blob: e74a16e4ced605111bd0f84154273c038bb0d9e3 [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 Dobriyana0ec95a2008-10-06 00:59:10 +040098#include <linux/proc_fs.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070099#include <linux/seq_file.h>
100#include <linux/notifier.h>
101#include <linux/kallsyms.h>
102#include <linux/cpu.h>
103#include <linux/sysctl.h>
104#include <linux/module.h>
Zhaolei02af61b2009-04-10 14:26:18 +0800105#include <linux/kmemtrace.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700106#include <linux/rcupdate.h>
Paulo Marques543537b2005-06-23 00:09:02 -0700107#include <linux/string.h>
Andrew Morton138ae662006-12-06 20:36:41 -0800108#include <linux/uaccess.h>
Christoph Lametere498be72005-09-09 13:03:32 -0700109#include <linux/nodemask.h>
Catalin Marinasd5cff632009-06-11 13:22:40 +0100110#include <linux/kmemleak.h>
Christoph Lameterdc85da12006-01-18 17:42:36 -0800111#include <linux/mempolicy.h>
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800112#include <linux/mutex.h>
Akinobu Mita8a8b6502006-12-08 02:39:44 -0800113#include <linux/fault-inject.h>
Ingo Molnare7eebaf2006-06-27 02:54:55 -0700114#include <linux/rtmutex.h>
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800115#include <linux/reciprocal_div.h>
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -0700116#include <linux/debugobjects.h>
Pekka Enbergc175eea2008-05-09 20:35:53 +0200117#include <linux/kmemcheck.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700118
Linus Torvalds1da177e2005-04-16 15:20:36 -0700119#include <asm/cacheflush.h>
120#include <asm/tlbflush.h>
121#include <asm/page.h>
122
123/*
Christoph Lameter50953fe2007-05-06 14:50:16 -0700124 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700125 * 0 for faster, smaller code (especially in the critical paths).
126 *
127 * STATS - 1 to collect stats for /proc/slabinfo.
128 * 0 for faster, smaller code (especially in the critical paths).
129 *
130 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
131 */
132
133#ifdef CONFIG_DEBUG_SLAB
134#define DEBUG 1
135#define STATS 1
136#define FORCED_DEBUG 1
137#else
138#define DEBUG 0
139#define STATS 0
140#define FORCED_DEBUG 0
141#endif
142
Linus Torvalds1da177e2005-04-16 15:20:36 -0700143/* Shouldn't this be in a header file somewhere? */
144#define BYTES_PER_WORD sizeof(void *)
David Woodhouse87a927c2007-07-04 21:26:44 -0400145#define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700146
Linus Torvalds1da177e2005-04-16 15:20:36 -0700147#ifndef ARCH_KMALLOC_MINALIGN
148/*
149 * Enforce a minimum alignment for the kmalloc caches.
150 * Usually, the kmalloc caches are cache_line_size() aligned, except when
151 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
152 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
David Woodhouseb46b8f12007-05-08 00:22:59 -0700153 * alignment larger than the alignment of a 64-bit integer.
154 * ARCH_KMALLOC_MINALIGN allows that.
155 * Note that increasing this value may disable some debug features.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700156 */
David Woodhouseb46b8f12007-05-08 00:22:59 -0700157#define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700158#endif
159
160#ifndef ARCH_SLAB_MINALIGN
161/*
162 * Enforce a minimum alignment for all caches.
163 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
164 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
165 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
166 * some debug features.
167 */
168#define ARCH_SLAB_MINALIGN 0
169#endif
170
171#ifndef ARCH_KMALLOC_FLAGS
172#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
173#endif
174
175/* Legal flag mask for kmem_cache_create(). */
176#if DEBUG
Christoph Lameter50953fe2007-05-06 14:50:16 -0700177# define CREATE_MASK (SLAB_RED_ZONE | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700178 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
Christoph Lameterac2b8982006-03-22 00:08:15 -0800179 SLAB_CACHE_DMA | \
Christoph Lameter5af60832007-05-06 14:49:56 -0700180 SLAB_STORE_USER | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700181 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -0700182 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
Pekka Enbergc175eea2008-05-09 20:35:53 +0200183 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700184#else
Christoph Lameterac2b8982006-03-22 00:08:15 -0800185# define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
Christoph Lameter5af60832007-05-06 14:49:56 -0700186 SLAB_CACHE_DMA | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700187 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -0700188 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
Pekka Enbergc175eea2008-05-09 20:35:53 +0200189 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700190#endif
191
192/*
193 * kmem_bufctl_t:
194 *
195 * Bufctl's are used for linking objs within a slab
196 * linked offsets.
197 *
198 * This implementation relies on "struct page" for locating the cache &
199 * slab an object belongs to.
200 * This allows the bufctl structure to be small (one int), but limits
201 * the number of objects a slab (not a cache) can contain when off-slab
202 * bufctls are used. The limit is the size of the largest general cache
203 * that does not use off-slab slabs.
204 * For 32bit archs with 4 kB pages, is this 56.
205 * This is not serious, as it is only for large objects, when it is unwise
206 * to have too many per slab.
207 * Note: This limit can be raised by introducing a general cache whose size
208 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
209 */
210
Kyle Moffettfa5b08d2005-09-03 15:55:03 -0700211typedef unsigned int kmem_bufctl_t;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700212#define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
213#define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
Al Viro871751e2006-03-25 03:06:39 -0800214#define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
215#define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700216
Linus Torvalds1da177e2005-04-16 15:20:36 -0700217/*
218 * struct slab
219 *
220 * Manages the objs in a slab. Placed either at the beginning of mem allocated
221 * for a slab, or allocated from an general cache.
222 * Slabs are chained into three list: fully used, partial, fully free slabs.
223 */
224struct slab {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800225 struct list_head list;
226 unsigned long colouroff;
227 void *s_mem; /* including colour offset */
228 unsigned int inuse; /* num of objs active in slab */
229 kmem_bufctl_t free;
230 unsigned short nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700231};
232
233/*
234 * struct slab_rcu
235 *
236 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
237 * arrange for kmem_freepages to be called via RCU. This is useful if
238 * we need to approach a kernel structure obliquely, from its address
239 * obtained without the usual locking. We can lock the structure to
240 * stabilize it and check it's still at the given address, only if we
241 * can be sure that the memory has not been meanwhile reused for some
242 * other kind of object (which our subsystem's lock might corrupt).
243 *
244 * rcu_read_lock before reading the address, then rcu_read_unlock after
245 * taking the spinlock within the structure expected at that address.
246 *
247 * We assume struct slab_rcu can overlay struct slab when destroying.
248 */
249struct slab_rcu {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800250 struct rcu_head head;
Pekka Enberg343e0d72006-02-01 03:05:50 -0800251 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800252 void *addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700253};
254
255/*
256 * struct array_cache
257 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700258 * Purpose:
259 * - LIFO ordering, to hand out cache-warm objects from _alloc
260 * - reduce the number of linked list operations
261 * - reduce spinlock operations
262 *
263 * The limit is stored in the per-cpu structure to reduce the data cache
264 * footprint.
265 *
266 */
267struct array_cache {
268 unsigned int avail;
269 unsigned int limit;
270 unsigned int batchcount;
271 unsigned int touched;
Christoph Lametere498be72005-09-09 13:03:32 -0700272 spinlock_t lock;
Robert P. J. Daybda5b652007-10-16 23:30:05 -0700273 void *entry[]; /*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800274 * Must have this definition in here for the proper
275 * alignment of array_cache. Also simplifies accessing
276 * the entries.
Andrew Mortona737b3e2006-03-22 00:08:11 -0800277 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700278};
279
Andrew Mortona737b3e2006-03-22 00:08:11 -0800280/*
281 * bootstrap: The caches do not work without cpuarrays anymore, but the
282 * cpuarrays are allocated from the generic caches...
Linus Torvalds1da177e2005-04-16 15:20:36 -0700283 */
284#define BOOT_CPUCACHE_ENTRIES 1
285struct arraycache_init {
286 struct array_cache cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800287 void *entries[BOOT_CPUCACHE_ENTRIES];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700288};
289
290/*
Christoph Lametere498be72005-09-09 13:03:32 -0700291 * The slab lists for all objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700292 */
293struct kmem_list3 {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800294 struct list_head slabs_partial; /* partial list first, better asm code */
295 struct list_head slabs_full;
296 struct list_head slabs_free;
297 unsigned long free_objects;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800298 unsigned int free_limit;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800299 unsigned int colour_next; /* Per-node cache coloring */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800300 spinlock_t list_lock;
301 struct array_cache *shared; /* shared per node */
302 struct array_cache **alien; /* on other nodes */
Christoph Lameter35386e32006-03-22 00:09:05 -0800303 unsigned long next_reap; /* updated without locking */
304 int free_touched; /* updated without locking */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700305};
306
Christoph Lametere498be72005-09-09 13:03:32 -0700307/*
308 * Need this for bootstrapping a per node allocator.
309 */
Pekka Enberg556a1692008-01-25 08:20:51 +0200310#define NUM_INIT_LISTS (3 * MAX_NUMNODES)
Christoph Lametere498be72005-09-09 13:03:32 -0700311struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
312#define CACHE_CACHE 0
Pekka Enberg556a1692008-01-25 08:20:51 +0200313#define SIZE_AC MAX_NUMNODES
314#define SIZE_L3 (2 * MAX_NUMNODES)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700315
Christoph Lametered11d9e2006-06-30 01:55:45 -0700316static int drain_freelist(struct kmem_cache *cache,
317 struct kmem_list3 *l3, int tofree);
318static void free_block(struct kmem_cache *cachep, void **objpp, int len,
319 int node);
Pekka Enberg83b519e2009-06-10 19:40:04 +0300320static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
David Howells65f27f32006-11-22 14:55:48 +0000321static void cache_reap(struct work_struct *unused);
Christoph Lametered11d9e2006-06-30 01:55:45 -0700322
Christoph Lametere498be72005-09-09 13:03:32 -0700323/*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800324 * This function must be completely optimized away if a constant is passed to
325 * it. Mostly the same as what is in linux/slab.h except it returns an index.
Christoph Lametere498be72005-09-09 13:03:32 -0700326 */
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700327static __always_inline int index_of(const size_t size)
Christoph Lametere498be72005-09-09 13:03:32 -0700328{
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800329 extern void __bad_size(void);
330
Christoph Lametere498be72005-09-09 13:03:32 -0700331 if (__builtin_constant_p(size)) {
332 int i = 0;
333
334#define CACHE(x) \
335 if (size <=x) \
336 return i; \
337 else \
338 i++;
Joe Perches1c61fc42008-03-05 13:58:17 -0800339#include <linux/kmalloc_sizes.h>
Christoph Lametere498be72005-09-09 13:03:32 -0700340#undef CACHE
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800341 __bad_size();
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700342 } else
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800343 __bad_size();
Christoph Lametere498be72005-09-09 13:03:32 -0700344 return 0;
345}
346
Ingo Molnare0a42722006-06-23 02:03:46 -0700347static int slab_early_init = 1;
348
Christoph Lametere498be72005-09-09 13:03:32 -0700349#define INDEX_AC index_of(sizeof(struct arraycache_init))
350#define INDEX_L3 index_of(sizeof(struct kmem_list3))
351
Pekka Enberg5295a742006-02-01 03:05:48 -0800352static void kmem_list3_init(struct kmem_list3 *parent)
Christoph Lametere498be72005-09-09 13:03:32 -0700353{
354 INIT_LIST_HEAD(&parent->slabs_full);
355 INIT_LIST_HEAD(&parent->slabs_partial);
356 INIT_LIST_HEAD(&parent->slabs_free);
357 parent->shared = NULL;
358 parent->alien = NULL;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800359 parent->colour_next = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700360 spin_lock_init(&parent->list_lock);
361 parent->free_objects = 0;
362 parent->free_touched = 0;
363}
364
Andrew Mortona737b3e2006-03-22 00:08:11 -0800365#define MAKE_LIST(cachep, listp, slab, nodeid) \
366 do { \
367 INIT_LIST_HEAD(listp); \
368 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
Christoph Lametere498be72005-09-09 13:03:32 -0700369 } while (0)
370
Andrew Mortona737b3e2006-03-22 00:08:11 -0800371#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
372 do { \
Christoph Lametere498be72005-09-09 13:03:32 -0700373 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
374 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
375 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
376 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700377
Linus Torvalds1da177e2005-04-16 15:20:36 -0700378#define CFLGS_OFF_SLAB (0x80000000UL)
379#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
380
381#define BATCHREFILL_LIMIT 16
Andrew Mortona737b3e2006-03-22 00:08:11 -0800382/*
383 * Optimization question: fewer reaps means less probability for unnessary
384 * cpucache drain/refill cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700385 *
Adrian Bunkdc6f3f22005-11-08 16:44:08 +0100386 * OTOH the cpuarrays can contain lots of objects,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700387 * which could lock up otherwise freeable slabs.
388 */
389#define REAPTIMEOUT_CPUC (2*HZ)
390#define REAPTIMEOUT_LIST3 (4*HZ)
391
392#if STATS
393#define STATS_INC_ACTIVE(x) ((x)->num_active++)
394#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
395#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
396#define STATS_INC_GROWN(x) ((x)->grown++)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700397#define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
Andrew Mortona737b3e2006-03-22 00:08:11 -0800398#define STATS_SET_HIGH(x) \
399 do { \
400 if ((x)->num_active > (x)->high_mark) \
401 (x)->high_mark = (x)->num_active; \
402 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700403#define STATS_INC_ERR(x) ((x)->errors++)
404#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
Christoph Lametere498be72005-09-09 13:03:32 -0700405#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700406#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800407#define STATS_SET_FREEABLE(x, i) \
408 do { \
409 if ((x)->max_freeable < i) \
410 (x)->max_freeable = i; \
411 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700412#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
413#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
414#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
415#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
416#else
417#define STATS_INC_ACTIVE(x) do { } while (0)
418#define STATS_DEC_ACTIVE(x) do { } while (0)
419#define STATS_INC_ALLOCED(x) do { } while (0)
420#define STATS_INC_GROWN(x) do { } while (0)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700421#define STATS_ADD_REAPED(x,y) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700422#define STATS_SET_HIGH(x) do { } while (0)
423#define STATS_INC_ERR(x) do { } while (0)
424#define STATS_INC_NODEALLOCS(x) do { } while (0)
Christoph Lametere498be72005-09-09 13:03:32 -0700425#define STATS_INC_NODEFREES(x) do { } while (0)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700426#define STATS_INC_ACOVERFLOW(x) do { } while (0)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800427#define STATS_SET_FREEABLE(x, i) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700428#define STATS_INC_ALLOCHIT(x) do { } while (0)
429#define STATS_INC_ALLOCMISS(x) do { } while (0)
430#define STATS_INC_FREEHIT(x) do { } while (0)
431#define STATS_INC_FREEMISS(x) do { } while (0)
432#endif
433
434#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -0700435
Andrew Mortona737b3e2006-03-22 00:08:11 -0800436/*
437 * memory layout of objects:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700438 * 0 : objp
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800439 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
Linus Torvalds1da177e2005-04-16 15:20:36 -0700440 * the end of an object is aligned with the end of the real
441 * allocation. Catches writes behind the end of the allocation.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800442 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700443 * redzone word.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800444 * cachep->obj_offset: The real object.
445 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
Andrew Mortona737b3e2006-03-22 00:08:11 -0800446 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
447 * [BYTES_PER_WORD long]
Linus Torvalds1da177e2005-04-16 15:20:36 -0700448 */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800449static int obj_offset(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700450{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800451 return cachep->obj_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700452}
453
Pekka Enberg343e0d72006-02-01 03:05:50 -0800454static int obj_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700455{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800456 return cachep->obj_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700457}
458
David Woodhouseb46b8f12007-05-08 00:22:59 -0700459static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700460{
461 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
David Woodhouseb46b8f12007-05-08 00:22:59 -0700462 return (unsigned long long*) (objp + obj_offset(cachep) -
463 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700464}
465
David Woodhouseb46b8f12007-05-08 00:22:59 -0700466static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700467{
468 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
469 if (cachep->flags & SLAB_STORE_USER)
David Woodhouseb46b8f12007-05-08 00:22:59 -0700470 return (unsigned long long *)(objp + cachep->buffer_size -
471 sizeof(unsigned long long) -
David Woodhouse87a927c2007-07-04 21:26:44 -0400472 REDZONE_ALIGN);
David Woodhouseb46b8f12007-05-08 00:22:59 -0700473 return (unsigned long long *) (objp + cachep->buffer_size -
474 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700475}
476
Pekka Enberg343e0d72006-02-01 03:05:50 -0800477static void **dbg_userword(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700478{
479 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800480 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700481}
482
483#else
484
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800485#define obj_offset(x) 0
486#define obj_size(cachep) (cachep->buffer_size)
David Woodhouseb46b8f12007-05-08 00:22:59 -0700487#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
488#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
Linus Torvalds1da177e2005-04-16 15:20:36 -0700489#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
490
491#endif
492
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +0300493#ifdef CONFIG_KMEMTRACE
494size_t slab_buffer_size(struct kmem_cache *cachep)
495{
496 return cachep->buffer_size;
497}
498EXPORT_SYMBOL(slab_buffer_size);
499#endif
500
Linus Torvalds1da177e2005-04-16 15:20:36 -0700501/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700502 * Do not go above this order unless 0 objects fit into the slab.
503 */
504#define BREAK_GFP_ORDER_HI 1
505#define BREAK_GFP_ORDER_LO 0
506static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
507
Andrew Mortona737b3e2006-03-22 00:08:11 -0800508/*
509 * Functions for storing/retrieving the cachep and or slab from the page
510 * allocator. These are used to find the slab an obj belongs to. With kfree(),
511 * these are used to find the cache which an obj belongs to.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700512 */
Pekka Enberg065d41c2005-11-13 16:06:46 -0800513static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
514{
515 page->lru.next = (struct list_head *)cache;
516}
517
518static inline struct kmem_cache *page_get_cache(struct page *page)
519{
Christoph Lameterd85f3382007-05-06 14:49:39 -0700520 page = compound_head(page);
Pekka Enbergddc2e812006-06-23 02:03:40 -0700521 BUG_ON(!PageSlab(page));
Pekka Enberg065d41c2005-11-13 16:06:46 -0800522 return (struct kmem_cache *)page->lru.next;
523}
524
525static inline void page_set_slab(struct page *page, struct slab *slab)
526{
527 page->lru.prev = (struct list_head *)slab;
528}
529
530static inline struct slab *page_get_slab(struct page *page)
531{
Pekka Enbergddc2e812006-06-23 02:03:40 -0700532 BUG_ON(!PageSlab(page));
Pekka Enberg065d41c2005-11-13 16:06:46 -0800533 return (struct slab *)page->lru.prev;
534}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700535
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800536static inline struct kmem_cache *virt_to_cache(const void *obj)
537{
Christoph Lameterb49af682007-05-06 14:49:41 -0700538 struct page *page = virt_to_head_page(obj);
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800539 return page_get_cache(page);
540}
541
542static inline struct slab *virt_to_slab(const void *obj)
543{
Christoph Lameterb49af682007-05-06 14:49:41 -0700544 struct page *page = virt_to_head_page(obj);
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800545 return page_get_slab(page);
546}
547
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800548static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
549 unsigned int idx)
550{
551 return slab->s_mem + cache->buffer_size * idx;
552}
553
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800554/*
555 * We want to avoid an expensive divide : (offset / cache->buffer_size)
556 * Using the fact that buffer_size is a constant for a particular cache,
557 * we can replace (offset / cache->buffer_size) by
558 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
559 */
560static inline unsigned int obj_to_index(const struct kmem_cache *cache,
561 const struct slab *slab, void *obj)
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800562{
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800563 u32 offset = (obj - slab->s_mem);
564 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800565}
566
Andrew Mortona737b3e2006-03-22 00:08:11 -0800567/*
568 * These are the default caches for kmalloc. Custom caches can have other sizes.
569 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700570struct cache_sizes malloc_sizes[] = {
571#define CACHE(x) { .cs_size = (x) },
572#include <linux/kmalloc_sizes.h>
573 CACHE(ULONG_MAX)
574#undef CACHE
575};
576EXPORT_SYMBOL(malloc_sizes);
577
578/* Must match cache_sizes above. Out of line to keep cache footprint low. */
579struct cache_names {
580 char *name;
581 char *name_dma;
582};
583
584static struct cache_names __initdata cache_names[] = {
585#define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
586#include <linux/kmalloc_sizes.h>
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800587 {NULL,}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700588#undef CACHE
589};
590
591static struct arraycache_init initarray_cache __initdata =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800592 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700593static struct arraycache_init initarray_generic =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800594 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700595
596/* internal cache of cache description objs */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800597static struct kmem_cache cache_cache = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800598 .batchcount = 1,
599 .limit = BOOT_CPUCACHE_ENTRIES,
600 .shared = 1,
Pekka Enberg343e0d72006-02-01 03:05:50 -0800601 .buffer_size = sizeof(struct kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800602 .name = "kmem_cache",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700603};
604
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700605#define BAD_ALIEN_MAGIC 0x01020304ul
606
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200607#ifdef CONFIG_LOCKDEP
608
609/*
610 * Slab sometimes uses the kmalloc slabs to store the slab headers
611 * for other slabs "off slab".
