blob: f46b65d124e5fabc39ba94653ff5fcf33bab89a5 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * linux/mm/slab.c
3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
5 *
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
7 *
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
10 *
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
13 *
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
21 *
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
27 *
28 * This means, that your constructor is used only for newly allocated
Simon Arlott183ff222007-10-20 01:27:18 +020029 * slabs and you must pass objects with the same initializations to
Linus Torvalds1da177e2005-04-16 15:20:36 -070030 * kmem_cache_free.
31 *
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
35 *
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
38 * partial slabs
39 * empty slabs with no allocated objects
40 *
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
43 *
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
46 *
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
52 *
Andrew Mortona737b3e2006-03-22 00:08:11 -080053 * The c_cpuarray may not be read with enabled local interrupts -
Linus Torvalds1da177e2005-04-16 15:20:36 -070054 * it's changed with a smp_call_function().
55 *
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
Pekka Enberg343e0d72006-02-01 03:05:50 -080058 * Several members in struct kmem_cache and struct slab never change, they
Linus Torvalds1da177e2005-04-16 15:20:36 -070059 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
63 *
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
66 * his patch.
67 *
68 * Further notes from the original documentation:
69 *
70 * 11 April '97. Started multi-threading - markhe
Ingo Molnarfc0abb12006-01-18 17:42:33 -080071 * The global cache-chain is protected by the mutex 'cache_chain_mutex'.
Linus Torvalds1da177e2005-04-16 15:20:36 -070072 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
75 *
76 * At present, each engine can be growing a cache. This should be blocked.
77 *
Christoph Lametere498be72005-09-09 13:03:32 -070078 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
83 *
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
Linus Torvalds1da177e2005-04-16 15:20:36 -070087 */
88
Linus Torvalds1da177e2005-04-16 15:20:36 -070089#include <linux/slab.h>
90#include <linux/mm.h>
Randy Dunlapc9cf5522006-06-27 02:53:52 -070091#include <linux/poison.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070092#include <linux/swap.h>
93#include <linux/cache.h>
94#include <linux/interrupt.h>
95#include <linux/init.h>
96#include <linux/compiler.h>
Paul Jackson101a5002006-03-24 03:16:07 -080097#include <linux/cpuset.h>
Alexey Dobriyana0ec95a82008-10-06 00:59:10 +040098#include <linux/proc_fs.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070099#include <linux/seq_file.h>
100#include <linux/notifier.h>
101#include <linux/kallsyms.h>
102#include <linux/cpu.h>
103#include <linux/sysctl.h>
104#include <linux/module.h>
Zhaolei02af61b2009-04-10 14:26:18 +0800105#include <linux/kmemtrace.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700106#include <linux/rcupdate.h>
Paulo Marques543537b2005-06-23 00:09:02 -0700107#include <linux/string.h>
Andrew Morton138ae662006-12-06 20:36:41 -0800108#include <linux/uaccess.h>
Christoph Lametere498be72005-09-09 13:03:32 -0700109#include <linux/nodemask.h>
Catalin Marinasd5cff632009-06-11 13:22:40 +0100110#include <linux/kmemleak.h>
Christoph Lameterdc85da12006-01-18 17:42:36 -0800111#include <linux/mempolicy.h>
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800112#include <linux/mutex.h>
Akinobu Mita8a8b6502006-12-08 02:39:44 -0800113#include <linux/fault-inject.h>
Ingo Molnare7eebaf2006-06-27 02:54:55 -0700114#include <linux/rtmutex.h>
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800115#include <linux/reciprocal_div.h>
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -0700116#include <linux/debugobjects.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700117
Linus Torvalds1da177e2005-04-16 15:20:36 -0700118#include <asm/cacheflush.h>
119#include <asm/tlbflush.h>
120#include <asm/page.h>
121
122/*
Christoph Lameter50953fe2007-05-06 14:50:16 -0700123 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700124 * 0 for faster, smaller code (especially in the critical paths).
125 *
126 * STATS - 1 to collect stats for /proc/slabinfo.
127 * 0 for faster, smaller code (especially in the critical paths).
