, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001731}
1732
1733static void kmem_rcu_free(struct rcu_head *head)
1734{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001735 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001736 struct kmem_cache *cachep = slab_rcu->cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001737
1738 kmem_freepages(cachep, slab_rcu->addr);
1739 if (OFF_SLAB(cachep))
1740 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1741}
1742
1743#if DEBUG
1744
1745#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001746static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001747 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001748{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001749 int size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001750
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001751 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001752
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001753 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001754 return;
1755
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001756 *addr++ = 0x12345678;
1757 *addr++ = caller;
1758 *addr++ = smp_processor_id();
1759 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001760 {
1761 unsigned long *sptr = &caller;
1762 unsigned long svalue;
1763
1764 while (!kstack_end(sptr)) {
1765 svalue = *sptr++;
1766 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001767 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001768 size -= sizeof(unsigned long);
1769 if (size <= sizeof(unsigned long))
1770 break;
1771 }
1772 }
1773
1774 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001775 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001776}
1777#endif
1778
Pekka Enberg343e0d72006-02-01 03:05:50 -08001779static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001780{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001781 int size = obj_size(cachep);
1782 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001783
1784 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001785 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001786}
1787
1788static void dump_line(char *data, int offset, int limit)
1789{
1790 int i;
Dave Jonesaa83aa42006-09-29 01:59:51 -07001791 unsigned char error = 0;
1792 int bad_count = 0;
1793
Linus Torvalds1da177e2005-04-16 15:20:36 -07001794 printk(KERN_ERR "%03x:", offset);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001795 for (i = 0; i < limit; i++) {
1796 if (data[offset + i] != POISON_FREE) {
1797 error = data[offset + i];
1798 bad_count++;
1799 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001800 printk(" %02x", (unsigned char)data[offset + i]);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001801 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001802 printk("\n");
Dave Jonesaa83aa42006-09-29 01:59:51 -07001803
1804 if (bad_count == 1) {
1805 error ^= POISON_FREE;
1806 if (!(error & (error - 1))) {
1807 printk(KERN_ERR "Single bit error detected. Probably "
1808 "bad RAM.\n");
1809#ifdef CONFIG_X86
1810 printk(KERN_ERR "Run memtest86+ or a similar memory "
1811 "test tool.\n");
1812#else
1813 printk(KERN_ERR "Run a memory test tool.\n");
1814#endif
1815 }
1816 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001817}
1818#endif
1819
1820#if DEBUG
1821
Pekka Enberg343e0d72006-02-01 03:05:50 -08001822static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001823{
1824 int i, size;
1825 char *realobj;
1826
1827 if (cachep->flags & SLAB_RED_ZONE) {
David Woodhouseb46b8f12007-05-08 00:22:59 -07001828 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001829 *dbg_redzone1(cachep, objp),
1830 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001831 }
1832
1833 if (cachep->flags & SLAB_STORE_USER) {
1834 printk(KERN_ERR "Last user: [<%p>]",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001835 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001836 print_symbol("(%s)",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001837 (unsigned long)*dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001838 printk("\n");
1839 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001840 realobj = (char *)objp + obj_offset(cachep);
1841 size = obj_size(cachep);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001842 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001843 int limit;
1844 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001845 if (i + limit > size)
1846 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001847 dump_line(realobj, i, limit);
1848 }
1849}
1850
Pekka Enberg343e0d72006-02-01 03:05:50 -08001851static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001852{
1853 char *realobj;
1854 int size, i;
1855 int lines = 0;
1856
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001857 realobj = (char *)objp + obj_offset(cachep);
1858 size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001859
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001860 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001861 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001862 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001863 exp = POISON_END;
1864 if (realobj[i] != exp) {
1865 int limit;
1866 /* Mismatch ! */
1867 /* Print header */
1868 if (lines == 0) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001869 printk(KERN_ERR
David Howellse94a40c2007-04-02 23:46:28 +01001870 "Slab corruption: %s start=%p, len=%d\n",
1871 cachep->name, realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001872 print_objinfo(cachep, objp, 0);
1873 }
1874 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001875 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001876 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001877 if (i + limit > size)
1878 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001879 dump_line(realobj, i, limit);
1880 i += 16;
1881 lines++;
1882 /* Limit to 5 lines */
1883 if (lines > 5)
1884 break;
1885 }
1886 }
1887 if (lines != 0) {
1888 /* Print some data about the neighboring objects, if they
1889 * exist:
1890 */
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08001891 struct slab *slabp = virt_to_slab(objp);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001892 unsigned int objnr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001893
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001894 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001895 if (objnr) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001896 objp = index_to_obj(cachep, slabp, objnr - 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001897 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001898 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001899 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001900 print_objinfo(cachep, objp, 2);
1901 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001902 if (objnr + 1 < cachep->num) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001903 objp = index_to_obj(cachep, slabp, objnr + 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001904 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001905 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001906 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001907 print_objinfo(cachep, objp, 2);
1908 }
1909 }
1910}
1911#endif
1912
Linus Torvalds1da177e2005-04-16 15:20:36 -07001913#if DEBUG
Rabin Vincente79aec22008-07-04 00:40:32 +05301914static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001915{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001916 int i;
1917 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001918 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001919
1920 if (cachep->flags & SLAB_POISON) {
1921#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08001922 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1923 OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001924 kernel_map_pages(virt_to_page(objp),
Andrew Mortona737b3e2006-03-22 00:08:11 -08001925 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001926 else
1927 check_poison_obj(cachep, objp);
1928#else
1929 check_poison_obj(cachep, objp);
1930#endif
1931 }
1932 if (cachep->flags & SLAB_RED_ZONE) {
1933 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1934 slab_error(cachep, "start of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001935 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001936 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1937 slab_error(cachep, "end of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001938 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001939 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001940 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001941}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001942#else
Rabin Vincente79aec22008-07-04 00:40:32 +05301943static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001944{
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001945}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001946#endif
1947
Randy Dunlap911851e2006-03-22 00:08:14 -08001948/**
1949 * slab_destroy - destroy and release all objects in a slab
1950 * @cachep: cache pointer being destroyed
1951 * @slabp: slab pointer being destroyed
1952 *
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001953 * Destroy all the objs in a slab, and release the mem back to the system.
Andrew Mortona737b3e2006-03-22 00:08:11 -08001954 * Before calling the slab must have been unlinked from the cache. The
1955 * cache-lock is not held/needed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001956 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001957static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001958{
1959 void *addr = slabp->s_mem - slabp->colouroff;
1960
Rabin Vincente79aec22008-07-04 00:40:32 +05301961 slab_destroy_debugcheck(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001962 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1963 struct slab_rcu *slab_rcu;
1964
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001965 slab_rcu = (struct slab_rcu *)slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001966 slab_rcu->cachep = cachep;
1967 slab_rcu->addr = addr;
1968 call_rcu(&slab_rcu->head, kmem_rcu_free);
1969 } else {
1970 kmem_freepages(cachep, addr);
Ingo Molnar873623d2006-07-13 14:44:38 +02001971 if (OFF_SLAB(cachep))
1972 kmem_cache_free(cachep->slabp_cache, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001973 }
1974}
1975
Christoph Lameter117f6eb2006-09-25 23:31:37 -07001976static void __kmem_cache_destroy(struct kmem_cache *cachep)
1977{
1978 int i;
1979 struct kmem_list3 *l3;
1980
1981 for_each_online_cpu(i)
1982 kfree(cachep->array[i]);
1983
1984 /* NUMA: free the list3 structures */
1985 for_each_online_node(i) {
1986 l3 = cachep->nodelists[i];
1987 if (l3) {
1988 kfree(l3->shared);
1989 free_alien_cache(l3->alien);
1990 kfree(l3);
1991 }
1992 }
1993 kmem_cache_free(&cache_cache, cachep);
1994}
1995
1996
Linus Torvalds1da177e2005-04-16 15:20:36 -07001997/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08001998 * calculate_slab_order - calculate size (page order) of slabs
1999 * @cachep: pointer to the cache that is being created
2000 * @size: size of objects to be created in this cache.
2001 * @align: required alignment for the objects.
2002 * @flags: slab allocation flags
2003 *
2004 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002005 *
2006 * This could be made much more intelligent. For now, try to avoid using
2007 * high order pages for slabs. When the gfp() functions are more friendly
2008 * towards high-order requests, this should be changed.
2009 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002010static size_t calculate_slab_order(struct kmem_cache *cachep,
Randy Dunlapee13d782006-02-01 03:05:53 -08002011 size_t size, size_t align, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002012{
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002013 unsigned long offslab_limit;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002014 size_t left_over = 0;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002015 int gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002016
Christoph Lameter0aa817f2007-05-16 22:11:01 -07002017 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002018 unsigned int num;
2019 size_t remainder;
2020
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002021 cache_estimate(gfporder, size, align, flags, &remainder, &num);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002022 if (!num)
2023 continue;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002024
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002025 if (flags & CFLGS_OFF_SLAB) {
2026 /*
2027 * Max number of objs-per-slab for caches which
2028 * use off-slab slabs. Needed to avoid a possible
2029 * looping condition in cache_grow().
2030 */
2031 offslab_limit = size - sizeof(struct slab);
2032 offslab_limit /= sizeof(kmem_bufctl_t);
2033
2034 if (num > offslab_limit)
2035 break;
2036 }
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002037
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002038 /* Found something acceptable - save it away */
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002039 cachep->num = num;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002040 cachep->gfporder = gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002041 left_over = remainder;
2042
2043 /*
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002044 * A VFS-reclaimable slab tends to have most allocations
2045 * as GFP_NOFS and we really don't want to have to be allocating
2046 * higher-order pages when we are unable to shrink dcache.
2047 */
2048 if (flags & SLAB_RECLAIM_ACCOUNT)
2049 break;
2050
2051 /*
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002052 * Large number of objects is good, but very large slabs are
2053 * currently bad for the gfp()s.
2054 */
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002055 if (gfporder >= slab_break_gfp_order)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002056 break;
2057
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002058 /*
2059 * Acceptable internal fragmentation?
2060 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002061 if (left_over * 8 <= (PAGE_SIZE << gfporder))
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002062 break;
2063 }
2064 return left_over;
2065}
2066
Sam Ravnborg38bdc322007-05-17 23:48:19 +02002067static int __init_refok setup_cpu_cache(struct kmem_cache *cachep)
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002068{
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002069 if (g_cpucache_up == FULL)
2070 return enable_cpucache(cachep);
2071
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002072 if (g_cpucache_up == NONE) {
2073 /*
2074 * Note: the first kmem_cache_create must create the cache
2075 * that's used by kmalloc(24), otherwise the creation of
2076 * further caches will BUG().
2077 */
2078 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2079
2080 /*
2081 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2082 * the first cache, then we need to set up all its list3s,
2083 * otherwise the creation of further caches will BUG().
2084 */
2085 set_up_list3s(cachep, SIZE_AC);
2086 if (INDEX_AC == INDEX_L3)
2087 g_cpucache_up = PARTIAL_L3;
2088 else
2089 g_cpucache_up = PARTIAL_AC;
2090 } else {
2091 cachep->array[smp_processor_id()] =
2092 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
2093
2094 if (g_cpucache_up == PARTIAL_AC) {
2095 set_up_list3s(cachep, SIZE_L3);
2096 g_cpucache_up = PARTIAL_L3;
2097 } else {
2098 int node;
Pekka Enberg556a1692008-01-25 08:20:51 +02002099 for_each_online_node(node) {
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002100 cachep->nodelists[node] =
2101 kmalloc_node(sizeof(struct kmem_list3),
2102 GFP_KERNEL, node);
2103 BUG_ON(!cachep->nodelists[node]);
2104 kmem_list3_init(cachep->nodelists[node]);
2105 }
2106 }
2107 }
2108 cachep->nodelists[numa_node_id()]->next_reap =
2109 jiffies + REAPTIMEOUT_LIST3 +
2110 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2111
2112 cpu_cache_get(cachep)->avail = 0;
2113 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2114 cpu_cache_get(cachep)->batchcount = 1;
2115 cpu_cache_get(cachep)->touched = 0;
2116 cachep->batchcount = 1;
2117 cachep->limit = BOOT_CPUCACHE_ENTRIES;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002118 return 0;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002119}
2120
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002121/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07002122 * kmem_cache_create - Create a cache.
2123 * @name: A string which is used in /proc/slabinfo to identify this cache.
2124 * @size: The size of objects to be created in this cache.
2125 * @align: The required alignment for the objects.
2126 * @flags: SLAB flags
2127 * @ctor: A constructor for the objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002128 *
2129 * Returns a ptr to the cache on success, NULL on failure.
2130 * Cannot be called within a int, but can be interrupted.