612 * The locking for this is tricky in that it nests within the locks
613 * of all other slabs in a few places; to deal with this special
614 * locking we put on-slab caches into a separate lock-class.
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700615 *
616 * We set lock class for alien array caches which are up during init.
617 * The lock annotation will be lost if all cpus of a node goes down and
618 * then comes back up during hotplug
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200619 */
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700620static struct lock_class_key on_slab_l3_key;
621static struct lock_class_key on_slab_alc_key;
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200622
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700623static inline void init_lock_keys(void)
624
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200625{
626 int q;
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700627 struct cache_sizes *s = malloc_sizes;
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200628
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700629 while (s->cs_size != ULONG_MAX) {
630 for_each_node(q) {
631 struct array_cache **alc;
632 int r;
633 struct kmem_list3 *l3 = s->cs_cachep->nodelists[q];
634 if (!l3 || OFF_SLAB(s->cs_cachep))
635 continue;
636 lockdep_set_class(&l3->list_lock, &on_slab_l3_key);
637 alc = l3->alien;
638 /*
639 * FIXME: This check for BAD_ALIEN_MAGIC
640 * should go away when common slab code is taught to
641 * work even without alien caches.
642 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
643 * for alloc_alien_cache,
644 */
645 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
646 continue;
647 for_each_node(r) {
648 if (alc[r])
649 lockdep_set_class(&alc[r]->lock,
650 &on_slab_alc_key);
651 }
652 }
653 s++;
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200654 }
655}
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200656#else
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700657static inline void init_lock_keys(void)
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200658{
659}
660#endif
661
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -0800662/*
Gautham R Shenoy95402b32008-01-25 21:08:02 +0100663 * Guard access to the cache-chain.
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -0800664 */
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800665static DEFINE_MUTEX(cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700666static struct list_head cache_chain;
667
668/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700669 * chicken and egg problem: delay the per-cpu array allocation
670 * until the general caches are up.
671 */
672static enum {
673 NONE,
Christoph Lametere498be72005-09-09 13:03:32 -0700674 PARTIAL_AC,
675 PARTIAL_L3,
Pekka Enberg8429db52009-06-12 15:58:59 +0300676 EARLY,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700677 FULL
678} g_cpucache_up;
679
Mike Kravetz39d24e62006-05-15 09:44:13 -0700680/*
681 * used by boot code to determine if it can use slab based allocator
682 */
683int slab_is_available(void)
684{
Pekka Enberg8429db52009-06-12 15:58:59 +0300685 return g_cpucache_up >= EARLY;
Mike Kravetz39d24e62006-05-15 09:44:13 -0700686}
687
David Howells52bad642006-11-22 14:54:01 +0000688static DEFINE_PER_CPU(struct delayed_work, reap_work);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700689
Pekka Enberg343e0d72006-02-01 03:05:50 -0800690static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700691{
692 return cachep->array[smp_processor_id()];
693}
694
Andrew Mortona737b3e2006-03-22 00:08:11 -0800695static inline struct kmem_cache *__find_general_cachep(size_t size,
696 gfp_t gfpflags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700697{
698 struct cache_sizes *csizep = malloc_sizes;
699
700#if DEBUG
701 /* This happens if someone tries to call
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800702 * kmem_cache_create(), or __kmalloc(), before
703 * the generic caches are initialized.
704 */
Alok Katariac7e43c72005-09-14 12:17:53 -0700705 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700706#endif
Christoph Lameter6cb8f912007-07-17 04:03:22 -0700707 if (!size)
708 return ZERO_SIZE_PTR;
709
Linus Torvalds1da177e2005-04-16 15:20:36 -0700710 while (size > csizep->cs_size)
711 csizep++;
712
713 /*
Martin Hicks0abf40c2005-09-03 15:54:54 -0700714 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
Linus Torvalds1da177e2005-04-16 15:20:36 -0700715 * has cs_{dma,}cachep==NULL. Thus no special case
716 * for large kmalloc calls required.
717 */
Christoph Lameter4b51d662007-02-10 01:43:10 -0800718#ifdef CONFIG_ZONE_DMA
Linus Torvalds1da177e2005-04-16 15:20:36 -0700719 if (unlikely(gfpflags & GFP_DMA))
720 return csizep->cs_dmacachep;
Christoph Lameter4b51d662007-02-10 01:43:10 -0800721#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700722 return csizep->cs_cachep;
723}
724
Adrian Bunkb2213852006-09-25 23:31:02 -0700725static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700726{
727 return __find_general_cachep(size, gfpflags);
728}
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700729
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800730static size_t slab_mgmt_size(size_t nr_objs, size_t align)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700731{
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800732 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
733}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700734
Andrew Mortona737b3e2006-03-22 00:08:11 -0800735/*
736 * Calculate the number of objects and left-over bytes for a given buffer size.
737 */
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800738static void cache_estimate(unsigned long gfporder, size_t buffer_size,
739 size_t align, int flags, size_t *left_over,
740 unsigned int *num)
741{
742 int nr_objs;
743 size_t mgmt_size;
744 size_t slab_size = PAGE_SIZE << gfporder;
745
746 /*
747 * The slab management structure can be either off the slab or
748 * on it. For the latter case, the memory allocated for a
749 * slab is used for:
750 *
751 * - The struct slab
752 * - One kmem_bufctl_t for each object
753 * - Padding to respect alignment of @align
754 * - @buffer_size bytes for each object
755 *
756 * If the slab management structure is off the slab, then the
757 * alignment will already be calculated into the size. Because
758 * the slabs are all pages aligned, the objects will be at the
759 * correct alignment when allocated.
760 */
761 if (flags & CFLGS_OFF_SLAB) {
762 mgmt_size = 0;
763 nr_objs = slab_size / buffer_size;
764
765 if (nr_objs > SLAB_LIMIT)
766 nr_objs = SLAB_LIMIT;
767 } else {
768 /*
769 * Ignore padding for the initial guess. The padding
770 * is at most @align-1 bytes, and @buffer_size is at
771 * least @align. In the worst case, this result will
772 * be one greater than the number of objects that fit
773 * into the memory allocation when taking the padding
774 * into account.
775 */
776 nr_objs = (slab_size - sizeof(struct slab)) /
777 (buffer_size + sizeof(kmem_bufctl_t));
778
779 /*
780 * This calculated number will be either the right
781 * amount, or one greater than what we want.
782 */
783 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
784 > slab_size)
785 nr_objs--;
786
787 if (nr_objs > SLAB_LIMIT)
788 nr_objs = SLAB_LIMIT;
789
790 mgmt_size = slab_mgmt_size(nr_objs, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700791 }
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800792 *num = nr_objs;
793 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700794}
795
Harvey Harrisond40cee22008-04-30 00:55:07 -0700796#define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700797
Andrew Mortona737b3e2006-03-22 00:08:11 -0800798static void __slab_error(const char *function, struct kmem_cache *cachep,
799 char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700800{
801 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800802 function, cachep->name, msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700803 dump_stack();
804}
805
Paul Menage3395ee02006-12-06 20:32:16 -0800806/*
807 * By default on NUMA we use alien caches to stage the freeing of
808 * objects allocated from other nodes. This causes massive memory
809 * inefficiencies when using fake NUMA setup to split memory into a
810 * large number of small nodes, so it can be disabled on the command
811 * line
812 */
813
814static int use_alien_caches __read_mostly = 1;
815static int __init noaliencache_setup(char *s)
816{
817 use_alien_caches = 0;
818 return 1;
819}
820__setup("noaliencache", noaliencache_setup);
821
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800822#ifdef CONFIG_NUMA
823/*
824 * Special reaping functions for NUMA systems called from cache_reap().
825 * These take care of doing round robin flushing of alien caches (containing
826 * objects freed on different nodes from which they were allocated) and the
827 * flushing of remote pcps by calling drain_node_pages.
828 */
829static DEFINE_PER_CPU(unsigned long, reap_node);
830
831static void init_reap_node(int cpu)
832{
833 int node;
834
835 node = next_node(cpu_to_node(cpu), node_online_map);
836 if (node == MAX_NUMNODES)
Paul Jackson442295c2006-03-22 00:09:11 -0800837 node = first_node(node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800838
Daniel Yeisley7f6b8872006-11-02 22:07:14 -0800839 per_cpu(reap_node, cpu) = node;
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800840}
841
842static void next_reap_node(void)
843{
844 int node = __get_cpu_var(reap_node);
845
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800846 node = next_node(node, node_online_map);
847 if (unlikely(node >= MAX_NUMNODES))
848 node = first_node(node_online_map);
849 __get_cpu_var(reap_node) = node;
850}
851
852#else
853#define init_reap_node(cpu) do { } while (0)
854#define next_reap_node(void) do { } while (0)
855#endif
856
Linus Torvalds1da177e2005-04-16 15:20:36 -0700857/*
858 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
859 * via the workqueue/eventd.
860 * Add the CPU number into the expiration time to minimize the possibility of
861 * the CPUs getting into lockstep and contending for the global cache chain
862 * lock.
863 */
Adrian Bunk897e6792007-07-15 23:38:20 -0700864static void __cpuinit start_cpu_timer(int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700865{
David Howells52bad642006-11-22 14:54:01 +0000866 struct delayed_work *reap_work = &per_cpu(reap_work, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700867
868 /*
869 * When this gets called from do_initcalls via cpucache_init(),
870 * init_workqueues() has already run, so keventd will be setup
871 * at that time.
872 */
David Howells52bad642006-11-22 14:54:01 +0000873 if (keventd_up() && reap_work->work.func == NULL) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800874 init_reap_node(cpu);
David Howells65f27f32006-11-22 14:55:48 +0000875 INIT_DELAYED_WORK(reap_work, cache_reap);
Arjan van de Ven2b284212006-12-10 02:21:28 -0800876 schedule_delayed_work_on(cpu, reap_work,
877 __round_jiffies_relative(HZ, cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700878 }
879}
880
Christoph Lametere498be72005-09-09 13:03:32 -0700881static struct array_cache *alloc_arraycache(int node, int entries,
Pekka Enberg83b519e2009-06-10 19:40:04 +0300882 int batchcount, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700883{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800884 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700885 struct array_cache *nc = NULL;
886
Pekka Enberg83b519e2009-06-10 19:40:04 +0300887 nc = kmalloc_node(memsize, gfp, node);
Catalin Marinasd5cff632009-06-11 13:22:40 +0100888 /*
889 * The array_cache structures contain pointers to free object.
890 * However, when such objects are allocated or transfered to another
891 * cache the pointers are not cleared and they could be counted as
892 * valid references during a kmemleak scan. Therefore, kmemleak must
893 * not scan such objects.
894 */
895 kmemleak_no_scan(nc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700896 if (nc) {
897 nc->avail = 0;
898 nc->limit = entries;
899 nc->batchcount = batchcount;
900 nc->touched = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700901 spin_lock_init(&nc->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700902 }
903 return nc;
904}
905
Christoph Lameter3ded1752006-03-25 03:06:44 -0800906/*
907 * Transfer objects in one arraycache to another.
908 * Locking must be handled by the caller.
909 *
910 * Return the number of entries transferred.
911 */
912static int transfer_objects(struct array_cache *to,
913 struct array_cache *from, unsigned int max)
914{
915 /* Figure out how many entries to transfer */
916 int nr = min(min(from->avail, max), to->limit - to->avail);
917
918 if (!nr)
919 return 0;
920
921 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
922 sizeof(void *) *nr);
923
924 from->avail -= nr;
925 to->avail += nr;
926 to->touched = 1;
927 return nr;
928}
929
Christoph Lameter765c4502006-09-27 01:50:08 -0700930#ifndef CONFIG_NUMA
931
932#define drain_alien_cache(cachep, alien) do { } while (0)
933#define reap_alien(cachep, l3) do { } while (0)
934
Pekka Enberg83b519e2009-06-10 19:40:04 +0300935static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
Christoph Lameter765c4502006-09-27 01:50:08 -0700936{
937 return (struct array_cache **)BAD_ALIEN_MAGIC;
938}
939
940static inline void free_alien_cache(struct array_cache **ac_ptr)
941{
942}
943
944static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
945{
946 return 0;
947}
948
949static inline void *alternate_node_alloc(struct kmem_cache *cachep,
950 gfp_t flags)
951{
952 return NULL;
953}
954
Christoph Hellwig8b98c162006-12-06 20:32:30 -0800955static inline void *____cache_alloc_node(struct kmem_cache *cachep,
Christoph Lameter765c4502006-09-27 01:50:08 -0700956 gfp_t flags, int nodeid)
957{
958 return NULL;
959}
960
961#else /* CONFIG_NUMA */
962
Christoph Hellwig8b98c162006-12-06 20:32:30 -0800963static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
Paul Jacksonc61afb12006-03-24 03:16:08 -0800964static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
Christoph Lameterdc85da12006-01-18 17:42:36 -0800965
Pekka Enberg83b519e2009-06-10 19:40:04 +0300966static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -0700967{
968 struct array_cache **ac_ptr;
Christoph Lameter8ef82862007-02-20 13:57:52 -0800969 int memsize = sizeof(void *) * nr_node_ids;
Christoph Lametere498be72005-09-09 13:03:32 -0700970 int i;
971
972 if (limit > 1)
973 limit = 12;
Pekka Enberg83b519e2009-06-10 19:40:04 +0300974 ac_ptr = kmalloc_node(memsize, gfp, node);
Christoph Lametere498be72005-09-09 13:03:32 -0700975 if (ac_ptr) {
976 for_each_node(i) {
977 if (i == node || !node_online(i)) {
978 ac_ptr[i] = NULL;
979 continue;
980 }
Pekka Enberg83b519e2009-06-10 19:40:04 +0300981 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
Christoph Lametere498be72005-09-09 13:03:32 -0700982 if (!ac_ptr[i]) {
Akinobu Mitacc550de2007-11-14 16:58:35 -0800983 for (i--; i >= 0; i--)
Christoph Lametere498be72005-09-09 13:03:32 -0700984 kfree(ac_ptr[i]);
985 kfree(ac_ptr);
986 return NULL;
987 }
988 }
989 }
990 return ac_ptr;
991}
992
Pekka Enberg5295a742006-02-01 03:05:48 -0800993static void free_alien_cache(struct array_cache **ac_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -0700994{
995 int i;
996
997 if (!ac_ptr)
998 return;
Christoph Lametere498be72005-09-09 13:03:32 -0700999 for_each_node(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001000 kfree(ac_ptr[i]);
Christoph Lametere498be72005-09-09 13:03:32 -07001001 kfree(ac_ptr);
1002}
1003
Pekka Enberg343e0d72006-02-01 03:05:50 -08001004static void __drain_alien_cache(struct kmem_cache *cachep,
Pekka Enberg5295a742006-02-01 03:05:48 -08001005 struct array_cache *ac, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07001006{
1007 struct kmem_list3 *rl3 = cachep->nodelists[node];
1008
1009 if (ac->avail) {
1010 spin_lock(&rl3->list_lock);
Christoph Lametere00946f2006-03-25 03:06:45 -08001011 /*
1012 * Stuff objects into the remote nodes shared array first.
1013 * That way we could avoid the overhead of putting the objects
1014 * into the free lists and getting them back later.
1015 */
shin, jacob693f7d32006-04-28 10:54:37 -05001016 if (rl3->shared)
1017 transfer_objects(rl3->shared, ac, ac->limit);
Christoph Lametere00946f2006-03-25 03:06:45 -08001018
Christoph Lameterff694162005-09-22 21:44:02 -07001019 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001020 ac->avail = 0;
1021 spin_unlock(&rl3->list_lock);
1022 }
1023}
1024
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001025/*
1026 * Called from cache_reap() to regularly drain alien caches round robin.
1027 */
1028static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1029{
1030 int node = __get_cpu_var(reap_node);
1031
1032 if (l3->alien) {
1033 struct array_cache *ac = l3->alien[node];
Christoph Lametere00946f2006-03-25 03:06:45 -08001034
1035 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001036 __drain_alien_cache(cachep, ac, node);
1037 spin_unlock_irq(&ac->lock);
1038 }
1039 }
1040}
1041
Andrew Mortona737b3e2006-03-22 00:08:11 -08001042static void drain_alien_cache(struct kmem_cache *cachep,
1043 struct array_cache **alien)
Christoph Lametere498be72005-09-09 13:03:32 -07001044{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001045 int i = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07001046 struct array_cache *ac;
1047 unsigned long flags;
1048
1049 for_each_online_node(i) {
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001050 ac = alien[i];
Christoph Lametere498be72005-09-09 13:03:32 -07001051 if (ac) {
1052 spin_lock_irqsave(&ac->lock, flags);
1053 __drain_alien_cache(cachep, ac, i);
1054 spin_unlock_irqrestore(&ac->lock, flags);
1055 }
1056 }
1057}
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001058
Ingo Molnar873623d2006-07-13 14:44:38 +02001059static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001060{
1061 struct slab *slabp = virt_to_slab(objp);
1062 int nodeid = slabp->nodeid;
1063 struct kmem_list3 *l3;
1064 struct array_cache *alien = NULL;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001065 int node;
1066
1067 node = numa_node_id();
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001068
1069 /*
1070 * Make sure we are not freeing a object from another node to the array
1071 * cache on this cpu.
1072 */
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001073 if (likely(slabp->nodeid == node))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001074 return 0;
1075
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001076 l3 = cachep->nodelists[node];
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001077 STATS_INC_NODEFREES(cachep);
1078 if (l3->alien && l3->alien[nodeid]) {
1079 alien = l3->alien[nodeid];
Ingo Molnar873623d2006-07-13 14:44:38 +02001080 spin_lock(&alien->lock);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001081 if (unlikely(alien->avail == alien->limit)) {
1082 STATS_INC_ACOVERFLOW(cachep);
1083 __drain_alien_cache(cachep, alien, nodeid);
1084 }
1085 alien->entry[alien->avail++] = objp;
1086 spin_unlock(&alien->lock);
1087 } else {
1088 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1089 free_block(cachep, &objp, 1, nodeid);
1090 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1091 }
1092 return 1;
1093}
Christoph Lametere498be72005-09-09 13:03:32 -07001094#endif
1095
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001096static void __cpuinit cpuup_canceled(long cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001097{
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001098 struct kmem_cache *cachep;
1099 struct kmem_list3 *l3 = NULL;
1100 int node = cpu_to_node(cpu);
Rusty Russella70f7302009-03-13 14:49:46 +10301101 const struct cpumask *mask = cpumask_of_node(node);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001102
1103 list_for_each_entry(cachep, &cache_chain, next) {
1104 struct array_cache *nc;
1105 struct array_cache *shared;
1106 struct array_cache **alien;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001107
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001108 /* cpu is dead; no one can alloc from it. */
1109 nc = cachep->array[cpu];
1110 cachep->array[cpu] = NULL;
1111 l3 = cachep->nodelists[node];
1112
1113 if (!l3)
1114 goto free_array_cache;
1115
1116 spin_lock_irq(&l3->list_lock);
1117
1118 /* Free limit for this kmem_list3 */
1119 l3->free_limit -= cachep->batchcount;
1120 if (nc)
1121 free_block(cachep, nc->entry, nc->avail, node);
1122
Mike Travisc5f59f02008-04-04 18:11:10 -07001123 if (!cpus_empty(*mask)) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001124 spin_unlock_irq(&l3->list_lock);
1125 goto free_array_cache;
1126 }
1127
1128 shared = l3->shared;
1129 if (shared) {
1130 free_block(cachep, shared->entry,
1131 shared->avail, node);
1132 l3->shared = NULL;
1133 }
1134
1135 alien = l3->alien;
1136 l3->alien = NULL;
1137
1138 spin_unlock_irq(&l3->list_lock);
1139
1140 kfree(shared);
1141 if (alien) {
1142 drain_alien_cache(cachep, alien);
1143 free_alien_cache(alien);
1144 }
1145free_array_cache:
1146 kfree(nc);
1147 }
1148 /*
1149 * In the previous loop, all the objects were freed to
1150 * the respective cache's slabs, now we can go ahead and
1151 * shrink each nodelist to its limit.