128 *
129 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
130 */
131
132#ifdef CONFIG_DEBUG_SLAB
133#define DEBUG 1
134#define STATS 1
135#define FORCED_DEBUG 1
136#else
137#define DEBUG 0
138#define STATS 0
139#define FORCED_DEBUG 0
140#endif
141
Linus Torvalds1da177e2005-04-16 15:20:36 -0700142/* Shouldn't this be in a header file somewhere? */
143#define BYTES_PER_WORD sizeof(void *)
David Woodhouse87a927c2007-07-04 21:26:44 -0400144#define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700145
Linus Torvalds1da177e2005-04-16 15:20:36 -0700146#ifndef ARCH_KMALLOC_MINALIGN
147/*
148 * Enforce a minimum alignment for the kmalloc caches.
149 * Usually, the kmalloc caches are cache_line_size() aligned, except when
150 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
151 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
David Woodhouseb46b8f12007-05-08 00:22:59 -0700152 * alignment larger than the alignment of a 64-bit integer.
153 * ARCH_KMALLOC_MINALIGN allows that.
154 * Note that increasing this value may disable some debug features.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700155 */
David Woodhouseb46b8f12007-05-08 00:22:59 -0700156#define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700157#endif
158
159#ifndef ARCH_SLAB_MINALIGN
160/*
161 * Enforce a minimum alignment for all caches.
162 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
163 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
164 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
165 * some debug features.
166 */
167#define ARCH_SLAB_MINALIGN 0
168#endif
169
170#ifndef ARCH_KMALLOC_FLAGS
171#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
172#endif
173
174/* Legal flag mask for kmem_cache_create(). */
175#if DEBUG
Christoph Lameter50953fe2007-05-06 14:50:16 -0700176# define CREATE_MASK (SLAB_RED_ZONE | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700177 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
Christoph Lameterac2b8982006-03-22 00:08:15 -0800178 SLAB_CACHE_DMA | \
Christoph Lameter5af60832007-05-06 14:49:56 -0700179 SLAB_STORE_USER | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700180 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -0700181 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
Catalin Marinasd5cff632009-06-11 13:22:40 +0100182 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700183#else
Christoph Lameterac2b8982006-03-22 00:08:15 -0800184# define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
Christoph Lameter5af60832007-05-06 14:49:56 -0700185 SLAB_CACHE_DMA | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700186 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -0700187 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
Catalin Marinasd5cff632009-06-11 13:22:40 +0100188 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700189#endif
190
191/*
192 * kmem_bufctl_t:
193 *
194 * Bufctl's are used for linking objs within a slab
195 * linked offsets.
196 *
197 * This implementation relies on "struct page" for locating the cache &
198 * slab an object belongs to.
199 * This allows the bufctl structure to be small (one int), but limits
200 * the number of objects a slab (not a cache) can contain when off-slab
201 * bufctls are used. The limit is the size of the largest general cache
202 * that does not use off-slab slabs.
203 * For 32bit archs with 4 kB pages, is this 56.
204 * This is not serious, as it is only for large objects, when it is unwise
205 * to have too many per slab.
206 * Note: This limit can be raised by introducing a general cache whose size
207 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
208 */
209
Kyle Moffettfa5b08d2005-09-03 15:55:03 -0700210typedef unsigned int kmem_bufctl_t;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700211#define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
212#define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
Al Viro871751e2006-03-25 03:06:39 -0800213#define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
214#define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700215
Linus Torvalds1da177e2005-04-16 15:20:36 -0700216/*
217 * struct slab
218 *
219 * Manages the objs in a slab. Placed either at the beginning of mem allocated
220 * for a slab, or allocated from an general cache.
221 * Slabs are chained into three list: fully used, partial, fully free slabs.
222 */
223struct slab {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800224 struct list_head list;
225 unsigned long colouroff;
226 void *s_mem; /* including colour offset */
227 unsigned int inuse; /* num of objs active in slab */
228 kmem_bufctl_t free;
229 unsigned short nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700230};
231
232/*
233 * struct slab_rcu
234 *
235 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
236 * arrange for kmem_freepages to be called via RCU. This is useful if
237 * we need to approach a kernel structure obliquely, from its address
238 * obtained without the usual locking. We can lock the structure to
239 * stabilize it and check it's still at the given address, only if we
240 * can be sure that the memory has not been meanwhile reused for some
241 * other kind of object (which our subsystem's lock might corrupt).