Paul Mundt20c2df82007-07-20 10:11:58 +09002131 * The @ctor is run when new pages are allocated by the cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002132 *
2133 * @name must be valid until the cache is destroyed. This implies that
Andrew Mortona737b3e2006-03-22 00:08:11 -08002134 * the module calling this has to destroy the cache before getting unloaded.
Catalin Marinas249da162008-11-21 12:56:22 +00002135 * Note that kmem_cache_name() is not guaranteed to return the same pointer,
2136 * therefore applications must manage it themselves.
Andrew Mortona737b3e2006-03-22 00:08:11 -08002137 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002138 * The flags are
2139 *
2140 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2141 * to catch references to uninitialised memory.
2142 *
2143 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2144 * for buffer overruns.
2145 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002146 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2147 * cacheline. This can be beneficial if you're counting cycles as closely
2148 * as davem.
2149 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002150struct kmem_cache *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002151kmem_cache_create (const char *name, size_t size, size_t align,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002152 unsigned long flags, void (*ctor)(void *))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002153{
2154 size_t left_over, slab_size, ralign;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07002155 struct kmem_cache *cachep = NULL, *pc;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002156
2157 /*
2158 * Sanity checks... these are all serious usage bugs.
2159 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002160 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
Paul Mundt20c2df82007-07-20 10:11:58 +09002161 size > KMALLOC_MAX_SIZE) {
Harvey Harrisond40cee22008-04-30 00:55:07 -07002162 printk(KERN_ERR "%s: Early error in slab %s\n", __func__,
Andrew Mortona737b3e2006-03-22 00:08:11 -08002163 name);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002164 BUG();
2165 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002166
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002167 /*
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002168 * We use cache_chain_mutex to ensure a consistent view of
Rusty Russell174596a2009-01-01 10:12:29 +10302169 * cpu_online_mask as well. Please see cpuup_callback
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002170 */
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002171 get_online_cpus();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002172 mutex_lock(&cache_chain_mutex);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002173
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07002174 list_for_each_entry(pc, &cache_chain, next) {
Andrew Morton4f12bb42005-11-07 00:58:00 -08002175 char tmp;
2176 int res;
2177
2178 /*
2179 * This happens when the module gets unloaded and doesn't
2180 * destroy its slab cache and no-one else reuses the vmalloc
2181 * area of the module. Print a warning.
2182 */
Andrew Morton138ae662006-12-06 20:36:41 -08002183 res = probe_kernel_address(pc->name, tmp);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002184 if (res) {
matzeb4169522007-05-06 14:49:52 -07002185 printk(KERN_ERR
2186 "SLAB: cache with size %d has lost its name\n",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002187 pc->buffer_size);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002188 continue;
2189 }
2190
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002191 if (!strcmp(pc->name, name)) {
matzeb4169522007-05-06 14:49:52 -07002192 printk(KERN_ERR
2193 "kmem_cache_create: duplicate cache %s\n", name);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002194 dump_stack();
2195 goto oops;
2196 }
2197 }
2198
Linus Torvalds1da177e2005-04-16 15:20:36 -07002199#if DEBUG
2200 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002201#if FORCED_DEBUG
2202 /*
2203 * Enable redzoning and last user accounting, except for caches with
2204 * large objects, if the increased size would increase the object size
2205 * above the next power of two: caches with object sizes just above a
2206 * power of two have a significant amount of internal fragmentation.
2207 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002208 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2209 2 * sizeof(unsigned long long)))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002210 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002211 if (!(flags & SLAB_DESTROY_BY_RCU))
2212 flags |= SLAB_POISON;
2213#endif
2214 if (flags & SLAB_DESTROY_BY_RCU)
2215 BUG_ON(flags & SLAB_POISON);
2216#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002217 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002218 * Always checks flags, a caller might be expecting debug support which
2219 * isn't available.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002220 */
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002221 BUG_ON(flags & ~CREATE_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002222
Andrew Mortona737b3e2006-03-22 00:08:11 -08002223 /*
2224 * Check that size is in terms of words. This is needed to avoid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002225 * unaligned accesses for some archs when redzoning is used, and makes
2226 * sure any on-slab bufctl's are also correctly aligned.
2227 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002228 if (size & (BYTES_PER_WORD - 1)) {
2229 size += (BYTES_PER_WORD - 1);
2230 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002231 }
2232
Andrew Mortona737b3e2006-03-22 00:08:11 -08002233 /* calculate the final buffer alignment: */
2234
Linus Torvalds1da177e2005-04-16 15:20:36 -07002235 /* 1) arch recommendation: can be overridden for debug */
2236 if (flags & SLAB_HWCACHE_ALIGN) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002237 /*
2238 * Default alignment: as specified by the arch code. Except if
2239 * an object is really small, then squeeze multiple objects into
2240 * one cacheline.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002241 */
2242 ralign = cache_line_size();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002243 while (size <= ralign / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002244 ralign /= 2;
2245 } else {
2246 ralign = BYTES_PER_WORD;
2247 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002248
2249 /*
David Woodhouse87a927c2007-07-04 21:26:44 -04002250 * Redzoning and user store require word alignment or possibly larger.
2251 * Note this will be overridden by architecture or caller mandated
2252 * alignment if either is greater than BYTES_PER_WORD.
Pekka Enbergca5f9702006-09-25 23:31:25 -07002253 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002254 if (flags & SLAB_STORE_USER)
2255 ralign = BYTES_PER_WORD;
2256
2257 if (flags & SLAB_RED_ZONE) {
2258 ralign = REDZONE_ALIGN;
2259 /* If redzoning, ensure that the second redzone is suitably
2260 * aligned, by adjusting the object size accordingly. */
2261 size += REDZONE_ALIGN - 1;
2262 size &= ~(REDZONE_ALIGN - 1);
2263 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002264
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002265 /* 2) arch mandated alignment */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002266 if (ralign < ARCH_SLAB_MINALIGN) {
2267 ralign = ARCH_SLAB_MINALIGN;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002268 }
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002269 /* 3) caller mandated alignment */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002270 if (ralign < align) {
2271 ralign = align;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002272 }
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002273 /* disable debug if necessary */
David Woodhouseb46b8f12007-05-08 00:22:59 -07002274 if (ralign > __alignof__(unsigned long long))
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002275 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002276 /*
Pekka Enbergca5f9702006-09-25 23:31:25 -07002277 * 4) Store it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002278 */
2279 align = ralign;
2280
2281 /* Get cache's description obj. */
Christoph Lametere94b1762006-12-06 20:33:17 -08002282 cachep = kmem_cache_zalloc(&cache_cache, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002283 if (!cachep)
Andrew Morton4f12bb42005-11-07 00:58:00 -08002284 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002285
2286#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002287 cachep->obj_size = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002288
Pekka Enbergca5f9702006-09-25 23:31:25 -07002289 /*
2290 * Both debugging options require word-alignment which is calculated
2291 * into align above.
2292 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002293 if (flags & SLAB_RED_ZONE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002294 /* add space for red zone words */
David Woodhouseb46b8f12007-05-08 00:22:59 -07002295 cachep->obj_offset += sizeof(unsigned long long);
2296 size += 2 * sizeof(unsigned long long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002297 }
2298 if (flags & SLAB_STORE_USER) {
Pekka Enbergca5f9702006-09-25 23:31:25 -07002299 /* user store requires one word storage behind the end of
David Woodhouse87a927c2007-07-04 21:26:44 -04002300 * the real object. But if the second red zone needs to be
2301 * aligned to 64 bits, we must allow that much space.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002302 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002303 if (flags & SLAB_RED_ZONE)
2304 size += REDZONE_ALIGN;
2305 else
2306 size += BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002307 }
2308#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002309 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002310 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2311 cachep->obj_offset += PAGE_SIZE - size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002312 size = PAGE_SIZE;
2313 }
2314#endif
2315#endif
2316
Ingo Molnare0a42722006-06-23 02:03:46 -07002317 /*
2318 * Determine if the slab management is 'on' or 'off' slab.
2319 * (bootstrapping cannot cope with offslab caches so don't do
2320 * it too early on.)
2321 */
2322 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002323 /*
2324 * Size is large, assume best to place the slab management obj
2325 * off-slab (should allow better packing of objs).
2326 */
2327 flags |= CFLGS_OFF_SLAB;
2328
2329 size = ALIGN(size, align);
2330
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002331 left_over = calculate_slab_order(cachep, size, align, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002332
2333 if (!cachep->num) {
matzeb4169522007-05-06 14:49:52 -07002334 printk(KERN_ERR
2335 "kmem_cache_create: couldn't create cache %s.\n", name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002336 kmem_cache_free(&cache_cache, cachep);
2337 cachep = NULL;
Andrew Morton4f12bb42005-11-07 00:58:00 -08002338 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002339 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002340 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2341 + sizeof(struct slab), align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002342
2343 /*
2344 * If the slab has been placed off-slab, and we have enough space then
2345 * move it on-slab. This is at the expense of any extra colouring.
2346 */
2347 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2348 flags &= ~CFLGS_OFF_SLAB;
2349 left_over -= slab_size;
2350 }
2351
2352 if (flags & CFLGS_OFF_SLAB) {
2353 /* really off slab. No need for manual alignment */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002354 slab_size =
2355 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002356 }
2357
2358 cachep->colour_off = cache_line_size();
2359 /* Offset must be a multiple of the alignment. */
2360 if (cachep->colour_off < align)
2361 cachep->colour_off = align;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002362 cachep->colour = left_over / cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002363 cachep->slab_size = slab_size;
2364 cachep->flags = flags;
2365 cachep->gfpflags = 0;
Christoph Lameter4b51d662007-02-10 01:43:10 -08002366 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002367 cachep->gfpflags |= GFP_DMA;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002368 cachep->buffer_size = size;
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08002369 cachep->reciprocal_buffer_size = reciprocal_value(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002370
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002371 if (flags & CFLGS_OFF_SLAB) {
Victor Fuscob2d55072005-09-10 00:26:36 -07002372 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002373 /*
2374 * This is a possibility for one of the malloc_sizes caches.
2375 * But since we go off slab only for object size greater than
2376 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2377 * this should not happen at all.
2378 * But leave a BUG_ON for some lucky dude.
2379 */
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002380 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002381 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002382 cachep->ctor = ctor;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002383 cachep->name = name;
2384
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002385 if (setup_cpu_cache(cachep)) {
2386 __kmem_cache_destroy(cachep);
2387 cachep = NULL;
2388 goto oops;
2389 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002390
Linus Torvalds1da177e2005-04-16 15:20:36 -07002391 /* cache setup completed, link it into the list */
2392 list_add(&cachep->next, &cache_chain);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002393oops:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002394 if (!cachep && (flags & SLAB_PANIC))
2395 panic("kmem_cache_create(): failed to create slab `%s'\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002396 name);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002397 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002398 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002399 return cachep;
2400}
2401EXPORT_SYMBOL(kmem_cache_create);
2402
2403#if DEBUG
2404static void check_irq_off(void)
2405{
2406 BUG_ON(!irqs_disabled());
2407}
2408
2409static void check_irq_on(void)
2410{
2411 BUG_ON(irqs_disabled());
2412}
2413
Pekka Enberg343e0d72006-02-01 03:05:50 -08002414static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002415{
2416#ifdef CONFIG_SMP
2417 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002418 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002419#endif
2420}
Christoph Lametere498be72005-09-09 13:03:32 -07002421
Pekka Enberg343e0d72006-02-01 03:05:50 -08002422static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002423{
2424#ifdef CONFIG_SMP
2425 check_irq_off();
2426 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2427#endif
2428}
2429
Linus Torvalds1da177e2005-04-16 15:20:36 -07002430#else
2431#define check_irq_off() do { } while(0)
2432#define check_irq_on() do { } while(0)
2433#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002434#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002435#endif
2436
Christoph Lameteraab22072006-03-22 00:09:06 -08002437static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2438 struct array_cache *ac,
2439 int force, int node);
2440
Linus Torvalds1da177e2005-04-16 15:20:36 -07002441static void do_drain(void *arg)
2442{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002443 struct kmem_cache *cachep = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002444 struct array_cache *ac;
Christoph Lameterff694162005-09-22 21:44:02 -07002445 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002446
2447 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002448 ac = cpu_cache_get(cachep);
Christoph Lameterff694162005-09-22 21:44:02 -07002449 spin_lock(&cachep->nodelists[node]->list_lock);
2450 free_block(cachep, ac->entry, ac->avail, node);
2451 spin_unlock(&cachep->nodelists[node]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002452 ac->avail = 0;
2453}
2454
Pekka Enberg343e0d72006-02-01 03:05:50 -08002455static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002456{
Christoph Lametere498be72005-09-09 13:03:32 -07002457 struct kmem_list3 *l3;
2458 int node;
2459
Jens Axboe15c8b6c2008-05-09 09:39:44 +02002460 on_each_cpu(do_drain, cachep, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002461 check_irq_on();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002462 for_each_online_node(node) {
Christoph Lametere498be72005-09-09 13:03:32 -07002463 l3 = cachep->nodelists[node];
Roland Dreiera4523a82006-05-15 11:41:00 -07002464 if (l3 && l3->alien)
2465 drain_alien_cache(cachep, l3->alien);
2466 }
2467
2468 for_each_online_node(node) {
2469 l3 = cachep->nodelists[node];
2470 if (l3)
Christoph Lameteraab22072006-03-22 00:09:06 -08002471 drain_array(cachep, l3, l3->shared, 1, node);
Christoph Lametere498be72005-09-09 13:03:32 -07002472 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002473}
2474
Christoph Lametered11d9e2006-06-30 01:55:45 -07002475/*
2476 * Remove slabs from the list of free slabs.