1152 */
1153 list_for_each_entry(cachep, &cache_chain, next) {
1154 l3 = cachep->nodelists[node];
1155 if (!l3)
1156 continue;
1157 drain_freelist(cachep, l3, l3->free_objects);
1158 }
1159}
1160
1161static int __cpuinit cpuup_prepare(long cpu)
1162{
Pekka Enberg343e0d72006-02-01 03:05:50 -08001163 struct kmem_cache *cachep;
Christoph Lametere498be72005-09-09 13:03:32 -07001164 struct kmem_list3 *l3 = NULL;
1165 int node = cpu_to_node(cpu);
David Howellsea02e3d2007-07-19 01:49:09 -07001166 const int memsize = sizeof(struct kmem_list3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001167
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001168 /*
1169 * We need to do this right in the beginning since
1170 * alloc_arraycache's are going to use this list.
1171 * kmalloc_node allows us to add the slab to the right
1172 * kmem_list3 and not this cpu's kmem_list3
1173 */
1174
1175 list_for_each_entry(cachep, &cache_chain, next) {
1176 /*
1177 * Set up the size64 kmemlist for cpu before we can
1178 * begin anything. Make sure some other cpu on this
1179 * node has not already allocated this
1180 */
1181 if (!cachep->nodelists[node]) {
1182 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1183 if (!l3)
1184 goto bad;
1185 kmem_list3_init(l3);
1186 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1187 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1188
1189 /*
1190 * The l3s don't come and go as CPUs come and
1191 * go. cache_chain_mutex is sufficient
1192 * protection here.
1193 */
1194 cachep->nodelists[node] = l3;
1195 }
1196
1197 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1198 cachep->nodelists[node]->free_limit =
1199 (1 + nr_cpus_node(node)) *
1200 cachep->batchcount + cachep->num;
1201 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1202 }
1203
1204 /*
1205 * Now we can go ahead with allocating the shared arrays and
1206 * array caches
1207 */
1208 list_for_each_entry(cachep, &cache_chain, next) {
1209 struct array_cache *nc;
1210 struct array_cache *shared = NULL;
1211 struct array_cache **alien = NULL;
1212
1213 nc = alloc_arraycache(node, cachep->limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03001214 cachep->batchcount, GFP_KERNEL);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001215 if (!nc)
1216 goto bad;
1217 if (cachep->shared) {
1218 shared = alloc_arraycache(node,
1219 cachep->shared * cachep->batchcount,
Pekka Enberg83b519e2009-06-10 19:40:04 +03001220 0xbaadf00d, GFP_KERNEL);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001221 if (!shared) {
1222 kfree(nc);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001223 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001224 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001225 }
1226 if (use_alien_caches) {
Pekka Enberg83b519e2009-06-10 19:40:04 +03001227 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001228 if (!alien) {
1229 kfree(shared);
1230 kfree(nc);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001231 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001232 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001233 }
1234 cachep->array[cpu] = nc;
1235 l3 = cachep->nodelists[node];
1236 BUG_ON(!l3);
1237
1238 spin_lock_irq(&l3->list_lock);
1239 if (!l3->shared) {
1240 /*
1241 * We are serialised from CPU_DEAD or
1242 * CPU_UP_CANCELLED by the cpucontrol lock
1243 */
1244 l3->shared = shared;
1245 shared = NULL;
1246 }
1247#ifdef CONFIG_NUMA
1248 if (!l3->alien) {
1249 l3->alien = alien;
1250 alien = NULL;
1251 }
1252#endif
1253 spin_unlock_irq(&l3->list_lock);
1254 kfree(shared);
1255 free_alien_cache(alien);
1256 }
1257 return 0;
1258bad:
Akinobu Mita12d00f62007-10-18 03:05:11 -07001259 cpuup_canceled(cpu);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001260 return -ENOMEM;
1261}
1262
1263static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1264 unsigned long action, void *hcpu)
1265{
1266 long cpu = (long)hcpu;
1267 int err = 0;
1268
Linus Torvalds1da177e2005-04-16 15:20:36 -07001269 switch (action) {
Heiko Carstens38c3bd92007-05-09 02:34:05 -07001270 case CPU_UP_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001271 case CPU_UP_PREPARE_FROZEN:
Gautham R Shenoy95402b32008-01-25 21:08:02 +01001272 mutex_lock(&cache_chain_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001273 err = cpuup_prepare(cpu);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01001274 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001275 break;
1276 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001277 case CPU_ONLINE_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001278 start_cpu_timer(cpu);
1279 break;
1280#ifdef CONFIG_HOTPLUG_CPU
Christoph Lameter5830c592007-05-09 02:34:22 -07001281 case CPU_DOWN_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001282 case CPU_DOWN_PREPARE_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001283 /*
1284 * Shutdown cache reaper. Note that the cache_chain_mutex is
1285 * held so that if cache_reap() is invoked it cannot do
1286 * anything expensive but will only modify reap_work
1287 * and reschedule the timer.
1288 */
1289 cancel_rearming_delayed_work(&per_cpu(reap_work, cpu));
1290 /* Now the cache_reaper is guaranteed to be not running. */
1291 per_cpu(reap_work, cpu).work.func = NULL;
1292 break;
1293 case CPU_DOWN_FAILED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001294 case CPU_DOWN_FAILED_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001295 start_cpu_timer(cpu);
1296 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001297 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001298 case CPU_DEAD_FROZEN:
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001299 /*
1300 * Even if all the cpus of a node are down, we don't free the
1301 * kmem_list3 of any cache. This to avoid a race between
1302 * cpu_down, and a kmalloc allocation from another cpu for
1303 * memory from the node of the cpu going down. The list3
1304 * structure is usually allocated from kmem_cache_create() and
1305 * gets destroyed at kmem_cache_destroy().
1306 */
Simon Arlott183ff222007-10-20 01:27:18 +02001307 /* fall through */
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08001308#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001309 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001310 case CPU_UP_CANCELED_FROZEN:
Gautham R Shenoy95402b32008-01-25 21:08:02 +01001311 mutex_lock(&cache_chain_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001312 cpuup_canceled(cpu);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001313 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001314 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001315 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001316 return err ? NOTIFY_BAD : NOTIFY_OK;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001317}
1318
Chandra Seetharaman74b85f32006-06-27 02:54:09 -07001319static struct notifier_block __cpuinitdata cpucache_notifier = {
1320 &cpuup_callback, NULL, 0
1321};
Linus Torvalds1da177e2005-04-16 15:20:36 -07001322
Christoph Lametere498be72005-09-09 13:03:32 -07001323/*
1324 * swap the static kmem_list3 with kmalloced memory
1325 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001326static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1327 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001328{
1329 struct kmem_list3 *ptr;
1330
Pekka Enberg83b519e2009-06-10 19:40:04 +03001331 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07001332 BUG_ON(!ptr);
1333
Christoph Lametere498be72005-09-09 13:03:32 -07001334 memcpy(ptr, list, sizeof(struct kmem_list3));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001335 /*
1336 * Do not assume that spinlocks can be initialized via memcpy:
1337 */
1338 spin_lock_init(&ptr->list_lock);
1339
Christoph Lametere498be72005-09-09 13:03:32 -07001340 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1341 cachep->nodelists[nodeid] = ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001342}
1343
Andrew Mortona737b3e2006-03-22 00:08:11 -08001344/*
Pekka Enberg556a1692008-01-25 08:20:51 +02001345 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1346 * size of kmem_list3.
1347 */
1348static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1349{
1350 int node;
1351
1352 for_each_online_node(node) {
1353 cachep->nodelists[node] = &initkmem_list3[index + node];
1354 cachep->nodelists[node]->next_reap = jiffies +
1355 REAPTIMEOUT_LIST3 +
1356 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1357 }
1358}
1359
1360/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08001361 * Initialisation. Called after the page allocator have been initialised and
1362 * before smp_init().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001363 */
1364void __init kmem_cache_init(void)
1365{
1366 size_t left_over;
1367 struct cache_sizes *sizes;
1368 struct cache_names *names;
Christoph Lametere498be72005-09-09 13:03:32 -07001369 int i;
Jack Steiner07ed76b2006-03-07 21:55:46 -08001370 int order;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001371 int node;
Christoph Lametere498be72005-09-09 13:03:32 -07001372
Mel Gormanb6e68bc2009-06-16 15:32:16 -07001373 if (num_possible_nodes() == 1)
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001374 use_alien_caches = 0;
1375
Christoph Lametere498be72005-09-09 13:03:32 -07001376 for (i = 0; i < NUM_INIT_LISTS; i++) {
1377 kmem_list3_init(&initkmem_list3[i]);
1378 if (i < MAX_NUMNODES)
1379 cache_cache.nodelists[i] = NULL;
1380 }
Pekka Enberg556a1692008-01-25 08:20:51 +02001381 set_up_list3s(&cache_cache, CACHE_CACHE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001382
1383 /*
1384 * Fragmentation resistance on low memory - only use bigger
1385 * page orders on machines with more than 32MB of memory.
1386 */
1387 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1388 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1389
Linus Torvalds1da177e2005-04-16 15:20:36 -07001390 /* Bootstrap is tricky, because several objects are allocated
1391 * from caches that do not exist yet:
Andrew Mortona737b3e2006-03-22 00:08:11 -08001392 * 1) initialize the cache_cache cache: it contains the struct
1393 * kmem_cache structures of all caches, except cache_cache itself:
1394 * cache_cache is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001395 * Initially an __init data area is used for the head array and the
1396 * kmem_list3 structures, it's replaced with a kmalloc allocated
1397 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001398 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001399 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001400 * An __init data area is used for the head array.
1401 * 3) Create the remaining kmalloc caches, with minimally sized
1402 * head arrays.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001403 * 4) Replace the __init data head arrays for cache_cache and the first
1404 * kmalloc cache with kmalloc allocated arrays.
Christoph Lametere498be72005-09-09 13:03:32 -07001405 * 5) Replace the __init data for kmem_list3 for cache_cache and
1406 * the other cache's with kmalloc allocated memory.
1407 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001408 */
1409
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001410 node = numa_node_id();
1411
Linus Torvalds1da177e2005-04-16 15:20:36 -07001412 /* 1) create the cache_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001413 INIT_LIST_HEAD(&cache_chain);
1414 list_add(&cache_cache.next, &cache_chain);
1415 cache_cache.colour_off = cache_line_size();
1416 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
Daniel Yeisleyec1f5ee2008-03-25 23:59:08 +02001417 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001418
Eric Dumazet8da34302007-05-06 14:49:29 -07001419 /*
1420 * struct kmem_cache size depends on nr_node_ids, which
1421 * can be less than MAX_NUMNODES.
1422 */
1423 cache_cache.buffer_size = offsetof(struct kmem_cache, nodelists) +
1424 nr_node_ids * sizeof(struct kmem_list3 *);
1425#if DEBUG
1426 cache_cache.obj_size = cache_cache.buffer_size;
1427#endif
Andrew Mortona737b3e2006-03-22 00:08:11 -08001428 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1429 cache_line_size());
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08001430 cache_cache.reciprocal_buffer_size =
1431 reciprocal_value(cache_cache.buffer_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001432
Jack Steiner07ed76b2006-03-07 21:55:46 -08001433 for (order = 0; order < MAX_ORDER; order++) {
1434 cache_estimate(order, cache_cache.buffer_size,
1435 cache_line_size(), 0, &left_over, &cache_cache.num);
1436 if (cache_cache.num)
1437 break;
1438 }
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02001439 BUG_ON(!cache_cache.num);
Jack Steiner07ed76b2006-03-07 21:55:46 -08001440 cache_cache.gfporder = order;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001441 cache_cache.colour = left_over / cache_cache.colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001442 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1443 sizeof(struct slab), cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001444
1445 /* 2+3) create the kmalloc caches */
1446 sizes = malloc_sizes;
1447 names = cache_names;
1448
Andrew Mortona737b3e2006-03-22 00:08:11 -08001449 /*
1450 * Initialize the caches that provide memory for the array cache and the
1451 * kmem_list3 structures first. Without this, further allocations will
1452 * bug.
Christoph Lametere498be72005-09-09 13:03:32 -07001453 */
1454
1455 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001456 sizes[INDEX_AC].cs_size,
1457 ARCH_KMALLOC_MINALIGN,
1458 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001459 NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07001460
Andrew Mortona737b3e2006-03-22 00:08:11 -08001461 if (INDEX_AC != INDEX_L3) {
Christoph Lametere498be72005-09-09 13:03:32 -07001462 sizes[INDEX_L3].cs_cachep =
Andrew Mortona737b3e2006-03-22 00:08:11 -08001463 kmem_cache_create(names[INDEX_L3].name,
1464 sizes[INDEX_L3].cs_size,
1465 ARCH_KMALLOC_MINALIGN,
1466 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001467 NULL);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001468 }
Christoph Lametere498be72005-09-09 13:03:32 -07001469
Ingo Molnare0a42722006-06-23 02:03:46 -07001470 slab_early_init = 0;
1471
Linus Torvalds1da177e2005-04-16 15:20:36 -07001472 while (sizes->cs_size != ULONG_MAX) {
Christoph Lametere498be72005-09-09 13:03:32 -07001473 /*
1474 * For performance, all the general caches are L1 aligned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001475 * This should be particularly beneficial on SMP boxes, as it
1476 * eliminates "false sharing".
1477 * Note for systems short on memory removing the alignment will
Christoph Lametere498be72005-09-09 13:03:32 -07001478 * allow tighter packing of the smaller caches.
1479 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001480 if (!sizes->cs_cachep) {
Christoph Lametere498be72005-09-09 13:03:32 -07001481 sizes->cs_cachep = kmem_cache_create(names->name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001482 sizes->cs_size,
1483 ARCH_KMALLOC_MINALIGN,
1484 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001485 NULL);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001486 }
Christoph Lameter4b51d662007-02-10 01:43:10 -08001487#ifdef CONFIG_ZONE_DMA
1488 sizes->cs_dmacachep = kmem_cache_create(
1489 names->name_dma,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001490 sizes->cs_size,
1491 ARCH_KMALLOC_MINALIGN,
1492 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1493 SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001494 NULL);
Christoph Lameter4b51d662007-02-10 01:43:10 -08001495#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001496 sizes++;
1497 names++;
1498 }
1499 /* 4) Replace the bootstrap head arrays */
1500 {
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001501 struct array_cache *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001502
Pekka Enberg83b519e2009-06-10 19:40:04 +03001503 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
Christoph Lametere498be72005-09-09 13:03:32 -07001504
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001505 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1506 memcpy(ptr, cpu_cache_get(&cache_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001507 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001508 /*
1509 * Do not assume that spinlocks can be initialized via memcpy:
1510 */
1511 spin_lock_init(&ptr->lock);
1512
Linus Torvalds1da177e2005-04-16 15:20:36 -07001513 cache_cache.array[smp_processor_id()] = ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001514
Pekka Enberg83b519e2009-06-10 19:40:04 +03001515 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
Christoph Lametere498be72005-09-09 13:03:32 -07001516
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001517 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001518 != &initarray_generic.cache);
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001519 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001520 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001521 /*
1522 * Do not assume that spinlocks can be initialized via memcpy:
1523 */
1524 spin_lock_init(&ptr->lock);
1525
Christoph Lametere498be72005-09-09 13:03:32 -07001526 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001527 ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001528 }
Christoph Lametere498be72005-09-09 13:03:32 -07001529 /* 5) Replace the bootstrap kmem_list3's */
1530 {
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001531 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001532
Mel Gorman9c09a952008-01-24 05:49:54 -08001533 for_each_online_node(nid) {
Daniel Yeisleyec1f5ee2008-03-25 23:59:08 +02001534 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid);
Pekka Enberg556a1692008-01-25 08:20:51 +02001535
Christoph Lametere498be72005-09-09 13:03:32 -07001536 init_list(malloc_sizes[INDEX_AC].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001537 &initkmem_list3[SIZE_AC + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001538
1539 if (INDEX_AC != INDEX_L3) {
1540 init_list(malloc_sizes[INDEX_L3].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001541 &initkmem_list3[SIZE_L3 + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001542 }
1543 }
1544 }
1545
Pekka Enberg8429db52009-06-12 15:58:59 +03001546 g_cpucache_up = EARLY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001547
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001548 /* Annotate slab for lockdep -- annotate the malloc caches */
1549 init_lock_keys();
Pekka Enberg8429db52009-06-12 15:58:59 +03001550}
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001551
Pekka Enberg8429db52009-06-12 15:58:59 +03001552void __init kmem_cache_init_late(void)
1553{
1554 struct kmem_cache *cachep;
1555
Pekka Enberg8429db52009-06-12 15:58:59 +03001556 /* 6) resize the head arrays to their final sizes */
1557 mutex_lock(&cache_chain_mutex);
1558 list_for_each_entry(cachep, &cache_chain, next)
1559 if (enable_cpucache(cachep, GFP_NOWAIT))
1560 BUG();
1561 mutex_unlock(&cache_chain_mutex);
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001562
Linus Torvalds1da177e2005-04-16 15:20:36 -07001563 /* Done! */
1564 g_cpucache_up = FULL;
1565
Andrew Mortona737b3e2006-03-22 00:08:11 -08001566 /*
1567 * Register a cpu startup notifier callback that initializes
1568 * cpu_cache_get for all new cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07001569 */
1570 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001571
Andrew Mortona737b3e2006-03-22 00:08:11 -08001572 /*
1573 * The reap timers are started later, with a module init call: That part
1574 * of the kernel is not yet operational.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001575 */
1576}
1577
1578static int __init cpucache_init(void)
1579{
1580 int cpu;
1581
Andrew Mortona737b3e2006-03-22 00:08:11 -08001582 /*
1583 * Register the timers that return unneeded pages to the page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -07001584 */
Christoph Lametere498be72005-09-09 13:03:32 -07001585 for_each_online_cpu(cpu)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001586 start_cpu_timer(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001587 return 0;
1588}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001589__initcall(cpucache_init);
1590
1591/*
1592 * Interface to system's page allocator. No need to hold the cache-lock.
1593 *
1594 * If we requested dmaable memory, we will get it. Even if we
1595 * did not request dmaable memory, we might get it, but that
1596 * would be relatively rare and ignorable.
1597 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001598static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001599{
1600 struct page *page;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001601 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001602 int i;
1603
Luke Yangd6fef9d2006-04-10 22:52:56 -07001604#ifndef CONFIG_MMU
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001605 /*
1606 * Nommu uses slab's for process anonymous memory allocations, and thus
1607 * requires __GFP_COMP to properly refcount higher order allocations
Luke Yangd6fef9d2006-04-10 22:52:56 -07001608 */
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001609 flags |= __GFP_COMP;
Luke Yangd6fef9d2006-04-10 22:52:56 -07001610#endif
Christoph Lameter765c4502006-09-27 01:50:08 -07001611
Christoph Lameter3c517a62006-12-06 20:33:29 -08001612 flags |= cachep->gfpflags;
Mel Gormane12ba742007-10-16 01:25:52 -07001613 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1614 flags |= __GFP_RECLAIMABLE;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001615
Linus Torvalds517d0862009-06-16 19:50:13 -07001616 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001617 if (!page)
1618 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001619
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001620 nr_pages = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001621 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
Christoph Lameter972d1a72006-09-25 23:31:51 -07001622 add_zone_page_state(page_zone(page),
1623 NR_SLAB_RECLAIMABLE, nr_pages);
1624 else
1625 add_zone_page_state(page_zone(page),
1626 NR_SLAB_UNRECLAIMABLE, nr_pages);
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001627 for (i = 0; i < nr_pages; i++)
1628 __SetPageSlab(page + i);
Pekka Enbergc175eea2008-05-09 20:35:53 +02001629
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001630 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1631 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1632
1633 if (cachep->ctor)
1634 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1635 else
1636 kmemcheck_mark_unallocated_pages(page, nr_pages);
1637 }
Pekka Enbergc175eea2008-05-09 20:35:53 +02001638
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001639 return page_address(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001640}
1641
1642/*
1643 * Interface to system's page release.