242 *
243 * rcu_read_lock before reading the address, then rcu_read_unlock after
244 * taking the spinlock within the structure expected at that address.
245 *
246 * We assume struct slab_rcu can overlay struct slab when destroying.
247 */
248struct slab_rcu {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800249 struct rcu_head head;
Pekka Enberg343e0d72006-02-01 03:05:50 -0800250 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800251 void *addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700252};
253
254/*
255 * struct array_cache
256 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700257 * Purpose:
258 * - LIFO ordering, to hand out cache-warm objects from _alloc
259 * - reduce the number of linked list operations
260 * - reduce spinlock operations
261 *
262 * The limit is stored in the per-cpu structure to reduce the data cache
263 * footprint.
264 *
265 */
266struct array_cache {
267 unsigned int avail;
268 unsigned int limit;
269 unsigned int batchcount;
270 unsigned int touched;
Christoph Lametere498be72005-09-09 13:03:32 -0700271 spinlock_t lock;
Robert P. J. Daybda5b652007-10-16 23:30:05 -0700272 void *entry[]; /*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800273 * Must have this definition in here for the proper
274 * alignment of array_cache. Also simplifies accessing
275 * the entries.
Andrew Mortona737b3e2006-03-22 00:08:11 -0800276 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700277};
278
Andrew Mortona737b3e2006-03-22 00:08:11 -0800279/*
280 * bootstrap: The caches do not work without cpuarrays anymore, but the
281 * cpuarrays are allocated from the generic caches...
Linus Torvalds1da177e2005-04-16 15:20:36 -0700282 */
283#define BOOT_CPUCACHE_ENTRIES 1
284struct arraycache_init {
285 struct array_cache cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800286 void *entries[BOOT_CPUCACHE_ENTRIES];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700287};
288
289/*
Christoph Lametere498be72005-09-09 13:03:32 -0700290 * The slab lists for all objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700291 */
292struct kmem_list3 {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800293 struct list_head slabs_partial; /* partial list first, better asm code */
294 struct list_head slabs_full;
295 struct list_head slabs_free;
296 unsigned long free_objects;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800297 unsigned int free_limit;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800298 unsigned int colour_next; /* Per-node cache coloring */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800299 spinlock_t list_lock;
300 struct array_cache *shared; /* shared per node */
301 struct array_cache **alien; /* on other nodes */
Christoph Lameter35386e32006-03-22 00:09:05 -0800302 unsigned long next_reap; /* updated without locking */
303 int free_touched; /* updated without locking */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700304};
305
Christoph Lametere498be72005-09-09 13:03:32 -0700306/*
307 * Need this for bootstrapping a per node allocator.
308 */
Pekka Enberg556a1692008-01-25 08:20:51 +0200309#define NUM_INIT_LISTS (3 * MAX_NUMNODES)
Christoph Lametere498be72005-09-09 13:03:32 -0700310struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
311#define CACHE_CACHE 0
Pekka Enberg556a1692008-01-25 08:20:51 +0200312#define SIZE_AC MAX_NUMNODES
313#define SIZE_L3 (2 * MAX_NUMNODES)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700314
Christoph Lametered11d9e2006-06-30 01:55:45 -0700315static int drain_freelist(struct kmem_cache *cache,
316 struct kmem_list3 *l3, int tofree);
317static void free_block(struct kmem_cache *cachep, void **objpp, int len,
318 int node);
Pekka Enberg83b519e2009-06-10 19:40:04 +0300319static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
David Howells65f27f32006-11-22 14:55:48 +0000320static void cache_reap(struct work_struct *unused);
Christoph Lametered11d9e2006-06-30 01:55:45 -0700321
Christoph Lametere498be72005-09-09 13:03:32 -0700322/*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800323 * This function must be completely optimized away if a constant is passed to
324 * it. Mostly the same as what is in linux/slab.h except it returns an index.