2477 * Specify the number of slabs to drain in tofree.
2478 *
2479 * Returns the actual number of slabs released.
2480 */
2481static int drain_freelist(struct kmem_cache *cache,
2482 struct kmem_list3 *l3, int tofree)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002483{
Christoph Lametered11d9e2006-06-30 01:55:45 -07002484 struct list_head *p;
2485 int nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002486 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002487
Christoph Lametered11d9e2006-06-30 01:55:45 -07002488 nr_freed = 0;
2489 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002490
Christoph Lametered11d9e2006-06-30 01:55:45 -07002491 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002492 p = l3->slabs_free.prev;
Christoph Lametered11d9e2006-06-30 01:55:45 -07002493 if (p == &l3->slabs_free) {
2494 spin_unlock_irq(&l3->list_lock);
2495 goto out;
2496 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002497
Christoph Lametered11d9e2006-06-30 01:55:45 -07002498 slabp = list_entry(p, struct slab, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002499#if DEBUG
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002500 BUG_ON(slabp->inuse);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002501#endif
2502 list_del(&slabp->list);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002503 /*
2504 * Safe to drop the lock. The slab is no longer linked
2505 * to the cache.
2506 */
2507 l3->free_objects -= cache->num;
Christoph Lametere498be72005-09-09 13:03:32 -07002508 spin_unlock_irq(&l3->list_lock);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002509 slab_destroy(cache, slabp);
2510 nr_freed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002511 }
Christoph Lametered11d9e2006-06-30 01:55:45 -07002512out:
2513 return nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002514}
2515
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002516/* Called with cache_chain_mutex held to protect against cpu hotplug */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002517static int __cache_shrink(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07002518{
2519 int ret = 0, i = 0;
2520 struct kmem_list3 *l3;
2521
2522 drain_cpu_caches(cachep);
2523
2524 check_irq_on();
2525 for_each_online_node(i) {
2526 l3 = cachep->nodelists[i];
Christoph Lametered11d9e2006-06-30 01:55:45 -07002527 if (!l3)
2528 continue;
2529
2530 drain_freelist(cachep, l3, l3->free_objects);
2531
2532 ret += !list_empty(&l3->slabs_full) ||
2533 !list_empty(&l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07002534 }
2535 return (ret ? 1 : 0);
2536}
2537
Linus Torvalds1da177e2005-04-16 15:20:36 -07002538/**
2539 * kmem_cache_shrink - Shrink a cache.
2540 * @cachep: The cache to shrink.
2541 *
2542 * Releases as many slabs as possible for a cache.
2543 * To help debugging, a zero exit status indicates all slabs were released.
2544 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002545int kmem_cache_shrink(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002546{
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002547 int ret;
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002548 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002549
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002550 get_online_cpus();
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002551 mutex_lock(&cache_chain_mutex);
2552 ret = __cache_shrink(cachep);
2553 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002554 put_online_cpus();
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002555 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002556}
2557EXPORT_SYMBOL(kmem_cache_shrink);
2558
2559/**
2560 * kmem_cache_destroy - delete a cache
2561 * @cachep: the cache to destroy
2562 *
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08002563 * Remove a &struct kmem_cache object from the slab cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002564 *
2565 * It is expected this function will be called by a module when it is
2566 * unloaded. This will remove the cache completely, and avoid a duplicate
2567 * cache being allocated each time a module is loaded and unloaded, if the
2568 * module doesn't have persistent in-kernel storage across loads and unloads.
2569 *
2570 * The cache must be empty before calling this function.
2571 *
2572 * The caller must guarantee that noone will allocate memory from the cache
2573 * during the kmem_cache_destroy().
2574 */
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002575void kmem_cache_destroy(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002576{
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002577 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002578
Linus Torvalds1da177e2005-04-16 15:20:36 -07002579 /* Find the cache in the chain of caches. */
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002580 get_online_cpus();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002581 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002582 /*
2583 * the chain is never empty, cache_cache is never destroyed
2584 */
2585 list_del(&cachep->next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002586 if (__cache_shrink(cachep)) {
2587 slab_error(cachep, "Can't free all objects");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002588 list_add(&cachep->next, &cache_chain);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002589 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002590 put_online_cpus();
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002591 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002592 }
2593
2594 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
Paul E. McKenneyfbd568a3e2005-05-01 08:59:04 -07002595 synchronize_rcu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002596
Christoph Lameter117f6eb2006-09-25 23:31:37 -07002597 __kmem_cache_destroy(cachep);
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002598 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002599 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002600}
2601EXPORT_SYMBOL(kmem_cache_destroy);
2602
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002603/*
2604 * Get the memory for a slab management obj.
2605 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2606 * always come from malloc_sizes caches. The slab descriptor cannot
2607 * come from the same cache which is getting created because,
2608 * when we are searching for an appropriate cache for these
2609 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2610 * If we are creating a malloc_sizes cache here it would not be visible to
2611 * kmem_find_general_cachep till the initialization is complete.
2612 * Hence we cannot have slabp_cache same as the original cache.
2613 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002614static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002615 int colour_off, gfp_t local_flags,
2616 int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002617{
2618 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002619
Linus Torvalds1da177e2005-04-16 15:20:36 -07002620 if (OFF_SLAB(cachep)) {
2621 /* Slab management obj is off-slab. */
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002622 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
Pekka Enberg8759ec52008-11-26 10:01:31 +02002623 local_flags, nodeid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002624 if (!slabp)
2625 return NULL;
2626 } else {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002627 slabp = objp + colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002628 colour_off += cachep->slab_size;
2629 }
2630 slabp->inuse = 0;
2631 slabp->colouroff = colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002632 slabp->s_mem = objp + colour_off;
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002633 slabp->nodeid = nodeid;
Marcin Slusarze51bfd02008-02-10 11:21:54 +01002634 slabp->free = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002635 return slabp;
2636}
2637
2638static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2639{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002640 return (kmem_bufctl_t *) (slabp + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002641}
2642
Pekka Enberg343e0d72006-02-01 03:05:50 -08002643static void cache_init_objs(struct kmem_cache *cachep,
Christoph Lametera35afb82007-05-16 22:10:57 -07002644 struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002645{
2646 int i;
2647
2648 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002649 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002650#if DEBUG
2651 /* need to poison the objs? */
2652 if (cachep->flags & SLAB_POISON)
2653 poison_obj(cachep, objp, POISON_FREE);
2654 if (cachep->flags & SLAB_STORE_USER)
2655 *dbg_userword(cachep, objp) = NULL;
2656
2657 if (cachep->flags & SLAB_RED_ZONE) {
2658 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2659 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2660 }
2661 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002662 * Constructors are not allowed to allocate memory from the same
2663 * cache which they are a constructor for. Otherwise, deadlock.
2664 * They must also be threaded.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002665 */
2666 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002667 cachep->ctor(objp + obj_offset(cachep));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002668
2669 if (cachep->flags & SLAB_RED_ZONE) {
2670 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2671 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002672 " end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002673 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2674 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002675 " start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002676 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08002677 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2678 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002679 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002680 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002681#else
2682 if (cachep->ctor)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002683 cachep->ctor(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002684#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002685 slab_bufctl(slabp)[i] = i + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002686 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002687 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002688}
2689
Pekka Enberg343e0d72006-02-01 03:05:50 -08002690static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002691{
Christoph Lameter4b51d662007-02-10 01:43:10 -08002692 if (CONFIG_ZONE_DMA_FLAG) {
2693 if (flags & GFP_DMA)
2694 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2695 else
2696 BUG_ON(cachep->gfpflags & GFP_DMA);
2697 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002698}
2699
Andrew Mortona737b3e2006-03-22 00:08:11 -08002700static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2701 int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002702{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002703 void *objp = index_to_obj(cachep, slabp, slabp->free);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002704 kmem_bufctl_t next;
2705
2706 slabp->inuse++;
2707 next = slab_bufctl(slabp)[slabp->free];
2708#if DEBUG
2709 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2710 WARN_ON(slabp->nodeid != nodeid);
2711#endif
2712 slabp->free = next;
2713
2714 return objp;
2715}
2716
Andrew Mortona737b3e2006-03-22 00:08:11 -08002717static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2718 void *objp, int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002719{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002720 unsigned int objnr = obj_to_index(cachep, slabp, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002721
2722#if DEBUG
2723 /* Verify that the slab belongs to the intended node */
2724 WARN_ON(slabp->nodeid != nodeid);
2725
Al Viro871751e2006-03-25 03:06:39 -08002726 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
Matthew Dobson78d382d2006-02-01 03:05:47 -08002727 printk(KERN_ERR "slab: double free detected in cache "
Andrew Mortona737b3e2006-03-22 00:08:11 -08002728 "'%s', objp %p\n", cachep->name, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002729 BUG();
2730 }
2731#endif
2732 slab_bufctl(slabp)[objnr] = slabp->free;
2733 slabp->free = objnr;
2734 slabp->inuse--;
2735}
2736
Pekka Enberg47768742006-06-23 02:03:07 -07002737/*
2738 * Map pages beginning at addr to the given cache and slab. This is required
2739 * for the slab allocator to be able to lookup the cache and slab of a
2740 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2741 */
2742static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2743 void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002744{
Pekka Enberg47768742006-06-23 02:03:07 -07002745 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002746 struct page *page;
2747
Pekka Enberg47768742006-06-23 02:03:07 -07002748 page = virt_to_page(addr);
Nick Piggin84097512006-03-22 00:08:34 -08002749
Pekka Enberg47768742006-06-23 02:03:07 -07002750 nr_pages = 1;
Nick Piggin84097512006-03-22 00:08:34 -08002751 if (likely(!PageCompound(page)))
Pekka Enberg47768742006-06-23 02:03:07 -07002752 nr_pages <<= cache->gfporder;
2753
Linus Torvalds1da177e2005-04-16 15:20:36 -07002754 do {
Pekka Enberg47768742006-06-23 02:03:07 -07002755 page_set_cache(page, cache);
2756 page_set_slab(page, slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002757 page++;
Pekka Enberg47768742006-06-23 02:03:07 -07002758 } while (--nr_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002759}
2760
2761/*
2762 * Grow (by 1) the number of slabs within a cache. This is called by
2763 * kmem_cache_alloc() when there are no active objs left in a cache.
2764 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002765static int cache_grow(struct kmem_cache *cachep,
2766 gfp_t flags, int nodeid, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002767{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002768 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002769 size_t offset;
2770 gfp_t local_flags;
Christoph Lametere498be72005-09-09 13:03:32 -07002771 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002772
Andrew Mortona737b3e2006-03-22 00:08:11 -08002773 /*
2774 * Be lazy and only check for valid flags here, keeping it out of the
2775 * critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002776 */
Christoph Lameter6cb06222007-10-16 01:25:41 -07002777 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2778 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002779
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002780 /* Take the l3 list lock to change the colour_next on this node */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002781 check_irq_off();
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002782 l3 = cachep->nodelists[nodeid];
2783 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002784
2785 /* Get colour for the slab, and cal the next value. */
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002786 offset = l3->colour_next;
2787 l3->colour_next++;
2788 if (l3->colour_next >= cachep->colour)
2789 l3->colour_next = 0;
2790 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002791
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002792 offset *= cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002793
2794 if (local_flags & __GFP_WAIT)
2795 local_irq_enable();
2796
2797 /*
2798 * The test for missing atomic flag is performed here, rather than
2799 * the more obvious place, simply to reduce the critical path length
2800 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2801 * will eventually be caught here (where it matters).
2802 */
2803 kmem_flagcheck(cachep, flags);
2804
Andrew Mortona737b3e2006-03-22 00:08:11 -08002805 /*
2806 * Get mem for the objs. Attempt to allocate a physical page from
2807 * 'nodeid'.