1644 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001645static void kmem_freepages(struct kmem_cache *cachep, void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001646{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001647 unsigned long i = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001648 struct page *page = virt_to_page(addr);
1649 const unsigned long nr_freed = i;
1650
Vegard Nossumb1eeab62008-11-25 16:55:53 +01001651 kmemcheck_free_shadow(page, cachep->gfporder);
Pekka Enbergc175eea2008-05-09 20:35:53 +02001652
Christoph Lameter972d1a72006-09-25 23:31:51 -07001653 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1654 sub_zone_page_state(page_zone(page),
1655 NR_SLAB_RECLAIMABLE, nr_freed);
1656 else
1657 sub_zone_page_state(page_zone(page),
1658 NR_SLAB_UNRECLAIMABLE, nr_freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001659 while (i--) {
Nick Pigginf205b2f2006-03-22 00:08:02 -08001660 BUG_ON(!PageSlab(page));
1661 __ClearPageSlab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001662 page++;
1663 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001664 if (current->reclaim_state)
1665 current->reclaim_state->reclaimed_slab += nr_freed;
1666 free_pages((unsigned long)addr, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001667}
1668
1669static void kmem_rcu_free(struct rcu_head *head)
1670{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001671 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001672 struct kmem_cache *cachep = slab_rcu->cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001673
1674 kmem_freepages(cachep, slab_rcu->addr);
1675 if (OFF_SLAB(cachep))
1676 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1677}
1678
1679#if DEBUG
1680
1681#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001682static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001683 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001684{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001685 int size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001686
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001687 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001688
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001689 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001690 return;
1691
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001692 *addr++ = 0x12345678;
1693 *addr++ = caller;
1694 *addr++ = smp_processor_id();
1695 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001696 {
1697 unsigned long *sptr = &caller;
1698 unsigned long svalue;
1699
1700 while (!kstack_end(sptr)) {
1701 svalue = *sptr++;
1702 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001703 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001704 size -= sizeof(unsigned long);
1705 if (size <= sizeof(unsigned long))
1706 break;
1707 }
1708 }
1709
1710 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001711 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001712}
1713#endif
1714
Pekka Enberg343e0d72006-02-01 03:05:50 -08001715static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001716{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001717 int size = obj_size(cachep);
1718 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001719
1720 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001721 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001722}
1723
1724static void dump_line(char *data, int offset, int limit)
1725{
1726 int i;
Dave Jonesaa83aa42006-09-29 01:59:51 -07001727 unsigned char error = 0;
1728 int bad_count = 0;
1729
Linus Torvalds1da177e2005-04-16 15:20:36 -07001730 printk(KERN_ERR "%03x:", offset);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001731 for (i = 0; i < limit; i++) {
1732 if (data[offset + i] != POISON_FREE) {
1733 error = data[offset + i];
1734 bad_count++;
1735 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001736 printk(" %02x", (unsigned char)data[offset + i]);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001737 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001738 printk("\n");
Dave Jonesaa83aa42006-09-29 01:59:51 -07001739
1740 if (bad_count == 1) {
1741 error ^= POISON_FREE;
1742 if (!(error & (error - 1))) {
1743 printk(KERN_ERR "Single bit error detected. Probably "
1744 "bad RAM.\n");
1745#ifdef CONFIG_X86
1746 printk(KERN_ERR "Run memtest86+ or a similar memory "
1747 "test tool.\n");
1748#else
1749 printk(KERN_ERR "Run a memory test tool.\n");
1750#endif
1751 }
1752 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001753}
1754#endif
1755
1756#if DEBUG
1757
Pekka Enberg343e0d72006-02-01 03:05:50 -08001758static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001759{
1760 int i, size;
1761 char *realobj;
1762
1763 if (cachep->flags & SLAB_RED_ZONE) {
David Woodhouseb46b8f12007-05-08 00:22:59 -07001764 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001765 *dbg_redzone1(cachep, objp),
1766 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001767 }
1768
1769 if (cachep->flags & SLAB_STORE_USER) {
1770 printk(KERN_ERR "Last user: [<%p>]",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001771 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001772 print_symbol("(%s)",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001773 (unsigned long)*dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001774 printk("\n");
1775 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001776 realobj = (char *)objp + obj_offset(cachep);
1777 size = obj_size(cachep);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001778 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001779 int limit;
1780 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001781 if (i + limit > size)
1782 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001783 dump_line(realobj, i, limit);
1784 }
1785}
1786
Pekka Enberg343e0d72006-02-01 03:05:50 -08001787static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001788{
1789 char *realobj;
1790 int size, i;
1791 int lines = 0;
1792
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001793 realobj = (char *)objp + obj_offset(cachep);
1794 size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001795
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001796 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001797 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001798 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001799 exp = POISON_END;
1800 if (realobj[i] != exp) {
1801 int limit;
1802 /* Mismatch ! */
1803 /* Print header */
1804 if (lines == 0) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001805 printk(KERN_ERR
David Howellse94a40c2007-04-02 23:46:28 +01001806 "Slab corruption: %s start=%p, len=%d\n",
1807 cachep->name, realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001808 print_objinfo(cachep, objp, 0);
1809 }
1810 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001811 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001812 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001813 if (i + limit > size)
1814 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001815 dump_line(realobj, i, limit);
1816 i += 16;
1817 lines++;
1818 /* Limit to 5 lines */
1819 if (lines > 5)
1820 break;
1821 }
1822 }
1823 if (lines != 0) {
1824 /* Print some data about the neighboring objects, if they
1825 * exist:
1826 */
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08001827 struct slab *slabp = virt_to_slab(objp);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001828 unsigned int objnr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001829
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001830 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001831 if (objnr) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001832 objp = index_to_obj(cachep, slabp, objnr - 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001833 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001834 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001835 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001836 print_objinfo(cachep, objp, 2);
1837 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001838 if (objnr + 1 < cachep->num) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001839 objp = index_to_obj(cachep, slabp, objnr + 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001840 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001841 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001842 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001843 print_objinfo(cachep, objp, 2);
1844 }
1845 }
1846}
1847#endif
1848
Linus Torvalds1da177e2005-04-16 15:20:36 -07001849#if DEBUG
Rabin Vincente79aec22008-07-04 00:40:32 +05301850static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001851{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001852 int i;
1853 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001854 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001855
1856 if (cachep->flags & SLAB_POISON) {
1857#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08001858 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1859 OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001860 kernel_map_pages(virt_to_page(objp),
Andrew Mortona737b3e2006-03-22 00:08:11 -08001861 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001862 else
1863 check_poison_obj(cachep, objp);
1864#else
1865 check_poison_obj(cachep, objp);
1866#endif
1867 }
1868 if (cachep->flags & SLAB_RED_ZONE) {
1869 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1870 slab_error(cachep, "start of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001871 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001872 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1873 slab_error(cachep, "end of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001874 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001875 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001876 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001877}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001878#else
Rabin Vincente79aec22008-07-04 00:40:32 +05301879static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001880{
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001881}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001882#endif
1883
Randy Dunlap911851e2006-03-22 00:08:14 -08001884/**
1885 * slab_destroy - destroy and release all objects in a slab
1886 * @cachep: cache pointer being destroyed
1887 * @slabp: slab pointer being destroyed
1888 *
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001889 * Destroy all the objs in a slab, and release the mem back to the system.
Andrew Mortona737b3e2006-03-22 00:08:11 -08001890 * Before calling the slab must have been unlinked from the cache. The
1891 * cache-lock is not held/needed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001892 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001893static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001894{
1895 void *addr = slabp->s_mem - slabp->colouroff;
1896
Rabin Vincente79aec22008-07-04 00:40:32 +05301897 slab_destroy_debugcheck(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001898 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1899 struct slab_rcu *slab_rcu;
1900
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001901 slab_rcu = (struct slab_rcu *)slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001902 slab_rcu->cachep = cachep;
1903 slab_rcu->addr = addr;
1904 call_rcu(&slab_rcu->head, kmem_rcu_free);
1905 } else {
1906 kmem_freepages(cachep, addr);
Ingo Molnar873623d2006-07-13 14:44:38 +02001907 if (OFF_SLAB(cachep))
1908 kmem_cache_free(cachep->slabp_cache, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001909 }
1910}
1911
Christoph Lameter117f6eb2006-09-25 23:31:37 -07001912static void __kmem_cache_destroy(struct kmem_cache *cachep)
1913{
1914 int i;
1915 struct kmem_list3 *l3;
1916
1917 for_each_online_cpu(i)
1918 kfree(cachep->array[i]);
1919
1920 /* NUMA: free the list3 structures */
1921 for_each_online_node(i) {
1922 l3 = cachep->nodelists[i];
1923 if (l3) {
1924 kfree(l3->shared);
1925 free_alien_cache(l3->alien);
1926 kfree(l3);
1927 }
1928 }
1929 kmem_cache_free(&cache_cache, cachep);
1930}
1931
1932
Linus Torvalds1da177e2005-04-16 15:20:36 -07001933/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08001934 * calculate_slab_order - calculate size (page order) of slabs
1935 * @cachep: pointer to the cache that is being created
1936 * @size: size of objects to be created in this cache.
1937 * @align: required alignment for the objects.
1938 * @flags: slab allocation flags
1939 *
1940 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001941 *
1942 * This could be made much more intelligent. For now, try to avoid using
1943 * high order pages for slabs. When the gfp() functions are more friendly
1944 * towards high-order requests, this should be changed.
1945 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001946static size_t calculate_slab_order(struct kmem_cache *cachep,
Randy Dunlapee13d782006-02-01 03:05:53 -08001947 size_t size, size_t align, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001948{
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001949 unsigned long offslab_limit;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001950 size_t left_over = 0;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001951 int gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001952
Christoph Lameter0aa817f2007-05-16 22:11:01 -07001953 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001954 unsigned int num;
1955 size_t remainder;
1956
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001957 cache_estimate(gfporder, size, align, flags, &remainder, &num);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001958 if (!num)
1959 continue;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001960
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001961 if (flags & CFLGS_OFF_SLAB) {
1962 /*
1963 * Max number of objs-per-slab for caches which
1964 * use off-slab slabs. Needed to avoid a possible
1965 * looping condition in cache_grow().
1966 */
1967 offslab_limit = size - sizeof(struct slab);
1968 offslab_limit /= sizeof(kmem_bufctl_t);
1969
1970 if (num > offslab_limit)
1971 break;
1972 }
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001973
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001974 /* Found something acceptable - save it away */
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001975 cachep->num = num;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001976 cachep->gfporder = gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001977 left_over = remainder;
1978
1979 /*
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08001980 * A VFS-reclaimable slab tends to have most allocations
1981 * as GFP_NOFS and we really don't want to have to be allocating
1982 * higher-order pages when we are unable to shrink dcache.
1983 */
1984 if (flags & SLAB_RECLAIM_ACCOUNT)
1985 break;
1986
1987 /*
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001988 * Large number of objects is good, but very large slabs are
1989 * currently bad for the gfp()s.
1990 */
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001991 if (gfporder >= slab_break_gfp_order)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001992 break;
1993
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001994 /*
1995 * Acceptable internal fragmentation?
1996 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001997 if (left_over * 8 <= (PAGE_SIZE << gfporder))
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001998 break;
1999 }
2000 return left_over;
2001}
2002
Pekka Enberg83b519e2009-06-10 19:40:04 +03002003static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002004{
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002005 if (g_cpucache_up == FULL)
Pekka Enberg83b519e2009-06-10 19:40:04 +03002006 return enable_cpucache(cachep, gfp);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002007
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002008 if (g_cpucache_up == NONE) {
2009 /*
2010 * Note: the first kmem_cache_create must create the cache
2011 * that's used by kmalloc(24), otherwise the creation of
2012 * further caches will BUG().
2013 */
2014 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2015
2016 /*
2017 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2018 * the first cache, then we need to set up all its list3s,
2019 * otherwise the creation of further caches will BUG().
2020 */
2021 set_up_list3s(cachep, SIZE_AC);
2022 if (INDEX_AC == INDEX_L3)
2023 g_cpucache_up = PARTIAL_L3;
2024 else
2025 g_cpucache_up = PARTIAL_AC;
2026 } else {
2027 cachep->array[smp_processor_id()] =
Pekka Enberg83b519e2009-06-10 19:40:04 +03002028 kmalloc(sizeof(struct arraycache_init), gfp);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002029
2030 if (g_cpucache_up == PARTIAL_AC) {
2031 set_up_list3s(cachep, SIZE_L3);
2032 g_cpucache_up = PARTIAL_L3;
2033 } else {
2034 int node;
Pekka Enberg556a1692008-01-25 08:20:51 +02002035 for_each_online_node(node) {
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002036 cachep->nodelists[node] =
2037 kmalloc_node(sizeof(struct kmem_list3),
Pekka Enbergeb91f1d2009-06-12 14:56:09 +03002038 gfp, node);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002039 BUG_ON(!cachep->nodelists[node]);
2040 kmem_list3_init(cachep->nodelists[node]);
2041 }
2042 }
2043 }
2044 cachep->nodelists[numa_node_id()]->next_reap =
2045 jiffies + REAPTIMEOUT_LIST3 +
2046 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2047
2048 cpu_cache_get(cachep)->avail = 0;
2049 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2050 cpu_cache_get(cachep)->batchcount = 1;
2051 cpu_cache_get(cachep)->touched = 0;
2052 cachep->batchcount = 1;
2053 cachep->limit = BOOT_CPUCACHE_ENTRIES;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002054 return 0;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002055}
2056
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002057/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07002058 * kmem_cache_create - Create a cache.
2059 * @name: A string which is used in /proc/slabinfo to identify this cache.
2060 * @size: The size of objects to be created in this cache.
2061 * @align: The required alignment for the objects.
2062 * @flags: SLAB flags
2063 * @ctor: A constructor for the objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002064 *
2065 * Returns a ptr to the cache on success, NULL on failure.
2066 * Cannot be called within a int, but can be interrupted.
Paul Mundt20c2df82007-07-20 10:11:58 +09002067 * The @ctor is run when new pages are allocated by the cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002068 *
2069 * @name must be valid until the cache is destroyed. This implies that
Andrew Mortona737b3e2006-03-22 00:08:11 -08002070 * the module calling this has to destroy the cache before getting unloaded.
Catalin Marinas249da162008-11-21 12:56:22 +00002071 * Note that kmem_cache_name() is not guaranteed to return the same pointer,
2072 * therefore applications must manage it themselves.
Andrew Mortona737b3e2006-03-22 00:08:11 -08002073 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002074 * The flags are
2075 *
2076 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2077 * to catch references to uninitialised memory.
2078 *
2079 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2080 * for buffer overruns.
2081 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002082 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2083 * cacheline. This can be beneficial if you're counting cycles as closely
2084 * as davem.
2085 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002086struct kmem_cache *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002087kmem_cache_create (const char *name, size_t size, size_t align,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002088 unsigned long flags, void (*ctor)(void *))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002089{
2090 size_t left_over, slab_size, ralign;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07002091 struct kmem_cache *cachep = NULL, *pc;
Pekka Enberg83b519e2009-06-10 19:40:04 +03002092 gfp_t gfp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002093
2094 /*
2095 * Sanity checks... these are all serious usage bugs.
2096 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002097 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
Paul Mundt20c2df82007-07-20 10:11:58 +09002098 size > KMALLOC_MAX_SIZE) {
Harvey Harrisond40cee22008-04-30 00:55:07 -07002099 printk(KERN_ERR "%s: Early error in slab %s\n", __func__,
Andrew Mortona737b3e2006-03-22 00:08:11 -08002100 name);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002101 BUG();
2102 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002103
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002104 /*
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002105 * We use cache_chain_mutex to ensure a consistent view of
Rusty Russell174596a2009-01-01 10:12:29 +10302106 * cpu_online_mask as well. Please see cpuup_callback
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002107 */
Pekka Enberg83b519e2009-06-10 19:40:04 +03002108 if (slab_is_available()) {
2109 get_online_cpus();
2110 mutex_lock(&cache_chain_mutex);
2111 }
Andrew Morton4f12bb42005-11-07 00:58:00 -08002112
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07002113 list_for_each_entry(pc, &cache_chain, next) {
Andrew Morton4f12bb42005-11-07 00:58:00 -08002114 char tmp;
2115 int res;
2116
2117 /*
2118 * This happens when the module gets unloaded and doesn't
2119 * destroy its slab cache and no-one else reuses the vmalloc
2120 * area of the module. Print a warning.
2121 */
Andrew Morton138ae662006-12-06 20:36:41 -08002122 res = probe_kernel_address(pc->name, tmp);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002123 if (res) {
matzeb4169522007-05-06 14:49:52 -07002124 printk(KERN_ERR
2125 "SLAB: cache with size %d has lost its name\n",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002126 pc->buffer_size);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002127 continue;
2128 }
2129
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002130 if (!strcmp(pc->name, name)) {
matzeb4169522007-05-06 14:49:52 -07002131 printk(KERN_ERR
2132 "kmem_cache_create: duplicate cache %s\n", name);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002133 dump_stack();
2134 goto oops;
2135 }
2136 }
2137
Linus Torvalds1da177e2005-04-16 15:20:36 -07002138#if DEBUG
2139 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002140#if FORCED_DEBUG
2141 /*
2142 * Enable redzoning and last user accounting, except for caches with
2143 * large objects, if the increased size would increase the object size
2144 * above the next power of two: caches with object sizes just above a
2145 * power of two have a significant amount of internal fragmentation.
2146 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002147 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2148 2 * sizeof(unsigned long long)))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002149 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002150 if (!(flags & SLAB_DESTROY_BY_RCU))
2151 flags |= SLAB_POISON;
2152#endif
2153 if (flags & SLAB_DESTROY_BY_RCU)
2154 BUG_ON(flags & SLAB_POISON);
2155#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002156 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002157 * Always checks flags, a caller might be expecting debug support which
2158 * isn't available.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002159 */
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002160 BUG_ON(flags & ~CREATE_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002161
Andrew Mortona737b3e2006-03-22 00:08:11 -08002162 /*
2163 * Check that size is in terms of words. This is needed to avoid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002164 * unaligned accesses for some archs when redzoning is used, and makes
2165 * sure any on-slab bufctl's are also correctly aligned.
2166 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002167 if (size & (BYTES_PER_WORD - 1)) {
2168 size += (BYTES_PER_WORD - 1);
2169 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002170 }
2171
Andrew Mortona737b3e2006-03-22 00:08:11 -08002172 /* calculate the final buffer alignment: */
2173
Linus Torvalds1da177e2005-04-16 15:20:36 -07002174 /* 1) arch recommendation: can be overridden for debug */
2175 if (flags & SLAB_HWCACHE_ALIGN) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002176 /*
2177 * Default alignment: as specified by the arch code. Except if
2178 * an object is really small, then squeeze multiple objects into
2179 * one cacheline.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002180 */
2181 ralign = cache_line_size();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002182 while (size <= ralign / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002183 ralign /= 2;
2184 } else {
2185 ralign = BYTES_PER_WORD;
2186 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002187
2188 /*
David Woodhouse87a927c2007-07-04 21:26:44 -04002189 * Redzoning and user store require word alignment or possibly larger.
2190 * Note this will be overridden by architecture or caller mandated
2191 * alignment if either is greater than BYTES_PER_WORD.
Pekka Enbergca5f9702006-09-25 23:31:25 -07002192 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002193 if (flags & SLAB_STORE_USER)
2194 ralign = BYTES_PER_WORD;
2195
2196 if (flags & SLAB_RED_ZONE) {
2197 ralign = REDZONE_ALIGN;
2198 /* If redzoning, ensure that the second redzone is suitably
2199 * aligned, by adjusting the object size accordingly. */
2200 size += REDZONE_ALIGN - 1;
2201 size &= ~(REDZONE_ALIGN - 1);
2202 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002203
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002204 /* 2) arch mandated alignment */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002205 if (ralign < ARCH_SLAB_MINALIGN) {
2206 ralign = ARCH_SLAB_MINALIGN;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002207 }
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002208 /* 3) caller mandated alignment */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002209 if (ralign < align) {
2210 ralign = align;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002211 }
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002212 /* disable debug if necessary */
David Woodhouseb46b8f12007-05-08 00:22:59 -07002213 if (ralign > __alignof__(unsigned long long))
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002214 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002215 /*
Pekka Enbergca5f9702006-09-25 23:31:25 -07002216 * 4) Store it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002217 */
2218 align = ralign;
2219
Pekka Enberg83b519e2009-06-10 19:40:04 +03002220 if (slab_is_available())
2221 gfp = GFP_KERNEL;
2222 else
2223 gfp = GFP_NOWAIT;
2224
Linus Torvalds1da177e2005-04-16 15:20:36 -07002225 /* Get cache's description obj. */
Pekka Enberg83b519e2009-06-10 19:40:04 +03002226 cachep = kmem_cache_zalloc(&cache_cache, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002227 if (!cachep)
Andrew Morton4f12bb42005-11-07 00:58:00 -08002228 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002229
2230#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002231 cachep->obj_size = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002232
Pekka Enbergca5f9702006-09-25 23:31:25 -07002233 /*
2234 * Both debugging options require word-alignment which is calculated
2235 * into align above.
2236 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002237 if (flags & SLAB_RED_ZONE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002238 /* add space for red zone words */
David Woodhouseb46b8f12007-05-08 00:22:59 -07002239 cachep->obj_offset += sizeof(unsigned long long);
2240 size += 2 * sizeof(unsigned long long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002241 }
2242 if (flags & SLAB_STORE_USER) {
Pekka Enbergca5f9702006-09-25 23:31:25 -07002243 /* user store requires one word storage behind the end of
David Woodhouse87a927c2007-07-04 21:26:44 -04002244 * the real object. But if the second red zone needs to be
2245 * aligned to 64 bits, we must allow that much space.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002246 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002247 if (flags & SLAB_RED_ZONE)
2248 size += REDZONE_ALIGN;
2249 else
2250 size += BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002251 }
2252#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002253 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002254 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2255 cachep->obj_offset += PAGE_SIZE - size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002256 size = PAGE_SIZE;
2257 }
2258#endif
2259#endif
2260
Ingo Molnare0a42722006-06-23 02:03:46 -07002261 /*
2262 * Determine if the slab management is 'on' or 'off' slab.