Christoph Lametere498be72005-09-09 13:03:32 -0700325 */
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700326static __always_inline int index_of(const size_t size)
Christoph Lametere498be72005-09-09 13:03:32 -0700327{
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800328 extern void __bad_size(void);
329
Christoph Lametere498be72005-09-09 13:03:32 -0700330 if (__builtin_constant_p(size)) {
331 int i = 0;
332
333#define CACHE(x) \
334 if (size <=x) \
335 return i; \
336 else \
337 i++;
Joe Perches1c61fc42008-03-05 13:58:17 -0800338#include <linux/kmalloc_sizes.h>
Christoph Lametere498be72005-09-09 13:03:32 -0700339#undef CACHE
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800340 __bad_size();
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700341 } else
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800342 __bad_size();
Christoph Lametere498be72005-09-09 13:03:32 -0700343 return 0;
344}
345
Ingo Molnare0a42722006-06-23 02:03:46 -0700346static int slab_early_init = 1;
347
Christoph Lametere498be72005-09-09 13:03:32 -0700348#define INDEX_AC index_of(sizeof(struct arraycache_init))
349#define INDEX_L3 index_of(sizeof(struct kmem_list3))
350
Pekka Enberg5295a742006-02-01 03:05:48 -0800351static void kmem_list3_init(struct kmem_list3 *parent)
Christoph Lametere498be72005-09-09 13:03:32 -0700352{
353 INIT_LIST_HEAD(&parent->slabs_full);
354 INIT_LIST_HEAD(&parent->slabs_partial);
355 INIT_LIST_HEAD(&parent->slabs_free);
356 parent->shared = NULL;
357 parent->alien = NULL;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800358 parent->colour_next = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700359 spin_lock_init(&parent->list_lock);
360 parent->free_objects = 0;
361 parent->free_touched = 0;
362}
363
Andrew Mortona737b3e2006-03-22 00:08:11 -0800364#define MAKE_LIST(cachep, listp, slab, nodeid) \
365 do { \
366 INIT_LIST_HEAD(listp); \
367 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
Christoph Lametere498be72005-09-09 13:03:32 -0700368 } while (0)
369
Andrew Mortona737b3e2006-03-22 00:08:11 -0800370#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
371 do { \
Christoph Lametere498be72005-09-09 13:03:32 -0700372 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
373 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
374 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
375 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700376
377/*
Pekka Enberg343e0d72006-02-01 03:05:50 -0800378 * struct kmem_cache
Linus Torvalds1da177e2005-04-16 15:20:36 -0700379 *
380 * manages a cache.
381 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800382
Pekka J Enberg2109a2d2005-11-07 00:58:01 -0800383struct kmem_cache {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700384/* 1) per-cpu data, touched during every alloc/free */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800385 struct array_cache *array[NR_CPUS];
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800386/* 2) Cache tunables. Protected by cache_chain_mutex */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800387 unsigned int batchcount;
388 unsigned int limit;
389 unsigned int shared;
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800390
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800391 unsigned int buffer_size;
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800392 u32 reciprocal_buffer_size;
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800393/* 3) touched by every alloc & free from the backend */
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800394
Andrew Mortona737b3e2006-03-22 00:08:11 -0800395 unsigned int flags; /* constant flags */
396 unsigned int num; /* # of objs per slab */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700397
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800398/* 4) cache_grow/shrink */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700399 /* order of pgs per slab (2^n) */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800400 unsigned int gfporder;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700401
402 /* force GFP flags, e.g. GFP_DMA */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800403 gfp_t gfpflags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700404
Andrew Mortona737b3e2006-03-22 00:08:11 -0800405 size_t colour; /* cache colouring range */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800406 unsigned int colour_off; /* colour offset */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800407 struct kmem_cache *slabp_cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800408 unsigned int slab_size;
Andrew Mortona737b3e2006-03-22 00:08:11 -0800409 unsigned int dflags; /* dynamic flags */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700410
411 /* constructor func */
Alexey Dobriyan51cc5062008-07-25 19:45:34 -0700412 void (*ctor)(void *obj);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700413
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800414/* 5) cache creation/removal */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800415 const char *name;
416 struct list_head next;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700417
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800418/* 6) statistics */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700419#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800420 unsigned long num_active;
421 unsigned long num_allocations;
422 unsigned long high_mark;
423 unsigned long grown;
424 unsigned long reaped;
425 unsigned long errors;
426 unsigned long max_freeable;
427 unsigned long node_allocs;
428 unsigned long node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700429 unsigned long node_overflow;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800430 atomic_t allochit;
431 atomic_t allocmiss;
432 atomic_t freehit;
433 atomic_t freemiss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700434#endif
435#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800436 /*
437 * If debugging is enabled, then the allocator can add additional
438 * fields and/or padding to every object. buffer_size contains the total
439 * object size including these internal fields, the following two
440 * variables contain the offset to the user object and its size.