Christoph Lametere498be72005-09-09 13:03:32 -07002808 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002809 if (!objp)
Andrew Mortonb8c1c5d2007-07-24 12:02:40 -07002810 objp = kmem_getpages(cachep, local_flags, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002811 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002812 goto failed;
2813
2814 /* Get slab management. */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002815 slabp = alloc_slabmgmt(cachep, objp, offset,
Christoph Lameter6cb06222007-10-16 01:25:41 -07002816 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002817 if (!slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002818 goto opps1;
2819
Pekka Enberg47768742006-06-23 02:03:07 -07002820 slab_map_pages(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002821
Christoph Lametera35afb82007-05-16 22:10:57 -07002822 cache_init_objs(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002823
2824 if (local_flags & __GFP_WAIT)
2825 local_irq_disable();
2826 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002827 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002828
2829 /* Make slab active. */
Christoph Lametere498be72005-09-09 13:03:32 -07002830 list_add_tail(&slabp->list, &(l3->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002831 STATS_INC_GROWN(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002832 l3->free_objects += cachep->num;
2833 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002834 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002835opps1:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002836 kmem_freepages(cachep, objp);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002837failed:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002838 if (local_flags & __GFP_WAIT)
2839 local_irq_disable();
2840 return 0;
2841}
2842
2843#if DEBUG
2844
2845/*
2846 * Perform extra freeing checks:
2847 * - detect bad pointers.
2848 * - POISON/RED_ZONE checking
Linus Torvalds1da177e2005-04-16 15:20:36 -07002849 */
2850static void kfree_debugcheck(const void *objp)
2851{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002852 if (!virt_addr_valid(objp)) {
2853 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002854 (unsigned long)objp);
2855 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002856 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002857}
2858
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002859static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2860{
David Woodhouseb46b8f12007-05-08 00:22:59 -07002861 unsigned long long redzone1, redzone2;
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002862
2863 redzone1 = *dbg_redzone1(cache, obj);
2864 redzone2 = *dbg_redzone2(cache, obj);
2865
2866 /*
2867 * Redzone is ok.
2868 */
2869 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2870 return;
2871
2872 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2873 slab_error(cache, "double free detected");
2874 else
2875 slab_error(cache, "memory outside object was overwritten");
2876
David Woodhouseb46b8f12007-05-08 00:22:59 -07002877 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002878 obj, redzone1, redzone2);
2879}
2880
Pekka Enberg343e0d72006-02-01 03:05:50 -08002881static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002882 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002883{
2884 struct page *page;
2885 unsigned int objnr;
2886 struct slab *slabp;
2887
Matthew Wilcox80cbd912007-11-29 12:05:13 -07002888 BUG_ON(virt_to_cache(objp) != cachep);
2889
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002890 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002891 kfree_debugcheck(objp);
Christoph Lameterb49af682007-05-06 14:49:41 -07002892 page = virt_to_head_page(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002893
Pekka Enberg065d41c2005-11-13 16:06:46 -08002894 slabp = page_get_slab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002895
2896 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002897 verify_redzone_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002898 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2899 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2900 }
2901 if (cachep->flags & SLAB_STORE_USER)
2902 *dbg_userword(cachep, objp) = caller;
2903
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002904 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002905
2906 BUG_ON(objnr >= cachep->num);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002907 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002908
Al Viro871751e2006-03-25 03:06:39 -08002909#ifdef CONFIG_DEBUG_SLAB_LEAK
2910 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2911#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002912 if (cachep->flags & SLAB_POISON) {
2913#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08002914 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002915 store_stackinfo(cachep, objp, (unsigned long)caller);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002916 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002917 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002918 } else {
2919 poison_obj(cachep, objp, POISON_FREE);
2920 }
2921#else
2922 poison_obj(cachep, objp, POISON_FREE);
2923#endif
2924 }
2925 return objp;
2926}
2927
Pekka Enberg343e0d72006-02-01 03:05:50 -08002928static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002929{
2930 kmem_bufctl_t i;
2931 int entries = 0;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002932
Linus Torvalds1da177e2005-04-16 15:20:36 -07002933 /* Check slab's freelist to see if this obj is there. */
2934 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2935 entries++;
2936 if (entries > cachep->num || i >= cachep->num)
2937 goto bad;
2938 }
2939 if (entries != cachep->num - slabp->inuse) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002940bad:
2941 printk(KERN_ERR "slab: Internal list corruption detected in "
2942 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2943 cachep->name, cachep->num, slabp, slabp->inuse);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002944 for (i = 0;
Linus Torvalds264132b2006-03-06 12:10:07 -08002945 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002946 i++) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002947 if (i % 16 == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002948 printk("\n%03x:", i);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002949 printk(" %02x", ((unsigned char *)slabp)[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002950 }
2951 printk("\n");
2952 BUG();
2953 }
2954}
2955#else
2956#define kfree_debugcheck(x) do { } while(0)
2957#define cache_free_debugcheck(x,objp,z) (objp)
2958#define check_slabp(x,y) do { } while(0)
2959#endif
2960
Pekka Enberg343e0d72006-02-01 03:05:50 -08002961static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002962{
2963 int batchcount;
2964 struct kmem_list3 *l3;
2965 struct array_cache *ac;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002966 int node;
2967
Andrew Mortona737b3e2006-03-22 00:08:11 -08002968retry:
Joe Korty6d2144d2008-03-05 15:04:59 -08002969 check_irq_off();
2970 node = numa_node_id();
2971 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002972 batchcount = ac->batchcount;
2973 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002974 /*
2975 * If there was little recent activity on this cache, then
2976 * perform only a partial refill. Otherwise we could generate
2977 * refill bouncing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002978 */
2979 batchcount = BATCHREFILL_LIMIT;
2980 }
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002981 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002982
Christoph Lametere498be72005-09-09 13:03:32 -07002983 BUG_ON(ac->avail > 0 || !l3);
2984 spin_lock(&l3->list_lock);
2985
Christoph Lameter3ded1752006-03-25 03:06:44 -08002986 /* See if we can refill from the shared array */
2987 if (l3->shared && transfer_objects(ac, l3->shared, batchcount))
2988 goto alloc_done;
2989
Linus Torvalds1da177e2005-04-16 15:20:36 -07002990 while (batchcount > 0) {
2991 struct list_head *entry;
2992 struct slab *slabp;
2993 /* Get slab alloc is to come from. */
2994 entry = l3->slabs_partial.next;
2995 if (entry == &l3->slabs_partial) {
2996 l3->free_touched = 1;
2997 entry = l3->slabs_free.next;
2998 if (entry == &l3->slabs_free)
2999 goto must_grow;
3000 }
3001
3002 slabp = list_entry(entry, struct slab, list);
3003 check_slabp(cachep, slabp);
3004 check_spinlock_acquired(cachep);
Pekka Enberg714b8172007-05-06 14:49:03 -07003005
3006 /*
3007 * The slab was either on partial or free list so
3008 * there must be at least one object available for
3009 * allocation.
3010 */
roel kluin249b9f32008-10-29 17:18:07 -04003011 BUG_ON(slabp->inuse >= cachep->num);
Pekka Enberg714b8172007-05-06 14:49:03 -07003012
Linus Torvalds1da177e2005-04-16 15:20:36 -07003013 while (slabp->inuse < cachep->num && batchcount--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003014 STATS_INC_ALLOCED(cachep);
3015 STATS_INC_ACTIVE(cachep);
3016 STATS_SET_HIGH(cachep);
3017
Matthew Dobson78d382d2006-02-01 03:05:47 -08003018 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07003019 node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003020 }
3021 check_slabp(cachep, slabp);
3022
3023 /* move slabp to correct slabp list: */
3024 list_del(&slabp->list);
3025 if (slabp->free == BUFCTL_END)
3026 list_add(&slabp->list, &l3->slabs_full);
3027 else
3028 list_add(&slabp->list, &l3->slabs_partial);
3029 }
3030
Andrew Mortona737b3e2006-03-22 00:08:11 -08003031must_grow:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003032 l3->free_objects -= ac->avail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003033alloc_done:
Christoph Lametere498be72005-09-09 13:03:32 -07003034 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003035
3036 if (unlikely(!ac->avail)) {
3037 int x;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003038 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07003039
Andrew Mortona737b3e2006-03-22 00:08:11 -08003040 /* cache_grow can reenable interrupts, then ac could change. */
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003041 ac = cpu_cache_get(cachep);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003042 if (!x && ac->avail == 0) /* no objects in sight? abort */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003043 return NULL;
3044
Andrew Mortona737b3e2006-03-22 00:08:11 -08003045 if (!ac->avail) /* objects refilled by interrupt? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003046 goto retry;
3047 }
3048 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07003049 return ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003050}
3051
Andrew Mortona737b3e2006-03-22 00:08:11 -08003052static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3053 gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003054{
3055 might_sleep_if(flags & __GFP_WAIT);
3056#if DEBUG
3057 kmem_flagcheck(cachep, flags);
3058#endif
3059}
3060
3061#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003062static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3063 gfp_t flags, void *objp, void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003064{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003065 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003066 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003067 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003068#ifdef CONFIG_DEBUG_PAGEALLOC
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003069 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003070 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003071 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003072 else
3073 check_poison_obj(cachep, objp);
3074#else
3075 check_poison_obj(cachep, objp);
3076#endif
3077 poison_obj(cachep, objp, POISON_INUSE);
3078 }
3079 if (cachep->flags & SLAB_STORE_USER)
3080 *dbg_userword(cachep, objp) = caller;
3081
3082 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003083 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3084 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3085 slab_error(cachep, "double free, or memory outside"
3086 " object was overwritten");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003087 printk(KERN_ERR
David Woodhouseb46b8f12007-05-08 00:22:59 -07003088 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08003089 objp, *dbg_redzone1(cachep, objp),
3090 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003091 }
3092 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3093 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3094 }
Al Viro871751e2006-03-25 03:06:39 -08003095#ifdef CONFIG_DEBUG_SLAB_LEAK
3096 {
3097 struct slab *slabp;
3098 unsigned objnr;
3099
Christoph Lameterb49af682007-05-06 14:49:41 -07003100 slabp = page_get_slab(virt_to_head_page(objp));
Al Viro871751e2006-03-25 03:06:39 -08003101 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3102 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3103 }
3104#endif
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003105 objp += obj_offset(cachep);
Christoph Lameter4f104932007-05-06 14:50:17 -07003106 if (cachep->ctor && cachep->flags & SLAB_POISON)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07003107 cachep->ctor(objp);
Kevin Hilmana44b56d2006-12-06 20:32:11 -08003108#if ARCH_SLAB_MINALIGN
3109 if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) {
3110 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3111 objp, ARCH_SLAB_MINALIGN);
3112 }
3113#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003114 return objp;
3115}
3116#else
3117#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3118#endif
3119
Akinobu Mita773ff602008-12-23 19:37:01 +09003120static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003121{
3122 if (cachep == &cache_cache)
Akinobu Mita773ff602008-12-23 19:37:01 +09003123 return false;
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003124
Akinobu Mita773ff602008-12-23 19:37:01 +09003125 return should_failslab(obj_size(cachep), flags);
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003126}
3127
Pekka Enberg343e0d72006-02-01 03:05:50 -08003128static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003129{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003130 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003131 struct array_cache *ac;
3132
Alok N Kataria5c382302005-09-27 21:45:46 -07003133 check_irq_off();
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003134
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003135 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003136 if (likely(ac->avail)) {
3137 STATS_INC_ALLOCHIT(cachep);
3138 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07003139 objp = ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003140 } else {
3141 STATS_INC_ALLOCMISS(cachep);
3142 objp = cache_alloc_refill(cachep, flags);
3143 }
Alok N Kataria5c382302005-09-27 21:45:46 -07003144 return objp;
3145}
3146
Christoph Lametere498be72005-09-09 13:03:32 -07003147#ifdef CONFIG_NUMA
3148/*
Paul Jacksonb2455392006-03-24 03:16:12 -08003149 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
Paul Jacksonc61afb12006-03-24 03:16:08 -08003150 *
3151 * If we are in_interrupt, then process context, including cpusets and
3152 * mempolicy, may not apply and should not be used for allocation policy.