2263 * (bootstrapping cannot cope with offslab caches so don't do
2264 * it too early on.)
2265 */
2266 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002267 /*
2268 * Size is large, assume best to place the slab management obj
2269 * off-slab (should allow better packing of objs).
2270 */
2271 flags |= CFLGS_OFF_SLAB;
2272
2273 size = ALIGN(size, align);
2274
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002275 left_over = calculate_slab_order(cachep, size, align, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002276
2277 if (!cachep->num) {
matzeb4169522007-05-06 14:49:52 -07002278 printk(KERN_ERR
2279 "kmem_cache_create: couldn't create cache %s.\n", name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002280 kmem_cache_free(&cache_cache, cachep);
2281 cachep = NULL;
Andrew Morton4f12bb42005-11-07 00:58:00 -08002282 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002283 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002284 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2285 + sizeof(struct slab), align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002286
2287 /*
2288 * If the slab has been placed off-slab, and we have enough space then
2289 * move it on-slab. This is at the expense of any extra colouring.
2290 */
2291 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2292 flags &= ~CFLGS_OFF_SLAB;
2293 left_over -= slab_size;
2294 }
2295
2296 if (flags & CFLGS_OFF_SLAB) {
2297 /* really off slab. No need for manual alignment */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002298 slab_size =
2299 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
Ron Lee67461362009-05-22 04:58:22 +09302300
2301#ifdef CONFIG_PAGE_POISONING
2302 /* If we're going to use the generic kernel_map_pages()
2303 * poisoning, then it's going to smash the contents of
2304 * the redzone and userword anyhow, so switch them off.
2305 */
2306 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2307 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2308#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002309 }
2310
2311 cachep->colour_off = cache_line_size();
2312 /* Offset must be a multiple of the alignment. */
2313 if (cachep->colour_off < align)
2314 cachep->colour_off = align;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002315 cachep->colour = left_over / cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002316 cachep->slab_size = slab_size;
2317 cachep->flags = flags;
2318 cachep->gfpflags = 0;
Christoph Lameter4b51d662007-02-10 01:43:10 -08002319 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002320 cachep->gfpflags |= GFP_DMA;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002321 cachep->buffer_size = size;
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08002322 cachep->reciprocal_buffer_size = reciprocal_value(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002323
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002324 if (flags & CFLGS_OFF_SLAB) {
Victor Fuscob2d55072005-09-10 00:26:36 -07002325 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002326 /*
2327 * This is a possibility for one of the malloc_sizes caches.
2328 * But since we go off slab only for object size greater than
2329 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2330 * this should not happen at all.
2331 * But leave a BUG_ON for some lucky dude.
2332 */
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002333 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002334 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002335 cachep->ctor = ctor;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002336 cachep->name = name;
2337
Pekka Enberg83b519e2009-06-10 19:40:04 +03002338 if (setup_cpu_cache(cachep, gfp)) {
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002339 __kmem_cache_destroy(cachep);
2340 cachep = NULL;
2341 goto oops;
2342 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002343
Linus Torvalds1da177e2005-04-16 15:20:36 -07002344 /* cache setup completed, link it into the list */
2345 list_add(&cachep->next, &cache_chain);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002346oops:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002347 if (!cachep && (flags & SLAB_PANIC))
2348 panic("kmem_cache_create(): failed to create slab `%s'\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002349 name);
Pekka Enberg83b519e2009-06-10 19:40:04 +03002350 if (slab_is_available()) {
2351 mutex_unlock(&cache_chain_mutex);
2352 put_online_cpus();
2353 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002354 return cachep;
2355}
2356EXPORT_SYMBOL(kmem_cache_create);
2357
2358#if DEBUG
2359static void check_irq_off(void)
2360{
2361 BUG_ON(!irqs_disabled());
2362}
2363
2364static void check_irq_on(void)
2365{
2366 BUG_ON(irqs_disabled());
2367}
2368
Pekka Enberg343e0d72006-02-01 03:05:50 -08002369static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002370{
2371#ifdef CONFIG_SMP
2372 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002373 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002374#endif
2375}
Christoph Lametere498be72005-09-09 13:03:32 -07002376
Pekka Enberg343e0d72006-02-01 03:05:50 -08002377static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002378{
2379#ifdef CONFIG_SMP
2380 check_irq_off();
2381 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2382#endif
2383}
2384
Linus Torvalds1da177e2005-04-16 15:20:36 -07002385#else
2386#define check_irq_off() do { } while(0)
2387#define check_irq_on() do { } while(0)
2388#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002389#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002390#endif
2391
Christoph Lameteraab22072006-03-22 00:09:06 -08002392static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2393 struct array_cache *ac,
2394 int force, int node);
2395
Linus Torvalds1da177e2005-04-16 15:20:36 -07002396static void do_drain(void *arg)
2397{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002398 struct kmem_cache *cachep = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002399 struct array_cache *ac;
Christoph Lameterff694162005-09-22 21:44:02 -07002400 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002401
2402 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002403 ac = cpu_cache_get(cachep);
Christoph Lameterff694162005-09-22 21:44:02 -07002404 spin_lock(&cachep->nodelists[node]->list_lock);
2405 free_block(cachep, ac->entry, ac->avail, node);
2406 spin_unlock(&cachep->nodelists[node]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002407 ac->avail = 0;
2408}
2409
Pekka Enberg343e0d72006-02-01 03:05:50 -08002410static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002411{
Christoph Lametere498be72005-09-09 13:03:32 -07002412 struct kmem_list3 *l3;
2413 int node;
2414
Jens Axboe15c8b6c2008-05-09 09:39:44 +02002415 on_each_cpu(do_drain, cachep, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002416 check_irq_on();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002417 for_each_online_node(node) {
Christoph Lametere498be72005-09-09 13:03:32 -07002418 l3 = cachep->nodelists[node];
Roland Dreiera4523a82006-05-15 11:41:00 -07002419 if (l3 && l3->alien)
2420 drain_alien_cache(cachep, l3->alien);
2421 }
2422
2423 for_each_online_node(node) {
2424 l3 = cachep->nodelists[node];
2425 if (l3)
Christoph Lameteraab22072006-03-22 00:09:06 -08002426 drain_array(cachep, l3, l3->shared, 1, node);
Christoph Lametere498be72005-09-09 13:03:32 -07002427 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002428}
2429
Christoph Lametered11d9e2006-06-30 01:55:45 -07002430/*
2431 * Remove slabs from the list of free slabs.
2432 * Specify the number of slabs to drain in tofree.
2433 *
2434 * Returns the actual number of slabs released.
2435 */
2436static int drain_freelist(struct kmem_cache *cache,
2437 struct kmem_list3 *l3, int tofree)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002438{
Christoph Lametered11d9e2006-06-30 01:55:45 -07002439 struct list_head *p;
2440 int nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002441 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002442
Christoph Lametered11d9e2006-06-30 01:55:45 -07002443 nr_freed = 0;
2444 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002445
Christoph Lametered11d9e2006-06-30 01:55:45 -07002446 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002447 p = l3->slabs_free.prev;
Christoph Lametered11d9e2006-06-30 01:55:45 -07002448 if (p == &l3->slabs_free) {
2449 spin_unlock_irq(&l3->list_lock);
2450 goto out;
2451 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002452
Christoph Lametered11d9e2006-06-30 01:55:45 -07002453 slabp = list_entry(p, struct slab, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002454#if DEBUG
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002455 BUG_ON(slabp->inuse);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002456#endif
2457 list_del(&slabp->list);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002458 /*
2459 * Safe to drop the lock. The slab is no longer linked
2460 * to the cache.
2461 */
2462 l3->free_objects -= cache->num;
Christoph Lametere498be72005-09-09 13:03:32 -07002463 spin_unlock_irq(&l3->list_lock);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002464 slab_destroy(cache, slabp);
2465 nr_freed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002466 }
Christoph Lametered11d9e2006-06-30 01:55:45 -07002467out:
2468 return nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002469}
2470
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002471/* Called with cache_chain_mutex held to protect against cpu hotplug */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002472static int __cache_shrink(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07002473{
2474 int ret = 0, i = 0;
2475 struct kmem_list3 *l3;
2476
2477 drain_cpu_caches(cachep);
2478
2479 check_irq_on();
2480 for_each_online_node(i) {
2481 l3 = cachep->nodelists[i];
Christoph Lametered11d9e2006-06-30 01:55:45 -07002482 if (!l3)
2483 continue;
2484
2485 drain_freelist(cachep, l3, l3->free_objects);
2486
2487 ret += !list_empty(&l3->slabs_full) ||
2488 !list_empty(&l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07002489 }
2490 return (ret ? 1 : 0);
2491}
2492
Linus Torvalds1da177e2005-04-16 15:20:36 -07002493/**
2494 * kmem_cache_shrink - Shrink a cache.
2495 * @cachep: The cache to shrink.
2496 *
2497 * Releases as many slabs as possible for a cache.
2498 * To help debugging, a zero exit status indicates all slabs were released.
2499 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002500int kmem_cache_shrink(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002501{
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002502 int ret;
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002503 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002504
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002505 get_online_cpus();
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002506 mutex_lock(&cache_chain_mutex);
2507 ret = __cache_shrink(cachep);
2508 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002509 put_online_cpus();
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002510 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002511}
2512EXPORT_SYMBOL(kmem_cache_shrink);
2513
2514/**
2515 * kmem_cache_destroy - delete a cache
2516 * @cachep: the cache to destroy
2517 *
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08002518 * Remove a &struct kmem_cache object from the slab cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002519 *
2520 * It is expected this function will be called by a module when it is
2521 * unloaded. This will remove the cache completely, and avoid a duplicate
2522 * cache being allocated each time a module is loaded and unloaded, if the
2523 * module doesn't have persistent in-kernel storage across loads and unloads.
2524 *
2525 * The cache must be empty before calling this function.
2526 *
2527 * The caller must guarantee that noone will allocate memory from the cache
2528 * during the kmem_cache_destroy().
2529 */
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002530void kmem_cache_destroy(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002531{
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002532 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002533
Linus Torvalds1da177e2005-04-16 15:20:36 -07002534 /* Find the cache in the chain of caches. */
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002535 get_online_cpus();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002536 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002537 /*
2538 * the chain is never empty, cache_cache is never destroyed
2539 */
2540 list_del(&cachep->next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002541 if (__cache_shrink(cachep)) {
2542 slab_error(cachep, "Can't free all objects");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002543 list_add(&cachep->next, &cache_chain);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002544 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002545 put_online_cpus();
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002546 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002547 }
2548
2549 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
Paul E. McKenneyfbd568a3e2005-05-01 08:59:04 -07002550 synchronize_rcu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002551
Christoph Lameter117f6eb2006-09-25 23:31:37 -07002552 __kmem_cache_destroy(cachep);
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002553 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002554 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002555}
2556EXPORT_SYMBOL(kmem_cache_destroy);
2557
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002558/*
2559 * Get the memory for a slab management obj.
2560 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2561 * always come from malloc_sizes caches. The slab descriptor cannot
2562 * come from the same cache which is getting created because,
2563 * when we are searching for an appropriate cache for these
2564 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2565 * If we are creating a malloc_sizes cache here it would not be visible to
2566 * kmem_find_general_cachep till the initialization is complete.
2567 * Hence we cannot have slabp_cache same as the original cache.
2568 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002569static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002570 int colour_off, gfp_t local_flags,
2571 int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002572{
2573 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002574
Linus Torvalds1da177e2005-04-16 15:20:36 -07002575 if (OFF_SLAB(cachep)) {
2576 /* Slab management obj is off-slab. */
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002577 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
Pekka Enberg8759ec52008-11-26 10:01:31 +02002578 local_flags, nodeid);
Catalin Marinasd5cff632009-06-11 13:22:40 +01002579 /*
2580 * If the first object in the slab is leaked (it's allocated
2581 * but no one has a reference to it), we want to make sure
2582 * kmemleak does not treat the ->s_mem pointer as a reference
2583 * to the object. Otherwise we will not report the leak.
2584 */
2585 kmemleak_scan_area(slabp, offsetof(struct slab, list),
2586 sizeof(struct list_head), local_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002587 if (!slabp)
2588 return NULL;
2589 } else {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002590 slabp = objp + colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002591 colour_off += cachep->slab_size;
2592 }
2593 slabp->inuse = 0;
2594 slabp->colouroff = colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002595 slabp->s_mem = objp + colour_off;
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002596 slabp->nodeid = nodeid;
Marcin Slusarze51bfd02008-02-10 11:21:54 +01002597 slabp->free = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002598 return slabp;
2599}
2600
2601static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2602{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002603 return (kmem_bufctl_t *) (slabp + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002604}
2605
Pekka Enberg343e0d72006-02-01 03:05:50 -08002606static void cache_init_objs(struct kmem_cache *cachep,
Christoph Lametera35afb82007-05-16 22:10:57 -07002607 struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002608{
2609 int i;
2610
2611 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002612 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002613#if DEBUG
2614 /* need to poison the objs? */
2615 if (cachep->flags & SLAB_POISON)
2616 poison_obj(cachep, objp, POISON_FREE);
2617 if (cachep->flags & SLAB_STORE_USER)
2618 *dbg_userword(cachep, objp) = NULL;
2619
2620 if (cachep->flags & SLAB_RED_ZONE) {
2621 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2622 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2623 }
2624 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002625 * Constructors are not allowed to allocate memory from the same
2626 * cache which they are a constructor for. Otherwise, deadlock.
2627 * They must also be threaded.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002628 */
2629 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002630 cachep->ctor(objp + obj_offset(cachep));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002631
2632 if (cachep->flags & SLAB_RED_ZONE) {
2633 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2634 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002635 " end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002636 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2637 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002638 " start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002639 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08002640 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2641 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002642 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002643 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002644#else
2645 if (cachep->ctor)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002646 cachep->ctor(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002647#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002648 slab_bufctl(slabp)[i] = i + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002649 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002650 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002651}
2652
Pekka Enberg343e0d72006-02-01 03:05:50 -08002653static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002654{
Christoph Lameter4b51d662007-02-10 01:43:10 -08002655 if (CONFIG_ZONE_DMA_FLAG) {
2656 if (flags & GFP_DMA)
2657 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2658 else
2659 BUG_ON(cachep->gfpflags & GFP_DMA);
2660 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002661}
2662
Andrew Mortona737b3e2006-03-22 00:08:11 -08002663static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2664 int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002665{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002666 void *objp = index_to_obj(cachep, slabp, slabp->free);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002667 kmem_bufctl_t next;
2668
2669 slabp->inuse++;
2670 next = slab_bufctl(slabp)[slabp->free];
2671#if DEBUG
2672 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2673 WARN_ON(slabp->nodeid != nodeid);
2674#endif
2675 slabp->free = next;
2676
2677 return objp;
2678}
2679
Andrew Mortona737b3e2006-03-22 00:08:11 -08002680static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2681 void *objp, int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002682{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002683 unsigned int objnr = obj_to_index(cachep, slabp, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002684
2685#if DEBUG
2686 /* Verify that the slab belongs to the intended node */
2687 WARN_ON(slabp->nodeid != nodeid);
2688
Al Viro871751e2006-03-25 03:06:39 -08002689 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
Matthew Dobson78d382d2006-02-01 03:05:47 -08002690 printk(KERN_ERR "slab: double free detected in cache "
Andrew Mortona737b3e2006-03-22 00:08:11 -08002691 "'%s', objp %p\n", cachep->name, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002692 BUG();
2693 }
2694#endif
2695 slab_bufctl(slabp)[objnr] = slabp->free;
2696 slabp->free = objnr;
2697 slabp->inuse--;
2698}
2699
Pekka Enberg47768742006-06-23 02:03:07 -07002700/*
2701 * Map pages beginning at addr to the given cache and slab. This is required
2702 * for the slab allocator to be able to lookup the cache and slab of a
2703 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2704 */
2705static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2706 void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002707{
Pekka Enberg47768742006-06-23 02:03:07 -07002708 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002709 struct page *page;
2710
Pekka Enberg47768742006-06-23 02:03:07 -07002711 page = virt_to_page(addr);
Nick Piggin84097512006-03-22 00:08:34 -08002712
Pekka Enberg47768742006-06-23 02:03:07 -07002713 nr_pages = 1;
Nick Piggin84097512006-03-22 00:08:34 -08002714 if (likely(!PageCompound(page)))
Pekka Enberg47768742006-06-23 02:03:07 -07002715 nr_pages <<= cache->gfporder;
2716
Linus Torvalds1da177e2005-04-16 15:20:36 -07002717 do {
Pekka Enberg47768742006-06-23 02:03:07 -07002718 page_set_cache(page, cache);
2719 page_set_slab(page, slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002720 page++;
Pekka Enberg47768742006-06-23 02:03:07 -07002721 } while (--nr_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002722}
2723
2724/*
2725 * Grow (by 1) the number of slabs within a cache. This is called by
2726 * kmem_cache_alloc() when there are no active objs left in a cache.
2727 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002728static int cache_grow(struct kmem_cache *cachep,
2729 gfp_t flags, int nodeid, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002730{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002731 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002732 size_t offset;
2733 gfp_t local_flags;
Christoph Lametere498be72005-09-09 13:03:32 -07002734 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002735
Andrew Mortona737b3e2006-03-22 00:08:11 -08002736 /*
2737 * Be lazy and only check for valid flags here, keeping it out of the
2738 * critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002739 */
Christoph Lameter6cb06222007-10-16 01:25:41 -07002740 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2741 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002742
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002743 /* Take the l3 list lock to change the colour_next on this node */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002744 check_irq_off();
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002745 l3 = cachep->nodelists[nodeid];
2746 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002747
2748 /* Get colour for the slab, and cal the next value. */
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002749 offset = l3->colour_next;
2750 l3->colour_next++;
2751 if (l3->colour_next >= cachep->colour)
2752 l3->colour_next = 0;
2753 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002754
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002755 offset *= cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002756
2757 if (local_flags & __GFP_WAIT)
2758 local_irq_enable();
2759
2760 /*
2761 * The test for missing atomic flag is performed here, rather than
2762 * the more obvious place, simply to reduce the critical path length
2763 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2764 * will eventually be caught here (where it matters).
2765 */
2766 kmem_flagcheck(cachep, flags);
2767
Andrew Mortona737b3e2006-03-22 00:08:11 -08002768 /*
2769 * Get mem for the objs. Attempt to allocate a physical page from
2770 * 'nodeid'.
Christoph Lametere498be72005-09-09 13:03:32 -07002771 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002772 if (!objp)
Andrew Mortonb8c1c5d2007-07-24 12:02:40 -07002773 objp = kmem_getpages(cachep, local_flags, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002774 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002775 goto failed;
2776
2777 /* Get slab management. */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002778 slabp = alloc_slabmgmt(cachep, objp, offset,
Christoph Lameter6cb06222007-10-16 01:25:41 -07002779 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002780 if (!slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002781 goto opps1;
2782
Pekka Enberg47768742006-06-23 02:03:07 -07002783 slab_map_pages(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002784
Christoph Lametera35afb82007-05-16 22:10:57 -07002785 cache_init_objs(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002786
2787 if (local_flags & __GFP_WAIT)
2788 local_irq_disable();
2789 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002790 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002791
2792 /* Make slab active. */
Christoph Lametere498be72005-09-09 13:03:32 -07002793 list_add_tail(&slabp->list, &(l3->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002794 STATS_INC_GROWN(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002795 l3->free_objects += cachep->num;
2796 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002797 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002798opps1:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002799 kmem_freepages(cachep, objp);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002800failed:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002801 if (local_flags & __GFP_WAIT)
2802 local_irq_disable();
2803 return 0;
2804}
2805
2806#if DEBUG
2807
2808/*
2809 * Perform extra freeing checks:
2810 * - detect bad pointers.
2811 * - POISON/RED_ZONE checking
Linus Torvalds1da177e2005-04-16 15:20:36 -07002812 */
2813static void kfree_debugcheck(const void *objp)
2814{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002815 if (!virt_addr_valid(objp)) {
2816 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002817 (unsigned long)objp);
2818 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002819 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002820}
2821
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002822static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2823{
David Woodhouseb46b8f12007-05-08 00:22:59 -07002824 unsigned long long redzone1, redzone2;
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002825
2826 redzone1 = *dbg_redzone1(cache, obj);
2827 redzone2 = *dbg_redzone2(cache, obj);
2828
2829 /*
2830 * Redzone is ok.