441 */
442 int obj_offset;
443 int obj_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700444#endif
Eric Dumazet8da34302007-05-06 14:49:29 -0700445 /*
446 * We put nodelists[] at the end of kmem_cache, because we want to size
447 * this array to nr_node_ids slots instead of MAX_NUMNODES
448 * (see kmem_cache_init())
449 * We still use [MAX_NUMNODES] and not [1] or [0] because cache_cache
450 * is statically defined, so we reserve the max number of nodes.
451 */
452 struct kmem_list3 *nodelists[MAX_NUMNODES];
453 /*
454 * Do not add fields after nodelists[]
455 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700456};
457
458#define CFLGS_OFF_SLAB (0x80000000UL)
459#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
460
461#define BATCHREFILL_LIMIT 16
Andrew Mortona737b3e2006-03-22 00:08:11 -0800462/*
463 * Optimization question: fewer reaps means less probability for unnessary
464 * cpucache drain/refill cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700465 *
Adrian Bunkdc6f3f22005-11-08 16:44:08 +0100466 * OTOH the cpuarrays can contain lots of objects,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700467 * which could lock up otherwise freeable slabs.
468 */
469#define REAPTIMEOUT_CPUC (2*HZ)
470#define REAPTIMEOUT_LIST3 (4*HZ)
471
472#if STATS
473#define STATS_INC_ACTIVE(x) ((x)->num_active++)
474#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
475#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
476#define STATS_INC_GROWN(x) ((x)->grown++)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700477#define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
Andrew Mortona737b3e2006-03-22 00:08:11 -0800478#define STATS_SET_HIGH(x) \
479 do { \
480 if ((x)->num_active > (x)->high_mark) \
481 (x)->high_mark = (x)->num_active; \
482 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700483#define STATS_INC_ERR(x) ((x)->errors++)
484#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
Christoph Lametere498be72005-09-09 13:03:32 -0700485#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700486#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800487#define STATS_SET_FREEABLE(x, i) \
488 do { \
489 if ((x)->max_freeable < i) \
490 (x)->max_freeable = i; \
491 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700492#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
493#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
494#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
495#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
496#else
497#define STATS_INC_ACTIVE(x) do { } while (0)
498#define STATS_DEC_ACTIVE(x) do { } while (0)
499#define STATS_INC_ALLOCED(x) do { } while (0)
500#define STATS_INC_GROWN(x) do { } while (0)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700501#define STATS_ADD_REAPED(x,y) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700502#define STATS_SET_HIGH(x) do { } while (0)
503#define STATS_INC_ERR(x) do { } while (0)
504#define STATS_INC_NODEALLOCS(x) do { } while (0)
Christoph Lametere498be72005-09-09 13:03:32 -0700505#define STATS_INC_NODEFREES(x) do { } while (0)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700506#define STATS_INC_ACOVERFLOW(x) do { } while (0)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800507#define STATS_SET_FREEABLE(x, i) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700508#define STATS_INC_ALLOCHIT(x) do { } while (0)
509#define STATS_INC_ALLOCMISS(x) do { } while (0)
510#define STATS_INC_FREEHIT(x) do { } while (0)
511#define STATS_INC_FREEMISS(x) do { } while (0)
512#endif
513
514#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -0700515
Andrew Mortona737b3e2006-03-22 00:08:11 -0800516/*
517 * memory layout of objects:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700518 * 0 : objp
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800519 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
Linus Torvalds1da177e2005-04-16 15:20:36 -0700520 * the end of an object is aligned with the end of the real
521 * allocation. Catches writes behind the end of the allocation.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800522 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700523 * redzone word.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800524 * cachep->obj_offset: The real object.