3153 */
3154static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3155{
3156 int nid_alloc, nid_here;
3157
Christoph Lameter765c4502006-09-27 01:50:08 -07003158 if (in_interrupt() || (flags & __GFP_THISNODE))
Paul Jacksonc61afb12006-03-24 03:16:08 -08003159 return NULL;
3160 nid_alloc = nid_here = numa_node_id();
3161 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3162 nid_alloc = cpuset_mem_spread_node();
3163 else if (current->mempolicy)
3164 nid_alloc = slab_node(current->mempolicy);
3165 if (nid_alloc != nid_here)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003166 return ____cache_alloc_node(cachep, flags, nid_alloc);
Paul Jacksonc61afb12006-03-24 03:16:08 -08003167 return NULL;
3168}
3169
3170/*
Christoph Lameter765c4502006-09-27 01:50:08 -07003171 * Fallback function if there was no memory available and no objects on a
Christoph Lameter3c517a62006-12-06 20:33:29 -08003172 * certain node and fall back is permitted. First we scan all the
3173 * available nodelists for available objects. If that fails then we
3174 * perform an allocation without specifying a node. This allows the page
3175 * allocator to do its reclaim / fallback magic. We then insert the
3176 * slab into the proper nodelist and then allocate from it.
Christoph Lameter765c4502006-09-27 01:50:08 -07003177 */
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003178static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
Christoph Lameter765c4502006-09-27 01:50:08 -07003179{
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003180 struct zonelist *zonelist;
3181 gfp_t local_flags;
Mel Gormandd1a2392008-04-28 02:12:17 -07003182 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07003183 struct zone *zone;
3184 enum zone_type high_zoneidx = gfp_zone(flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003185 void *obj = NULL;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003186 int nid;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003187
3188 if (flags & __GFP_THISNODE)
3189 return NULL;
3190
Mel Gorman0e884602008-04-28 02:12:14 -07003191 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
Christoph Lameter6cb06222007-10-16 01:25:41 -07003192 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Christoph Lameter765c4502006-09-27 01:50:08 -07003193
Christoph Lameter3c517a62006-12-06 20:33:29 -08003194retry:
3195 /*
3196 * Look through allowed nodes for objects available
3197 * from existing per node queues.
3198 */
Mel Gorman54a6eb52008-04-28 02:12:16 -07003199 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3200 nid = zone_to_nid(zone);
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003201
Mel Gorman54a6eb52008-04-28 02:12:16 -07003202 if (cpuset_zone_allowed_hardwall(zone, flags) &&
Christoph Lameter3c517a62006-12-06 20:33:29 -08003203 cache->nodelists[nid] &&
Christoph Lameter481c5342008-06-21 16:46:35 -07003204 cache->nodelists[nid]->free_objects) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003205 obj = ____cache_alloc_node(cache,
3206 flags | GFP_THISNODE, nid);
Christoph Lameter481c5342008-06-21 16:46:35 -07003207 if (obj)
3208 break;
3209 }
Christoph Lameter3c517a62006-12-06 20:33:29 -08003210 }
3211
Christoph Lametercfce6602007-05-06 14:50:17 -07003212 if (!obj) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003213 /*
3214 * This allocation will be performed within the constraints
3215 * of the current cpuset / memory policy requirements.
3216 * We may trigger various forms of reclaim on the allowed
3217 * set and go into memory reserves if necessary.
3218 */
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003219 if (local_flags & __GFP_WAIT)
3220 local_irq_enable();
3221 kmem_flagcheck(cache, flags);
Christoph Lameter9ac33b22008-03-04 12:24:22 -08003222 obj = kmem_getpages(cache, local_flags, -1);
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003223 if (local_flags & __GFP_WAIT)
3224 local_irq_disable();
Christoph Lameter3c517a62006-12-06 20:33:29 -08003225 if (obj) {
3226 /*
3227 * Insert into the appropriate per node queues
3228 */
3229 nid = page_to_nid(virt_to_page(obj));
3230 if (cache_grow(cache, flags, nid, obj)) {
3231 obj = ____cache_alloc_node(cache,
3232 flags | GFP_THISNODE, nid);
3233 if (!obj)
3234 /*
3235 * Another processor may allocate the
3236 * objects in the slab since we are
3237 * not holding any locks.
3238 */
3239 goto retry;
3240 } else {
Hugh Dickinsb6a60452007-01-05 16:36:36 -08003241 /* cache_grow already freed obj */
Christoph Lameter3c517a62006-12-06 20:33:29 -08003242 obj = NULL;
3243 }
3244 }
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003245 }
Christoph Lameter765c4502006-09-27 01:50:08 -07003246 return obj;
3247}
3248
3249/*
Christoph Lametere498be72005-09-09 13:03:32 -07003250 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07003251 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003252static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003253 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07003254{
3255 struct list_head *entry;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003256 struct slab *slabp;
3257 struct kmem_list3 *l3;
3258 void *obj;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003259 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003260
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003261 l3 = cachep->nodelists[nodeid];
3262 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07003263
Andrew Mortona737b3e2006-03-22 00:08:11 -08003264retry:
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003265 check_irq_off();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003266 spin_lock(&l3->list_lock);
3267 entry = l3->slabs_partial.next;
3268 if (entry == &l3->slabs_partial) {
3269 l3->free_touched = 1;
3270 entry = l3->slabs_free.next;
3271 if (entry == &l3->slabs_free)
3272 goto must_grow;
3273 }
Christoph Lametere498be72005-09-09 13:03:32 -07003274
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003275 slabp = list_entry(entry, struct slab, list);
3276 check_spinlock_acquired_node(cachep, nodeid);
3277 check_slabp(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07003278
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003279 STATS_INC_NODEALLOCS(cachep);
3280 STATS_INC_ACTIVE(cachep);
3281 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003282
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003283 BUG_ON(slabp->inuse == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07003284
Matthew Dobson78d382d2006-02-01 03:05:47 -08003285 obj = slab_get_obj(cachep, slabp, nodeid);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003286 check_slabp(cachep, slabp);
3287 l3->free_objects--;
3288 /* move slabp to correct slabp list: */
3289 list_del(&slabp->list);
Christoph Lametere498be72005-09-09 13:03:32 -07003290
Andrew Mortona737b3e2006-03-22 00:08:11 -08003291 if (slabp->free == BUFCTL_END)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003292 list_add(&slabp->list, &l3->slabs_full);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003293 else
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003294 list_add(&slabp->list, &l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07003295
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003296 spin_unlock(&l3->list_lock);
3297 goto done;
Christoph Lametere498be72005-09-09 13:03:32 -07003298
Andrew Mortona737b3e2006-03-22 00:08:11 -08003299must_grow:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003300 spin_unlock(&l3->list_lock);
Christoph Lameter3c517a62006-12-06 20:33:29 -08003301 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
Christoph Lameter765c4502006-09-27 01:50:08 -07003302 if (x)
3303 goto retry;
Christoph Lametere498be72005-09-09 13:03:32 -07003304
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003305 return fallback_alloc(cachep, flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003306
Andrew Mortona737b3e2006-03-22 00:08:11 -08003307done:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003308 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07003309}
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003310
3311/**
3312 * kmem_cache_alloc_node - Allocate an object on the specified node
3313 * @cachep: The cache to allocate from.
3314 * @flags: See kmalloc().
3315 * @nodeid: node number of the target node.
3316 * @caller: return address of caller, used for debug information
3317 *
3318 * Identical to kmem_cache_alloc but it will allocate memory on the given
3319 * node, which can improve the performance for cpu bound structures.
3320 *
3321 * Fallback to other node is possible if __GFP_THISNODE is not set.
3322 */
3323static __always_inline void *
3324__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3325 void *caller)
3326{
3327 unsigned long save_flags;
3328 void *ptr;
3329
Nick Piggincf40bd12009-01-21 08:12:39 +01003330 lockdep_trace_alloc(flags);
3331
Akinobu Mita773ff602008-12-23 19:37:01 +09003332 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003333 return NULL;
3334
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003335 cache_alloc_debugcheck_before(cachep, flags);
3336 local_irq_save(save_flags);
3337
3338 if (unlikely(nodeid == -1))
3339 nodeid = numa_node_id();
3340
3341 if (unlikely(!cachep->nodelists[nodeid])) {
3342 /* Node not bootstrapped yet */
3343 ptr = fallback_alloc(cachep, flags);
3344 goto out;
3345 }
3346
3347 if (nodeid == numa_node_id()) {
3348 /*
3349 * Use the locally cached objects if possible.
3350 * However ____cache_alloc does not allow fallback
3351 * to other nodes. It may fail while we still have
3352 * objects on other nodes available.
3353 */
3354 ptr = ____cache_alloc(cachep, flags);
3355 if (ptr)
3356 goto out;
3357 }
3358 /* ___cache_alloc_node can fall back to other nodes */
3359 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3360 out:
3361 local_irq_restore(save_flags);
3362 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3363
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003364 if (unlikely((flags & __GFP_ZERO) && ptr))
3365 memset(ptr, 0, obj_size(cachep));
3366
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003367 return ptr;
3368}
3369
3370static __always_inline void *
3371__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3372{
3373 void *objp;
3374
3375 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3376 objp = alternate_node_alloc(cache, flags);
3377 if (objp)
3378 goto out;
3379 }
3380 objp = ____cache_alloc(cache, flags);
3381
3382 /*
3383 * We may just have run out of memory on the local node.
3384 * ____cache_alloc_node() knows how to locate memory on other nodes
3385 */
3386 if (!objp)
3387 objp = ____cache_alloc_node(cache, flags, numa_node_id());
3388
3389 out:
3390 return objp;
3391}
3392#else
3393
3394static __always_inline void *
3395__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3396{
3397 return ____cache_alloc(cachep, flags);
3398}
3399
3400#endif /* CONFIG_NUMA */
3401
3402static __always_inline void *
3403__cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3404{
3405 unsigned long save_flags;
3406 void *objp;
3407
Nick Piggincf40bd12009-01-21 08:12:39 +01003408 lockdep_trace_alloc(flags);
3409
Akinobu Mita773ff602008-12-23 19:37:01 +09003410 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003411 return NULL;
3412
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003413 cache_alloc_debugcheck_before(cachep, flags);
3414 local_irq_save(save_flags);
3415 objp = __do_cache_alloc(cachep, flags);
3416 local_irq_restore(save_flags);
3417 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3418 prefetchw(objp);
3419
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003420 if (unlikely((flags & __GFP_ZERO) && objp))
3421 memset(objp, 0, obj_size(cachep));
3422
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003423 return objp;
3424}
Christoph Lametere498be72005-09-09 13:03:32 -07003425
3426/*
3427 * Caller needs to acquire correct kmem_list's list_lock
3428 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003429static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003430 int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003431{
3432 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07003433 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003434
3435 for (i = 0; i < nr_objects; i++) {
3436 void *objp = objpp[i];
3437 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003438
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08003439 slabp = virt_to_slab(objp);
Christoph Lameterff694162005-09-22 21:44:02 -07003440 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003441 list_del(&slabp->list);
Christoph Lameterff694162005-09-22 21:44:02 -07003442 check_spinlock_acquired_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003443 check_slabp(cachep, slabp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08003444 slab_put_obj(cachep, slabp, objp, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003445 STATS_DEC_ACTIVE(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003446 l3->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003447 check_slabp(cachep, slabp);
3448
3449 /* fixup slab chains */
3450 if (slabp->inuse == 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07003451 if (l3->free_objects > l3->free_limit) {
3452 l3->free_objects -= cachep->num;
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07003453 /* No need to drop any previously held
3454 * lock here, even if we have a off-slab slab
3455 * descriptor it is guaranteed to come from
3456 * a different cache, refer to comments before
3457 * alloc_slabmgmt.
3458 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003459 slab_destroy(cachep, slabp);
3460 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07003461 list_add(&slabp->list, &l3->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003462 }
3463 } else {
3464 /* Unconditionally move a slab to the end of the
3465 * partial list on free - maximum time for the
3466 * other objects to be freed, too.
3467 */
Christoph Lametere498be72005-09-09 13:03:32 -07003468 list_add_tail(&slabp->list, &l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003469 }
3470 }
3471}
3472
Pekka Enberg343e0d72006-02-01 03:05:50 -08003473static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003474{
3475 int batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07003476 struct kmem_list3 *l3;
Christoph Lameterff694162005-09-22 21:44:02 -07003477 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003478
3479 batchcount = ac->batchcount;
3480#if DEBUG
3481 BUG_ON(!batchcount || batchcount > ac->avail);
3482#endif
3483 check_irq_off();
Christoph Lameterff694162005-09-22 21:44:02 -07003484 l3 = cachep->nodelists[node];
Ingo Molnar873623d2006-07-13 14:44:38 +02003485 spin_lock(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003486 if (l3->shared) {
3487 struct array_cache *shared_array = l3->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003488 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003489 if (max) {
3490 if (batchcount > max)
3491 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07003492 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003493 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003494 shared_array->avail += batchcount;
3495 goto free_done;
3496 }
3497 }
3498
Christoph Lameterff694162005-09-22 21:44:02 -07003499 free_block(cachep, ac->entry, batchcount, node);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003500free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003501#if STATS
3502 {
3503 int i = 0;
3504 struct list_head *p;
3505
Christoph Lametere498be72005-09-09 13:03:32 -07003506 p = l3->slabs_free.next;
3507 while (p != &(l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003508 struct slab *slabp;
3509
3510 slabp = list_entry(p, struct slab, list);
3511 BUG_ON(slabp->inuse);
3512
3513 i++;
3514 p = p->next;
3515 }
3516 STATS_SET_FREEABLE(cachep, i);
3517 }
3518#endif
Christoph Lametere498be72005-09-09 13:03:32 -07003519 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003520 ac->avail -= batchcount;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003521 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003522}
3523
3524/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08003525 * Release an obj back to its cache. If the obj has a constructed state, it must
3526 * be in this state _before_ it is released. Called with disabled ints.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003527 */
Ingo Molnar873623d2006-07-13 14:44:38 +02003528static inline void __cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003529{
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003530 struct array_cache *ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003531
3532 check_irq_off();
3533 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3534
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -07003535 /*
3536 * Skip calling cache_free_alien() when the platform is not numa.