2831 */
2832 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2833 return;
2834
2835 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2836 slab_error(cache, "double free detected");
2837 else
2838 slab_error(cache, "memory outside object was overwritten");
2839
David Woodhouseb46b8f12007-05-08 00:22:59 -07002840 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002841 obj, redzone1, redzone2);
2842}
2843
Pekka Enberg343e0d72006-02-01 03:05:50 -08002844static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002845 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002846{
2847 struct page *page;
2848 unsigned int objnr;
2849 struct slab *slabp;
2850
Matthew Wilcox80cbd912007-11-29 12:05:13 -07002851 BUG_ON(virt_to_cache(objp) != cachep);
2852
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002853 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002854 kfree_debugcheck(objp);
Christoph Lameterb49af682007-05-06 14:49:41 -07002855 page = virt_to_head_page(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002856
Pekka Enberg065d41c2005-11-13 16:06:46 -08002857 slabp = page_get_slab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002858
2859 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002860 verify_redzone_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002861 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2862 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2863 }
2864 if (cachep->flags & SLAB_STORE_USER)
2865 *dbg_userword(cachep, objp) = caller;
2866
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002867 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002868
2869 BUG_ON(objnr >= cachep->num);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002870 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002871
Al Viro871751e2006-03-25 03:06:39 -08002872#ifdef CONFIG_DEBUG_SLAB_LEAK
2873 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2874#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002875 if (cachep->flags & SLAB_POISON) {
2876#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08002877 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002878 store_stackinfo(cachep, objp, (unsigned long)caller);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002879 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002880 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002881 } else {
2882 poison_obj(cachep, objp, POISON_FREE);
2883 }
2884#else
2885 poison_obj(cachep, objp, POISON_FREE);
2886#endif
2887 }
2888 return objp;
2889}
2890
Pekka Enberg343e0d72006-02-01 03:05:50 -08002891static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002892{
2893 kmem_bufctl_t i;
2894 int entries = 0;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002895
Linus Torvalds1da177e2005-04-16 15:20:36 -07002896 /* Check slab's freelist to see if this obj is there. */
2897 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2898 entries++;
2899 if (entries > cachep->num || i >= cachep->num)
2900 goto bad;
2901 }
2902 if (entries != cachep->num - slabp->inuse) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002903bad:
2904 printk(KERN_ERR "slab: Internal list corruption detected in "
2905 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2906 cachep->name, cachep->num, slabp, slabp->inuse);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002907 for (i = 0;
Linus Torvalds264132b2006-03-06 12:10:07 -08002908 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002909 i++) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002910 if (i % 16 == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002911 printk("\n%03x:", i);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002912 printk(" %02x", ((unsigned char *)slabp)[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002913 }
2914 printk("\n");
2915 BUG();
2916 }
2917}
2918#else
2919#define kfree_debugcheck(x) do { } while(0)
2920#define cache_free_debugcheck(x,objp,z) (objp)
2921#define check_slabp(x,y) do { } while(0)
2922#endif
2923
Pekka Enberg343e0d72006-02-01 03:05:50 -08002924static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002925{
2926 int batchcount;
2927 struct kmem_list3 *l3;
2928 struct array_cache *ac;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002929 int node;
2930
Andrew Mortona737b3e2006-03-22 00:08:11 -08002931retry:
Joe Korty6d2144d2008-03-05 15:04:59 -08002932 check_irq_off();
2933 node = numa_node_id();
2934 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002935 batchcount = ac->batchcount;
2936 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002937 /*
2938 * If there was little recent activity on this cache, then
2939 * perform only a partial refill. Otherwise we could generate
2940 * refill bouncing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002941 */
2942 batchcount = BATCHREFILL_LIMIT;
2943 }
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002944 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002945
Christoph Lametere498be72005-09-09 13:03:32 -07002946 BUG_ON(ac->avail > 0 || !l3);
2947 spin_lock(&l3->list_lock);
2948
Christoph Lameter3ded1752006-03-25 03:06:44 -08002949 /* See if we can refill from the shared array */
2950 if (l3->shared && transfer_objects(ac, l3->shared, batchcount))
2951 goto alloc_done;
2952
Linus Torvalds1da177e2005-04-16 15:20:36 -07002953 while (batchcount > 0) {
2954 struct list_head *entry;
2955 struct slab *slabp;
2956 /* Get slab alloc is to come from. */
2957 entry = l3->slabs_partial.next;
2958 if (entry == &l3->slabs_partial) {
2959 l3->free_touched = 1;
2960 entry = l3->slabs_free.next;
2961 if (entry == &l3->slabs_free)
2962 goto must_grow;
2963 }
2964
2965 slabp = list_entry(entry, struct slab, list);
2966 check_slabp(cachep, slabp);
2967 check_spinlock_acquired(cachep);
Pekka Enberg714b81712007-05-06 14:49:03 -07002968
2969 /*
2970 * The slab was either on partial or free list so
2971 * there must be at least one object available for
2972 * allocation.
2973 */
roel kluin249b9f32008-10-29 17:18:07 -04002974 BUG_ON(slabp->inuse >= cachep->num);
Pekka Enberg714b81712007-05-06 14:49:03 -07002975
Linus Torvalds1da177e2005-04-16 15:20:36 -07002976 while (slabp->inuse < cachep->num && batchcount--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002977 STATS_INC_ALLOCED(cachep);
2978 STATS_INC_ACTIVE(cachep);
2979 STATS_SET_HIGH(cachep);
2980
Matthew Dobson78d382d2006-02-01 03:05:47 -08002981 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002982 node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002983 }
2984 check_slabp(cachep, slabp);
2985
2986 /* move slabp to correct slabp list: */
2987 list_del(&slabp->list);
2988 if (slabp->free == BUFCTL_END)
2989 list_add(&slabp->list, &l3->slabs_full);
2990 else
2991 list_add(&slabp->list, &l3->slabs_partial);
2992 }
2993
Andrew Mortona737b3e2006-03-22 00:08:11 -08002994must_grow:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002995 l3->free_objects -= ac->avail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002996alloc_done:
Christoph Lametere498be72005-09-09 13:03:32 -07002997 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002998
2999 if (unlikely(!ac->avail)) {
3000 int x;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003001 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07003002
Andrew Mortona737b3e2006-03-22 00:08:11 -08003003 /* cache_grow can reenable interrupts, then ac could change. */
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003004 ac = cpu_cache_get(cachep);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003005 if (!x && ac->avail == 0) /* no objects in sight? abort */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003006 return NULL;
3007
Andrew Mortona737b3e2006-03-22 00:08:11 -08003008 if (!ac->avail) /* objects refilled by interrupt? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003009 goto retry;
3010 }
3011 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07003012 return ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003013}
3014
Andrew Mortona737b3e2006-03-22 00:08:11 -08003015static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3016 gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003017{
3018 might_sleep_if(flags & __GFP_WAIT);
3019#if DEBUG
3020 kmem_flagcheck(cachep, flags);
3021#endif
3022}
3023
3024#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003025static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3026 gfp_t flags, void *objp, void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003027{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003028 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003029 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003030 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003031#ifdef CONFIG_DEBUG_PAGEALLOC
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003032 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003033 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003034 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003035 else
3036 check_poison_obj(cachep, objp);
3037#else
3038 check_poison_obj(cachep, objp);
3039#endif
3040 poison_obj(cachep, objp, POISON_INUSE);
3041 }
3042 if (cachep->flags & SLAB_STORE_USER)
3043 *dbg_userword(cachep, objp) = caller;
3044
3045 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003046 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3047 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3048 slab_error(cachep, "double free, or memory outside"
3049 " object was overwritten");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003050 printk(KERN_ERR
David Woodhouseb46b8f12007-05-08 00:22:59 -07003051 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08003052 objp, *dbg_redzone1(cachep, objp),
3053 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003054 }
3055 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3056 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3057 }
Al Viro871751e2006-03-25 03:06:39 -08003058#ifdef CONFIG_DEBUG_SLAB_LEAK
3059 {
3060 struct slab *slabp;
3061 unsigned objnr;
3062
Christoph Lameterb49af682007-05-06 14:49:41 -07003063 slabp = page_get_slab(virt_to_head_page(objp));
Al Viro871751e2006-03-25 03:06:39 -08003064 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3065 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3066 }
3067#endif
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003068 objp += obj_offset(cachep);
Christoph Lameter4f104932007-05-06 14:50:17 -07003069 if (cachep->ctor && cachep->flags & SLAB_POISON)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07003070 cachep->ctor(objp);
Kevin Hilmana44b56d2006-12-06 20:32:11 -08003071#if ARCH_SLAB_MINALIGN
3072 if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) {
3073 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3074 objp, ARCH_SLAB_MINALIGN);
3075 }
3076#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003077 return objp;
3078}
3079#else
3080#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3081#endif
3082
Akinobu Mita773ff602008-12-23 19:37:01 +09003083static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003084{
3085 if (cachep == &cache_cache)
Akinobu Mita773ff602008-12-23 19:37:01 +09003086 return false;
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003087
Akinobu Mita773ff602008-12-23 19:37:01 +09003088 return should_failslab(obj_size(cachep), flags);
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003089}
3090
Pekka Enberg343e0d72006-02-01 03:05:50 -08003091static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003092{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003093 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003094 struct array_cache *ac;
3095
Alok N Kataria5c382302005-09-27 21:45:46 -07003096 check_irq_off();
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003097
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003098 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003099 if (likely(ac->avail)) {
3100 STATS_INC_ALLOCHIT(cachep);
3101 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07003102 objp = ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003103 } else {
3104 STATS_INC_ALLOCMISS(cachep);
3105 objp = cache_alloc_refill(cachep, flags);
3106 }
Catalin Marinasd5cff632009-06-11 13:22:40 +01003107 /*
3108 * To avoid a false negative, if an object that is in one of the
3109 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3110 * treat the array pointers as a reference to the object.
3111 */
3112 kmemleak_erase(&ac->entry[ac->avail]);
Alok N Kataria5c382302005-09-27 21:45:46 -07003113 return objp;
3114}
3115
Christoph Lametere498be72005-09-09 13:03:32 -07003116#ifdef CONFIG_NUMA
3117/*
Paul Jacksonb2455392006-03-24 03:16:12 -08003118 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
Paul Jacksonc61afb12006-03-24 03:16:08 -08003119 *
3120 * If we are in_interrupt, then process context, including cpusets and
3121 * mempolicy, may not apply and should not be used for allocation policy.
3122 */
3123static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3124{
3125 int nid_alloc, nid_here;
3126
Christoph Lameter765c4502006-09-27 01:50:08 -07003127 if (in_interrupt() || (flags & __GFP_THISNODE))
Paul Jacksonc61afb12006-03-24 03:16:08 -08003128 return NULL;
3129 nid_alloc = nid_here = numa_node_id();
3130 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3131 nid_alloc = cpuset_mem_spread_node();
3132 else if (current->mempolicy)
3133 nid_alloc = slab_node(current->mempolicy);
3134 if (nid_alloc != nid_here)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003135 return ____cache_alloc_node(cachep, flags, nid_alloc);
Paul Jacksonc61afb12006-03-24 03:16:08 -08003136 return NULL;
3137}
3138
3139/*
Christoph Lameter765c4502006-09-27 01:50:08 -07003140 * Fallback function if there was no memory available and no objects on a
Christoph Lameter3c517a62006-12-06 20:33:29 -08003141 * certain node and fall back is permitted. First we scan all the
3142 * available nodelists for available objects. If that fails then we
3143 * perform an allocation without specifying a node. This allows the page
3144 * allocator to do its reclaim / fallback magic. We then insert the
3145 * slab into the proper nodelist and then allocate from it.
Christoph Lameter765c4502006-09-27 01:50:08 -07003146 */
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003147static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
Christoph Lameter765c4502006-09-27 01:50:08 -07003148{
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003149 struct zonelist *zonelist;
3150 gfp_t local_flags;
Mel Gormandd1a2392008-04-28 02:12:17 -07003151 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07003152 struct zone *zone;
3153 enum zone_type high_zoneidx = gfp_zone(flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003154 void *obj = NULL;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003155 int nid;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003156
3157 if (flags & __GFP_THISNODE)
3158 return NULL;
3159
Mel Gorman0e884602008-04-28 02:12:14 -07003160 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
Christoph Lameter6cb06222007-10-16 01:25:41 -07003161 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Christoph Lameter765c4502006-09-27 01:50:08 -07003162
Christoph Lameter3c517a62006-12-06 20:33:29 -08003163retry:
3164 /*
3165 * Look through allowed nodes for objects available
3166 * from existing per node queues.
3167 */
Mel Gorman54a6eb52008-04-28 02:12:16 -07003168 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3169 nid = zone_to_nid(zone);
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003170
Mel Gorman54a6eb52008-04-28 02:12:16 -07003171 if (cpuset_zone_allowed_hardwall(zone, flags) &&
Christoph Lameter3c517a62006-12-06 20:33:29 -08003172 cache->nodelists[nid] &&
Christoph Lameter481c5342008-06-21 16:46:35 -07003173 cache->nodelists[nid]->free_objects) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003174 obj = ____cache_alloc_node(cache,
3175 flags | GFP_THISNODE, nid);
Christoph Lameter481c5342008-06-21 16:46:35 -07003176 if (obj)
3177 break;
3178 }
Christoph Lameter3c517a62006-12-06 20:33:29 -08003179 }
3180
Christoph Lametercfce6602007-05-06 14:50:17 -07003181 if (!obj) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003182 /*
3183 * This allocation will be performed within the constraints
3184 * of the current cpuset / memory policy requirements.
3185 * We may trigger various forms of reclaim on the allowed
3186 * set and go into memory reserves if necessary.
3187 */
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003188 if (local_flags & __GFP_WAIT)
3189 local_irq_enable();
3190 kmem_flagcheck(cache, flags);
Mel Gorman6484eb32009-06-16 15:31:54 -07003191 obj = kmem_getpages(cache, local_flags, numa_node_id());
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003192 if (local_flags & __GFP_WAIT)
3193 local_irq_disable();
Christoph Lameter3c517a62006-12-06 20:33:29 -08003194 if (obj) {
3195 /*
3196 * Insert into the appropriate per node queues
3197 */
3198 nid = page_to_nid(virt_to_page(obj));
3199 if (cache_grow(cache, flags, nid, obj)) {
3200 obj = ____cache_alloc_node(cache,
3201 flags | GFP_THISNODE, nid);
3202 if (!obj)
3203 /*
3204 * Another processor may allocate the
3205 * objects in the slab since we are
3206 * not holding any locks.
3207 */
3208 goto retry;
3209 } else {
Hugh Dickinsb6a60452007-01-05 16:36:36 -08003210 /* cache_grow already freed obj */
Christoph Lameter3c517a62006-12-06 20:33:29 -08003211 obj = NULL;
3212 }
3213 }
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003214 }
Christoph Lameter765c4502006-09-27 01:50:08 -07003215 return obj;
3216}
3217
3218/*
Christoph Lametere498be72005-09-09 13:03:32 -07003219 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07003220 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003221static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003222 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07003223{
3224 struct list_head *entry;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003225 struct slab *slabp;
3226 struct kmem_list3 *l3;
3227 void *obj;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003228 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003229
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003230 l3 = cachep->nodelists[nodeid];
3231 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07003232
Andrew Mortona737b3e2006-03-22 00:08:11 -08003233retry:
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003234 check_irq_off();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003235 spin_lock(&l3->list_lock);
3236 entry = l3->slabs_partial.next;
3237 if (entry == &l3->slabs_partial) {
3238 l3->free_touched = 1;
3239 entry = l3->slabs_free.next;
3240 if (entry == &l3->slabs_free)
3241 goto must_grow;
3242 }
Christoph Lametere498be72005-09-09 13:03:32 -07003243
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003244 slabp = list_entry(entry, struct slab, list);
3245 check_spinlock_acquired_node(cachep, nodeid);
3246 check_slabp(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07003247
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003248 STATS_INC_NODEALLOCS(cachep);
3249 STATS_INC_ACTIVE(cachep);
3250 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003251
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003252 BUG_ON(slabp->inuse == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07003253
Matthew Dobson78d382d2006-02-01 03:05:47 -08003254 obj = slab_get_obj(cachep, slabp, nodeid);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003255 check_slabp(cachep, slabp);
3256 l3->free_objects--;
3257 /* move slabp to correct slabp list: */
3258 list_del(&slabp->list);
Christoph Lametere498be72005-09-09 13:03:32 -07003259
Andrew Mortona737b3e2006-03-22 00:08:11 -08003260 if (slabp->free == BUFCTL_END)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003261 list_add(&slabp->list, &l3->slabs_full);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003262 else
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003263 list_add(&slabp->list, &l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07003264
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003265 spin_unlock(&l3->list_lock);
3266 goto done;
Christoph Lametere498be72005-09-09 13:03:32 -07003267
Andrew Mortona737b3e2006-03-22 00:08:11 -08003268must_grow:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003269 spin_unlock(&l3->list_lock);
Christoph Lameter3c517a62006-12-06 20:33:29 -08003270 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
Christoph Lameter765c4502006-09-27 01:50:08 -07003271 if (x)
3272 goto retry;
Christoph Lametere498be72005-09-09 13:03:32 -07003273
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003274 return fallback_alloc(cachep, flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003275
Andrew Mortona737b3e2006-03-22 00:08:11 -08003276done:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003277 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07003278}
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003279
3280/**
3281 * kmem_cache_alloc_node - Allocate an object on the specified node
3282 * @cachep: The cache to allocate from.
3283 * @flags: See kmalloc().
3284 * @nodeid: node number of the target node.
3285 * @caller: return address of caller, used for debug information
3286 *
3287 * Identical to kmem_cache_alloc but it will allocate memory on the given
3288 * node, which can improve the performance for cpu bound structures.
3289 *
3290 * Fallback to other node is possible if __GFP_THISNODE is not set.
3291 */
3292static __always_inline void *
3293__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3294 void *caller)
3295{
3296 unsigned long save_flags;
3297 void *ptr;
3298
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003299 flags &= gfp_allowed_mask;
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003300
Nick Piggincf40bd12009-01-21 08:12:39 +01003301 lockdep_trace_alloc(flags);
3302
Akinobu Mita773ff602008-12-23 19:37:01 +09003303 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003304 return NULL;
3305
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003306 cache_alloc_debugcheck_before(cachep, flags);
3307 local_irq_save(save_flags);
3308
3309 if (unlikely(nodeid == -1))
3310 nodeid = numa_node_id();
3311
3312 if (unlikely(!cachep->nodelists[nodeid])) {
3313 /* Node not bootstrapped yet */
3314 ptr = fallback_alloc(cachep, flags);
3315 goto out;
3316 }
3317
3318 if (nodeid == numa_node_id()) {
3319 /*
3320 * Use the locally cached objects if possible.
3321 * However ____cache_alloc does not allow fallback
3322 * to other nodes. It may fail while we still have
3323 * objects on other nodes available.
3324 */
3325 ptr = ____cache_alloc(cachep, flags);
3326 if (ptr)
3327 goto out;
3328 }
3329 /* ___cache_alloc_node can fall back to other nodes */
3330 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3331 out:
3332 local_irq_restore(save_flags);
3333 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
Catalin Marinasd5cff632009-06-11 13:22:40 +01003334 kmemleak_alloc_recursive(ptr, obj_size(cachep), 1, cachep->flags,
3335 flags);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003336
Pekka Enbergc175eea2008-05-09 20:35:53 +02003337 if (likely(ptr))
3338 kmemcheck_slab_alloc(cachep, flags, ptr, obj_size(cachep));
3339
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003340 if (unlikely((flags & __GFP_ZERO) && ptr))
3341 memset(ptr, 0, obj_size(cachep));
3342
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003343 return ptr;
3344}
3345
3346static __always_inline void *
3347__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3348{
3349 void *objp;
3350
3351 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3352 objp = alternate_node_alloc(cache, flags);
3353 if (objp)
3354 goto out;
3355 }
3356 objp = ____cache_alloc(cache, flags);
3357
3358 /*
3359 * We may just have run out of memory on the local node.