3537 * This will avoid cache misses that happen while accessing slabp (which
3538 * is per page memory reference) to get nodeid. Instead use a global
3539 * variable to skip the call, which is mostly likely to be present in
3540 * the cache.
3541 */
3542 if (numa_platform && cache_free_alien(cachep, objp))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07003543 return;
Christoph Lametere498be72005-09-09 13:03:32 -07003544
Linus Torvalds1da177e2005-04-16 15:20:36 -07003545 if (likely(ac->avail < ac->limit)) {
3546 STATS_INC_FREEHIT(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003547 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003548 return;
3549 } else {
3550 STATS_INC_FREEMISS(cachep);
3551 cache_flusharray(cachep, ac);
Christoph Lametere498be72005-09-09 13:03:32 -07003552 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003553 }
3554}
3555
3556/**
3557 * kmem_cache_alloc - Allocate an object
3558 * @cachep: The cache to allocate from.
3559 * @flags: See kmalloc().
3560 *
3561 * Allocate an object from this cache. The flags are only relevant
3562 * if the cache has no available objects.
3563 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003564void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003565{
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003566 void *ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3567
Eduard - Gabriel Munteanuca2b84c2009-03-23 15:12:24 +02003568 trace_kmem_cache_alloc(_RET_IP_, ret,
3569 obj_size(cachep), cachep->buffer_size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003570
3571 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003572}
3573EXPORT_SYMBOL(kmem_cache_alloc);
3574
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003575#ifdef CONFIG_KMEMTRACE
3576void *kmem_cache_alloc_notrace(struct kmem_cache *cachep, gfp_t flags)
3577{
3578 return __cache_alloc(cachep, flags, __builtin_return_address(0));
3579}
3580EXPORT_SYMBOL(kmem_cache_alloc_notrace);
3581#endif
3582
Linus Torvalds1da177e2005-04-16 15:20:36 -07003583/**
Randy Dunlap76824862008-03-19 17:00:40 -07003584 * kmem_ptr_validate - check if an untrusted pointer might be a slab entry.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003585 * @cachep: the cache we're checking against
3586 * @ptr: pointer to validate
3587 *
Randy Dunlap76824862008-03-19 17:00:40 -07003588 * This verifies that the untrusted pointer looks sane;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003589 * it is _not_ a guarantee that the pointer is actually
3590 * part of the slab cache in question, but it at least
3591 * validates that the pointer can be dereferenced and
3592 * looks half-way sane.
3593 *
3594 * Currently only used for dentry validation.
3595 */
Christoph Lameterb7f869a2006-12-22 01:06:44 -08003596int kmem_ptr_validate(struct kmem_cache *cachep, const void *ptr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003597{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003598 unsigned long addr = (unsigned long)ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003599 unsigned long min_addr = PAGE_OFFSET;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003600 unsigned long align_mask = BYTES_PER_WORD - 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003601 unsigned long size = cachep->buffer_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003602 struct page *page;
3603
3604 if (unlikely(addr < min_addr))
3605 goto out;
3606 if (unlikely(addr > (unsigned long)high_memory - size))
3607 goto out;
3608 if (unlikely(addr & align_mask))
3609 goto out;
3610 if (unlikely(!kern_addr_valid(addr)))
3611 goto out;
3612 if (unlikely(!kern_addr_valid(addr + size - 1)))
3613 goto out;
3614 page = virt_to_page(ptr);
3615 if (unlikely(!PageSlab(page)))
3616 goto out;
Pekka Enberg065d41c2005-11-13 16:06:46 -08003617 if (unlikely(page_get_cache(page) != cachep))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003618 goto out;
3619 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003620out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003621 return 0;
3622}
3623
3624#ifdef CONFIG_NUMA
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003625void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3626{
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003627 void *ret = __cache_alloc_node(cachep, flags, nodeid,
3628 __builtin_return_address(0));
3629
Eduard - Gabriel Munteanuca2b84c2009-03-23 15:12:24 +02003630 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3631 obj_size(cachep), cachep->buffer_size,
3632 flags, nodeid);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003633
3634 return ret;
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003635}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003636EXPORT_SYMBOL(kmem_cache_alloc_node);
3637
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003638#ifdef CONFIG_KMEMTRACE
3639void *kmem_cache_alloc_node_notrace(struct kmem_cache *cachep,
3640 gfp_t flags,
3641 int nodeid)
3642{
3643 return __cache_alloc_node(cachep, flags, nodeid,
3644 __builtin_return_address(0));
3645}
3646EXPORT_SYMBOL(kmem_cache_alloc_node_notrace);
3647#endif
3648
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003649static __always_inline void *
3650__do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003651{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003652 struct kmem_cache *cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003653 void *ret;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003654
3655 cachep = kmem_find_general_cachep(size, flags);
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003656 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3657 return cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003658 ret = kmem_cache_alloc_node_notrace(cachep, flags, node);
3659
Eduard - Gabriel Munteanuca2b84c2009-03-23 15:12:24 +02003660 trace_kmalloc_node((unsigned long) caller, ret,
3661 size, cachep->buffer_size, flags, node);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003662
3663 return ret;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003664}
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003665
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003666#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_KMEMTRACE)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003667void *__kmalloc_node(size_t size, gfp_t flags, int node)
3668{
3669 return __do_kmalloc_node(size, flags, node,
3670 __builtin_return_address(0));
3671}
Christoph Hellwigdbe5e692006-09-25 23:31:36 -07003672EXPORT_SYMBOL(__kmalloc_node);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003673
3674void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003675 int node, unsigned long caller)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003676{
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003677 return __do_kmalloc_node(size, flags, node, (void *)caller);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003678}
3679EXPORT_SYMBOL(__kmalloc_node_track_caller);
3680#else
3681void *__kmalloc_node(size_t size, gfp_t flags, int node)
3682{
3683 return __do_kmalloc_node(size, flags, node, NULL);
3684}
3685EXPORT_SYMBOL(__kmalloc_node);
3686#endif /* CONFIG_DEBUG_SLAB */
3687#endif /* CONFIG_NUMA */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003688
3689/**
Paul Drynoff800590f2006-06-23 02:03:48 -07003690 * __do_kmalloc - allocate memory
Linus Torvalds1da177e2005-04-16 15:20:36 -07003691 * @size: how many bytes of memory are required.
Paul Drynoff800590f2006-06-23 02:03:48 -07003692 * @flags: the type of memory to allocate (see kmalloc).
Randy Dunlap911851e2006-03-22 00:08:14 -08003693 * @caller: function caller for debug tracking of the caller
Linus Torvalds1da177e2005-04-16 15:20:36 -07003694 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003695static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3696 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003697{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003698 struct kmem_cache *cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003699 void *ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003700
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003701 /* If you want to save a few bytes .text space: replace
3702 * __ with kmem_.
3703 * Then kmalloc uses the uninlined functions instead of the inline
3704 * functions.
3705 */
3706 cachep = __find_general_cachep(size, flags);
Linus Torvaldsa5c96d82007-07-19 13:17:15 -07003707 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3708 return cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003709 ret = __cache_alloc(cachep, flags, caller);
3710
Eduard - Gabriel Munteanuca2b84c2009-03-23 15:12:24 +02003711 trace_kmalloc((unsigned long) caller, ret,
3712 size, cachep->buffer_size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003713
3714 return ret;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003715}
3716
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003717
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003718#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_KMEMTRACE)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003719void *__kmalloc(size_t size, gfp_t flags)
3720{
Al Viro871751e2006-03-25 03:06:39 -08003721 return __do_kmalloc(size, flags, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003722}
3723EXPORT_SYMBOL(__kmalloc);
3724
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003725void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003726{
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003727 return __do_kmalloc(size, flags, (void *)caller);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003728}
3729EXPORT_SYMBOL(__kmalloc_track_caller);
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003730
3731#else
3732void *__kmalloc(size_t size, gfp_t flags)
3733{
3734 return __do_kmalloc(size, flags, NULL);
3735}
3736EXPORT_SYMBOL(__kmalloc);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003737#endif
3738
Linus Torvalds1da177e2005-04-16 15:20:36 -07003739/**
3740 * kmem_cache_free - Deallocate an object
3741 * @cachep: The cache the allocation was from.
3742 * @objp: The previously allocated object.
3743 *
3744 * Free an object which was previously allocated from this
3745 * cache.
3746 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003747void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003748{
3749 unsigned long flags;
3750
3751 local_irq_save(flags);
Ingo Molnar898552c2007-02-10 01:44:57 -08003752 debug_check_no_locks_freed(objp, obj_size(cachep));
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07003753 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3754 debug_check_no_obj_freed(objp, obj_size(cachep));
Ingo Molnar873623d2006-07-13 14:44:38 +02003755 __cache_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003756 local_irq_restore(flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003757
Eduard - Gabriel Munteanuca2b84c2009-03-23 15:12:24 +02003758 trace_kmem_cache_free(_RET_IP_, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003759}
3760EXPORT_SYMBOL(kmem_cache_free);
3761
3762/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003763 * kfree - free previously allocated memory
3764 * @objp: pointer returned by kmalloc.
3765 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003766 * If @objp is NULL, no operation is performed.
3767 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003768 * Don't free memory not originally allocated by kmalloc()
3769 * or you will run into trouble.
3770 */
3771void kfree(const void *objp)
3772{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003773 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003774 unsigned long flags;
3775
Pekka Enberg2121db72009-03-25 11:05:57 +02003776 trace_kfree(_RET_IP_, objp);
3777
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003778 if (unlikely(ZERO_OR_NULL_PTR(objp)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003779 return;
3780 local_irq_save(flags);
3781 kfree_debugcheck(objp);
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08003782 c = virt_to_cache(objp);
Ingo Molnarf9b84042006-06-27 02:54:49 -07003783 debug_check_no_locks_freed(objp, obj_size(c));
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07003784 debug_check_no_obj_freed(objp, obj_size(c));
Ingo Molnar873623d2006-07-13 14:44:38 +02003785 __cache_free(c, (void *)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003786 local_irq_restore(flags);
3787}
3788EXPORT_SYMBOL(kfree);
3789
Pekka Enberg343e0d72006-02-01 03:05:50 -08003790unsigned int kmem_cache_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003791{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003792 return obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003793}
3794EXPORT_SYMBOL(kmem_cache_size);
3795
Pekka Enberg343e0d72006-02-01 03:05:50 -08003796const char *kmem_cache_name(struct kmem_cache *cachep)
Arnaldo Carvalho de Melo19449722005-06-18 22:46:19 -07003797{
3798 return cachep->name;
3799}
3800EXPORT_SYMBOL_GPL(kmem_cache_name);
3801
Christoph Lametere498be72005-09-09 13:03:32 -07003802/*
Simon Arlott183ff222007-10-20 01:27:18 +02003803 * This initializes kmem_list3 or resizes various caches for all nodes.
Christoph Lametere498be72005-09-09 13:03:32 -07003804 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003805static int alloc_kmemlist(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07003806{
3807 int node;
3808 struct kmem_list3 *l3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003809 struct array_cache *new_shared;
Paul Menage3395ee02006-12-06 20:32:16 -08003810 struct array_cache **new_alien = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003811
Mel Gorman9c09a952008-01-24 05:49:54 -08003812 for_each_online_node(node) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003813
Paul Menage3395ee02006-12-06 20:32:16 -08003814 if (use_alien_caches) {
3815 new_alien = alloc_alien_cache(node, cachep->limit);
3816 if (!new_alien)
3817 goto fail;
3818 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003819
Eric Dumazet63109842007-05-06 14:49:28 -07003820 new_shared = NULL;
3821 if (cachep->shared) {
3822 new_shared = alloc_arraycache(node,
Christoph Lameter0718dc22006-03-25 03:06:47 -08003823 cachep->shared*cachep->batchcount,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003824 0xbaadf00d);
Eric Dumazet63109842007-05-06 14:49:28 -07003825 if (!new_shared) {
3826 free_alien_cache(new_alien);
3827 goto fail;
3828 }
Christoph Lameter0718dc22006-03-25 03:06:47 -08003829 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003830
Andrew Mortona737b3e2006-03-22 00:08:11 -08003831 l3 = cachep->nodelists[node];
3832 if (l3) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003833 struct array_cache *shared = l3->shared;
3834
Christoph Lametere498be72005-09-09 13:03:32 -07003835 spin_lock_irq(&l3->list_lock);
3836
Christoph Lametercafeb022006-03-25 03:06:46 -08003837 if (shared)
Christoph Lameter0718dc22006-03-25 03:06:47 -08003838 free_block(cachep, shared->entry,
3839 shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07003840
Christoph Lametercafeb022006-03-25 03:06:46 -08003841 l3->shared = new_shared;
3842 if (!l3->alien) {
Christoph Lametere498be72005-09-09 13:03:32 -07003843 l3->alien = new_alien;
3844 new_alien = NULL;
3845 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003846 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003847 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003848 spin_unlock_irq(&l3->list_lock);
Christoph Lametercafeb022006-03-25 03:06:46 -08003849 kfree(shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003850 free_alien_cache(new_alien);
3851 continue;
3852 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08003853 l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node);
Christoph Lameter0718dc22006-03-25 03:06:47 -08003854 if (!l3) {
3855 free_alien_cache(new_alien);
3856 kfree(new_shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003857 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003858 }
Christoph Lametere498be72005-09-09 13:03:32 -07003859
3860 kmem_list3_init(l3);
3861 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Andrew Mortona737b3e2006-03-22 00:08:11 -08003862 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003863 l3->shared = new_shared;
Christoph Lametere498be72005-09-09 13:03:32 -07003864 l3->alien = new_alien;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003865 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003866 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003867 cachep->nodelists[node] = l3;
3868 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003869 return 0;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003870
Andrew Mortona737b3e2006-03-22 00:08:11 -08003871fail:
Christoph Lameter0718dc22006-03-25 03:06:47 -08003872 if (!cachep->next.next) {
3873 /* Cache is not active yet. Roll back what we did */
3874 node--;
3875 while (node >= 0) {
3876 if (cachep->nodelists[node]) {
3877 l3 = cachep->nodelists[node];
3878
3879 kfree(l3->shared);
3880 free_alien_cache(l3->alien);
3881 kfree(l3);
3882 cachep->nodelists[node] = NULL;
3883 }
3884 node--;
3885 }
3886 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003887 return -ENOMEM;
Christoph Lametere498be72005-09-09 13:03:32 -07003888}
3889
Linus Torvalds1da177e2005-04-16 15:20:36 -07003890struct ccupdate_struct {
Pekka Enberg343e0d72006-02-01 03:05:50 -08003891 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003892 struct array_cache *new[NR_CPUS];
3893};
3894
3895static void do_ccupdate_local(void *info)
3896{
Andrew Mortona737b3e2006-03-22 00:08:11 -08003897 struct ccupdate_struct *new = info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003898 struct array_cache *old;
3899
3900 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003901 old = cpu_cache_get(new->cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003902
Linus Torvalds1da177e2005-04-16 15:20:36 -07003903 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3904 new->new[smp_processor_id()] = old;
3905}
3906
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003907/* Always called with the cache_chain_mutex held */
Andrew Mortona737b3e2006-03-22 00:08:11 -08003908static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3909 int batchcount, int shared)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003910{
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003911 struct ccupdate_struct *new;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003912 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003913
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003914 new = kzalloc(sizeof(*new), GFP_KERNEL);
3915 if (!new)
3916 return -ENOMEM;
3917
Christoph Lametere498be72005-09-09 13:03:32 -07003918 for_each_online_cpu(i) {
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003919 new->new[i] = alloc_arraycache(cpu_to_node(i), limit,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003920 batchcount);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003921 if (!new->new[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003922 for (i--; i >= 0; i--)
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003923 kfree(new->new[i]);
3924 kfree(new);
Christoph Lametere498be72005-09-09 13:03:32 -07003925 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003926 }
3927 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003928 new->cachep = cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003929
Jens Axboe15c8b6c2008-05-09 09:39:44 +02003930 on_each_cpu(do_ccupdate_local, (void *)new, 1);
Christoph Lametere498be72005-09-09 13:03:32 -07003931
Linus Torvalds1da177e2005-04-16 15:20:36 -07003932 check_irq_on();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003933 cachep->batchcount = batchcount;
3934 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07003935 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003936
Christoph Lametere498be72005-09-09 13:03:32 -07003937 for_each_online_cpu(i) {
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003938 struct array_cache *ccold = new->new[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003939 if (!ccold)
3940 continue;
Christoph Lametere498be72005-09-09 13:03:32 -07003941 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -07003942 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
Christoph Lametere498be72005-09-09 13:03:32 -07003943 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003944 kfree(ccold);
3945 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003946 kfree(new);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003947 return alloc_kmemlist(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003948}
3949
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003950/* Called with cache_chain_mutex held always */
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003951static int enable_cpucache(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003952{
3953 int err;
3954 int limit, shared;
3955
Andrew Mortona737b3e2006-03-22 00:08:11 -08003956 /*
3957 * The head array serves three purposes:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003958 * - create a LIFO ordering, i.e. return objects that are cache-warm
3959 * - reduce the number of spinlock operations.
Andrew Mortona737b3e2006-03-22 00:08:11 -08003960 * - reduce the number of linked list operations on the slab and
Linus Torvalds1da177e2005-04-16 15:20:36 -07003961 * bufctl chains: array operations are cheaper.
3962 * The numbers are guessed, we should auto-tune as described by
3963 * Bonwick.
3964 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003965 if (cachep->buffer_size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003966 limit = 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003967 else if (cachep->buffer_size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003968 limit = 8;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003969 else if (cachep->buffer_size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003970 limit = 24;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003971 else if (cachep->buffer_size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003972 limit = 54;
3973 else
3974 limit = 120;
3975
Andrew Mortona737b3e2006-03-22 00:08:11 -08003976 /*
3977 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
Linus Torvalds1da177e2005-04-16 15:20:36 -07003978 * allocation behaviour: Most allocs on one cpu, most free operations
3979 * on another cpu. For these cases, an efficient object passing between
3980 * cpus is necessary. This is provided by a shared array. The array
3981 * replaces Bonwick's magazine layer.
3982 * On uniprocessor, it's functionally equivalent (but less efficient)
3983 * to a larger limit. Thus disabled by default.
3984 */
3985 shared = 0;
Eric Dumazet364fbb22007-05-06 14:49:27 -07003986 if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003987 shared = 8;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003988
3989#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003990 /*
3991 * With debugging enabled, large batchcount lead to excessively long
3992 * periods with disabled local interrupts. Limit the batchcount
Linus Torvalds1da177e2005-04-16 15:20:36 -07003993 */
3994 if (limit > 32)
3995 limit = 32;
3996#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003997 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003998 if (err)
3999 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004000 cachep->name, -err);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07004001 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004002}
4003
Christoph Lameter1b552532006-03-22 00:09:07 -08004004/*
4005 * Drain an array if it contains any elements taking the l3 lock only if
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004006 * necessary. Note that the l3 listlock also protects the array_cache
4007 * if drain_array() is used on the shared array.
Christoph Lameter1b552532006-03-22 00:09:07 -08004008 */
4009void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
4010 struct array_cache *ac, int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004011{
4012 int tofree;
4013
Christoph Lameter1b552532006-03-22 00:09:07 -08004014 if (!ac || !ac->avail)
4015 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004016 if (ac->touched && !force) {
4017 ac->touched = 0;
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004018 } else {
Christoph Lameter1b552532006-03-22 00:09:07 -08004019 spin_lock_irq(&l3->list_lock);
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004020 if (ac->avail) {
4021 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4022 if (tofree > ac->avail)
4023 tofree = (ac->avail + 1) / 2;
4024 free_block(cachep, ac->entry, tofree, node);
4025 ac->avail -= tofree;
4026 memmove(ac->entry, &(ac->entry[tofree]),
4027 sizeof(void *) * ac->avail);
4028 }
Christoph Lameter1b552532006-03-22 00:09:07 -08004029 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004030 }
4031}
4032
4033/**
4034 * cache_reap - Reclaim memory from caches.
Randy Dunlap05fb6bf2007-02-28 20:12:13 -08004035 * @w: work descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07004036 *
4037 * Called from workqueue/eventd every few seconds.
4038 * Purpose:
4039 * - clear the per-cpu caches for this CPU.
4040 * - return freeable pages to the main free memory pool.
4041 *
Andrew Mortona737b3e2006-03-22 00:08:11 -08004042 * If we cannot acquire the cache chain mutex then just give up - we'll try
4043 * again on the next iteration.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004044 */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004045static void cache_reap(struct work_struct *w)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004046{
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004047 struct kmem_cache *searchp;
Christoph Lametere498be72005-09-09 13:03:32 -07004048 struct kmem_list3 *l3;
Christoph Lameteraab22072006-03-22 00:09:06 -08004049 int node = numa_node_id();
Jean Delvarebf6aede2009-04-02 16:56:54 -07004050 struct delayed_work *work = to_delayed_work(w);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004051
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004052 if (!mutex_trylock(&cache_chain_mutex))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004053 /* Give up. Setup the next iteration. */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004054 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004055
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004056 list_for_each_entry(searchp, &cache_chain, next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004057 check_irq_on();
4058
Christoph Lameter35386e32006-03-22 00:09:05 -08004059 /*
4060 * We only take the l3 lock if absolutely necessary and we
4061 * have established with reasonable certainty that
4062 * we can do some work if the lock was obtained.
4063 */
Christoph Lameteraab22072006-03-22 00:09:06 -08004064 l3 = searchp->nodelists[node];
Christoph Lameter35386e32006-03-22 00:09:05 -08004065
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004066 reap_alien(searchp, l3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004067
Christoph Lameteraab22072006-03-22 00:09:06 -08004068 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004069
Christoph Lameter35386e32006-03-22 00:09:05 -08004070 /*
4071 * These are racy checks but it does not matter
4072 * if we skip one check or scan twice.
4073 */
Christoph Lametere498be72005-09-09 13:03:32 -07004074 if (time_after(l3->next_reap, jiffies))
Christoph Lameter35386e32006-03-22 00:09:05 -08004075 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004076
Christoph Lametere498be72005-09-09 13:03:32 -07004077 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004078
Christoph Lameteraab22072006-03-22 00:09:06 -08004079 drain_array(searchp, l3, l3->shared, 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004080
Christoph Lametered11d9e2006-06-30 01:55:45 -07004081 if (l3->free_touched)
Christoph Lametere498be72005-09-09 13:03:32 -07004082 l3->free_touched = 0;
Christoph Lametered11d9e2006-06-30 01:55:45 -07004083 else {
4084 int freed;
4085
4086 freed = drain_freelist(searchp, l3, (l3->free_limit +
4087 5 * searchp->num - 1) / (5 * searchp->num));
4088 STATS_ADD_REAPED(searchp, freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004089 }
Christoph Lameter35386e32006-03-22 00:09:05 -08004090next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004091 cond_resched();
4092 }
4093 check_irq_on();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004094 mutex_unlock(&cache_chain_mutex);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004095 next_reap_node();
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004096out:
Andrew Mortona737b3e2006-03-22 00:08:11 -08004097 /* Set up the next iteration */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004098 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004099}
4100
Linus Torvalds158a9622008-01-02 13:04:48 -08004101#ifdef CONFIG_SLABINFO
Linus Torvalds1da177e2005-04-16 15:20:36 -07004102
Pekka Enberg85289f92006-01-08 01:00:36 -08004103static void print_slabinfo_header(struct seq_file *m)
4104{
4105 /*
4106 * Output format version, so at least we can change it
4107 * without _too_ many complaints.