3360 * ____cache_alloc_node() knows how to locate memory on other nodes
3361 */
3362 if (!objp)
3363 objp = ____cache_alloc_node(cache, flags, numa_node_id());
3364
3365 out:
3366 return objp;
3367}
3368#else
3369
3370static __always_inline void *
3371__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3372{
3373 return ____cache_alloc(cachep, flags);
3374}
3375
3376#endif /* CONFIG_NUMA */
3377
3378static __always_inline void *
3379__cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3380{
3381 unsigned long save_flags;
3382 void *objp;
3383
Benjamin Herrenschmidtdcce2842009-06-18 13:24:12 +10003384 flags &= gfp_allowed_mask;
Pekka Enberg7e85ee02009-06-12 14:03:06 +03003385
Nick Piggincf40bd12009-01-21 08:12:39 +01003386 lockdep_trace_alloc(flags);
3387
Akinobu Mita773ff602008-12-23 19:37:01 +09003388 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003389 return NULL;
3390
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003391 cache_alloc_debugcheck_before(cachep, flags);
3392 local_irq_save(save_flags);
3393 objp = __do_cache_alloc(cachep, flags);
3394 local_irq_restore(save_flags);
3395 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
Catalin Marinasd5cff632009-06-11 13:22:40 +01003396 kmemleak_alloc_recursive(objp, obj_size(cachep), 1, cachep->flags,
3397 flags);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003398 prefetchw(objp);
3399
Pekka Enbergc175eea2008-05-09 20:35:53 +02003400 if (likely(objp))
3401 kmemcheck_slab_alloc(cachep, flags, objp, obj_size(cachep));
3402
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003403 if (unlikely((flags & __GFP_ZERO) && objp))
3404 memset(objp, 0, obj_size(cachep));
3405
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003406 return objp;
3407}
Christoph Lametere498be72005-09-09 13:03:32 -07003408
3409/*
3410 * Caller needs to acquire correct kmem_list's list_lock
3411 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003412static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003413 int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003414{
3415 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07003416 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003417
3418 for (i = 0; i < nr_objects; i++) {
3419 void *objp = objpp[i];
3420 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003421
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003422 slabp = virt_to_slab(objp);
Christoph Lameterff694162005-09-22 21:44:02 -07003423 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003424 list_del(&slabp->list);
Christoph Lameterff694162005-09-22 21:44:02 -07003425 check_spinlock_acquired_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003426 check_slabp(cachep, slabp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08003427 slab_put_obj(cachep, slabp, objp, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003428 STATS_DEC_ACTIVE(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003429 l3->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003430 check_slabp(cachep, slabp);
3431
3432 /* fixup slab chains */
3433 if (slabp->inuse == 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07003434 if (l3->free_objects > l3->free_limit) {
3435 l3->free_objects -= cachep->num;
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07003436 /* No need to drop any previously held
3437 * lock here, even if we have a off-slab slab
3438 * descriptor it is guaranteed to come from
3439 * a different cache, refer to comments before
3440 * alloc_slabmgmt.
3441 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003442 slab_destroy(cachep, slabp);
3443 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07003444 list_add(&slabp->list, &l3->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003445 }
3446 } else {
3447 /* Unconditionally move a slab to the end of the
3448 * partial list on free - maximum time for the
3449 * other objects to be freed, too.
3450 */
Christoph Lametere498be72005-09-09 13:03:32 -07003451 list_add_tail(&slabp->list, &l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003452 }
3453 }
3454}
3455
Pekka Enberg343e0d72006-02-01 03:05:50 -08003456static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003457{
3458 int batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07003459 struct kmem_list3 *l3;
Christoph Lameterff694162005-09-22 21:44:02 -07003460 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003461
3462 batchcount = ac->batchcount;
3463#if DEBUG
3464 BUG_ON(!batchcount || batchcount > ac->avail);
3465#endif
3466 check_irq_off();
Christoph Lameterff694162005-09-22 21:44:02 -07003467 l3 = cachep->nodelists[node];
Ingo Molnar873623d2006-07-13 14:44:38 +02003468 spin_lock(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003469 if (l3->shared) {
3470 struct array_cache *shared_array = l3->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003471 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003472 if (max) {
3473 if (batchcount > max)
3474 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07003475 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003476 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003477 shared_array->avail += batchcount;
3478 goto free_done;
3479 }
3480 }
3481
Christoph Lameterff694162005-09-22 21:44:02 -07003482 free_block(cachep, ac->entry, batchcount, node);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003483free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003484#if STATS
3485 {
3486 int i = 0;
3487 struct list_head *p;
3488
Christoph Lametere498be72005-09-09 13:03:32 -07003489 p = l3->slabs_free.next;
3490 while (p != &(l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003491 struct slab *slabp;
3492
3493 slabp = list_entry(p, struct slab, list);
3494 BUG_ON(slabp->inuse);
3495
3496 i++;
3497 p = p->next;
3498 }
3499 STATS_SET_FREEABLE(cachep, i);
3500 }
3501#endif
Christoph Lametere498be72005-09-09 13:03:32 -07003502 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003503 ac->avail -= batchcount;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003504 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003505}
3506
3507/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08003508 * Release an obj back to its cache. If the obj has a constructed state, it must
3509 * be in this state _before_ it is released. Called with disabled ints.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003510 */
Ingo Molnar873623d2006-07-13 14:44:38 +02003511static inline void __cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003512{
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003513 struct array_cache *ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003514
3515 check_irq_off();
Catalin Marinasd5cff632009-06-11 13:22:40 +01003516 kmemleak_free_recursive(objp, cachep->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003517 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3518
Pekka Enbergc175eea2008-05-09 20:35:53 +02003519 kmemcheck_slab_free(cachep, objp, obj_size(cachep));
3520
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -07003521 /*
3522 * Skip calling cache_free_alien() when the platform is not numa.
3523 * This will avoid cache misses that happen while accessing slabp (which
3524 * is per page memory reference) to get nodeid. Instead use a global
3525 * variable to skip the call, which is mostly likely to be present in
3526 * the cache.
3527 */
Mel Gormanb6e68bc2009-06-16 15:32:16 -07003528 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07003529 return;
Christoph Lametere498be72005-09-09 13:03:32 -07003530
Linus Torvalds1da177e2005-04-16 15:20:36 -07003531 if (likely(ac->avail < ac->limit)) {
3532 STATS_INC_FREEHIT(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003533 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003534 return;
3535 } else {
3536 STATS_INC_FREEMISS(cachep);
3537 cache_flusharray(cachep, ac);
Christoph Lametere498be72005-09-09 13:03:32 -07003538 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003539 }
3540}
3541
3542/**
3543 * kmem_cache_alloc - Allocate an object
3544 * @cachep: The cache to allocate from.
3545 * @flags: See kmalloc().
3546 *
3547 * Allocate an object from this cache. The flags are only relevant
3548 * if the cache has no available objects.
3549 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003550void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003551{
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003552 void *ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3553
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003554 trace_kmem_cache_alloc(_RET_IP_, ret,
3555 obj_size(cachep), cachep->buffer_size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003556
3557 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003558}
3559EXPORT_SYMBOL(kmem_cache_alloc);
3560
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003561#ifdef CONFIG_KMEMTRACE
3562void *kmem_cache_alloc_notrace(struct kmem_cache *cachep, gfp_t flags)
3563{
3564 return __cache_alloc(cachep, flags, __builtin_return_address(0));
3565}
3566EXPORT_SYMBOL(kmem_cache_alloc_notrace);
3567#endif
3568
Linus Torvalds1da177e2005-04-16 15:20:36 -07003569/**
Randy Dunlap76824862008-03-19 17:00:40 -07003570 * kmem_ptr_validate - check if an untrusted pointer might be a slab entry.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003571 * @cachep: the cache we're checking against
3572 * @ptr: pointer to validate
3573 *
Randy Dunlap76824862008-03-19 17:00:40 -07003574 * This verifies that the untrusted pointer looks sane;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003575 * it is _not_ a guarantee that the pointer is actually
3576 * part of the slab cache in question, but it at least
3577 * validates that the pointer can be dereferenced and
3578 * looks half-way sane.
3579 *
3580 * Currently only used for dentry validation.
3581 */
Christoph Lameterb7f869a22006-12-22 01:06:44 -08003582int kmem_ptr_validate(struct kmem_cache *cachep, const void *ptr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003583{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003584 unsigned long addr = (unsigned long)ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003585 unsigned long min_addr = PAGE_OFFSET;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003586 unsigned long align_mask = BYTES_PER_WORD - 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003587 unsigned long size = cachep->buffer_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003588 struct page *page;
3589
3590 if (unlikely(addr < min_addr))
3591 goto out;
3592 if (unlikely(addr > (unsigned long)high_memory - size))
3593 goto out;
3594 if (unlikely(addr & align_mask))
3595 goto out;
3596 if (unlikely(!kern_addr_valid(addr)))
3597 goto out;
3598 if (unlikely(!kern_addr_valid(addr + size - 1)))
3599 goto out;
3600 page = virt_to_page(ptr);
3601 if (unlikely(!PageSlab(page)))
3602 goto out;
Pekka Enberg065d41c2005-11-13 16:06:46 -08003603 if (unlikely(page_get_cache(page) != cachep))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003604 goto out;
3605 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003606out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003607 return 0;
3608}
3609
3610#ifdef CONFIG_NUMA
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003611void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3612{
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003613 void *ret = __cache_alloc_node(cachep, flags, nodeid,
3614 __builtin_return_address(0));
3615
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003616 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3617 obj_size(cachep), cachep->buffer_size,
3618 flags, nodeid);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003619
3620 return ret;
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003621}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003622EXPORT_SYMBOL(kmem_cache_alloc_node);
3623
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003624#ifdef CONFIG_KMEMTRACE
3625void *kmem_cache_alloc_node_notrace(struct kmem_cache *cachep,
3626 gfp_t flags,
3627 int nodeid)
3628{
3629 return __cache_alloc_node(cachep, flags, nodeid,
3630 __builtin_return_address(0));
3631}
3632EXPORT_SYMBOL(kmem_cache_alloc_node_notrace);
3633#endif
3634
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003635static __always_inline void *
3636__do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003637{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003638 struct kmem_cache *cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003639 void *ret;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003640
3641 cachep = kmem_find_general_cachep(size, flags);
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003642 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3643 return cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003644 ret = kmem_cache_alloc_node_notrace(cachep, flags, node);
3645
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003646 trace_kmalloc_node((unsigned long) caller, ret,
3647 size, cachep->buffer_size, flags, node);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003648
3649 return ret;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003650}
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003651
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003652#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_KMEMTRACE)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003653void *__kmalloc_node(size_t size, gfp_t flags, int node)
3654{
3655 return __do_kmalloc_node(size, flags, node,
3656 __builtin_return_address(0));
3657}
Christoph Hellwigdbe5e692006-09-25 23:31:36 -07003658EXPORT_SYMBOL(__kmalloc_node);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003659
3660void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003661 int node, unsigned long caller)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003662{
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003663 return __do_kmalloc_node(size, flags, node, (void *)caller);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003664}
3665EXPORT_SYMBOL(__kmalloc_node_track_caller);
3666#else
3667void *__kmalloc_node(size_t size, gfp_t flags, int node)
3668{
3669 return __do_kmalloc_node(size, flags, node, NULL);
3670}
3671EXPORT_SYMBOL(__kmalloc_node);
3672#endif /* CONFIG_DEBUG_SLAB */
3673#endif /* CONFIG_NUMA */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003674
3675/**
Paul Drynoff800590f2006-06-23 02:03:48 -07003676 * __do_kmalloc - allocate memory
Linus Torvalds1da177e2005-04-16 15:20:36 -07003677 * @size: how many bytes of memory are required.
Paul Drynoff800590f2006-06-23 02:03:48 -07003678 * @flags: the type of memory to allocate (see kmalloc).
Randy Dunlap911851e2006-03-22 00:08:14 -08003679 * @caller: function caller for debug tracking of the caller
Linus Torvalds1da177e2005-04-16 15:20:36 -07003680 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003681static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3682 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003683{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003684 struct kmem_cache *cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003685 void *ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003686
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003687 /* If you want to save a few bytes .text space: replace
3688 * __ with kmem_.
3689 * Then kmalloc uses the uninlined functions instead of the inline
3690 * functions.
3691 */
3692 cachep = __find_general_cachep(size, flags);
Linus Torvaldsa5c96d82007-07-19 13:17:15 -07003693 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3694 return cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003695 ret = __cache_alloc(cachep, flags, caller);
3696
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003697 trace_kmalloc((unsigned long) caller, ret,
3698 size, cachep->buffer_size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003699
3700 return ret;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003701}
3702
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003703
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003704#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_KMEMTRACE)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003705void *__kmalloc(size_t size, gfp_t flags)
3706{
Al Viro871751e2006-03-25 03:06:39 -08003707 return __do_kmalloc(size, flags, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003708}
3709EXPORT_SYMBOL(__kmalloc);
3710
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003711void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003712{
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003713 return __do_kmalloc(size, flags, (void *)caller);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003714}
3715EXPORT_SYMBOL(__kmalloc_track_caller);
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003716
3717#else
3718void *__kmalloc(size_t size, gfp_t flags)
3719{
3720 return __do_kmalloc(size, flags, NULL);
3721}
3722EXPORT_SYMBOL(__kmalloc);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003723#endif
3724
Linus Torvalds1da177e2005-04-16 15:20:36 -07003725/**
3726 * kmem_cache_free - Deallocate an object
3727 * @cachep: The cache the allocation was from.
3728 * @objp: The previously allocated object.
3729 *
3730 * Free an object which was previously allocated from this
3731 * cache.
3732 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003733void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003734{
3735 unsigned long flags;
3736
3737 local_irq_save(flags);
Ingo Molnar898552c2007-02-10 01:44:57 -08003738 debug_check_no_locks_freed(objp, obj_size(cachep));
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07003739 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3740 debug_check_no_obj_freed(objp, obj_size(cachep));
Ingo Molnar873623d2006-07-13 14:44:38 +02003741 __cache_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003742 local_irq_restore(flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003743
Eduard - Gabriel Munteanuca2b84cb2009-03-23 15:12:24 +02003744 trace_kmem_cache_free(_RET_IP_, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003745}
3746EXPORT_SYMBOL(kmem_cache_free);
3747
3748/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003749 * kfree - free previously allocated memory
3750 * @objp: pointer returned by kmalloc.
3751 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003752 * If @objp is NULL, no operation is performed.
3753 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003754 * Don't free memory not originally allocated by kmalloc()
3755 * or you will run into trouble.
3756 */
3757void kfree(const void *objp)
3758{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003759 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003760 unsigned long flags;
3761
Pekka Enberg2121db72009-03-25 11:05:57 +02003762 trace_kfree(_RET_IP_, objp);
3763
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003764 if (unlikely(ZERO_OR_NULL_PTR(objp)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003765 return;
3766 local_irq_save(flags);
3767 kfree_debugcheck(objp);
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003768 c = virt_to_cache(objp);
Ingo Molnarf9b84042006-06-27 02:54:49 -07003769 debug_check_no_locks_freed(objp, obj_size(c));
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07003770 debug_check_no_obj_freed(objp, obj_size(c));
Ingo Molnar873623d2006-07-13 14:44:38 +02003771 __cache_free(c, (void *)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003772 local_irq_restore(flags);
3773}
3774EXPORT_SYMBOL(kfree);
3775
Pekka Enberg343e0d72006-02-01 03:05:50 -08003776unsigned int kmem_cache_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003777{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003778 return obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003779}
3780EXPORT_SYMBOL(kmem_cache_size);
3781
Pekka Enberg343e0d72006-02-01 03:05:50 -08003782const char *kmem_cache_name(struct kmem_cache *cachep)
Arnaldo Carvalho de Melo19449722005-06-18 22:46:19 -07003783{
3784 return cachep->name;
3785}
3786EXPORT_SYMBOL_GPL(kmem_cache_name);
3787
Christoph Lametere498be72005-09-09 13:03:32 -07003788/*
Simon Arlott183ff222007-10-20 01:27:18 +02003789 * This initializes kmem_list3 or resizes various caches for all nodes.
Christoph Lametere498be72005-09-09 13:03:32 -07003790 */
Pekka Enberg83b519e2009-06-10 19:40:04 +03003791static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -07003792{
3793 int node;
3794 struct kmem_list3 *l3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003795 struct array_cache *new_shared;
Paul Menage3395ee02006-12-06 20:32:16 -08003796 struct array_cache **new_alien = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003797
Mel Gorman9c09a952008-01-24 05:49:54 -08003798 for_each_online_node(node) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003799
Paul Menage3395ee02006-12-06 20:32:16 -08003800 if (use_alien_caches) {
Pekka Enberg83b519e2009-06-10 19:40:04 +03003801 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
Paul Menage3395ee02006-12-06 20:32:16 -08003802 if (!new_alien)
3803 goto fail;
3804 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003805
Eric Dumazet63109842007-05-06 14:49:28 -07003806 new_shared = NULL;
3807 if (cachep->shared) {
3808 new_shared = alloc_arraycache(node,
Christoph Lameter0718dc22006-03-25 03:06:47 -08003809 cachep->shared*cachep->batchcount,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003810 0xbaadf00d, gfp);
Eric Dumazet63109842007-05-06 14:49:28 -07003811 if (!new_shared) {
3812 free_alien_cache(new_alien);
3813 goto fail;
3814 }
Christoph Lameter0718dc22006-03-25 03:06:47 -08003815 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003816
Andrew Mortona737b3e2006-03-22 00:08:11 -08003817 l3 = cachep->nodelists[node];
3818 if (l3) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003819 struct array_cache *shared = l3->shared;
3820
Christoph Lametere498be72005-09-09 13:03:32 -07003821 spin_lock_irq(&l3->list_lock);
3822
Christoph Lametercafeb022006-03-25 03:06:46 -08003823 if (shared)
Christoph Lameter0718dc22006-03-25 03:06:47 -08003824 free_block(cachep, shared->entry,
3825 shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07003826
Christoph Lametercafeb022006-03-25 03:06:46 -08003827 l3->shared = new_shared;
3828 if (!l3->alien) {
Christoph Lametere498be72005-09-09 13:03:32 -07003829 l3->alien = new_alien;
3830 new_alien = NULL;
3831 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003832 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003833 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003834 spin_unlock_irq(&l3->list_lock);
Christoph Lametercafeb022006-03-25 03:06:46 -08003835 kfree(shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003836 free_alien_cache(new_alien);
3837 continue;
3838 }
Pekka Enberg83b519e2009-06-10 19:40:04 +03003839 l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node);
Christoph Lameter0718dc22006-03-25 03:06:47 -08003840 if (!l3) {
3841 free_alien_cache(new_alien);
3842 kfree(new_shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003843 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003844 }
Christoph Lametere498be72005-09-09 13:03:32 -07003845
3846 kmem_list3_init(l3);
3847 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Andrew Mortona737b3e2006-03-22 00:08:11 -08003848 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003849 l3->shared = new_shared;
Christoph Lametere498be72005-09-09 13:03:32 -07003850 l3->alien = new_alien;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003851 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003852 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003853 cachep->nodelists[node] = l3;
3854 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003855 return 0;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003856
Andrew Mortona737b3e2006-03-22 00:08:11 -08003857fail:
Christoph Lameter0718dc22006-03-25 03:06:47 -08003858 if (!cachep->next.next) {
3859 /* Cache is not active yet. Roll back what we did */
3860 node--;
3861 while (node >= 0) {
3862 if (cachep->nodelists[node]) {
3863 l3 = cachep->nodelists[node];
3864
3865 kfree(l3->shared);
3866 free_alien_cache(l3->alien);
3867 kfree(l3);
3868 cachep->nodelists[node] = NULL;
3869 }
3870 node--;
3871 }
3872 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003873 return -ENOMEM;
Christoph Lametere498be72005-09-09 13:03:32 -07003874}
3875
Linus Torvalds1da177e2005-04-16 15:20:36 -07003876struct ccupdate_struct {
Pekka Enberg343e0d72006-02-01 03:05:50 -08003877 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003878 struct array_cache *new[NR_CPUS];
3879};
3880
3881static void do_ccupdate_local(void *info)
3882{
Andrew Mortona737b3e2006-03-22 00:08:11 -08003883 struct ccupdate_struct *new = info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003884 struct array_cache *old;
3885
3886 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003887 old = cpu_cache_get(new->cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003888
Linus Torvalds1da177e2005-04-16 15:20:36 -07003889 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3890 new->new[smp_processor_id()] = old;
3891}
3892
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003893/* Always called with the cache_chain_mutex held */
Andrew Mortona737b3e2006-03-22 00:08:11 -08003894static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003895 int batchcount, int shared, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003896{
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003897 struct ccupdate_struct *new;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003898 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003899
Pekka Enberg83b519e2009-06-10 19:40:04 +03003900 new = kzalloc(sizeof(*new), gfp);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003901 if (!new)
3902 return -ENOMEM;
3903
Christoph Lametere498be72005-09-09 13:03:32 -07003904 for_each_online_cpu(i) {
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003905 new->new[i] = alloc_arraycache(cpu_to_node(i), limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003906 batchcount, gfp);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003907 if (!new->new[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003908 for (i--; i >= 0; i--)
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003909 kfree(new->new[i]);
3910 kfree(new);
Christoph Lametere498be72005-09-09 13:03:32 -07003911 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003912 }
3913 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003914 new->cachep = cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003915
Jens Axboe15c8b6c2008-05-09 09:39:44 +02003916 on_each_cpu(do_ccupdate_local, (void *)new, 1);
Christoph Lametere498be72005-09-09 13:03:32 -07003917
Linus Torvalds1da177e2005-04-16 15:20:36 -07003918 check_irq_on();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003919 cachep->batchcount = batchcount;
3920 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07003921 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003922
Christoph Lametere498be72005-09-09 13:03:32 -07003923 for_each_online_cpu(i) {
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003924 struct array_cache *ccold = new->new[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003925 if (!ccold)
3926 continue;
Christoph Lametere498be72005-09-09 13:03:32 -07003927 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -07003928 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
Christoph Lametere498be72005-09-09 13:03:32 -07003929 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003930 kfree(ccold);
3931 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003932 kfree(new);
Pekka Enberg83b519e2009-06-10 19:40:04 +03003933 return alloc_kmemlist(cachep, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003934}
3935
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003936/* Called with cache_chain_mutex held always */
Pekka Enberg83b519e2009-06-10 19:40:04 +03003937static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003938{
3939 int err;
3940 int limit, shared;
3941
Andrew Mortona737b3e2006-03-22 00:08:11 -08003942 /*
3943 * The head array serves three purposes:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003944 * - create a LIFO ordering, i.e. return objects that are cache-warm
3945 * - reduce the number of spinlock operations.