4108 */
4109#if STATS
4110 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4111#else
4112 seq_puts(m, "slabinfo - version: 2.1\n");
4113#endif
4114 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4115 "<objperslab> <pagesperslab>");
4116 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4117 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4118#if STATS
4119 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004120 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
Pekka Enberg85289f92006-01-08 01:00:36 -08004121 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4122#endif
4123 seq_putc(m, '\n');
4124}
4125
Linus Torvalds1da177e2005-04-16 15:20:36 -07004126static void *s_start(struct seq_file *m, loff_t *pos)
4127{
4128 loff_t n = *pos;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004129
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004130 mutex_lock(&cache_chain_mutex);
Pekka Enberg85289f92006-01-08 01:00:36 -08004131 if (!n)
4132 print_slabinfo_header(m);
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004133
4134 return seq_list_start(&cache_chain, *pos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004135}
4136
4137static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4138{
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004139 return seq_list_next(p, &cache_chain, pos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004140}
4141
4142static void s_stop(struct seq_file *m, void *p)
4143{
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004144 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004145}
4146
4147static int s_show(struct seq_file *m, void *p)
4148{
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004149 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004150 struct slab *slabp;
4151 unsigned long active_objs;
4152 unsigned long num_objs;
4153 unsigned long active_slabs = 0;
4154 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004155 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004156 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07004157 int node;
4158 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004159
Linus Torvalds1da177e2005-04-16 15:20:36 -07004160 active_objs = 0;
4161 num_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004162 for_each_online_node(node) {
4163 l3 = cachep->nodelists[node];
4164 if (!l3)
4165 continue;
4166
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004167 check_irq_on();
4168 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07004169
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004170 list_for_each_entry(slabp, &l3->slabs_full, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004171 if (slabp->inuse != cachep->num && !error)
4172 error = "slabs_full accounting error";
4173 active_objs += cachep->num;
4174 active_slabs++;
4175 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004176 list_for_each_entry(slabp, &l3->slabs_partial, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004177 if (slabp->inuse == cachep->num && !error)
4178 error = "slabs_partial inuse accounting error";
4179 if (!slabp->inuse && !error)
4180 error = "slabs_partial/inuse accounting error";
4181 active_objs += slabp->inuse;
4182 active_slabs++;
4183 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004184 list_for_each_entry(slabp, &l3->slabs_free, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004185 if (slabp->inuse && !error)
4186 error = "slabs_free/inuse accounting error";
4187 num_slabs++;
4188 }
4189 free_objects += l3->free_objects;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08004190 if (l3->shared)
4191 shared_avail += l3->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07004192
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004193 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004194 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004195 num_slabs += active_slabs;
4196 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07004197 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004198 error = "free_objects accounting error";
4199
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004200 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004201 if (error)
4202 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4203
4204 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004205 name, active_objs, num_objs, cachep->buffer_size,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004206 cachep->num, (1 << cachep->gfporder));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004207 seq_printf(m, " : tunables %4u %4u %4u",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004208 cachep->limit, cachep->batchcount, cachep->shared);
Christoph Lametere498be72005-09-09 13:03:32 -07004209 seq_printf(m, " : slabdata %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004210 active_slabs, num_slabs, shared_avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004211#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004212 { /* list3 stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004213 unsigned long high = cachep->high_mark;
4214 unsigned long allocs = cachep->num_allocations;
4215 unsigned long grown = cachep->grown;
4216 unsigned long reaped = cachep->reaped;
4217 unsigned long errors = cachep->errors;
4218 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004219 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07004220 unsigned long node_frees = cachep->node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004221 unsigned long overflows = cachep->node_overflow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004222
Christoph Lametere498be72005-09-09 13:03:32 -07004223 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004224 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown,
Andrew Mortona737b3e2006-03-22 00:08:11 -08004225 reaped, errors, max_freeable, node_allocs,
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004226 node_frees, overflows);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004227 }
4228 /* cpu stats */
4229 {
4230 unsigned long allochit = atomic_read(&cachep->allochit);
4231 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4232 unsigned long freehit = atomic_read(&cachep->freehit);
4233 unsigned long freemiss = atomic_read(&cachep->freemiss);
4234
4235 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004236 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004237 }
4238#endif
4239 seq_putc(m, '\n');
Linus Torvalds1da177e2005-04-16 15:20:36 -07004240 return 0;
4241}
4242
4243/*
4244 * slabinfo_op - iterator that generates /proc/slabinfo
4245 *
4246 * Output layout:
4247 * cache-name
4248 * num-active-objs
4249 * total-objs
4250 * object size
4251 * num-active-slabs
4252 * total-slabs
4253 * num-pages-per-slab
4254 * + further values on SMP and with statistics enabled
4255 */
4256
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04004257static const struct seq_operations slabinfo_op = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004258 .start = s_start,
4259 .next = s_next,
4260 .stop = s_stop,
4261 .show = s_show,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004262};
4263
4264#define MAX_SLABINFO_WRITE 128
4265/**
4266 * slabinfo_write - Tuning for the slab allocator
4267 * @file: unused
4268 * @buffer: user buffer
4269 * @count: data length
4270 * @ppos: unused
4271 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004272ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4273 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004274{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004275 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004276 int limit, batchcount, shared, res;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004277 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004278
Linus Torvalds1da177e2005-04-16 15:20:36 -07004279 if (count > MAX_SLABINFO_WRITE)
4280 return -EINVAL;
4281 if (copy_from_user(&kbuf, buffer, count))
4282 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004283 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07004284
4285 tmp = strchr(kbuf, ' ');
4286 if (!tmp)
4287 return -EINVAL;
4288 *tmp = '\0';
4289 tmp++;
4290 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4291 return -EINVAL;
4292
4293 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004294 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004295 res = -EINVAL;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004296 list_for_each_entry(cachep, &cache_chain, next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004297 if (!strcmp(cachep->name, kbuf)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08004298 if (limit < 1 || batchcount < 1 ||
4299 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07004300 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004301 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07004302 res = do_tune_cpucache(cachep, limit,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004303 batchcount, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004304 }
4305 break;
4306 }
4307 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004308 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004309 if (res >= 0)
4310 res = count;
4311 return res;
4312}
Al Viro871751e2006-03-25 03:06:39 -08004313
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04004314static int slabinfo_open(struct inode *inode, struct file *file)
4315{
4316 return seq_open(file, &slabinfo_op);
4317}
4318
4319static const struct file_operations proc_slabinfo_operations = {
4320 .open = slabinfo_open,
4321 .read = seq_read,
4322 .write = slabinfo_write,
4323 .llseek = seq_lseek,
4324 .release = seq_release,
4325};
4326
Al Viro871751e2006-03-25 03:06:39 -08004327#ifdef CONFIG_DEBUG_SLAB_LEAK
4328
4329static void *leaks_start(struct seq_file *m, loff_t *pos)
4330{
Al Viro871751e2006-03-25 03:06:39 -08004331 mutex_lock(&cache_chain_mutex);
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004332 return seq_list_start(&cache_chain, *pos);
Al Viro871751e2006-03-25 03:06:39 -08004333}
4334
4335static inline int add_caller(unsigned long *n, unsigned long v)
4336{
4337 unsigned long *p;
4338 int l;
4339 if (!v)
4340 return 1;
4341 l = n[1];
4342 p = n + 2;
4343 while (l) {
4344 int i = l/2;
4345 unsigned long *q = p + 2 * i;
4346 if (*q == v) {
4347 q[1]++;
4348 return 1;
4349 }
4350 if (*q > v) {
4351 l = i;
4352 } else {
4353 p = q + 2;
4354 l -= i + 1;
4355 }
4356 }
4357 if (++n[1] == n[0])
4358 return 0;
4359 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4360 p[0] = v;
4361 p[1] = 1;
4362 return 1;
4363}
4364
4365static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4366{
4367 void *p;
4368 int i;
4369 if (n[0] == n[1])
4370 return;
4371 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4372 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4373 continue;
4374 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4375 return;
4376 }
4377}
4378
4379static void show_symbol(struct seq_file *m, unsigned long address)
4380{
4381#ifdef CONFIG_KALLSYMS
Al Viro871751e2006-03-25 03:06:39 -08004382 unsigned long offset, size;
Tejun Heo9281ace2007-07-17 04:03:51 -07004383 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
Al Viro871751e2006-03-25 03:06:39 -08004384
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004385 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
Al Viro871751e2006-03-25 03:06:39 -08004386 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004387 if (modname[0])
Al Viro871751e2006-03-25 03:06:39 -08004388 seq_printf(m, " [%s]", modname);
4389 return;
4390 }
4391#endif
4392 seq_printf(m, "%p", (void *)address);
4393}
4394
4395static int leaks_show(struct seq_file *m, void *p)
4396{
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004397 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
Al Viro871751e2006-03-25 03:06:39 -08004398 struct slab *slabp;
4399 struct kmem_list3 *l3;
4400 const char *name;
4401 unsigned long *n = m->private;
4402 int node;
4403 int i;
4404
4405 if (!(cachep->flags & SLAB_STORE_USER))
4406 return 0;
4407 if (!(cachep->flags & SLAB_RED_ZONE))
4408 return 0;
4409
4410 /* OK, we can do it */
4411
4412 n[1] = 0;
4413
4414 for_each_online_node(node) {
4415 l3 = cachep->nodelists[node];
4416 if (!l3)
4417 continue;
4418
4419 check_irq_on();
4420 spin_lock_irq(&l3->list_lock);
4421
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004422 list_for_each_entry(slabp, &l3->slabs_full, list)
Al Viro871751e2006-03-25 03:06:39 -08004423 handle_slab(n, cachep, slabp);
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004424 list_for_each_entry(slabp, &l3->slabs_partial, list)
Al Viro871751e2006-03-25 03:06:39 -08004425 handle_slab(n, cachep, slabp);
Al Viro871751e2006-03-25 03:06:39 -08004426 spin_unlock_irq(&l3->list_lock);
4427 }
4428 name = cachep->name;
4429 if (n[0] == n[1]) {
4430 /* Increase the buffer size */
4431 mutex_unlock(&cache_chain_mutex);
4432 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4433 if (!m->private) {
4434 /* Too bad, we are really out */
4435 m->private = n;
4436 mutex_lock(&cache_chain_mutex);
4437 return -ENOMEM;
4438 }
4439 *(unsigned long *)m->private = n[0] * 2;
4440 kfree(n);
4441 mutex_lock(&cache_chain_mutex);
4442 /* Now make sure this entry will be retried */
4443 m->count = m->size;
4444 return 0;
4445 }
4446 for (i = 0; i < n[1]; i++) {
4447 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4448 show_symbol(m, n[2*i+2]);
4449 seq_putc(m, '\n');
4450 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004451
Al Viro871751e2006-03-25 03:06:39 -08004452 return 0;
4453}
4454
Alexey Dobriyana0ec95a82008-10-06 00:59:10 +04004455static const struct seq_operations slabstats_op = {
Al Viro871751e2006-03-25 03:06:39 -08004456 .start = leaks_start,
4457 .next = s_next,
4458 .stop = s_stop,
4459 .show = leaks_show,
4460};
Alexey Dobriyana0ec95a82008-10-06 00:59:10 +04004461
4462static int slabstats_open(struct inode *inode, struct file *file)
4463{
4464 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4465 int ret = -ENOMEM;
4466 if (n) {
4467 ret = seq_open(file, &slabstats_op);
4468 if (!ret) {
4469 struct seq_file *m = file->private_data;
4470 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4471 m->private = n;
4472 n = NULL;
4473 }
4474 kfree(n);
4475 }
4476 return ret;
4477}
4478
4479static const struct file_operations proc_slabstats_operations = {
4480 .open = slabstats_open,
4481 .read = seq_read,
4482 .llseek = seq_lseek,
4483 .release = seq_release_private,
4484};
Al Viro871751e2006-03-25 03:06:39 -08004485#endif
Alexey Dobriyana0ec95a82008-10-06 00:59:10 +04004486
4487static int __init slab_proc_init(void)
4488{
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04004489 proc_create("slabinfo",S_IWUSR|S_IRUGO,NULL,&proc_slabinfo_operations);
Alexey Dobriyana0ec95a82008-10-06 00:59:10 +04004490#ifdef CONFIG_DEBUG_SLAB_LEAK
4491 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4492#endif
4493 return 0;
4494}
4495module_init(slab_proc_init);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004496#endif
4497
Manfred Spraul00e145b2005-09-03 15:55:07 -07004498/**
4499 * ksize - get the actual amount of memory allocated for a given object
4500 * @objp: Pointer to the object
4501 *
4502 * kmalloc may internally round up allocations and return more memory
4503 * than requested. ksize() can be used to determine the actual amount of
4504 * memory allocated. The caller may use this additional memory, even though
4505 * a smaller amount of memory was initially specified with the kmalloc call.
4506 * The caller must guarantee that objp points to a valid object previously
4507 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4508 * must not be freed during the duration of the call.
4509 */
Pekka Enbergfd76bab2007-05-06 14:48:40 -07004510size_t ksize(const void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004511{
Christoph Lameteref8b4522007-10-16 01:24:46 -07004512 BUG_ON(!objp);
4513 if (unlikely(objp == ZERO_SIZE_PTR))
Manfred Spraul00e145b2005-09-03 15:55:07 -07004514 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004515
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08004516 return obj_size(virt_to_cache(objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004517}
Kirill A. Shutemovb1aabec2009-02-10 15:21:44 +02004518EXPORT_SYMBOL(ksize);