Andrew Mortona737b3e2006-03-22 00:08:11 -08003946 * - reduce the number of linked list operations on the slab and
Linus Torvalds1da177e2005-04-16 15:20:36 -07003947 * bufctl chains: array operations are cheaper.
3948 * The numbers are guessed, we should auto-tune as described by
3949 * Bonwick.
3950 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003951 if (cachep->buffer_size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003952 limit = 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003953 else if (cachep->buffer_size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003954 limit = 8;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003955 else if (cachep->buffer_size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003956 limit = 24;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003957 else if (cachep->buffer_size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003958 limit = 54;
3959 else
3960 limit = 120;
3961
Andrew Mortona737b3e2006-03-22 00:08:11 -08003962 /*
3963 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
Linus Torvalds1da177e2005-04-16 15:20:36 -07003964 * allocation behaviour: Most allocs on one cpu, most free operations
3965 * on another cpu. For these cases, an efficient object passing between
3966 * cpus is necessary. This is provided by a shared array. The array
3967 * replaces Bonwick's magazine layer.
3968 * On uniprocessor, it's functionally equivalent (but less efficient)
3969 * to a larger limit. Thus disabled by default.
3970 */
3971 shared = 0;
Eric Dumazet364fbb22007-05-06 14:49:27 -07003972 if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003973 shared = 8;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003974
3975#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003976 /*
3977 * With debugging enabled, large batchcount lead to excessively long
3978 * periods with disabled local interrupts. Limit the batchcount
Linus Torvalds1da177e2005-04-16 15:20:36 -07003979 */
3980 if (limit > 32)
3981 limit = 32;
3982#endif
Pekka Enberg83b519e2009-06-10 19:40:04 +03003983 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003984 if (err)
3985 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003986 cachep->name, -err);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003987 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003988}
3989
Christoph Lameter1b552532006-03-22 00:09:07 -08003990/*
3991 * Drain an array if it contains any elements taking the l3 lock only if
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003992 * necessary. Note that the l3 listlock also protects the array_cache
3993 * if drain_array() is used on the shared array.
Christoph Lameter1b552532006-03-22 00:09:07 -08003994 */
3995void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
3996 struct array_cache *ac, int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003997{
3998 int tofree;
3999
Christoph Lameter1b552532006-03-22 00:09:07 -08004000 if (!ac || !ac->avail)
4001 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004002 if (ac->touched && !force) {
4003 ac->touched = 0;
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004004 } else {
Christoph Lameter1b552532006-03-22 00:09:07 -08004005 spin_lock_irq(&l3->list_lock);
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004006 if (ac->avail) {
4007 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4008 if (tofree > ac->avail)
4009 tofree = (ac->avail + 1) / 2;
4010 free_block(cachep, ac->entry, tofree, node);
4011 ac->avail -= tofree;
4012 memmove(ac->entry, &(ac->entry[tofree]),
4013 sizeof(void *) * ac->avail);
4014 }
Christoph Lameter1b552532006-03-22 00:09:07 -08004015 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004016 }
4017}
4018
4019/**
4020 * cache_reap - Reclaim memory from caches.
Randy Dunlap05fb6bf2007-02-28 20:12:13 -08004021 * @w: work descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07004022 *
4023 * Called from workqueue/eventd every few seconds.
4024 * Purpose:
4025 * - clear the per-cpu caches for this CPU.
4026 * - return freeable pages to the main free memory pool.
4027 *
Andrew Mortona737b3e2006-03-22 00:08:11 -08004028 * If we cannot acquire the cache chain mutex then just give up - we'll try
4029 * again on the next iteration.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004030 */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004031static void cache_reap(struct work_struct *w)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004032{
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004033 struct kmem_cache *searchp;
Christoph Lametere498be72005-09-09 13:03:32 -07004034 struct kmem_list3 *l3;
Christoph Lameteraab22072006-03-22 00:09:06 -08004035 int node = numa_node_id();
Jean Delvarebf6aede2009-04-02 16:56:54 -07004036 struct delayed_work *work = to_delayed_work(w);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004037
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004038 if (!mutex_trylock(&cache_chain_mutex))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004039 /* Give up. Setup the next iteration. */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004040 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004041
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004042 list_for_each_entry(searchp, &cache_chain, next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004043 check_irq_on();
4044
Christoph Lameter35386e32006-03-22 00:09:05 -08004045 /*
4046 * We only take the l3 lock if absolutely necessary and we
4047 * have established with reasonable certainty that
4048 * we can do some work if the lock was obtained.
4049 */
Christoph Lameteraab22072006-03-22 00:09:06 -08004050 l3 = searchp->nodelists[node];
Christoph Lameter35386e32006-03-22 00:09:05 -08004051
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004052 reap_alien(searchp, l3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004053
Christoph Lameteraab22072006-03-22 00:09:06 -08004054 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004055
Christoph Lameter35386e32006-03-22 00:09:05 -08004056 /*
4057 * These are racy checks but it does not matter
4058 * if we skip one check or scan twice.
4059 */
Christoph Lametere498be72005-09-09 13:03:32 -07004060 if (time_after(l3->next_reap, jiffies))
Christoph Lameter35386e32006-03-22 00:09:05 -08004061 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004062
Christoph Lametere498be72005-09-09 13:03:32 -07004063 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004064
Christoph Lameteraab22072006-03-22 00:09:06 -08004065 drain_array(searchp, l3, l3->shared, 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004066
Christoph Lametered11d9e2006-06-30 01:55:45 -07004067 if (l3->free_touched)
Christoph Lametere498be72005-09-09 13:03:32 -07004068 l3->free_touched = 0;
Christoph Lametered11d9e2006-06-30 01:55:45 -07004069 else {
4070 int freed;
4071
4072 freed = drain_freelist(searchp, l3, (l3->free_limit +
4073 5 * searchp->num - 1) / (5 * searchp->num));
4074 STATS_ADD_REAPED(searchp, freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004075 }
Christoph Lameter35386e32006-03-22 00:09:05 -08004076next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004077 cond_resched();
4078 }
4079 check_irq_on();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004080 mutex_unlock(&cache_chain_mutex);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004081 next_reap_node();
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004082out:
Andrew Mortona737b3e2006-03-22 00:08:11 -08004083 /* Set up the next iteration */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004084 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004085}
4086
Linus Torvalds158a9622008-01-02 13:04:48 -08004087#ifdef CONFIG_SLABINFO
Linus Torvalds1da177e2005-04-16 15:20:36 -07004088
Pekka Enberg85289f92006-01-08 01:00:36 -08004089static void print_slabinfo_header(struct seq_file *m)
4090{
4091 /*
4092 * Output format version, so at least we can change it
4093 * without _too_ many complaints.
4094 */
4095#if STATS
4096 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4097#else
4098 seq_puts(m, "slabinfo - version: 2.1\n");
4099#endif
4100 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4101 "<objperslab> <pagesperslab>");
4102 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4103 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4104#if STATS
4105 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004106 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
Pekka Enberg85289f92006-01-08 01:00:36 -08004107 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4108#endif
4109 seq_putc(m, '\n');
4110}
4111
Linus Torvalds1da177e2005-04-16 15:20:36 -07004112static void *s_start(struct seq_file *m, loff_t *pos)
4113{
4114 loff_t n = *pos;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004115
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004116 mutex_lock(&cache_chain_mutex);
Pekka Enberg85289f92006-01-08 01:00:36 -08004117 if (!n)
4118 print_slabinfo_header(m);
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004119
4120 return seq_list_start(&cache_chain, *pos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004121}
4122
4123static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4124{
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004125 return seq_list_next(p, &cache_chain, pos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004126}
4127
4128static void s_stop(struct seq_file *m, void *p)
4129{
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004130 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004131}
4132
4133static int s_show(struct seq_file *m, void *p)
4134{
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004135 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004136 struct slab *slabp;
4137 unsigned long active_objs;
4138 unsigned long num_objs;
4139 unsigned long active_slabs = 0;
4140 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004141 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004142 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07004143 int node;
4144 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004145
Linus Torvalds1da177e2005-04-16 15:20:36 -07004146 active_objs = 0;
4147 num_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004148 for_each_online_node(node) {
4149 l3 = cachep->nodelists[node];
4150 if (!l3)
4151 continue;
4152
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004153 check_irq_on();
4154 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07004155
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004156 list_for_each_entry(slabp, &l3->slabs_full, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004157 if (slabp->inuse != cachep->num && !error)
4158 error = "slabs_full accounting error";
4159 active_objs += cachep->num;
4160 active_slabs++;
4161 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004162 list_for_each_entry(slabp, &l3->slabs_partial, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004163 if (slabp->inuse == cachep->num && !error)
4164 error = "slabs_partial inuse accounting error";
4165 if (!slabp->inuse && !error)
4166 error = "slabs_partial/inuse accounting error";
4167 active_objs += slabp->inuse;
4168 active_slabs++;
4169 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004170 list_for_each_entry(slabp, &l3->slabs_free, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004171 if (slabp->inuse && !error)
4172 error = "slabs_free/inuse accounting error";
4173 num_slabs++;
4174 }
4175 free_objects += l3->free_objects;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08004176 if (l3->shared)
4177 shared_avail += l3->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07004178
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004179 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004180 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004181 num_slabs += active_slabs;
4182 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07004183 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004184 error = "free_objects accounting error";
4185
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004186 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004187 if (error)
4188 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4189
4190 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004191 name, active_objs, num_objs, cachep->buffer_size,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004192 cachep->num, (1 << cachep->gfporder));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004193 seq_printf(m, " : tunables %4u %4u %4u",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004194 cachep->limit, cachep->batchcount, cachep->shared);
Christoph Lametere498be72005-09-09 13:03:32 -07004195 seq_printf(m, " : slabdata %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004196 active_slabs, num_slabs, shared_avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004197#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004198 { /* list3 stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004199 unsigned long high = cachep->high_mark;
4200 unsigned long allocs = cachep->num_allocations;
4201 unsigned long grown = cachep->grown;
4202 unsigned long reaped = cachep->reaped;
4203 unsigned long errors = cachep->errors;
4204 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004205 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07004206 unsigned long node_frees = cachep->node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004207 unsigned long overflows = cachep->node_overflow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004208
Christoph Lametere498be72005-09-09 13:03:32 -07004209 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004210 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown,
Andrew Mortona737b3e2006-03-22 00:08:11 -08004211 reaped, errors, max_freeable, node_allocs,
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004212 node_frees, overflows);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004213 }
4214 /* cpu stats */
4215 {
4216 unsigned long allochit = atomic_read(&cachep->allochit);
4217 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4218 unsigned long freehit = atomic_read(&cachep->freehit);
4219 unsigned long freemiss = atomic_read(&cachep->freemiss);
4220
4221 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004222 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004223 }
4224#endif
4225 seq_putc(m, '\n');
Linus Torvalds1da177e2005-04-16 15:20:36 -07004226 return 0;
4227}
4228
4229/*
4230 * slabinfo_op - iterator that generates /proc/slabinfo
4231 *
4232 * Output layout:
4233 * cache-name
4234 * num-active-objs
4235 * total-objs
4236 * object size
4237 * num-active-slabs
4238 * total-slabs
4239 * num-pages-per-slab
4240 * + further values on SMP and with statistics enabled
4241 */
4242
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04004243static const struct seq_operations slabinfo_op = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004244 .start = s_start,
4245 .next = s_next,
4246 .stop = s_stop,
4247 .show = s_show,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004248};
4249
4250#define MAX_SLABINFO_WRITE 128
4251/**
4252 * slabinfo_write - Tuning for the slab allocator
4253 * @file: unused
4254 * @buffer: user buffer
4255 * @count: data length
4256 * @ppos: unused
4257 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004258ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4259 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004260{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004261 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004262 int limit, batchcount, shared, res;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004263 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004264
Linus Torvalds1da177e2005-04-16 15:20:36 -07004265 if (count > MAX_SLABINFO_WRITE)
4266 return -EINVAL;
4267 if (copy_from_user(&kbuf, buffer, count))
4268 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004269 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07004270
4271 tmp = strchr(kbuf, ' ');
4272 if (!tmp)
4273 return -EINVAL;
4274 *tmp = '\0';
4275 tmp++;
4276 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4277 return -EINVAL;
4278
4279 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004280 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004281 res = -EINVAL;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004282 list_for_each_entry(cachep, &cache_chain, next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004283 if (!strcmp(cachep->name, kbuf)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08004284 if (limit < 1 || batchcount < 1 ||
4285 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07004286 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004287 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07004288 res = do_tune_cpucache(cachep, limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03004289 batchcount, shared,
4290 GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004291 }
4292 break;
4293 }
4294 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004295 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004296 if (res >= 0)
4297 res = count;
4298 return res;
4299}
Al Viro871751e2006-03-25 03:06:39 -08004300
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04004301static int slabinfo_open(struct inode *inode, struct file *file)
4302{
4303 return seq_open(file, &slabinfo_op);
4304}
4305
4306static const struct file_operations proc_slabinfo_operations = {
4307 .open = slabinfo_open,
4308 .read = seq_read,
4309 .write = slabinfo_write,
4310 .llseek = seq_lseek,
4311 .release = seq_release,
4312};
4313
Al Viro871751e2006-03-25 03:06:39 -08004314#ifdef CONFIG_DEBUG_SLAB_LEAK
4315
4316static void *leaks_start(struct seq_file *m, loff_t *pos)
4317{
Al Viro871751e2006-03-25 03:06:39 -08004318 mutex_lock(&cache_chain_mutex);
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004319 return seq_list_start(&cache_chain, *pos);
Al Viro871751e2006-03-25 03:06:39 -08004320}
4321
4322static inline int add_caller(unsigned long *n, unsigned long v)
4323{
4324 unsigned long *p;
4325 int l;
4326 if (!v)
4327 return 1;
4328 l = n[1];
4329 p = n + 2;
4330 while (l) {
4331 int i = l/2;
4332 unsigned long *q = p + 2 * i;
4333 if (*q == v) {
4334 q[1]++;
4335 return 1;
4336 }
4337 if (*q > v) {
4338 l = i;
4339 } else {
4340 p = q + 2;
4341 l -= i + 1;
4342 }
4343 }
4344 if (++n[1] == n[0])
4345 return 0;
4346 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4347 p[0] = v;
4348 p[1] = 1;
4349 return 1;
4350}
4351
4352static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4353{
4354 void *p;
4355 int i;
4356 if (n[0] == n[1])
4357 return;
4358 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4359 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4360 continue;
4361 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4362 return;
4363 }
4364}
4365
4366static void show_symbol(struct seq_file *m, unsigned long address)
4367{
4368#ifdef CONFIG_KALLSYMS
Al Viro871751e2006-03-25 03:06:39 -08004369 unsigned long offset, size;
Tejun Heo9281ace2007-07-17 04:03:51 -07004370 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
Al Viro871751e2006-03-25 03:06:39 -08004371
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004372 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
Al Viro871751e2006-03-25 03:06:39 -08004373 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004374 if (modname[0])
Al Viro871751e2006-03-25 03:06:39 -08004375 seq_printf(m, " [%s]", modname);
4376 return;
4377 }
4378#endif
4379 seq_printf(m, "%p", (void *)address);
4380}
4381
4382static int leaks_show(struct seq_file *m, void *p)
4383{
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004384 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
Al Viro871751e2006-03-25 03:06:39 -08004385 struct slab *slabp;
4386 struct kmem_list3 *l3;
4387 const char *name;
4388 unsigned long *n = m->private;
4389 int node;
4390 int i;
4391
4392 if (!(cachep->flags & SLAB_STORE_USER))
4393 return 0;
4394 if (!(cachep->flags & SLAB_RED_ZONE))
4395 return 0;
4396
4397 /* OK, we can do it */
4398
4399 n[1] = 0;
4400
4401 for_each_online_node(node) {
4402 l3 = cachep->nodelists[node];
4403 if (!l3)
4404 continue;
4405
4406 check_irq_on();
4407 spin_lock_irq(&l3->list_lock);
4408
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004409 list_for_each_entry(slabp, &l3->slabs_full, list)
Al Viro871751e2006-03-25 03:06:39 -08004410 handle_slab(n, cachep, slabp);
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004411 list_for_each_entry(slabp, &l3->slabs_partial, list)
Al Viro871751e2006-03-25 03:06:39 -08004412 handle_slab(n, cachep, slabp);
Al Viro871751e2006-03-25 03:06:39 -08004413 spin_unlock_irq(&l3->list_lock);
4414 }
4415 name = cachep->name;
4416 if (n[0] == n[1]) {
4417 /* Increase the buffer size */
4418 mutex_unlock(&cache_chain_mutex);
4419 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4420 if (!m->private) {
4421 /* Too bad, we are really out */
4422 m->private = n;
4423 mutex_lock(&cache_chain_mutex);
4424 return -ENOMEM;
4425 }
4426 *(unsigned long *)m->private = n[0] * 2;
4427 kfree(n);
4428 mutex_lock(&cache_chain_mutex);
4429 /* Now make sure this entry will be retried */
4430 m->count = m->size;
4431 return 0;
4432 }
4433 for (i = 0; i < n[1]; i++) {
4434 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4435 show_symbol(m, n[2*i+2]);
4436 seq_putc(m, '\n');
4437 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004438
Al Viro871751e2006-03-25 03:06:39 -08004439 return 0;
4440}
4441
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004442static const struct seq_operations slabstats_op = {
Al Viro871751e2006-03-25 03:06:39 -08004443 .start = leaks_start,
4444 .next = s_next,
4445 .stop = s_stop,
4446 .show = leaks_show,
4447};
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004448
4449static int slabstats_open(struct inode *inode, struct file *file)
4450{
4451 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4452 int ret = -ENOMEM;
4453 if (n) {
4454 ret = seq_open(file, &slabstats_op);
4455 if (!ret) {
4456 struct seq_file *m = file->private_data;
4457 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4458 m->private = n;
4459 n = NULL;
4460 }
4461 kfree(n);
4462 }
4463 return ret;
4464}
4465
4466static const struct file_operations proc_slabstats_operations = {
4467 .open = slabstats_open,
4468 .read = seq_read,
4469 .llseek = seq_lseek,
4470 .release = seq_release_private,
4471};
Al Viro871751e2006-03-25 03:06:39 -08004472#endif
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004473
4474static int __init slab_proc_init(void)
4475{
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04004476 proc_create("slabinfo",S_IWUSR|S_IRUGO,NULL,&proc_slabinfo_operations);
Alexey Dobriyana0ec95a2008-10-06 00:59:10 +04004477#ifdef CONFIG_DEBUG_SLAB_LEAK
4478 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4479#endif
4480 return 0;
4481}
4482module_init(slab_proc_init);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004483#endif
4484
Manfred Spraul00e145b2005-09-03 15:55:07 -07004485/**
4486 * ksize - get the actual amount of memory allocated for a given object
4487 * @objp: Pointer to the object
4488 *
4489 * kmalloc may internally round up allocations and return more memory
4490 * than requested. ksize() can be used to determine the actual amount of
4491 * memory allocated. The caller may use this additional memory, even though
4492 * a smaller amount of memory was initially specified with the kmalloc call.
4493 * The caller must guarantee that objp points to a valid object previously
4494 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4495 * must not be freed during the duration of the call.
4496 */
Pekka Enbergfd76bab2007-05-06 14:48:40 -07004497size_t ksize(const void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004498{
Christoph Lameteref8b4522007-10-16 01:24:46 -07004499 BUG_ON(!objp);
4500 if (unlikely(objp == ZERO_SIZE_PTR))
Manfred Spraul00e145b2005-09-03 15:55:07 -07004501 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004502
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08004503 return obj_size(virt_to_cache(objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004504}
Kirill A. Shutemovb1aabec2009-02-10 15:21:44 +02004505EXPORT_SYMBOL(ksize);