525 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
Andrew Mortona737b3e2006-03-22 00:08:11 -0800526 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
527 * [BYTES_PER_WORD long]
Linus Torvalds1da177e2005-04-16 15:20:36 -0700528 */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800529static int obj_offset(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700530{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800531 return cachep->obj_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700532}
533
Pekka Enberg343e0d72006-02-01 03:05:50 -0800534static int obj_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700535{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800536 return cachep->obj_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700537}
538
David Woodhouseb46b8f12007-05-08 00:22:59 -0700539static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700540{
541 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
David Woodhouseb46b8f12007-05-08 00:22:59 -0700542 return (unsigned long long*) (objp + obj_offset(cachep) -
543 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700544}
545
David Woodhouseb46b8f12007-05-08 00:22:59 -0700546static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700547{
548 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
549 if (cachep->flags & SLAB_STORE_USER)
David Woodhouseb46b8f12007-05-08 00:22:59 -0700550 return (unsigned long long *)(objp + cachep->buffer_size -
551 sizeof(unsigned long long) -
David Woodhouse87a927c2007-07-04 21:26:44 -0400552 REDZONE_ALIGN);
David Woodhouseb46b8f12007-05-08 00:22:59 -0700553 return (unsigned long long *) (objp + cachep->buffer_size -
554 sizeof(unsigned long long));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700555}
556
Pekka Enberg343e0d72006-02-01 03:05:50 -0800557static void **dbg_userword(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700558{
559 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800560 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700561}
562
563#else
564
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800565#define obj_offset(x) 0
566#define obj_size(cachep) (cachep->buffer_size)
David Woodhouseb46b8f12007-05-08 00:22:59 -0700567#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
568#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
Linus Torvalds1da177e2005-04-16 15:20:36 -0700569#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
570
571#endif
572
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +0300573#ifdef CONFIG_KMEMTRACE
574size_t slab_buffer_size(struct kmem_cache *cachep)
575{
576 return cachep->buffer_size;
577}
578EXPORT_SYMBOL(slab_buffer_size);
579#endif
580
Linus Torvalds1da177e2005-04-16 15:20:36 -0700581/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700582 * Do not go above this order unless 0 objects fit into the slab.
583 */
584#define BREAK_GFP_ORDER_HI 1
585#define BREAK_GFP_ORDER_LO 0
586static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
587
Andrew Mortona737b3e2006-03-22 00:08:11 -0800588/*
589 * Functions for storing/retrieving the cachep and or slab from the page
590 * allocator. These are used to find the slab an obj belongs to. With kfree(),
591 * these are used to find the cache which an obj belongs to.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700592 */
Pekka Enberg065d41c2005-11-13 16:06:46 -0800593static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
594{
595 page->lru.next = (struct list_head *)cache;
596}
597
598static inline struct kmem_cache *page_get_cache(struct page *page)
599{
Christoph Lameterd85f3382007-05-06 14:49:39 -0700600 page = compound_head(page);
Pekka Enbergddc2e812006-06-23 02:03:40 -0700601 BUG_ON(!PageSlab(page));
Pekka Enberg065d41c2005-11-13 16:06:46 -0800602 return (struct kmem_cache *)page->lru.next;
603}
604
605static inline void page_set_slab(struct page *page, struct slab *slab)
606{
607 page->lru.prev = (struct list_head *)slab;
608}
609
610static inline struct slab *page_get_slab(struct page *page)
611{
Pekka Enbergddc2e812006-06-23 02:03:40 -0700612 BUG_ON(!PageSlab(page));
Pekka Enberg065d41c2005-11-13 16:06:46 -0800613 return (struct slab *)page->lru.prev;
614}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700615
Pekka Enberg6ed5eb22006-02-01 03:05:49 -0800616static inline struct kmem_cache *virt_to_cache(const void *obj)
617{
Christoph Lameterb49af682007-05-06 14:49:41 -0700618 struct page *page = virt_to_head_page(obj);
Pekka Enberg6ed5eb22006-02-01 03:05:49 -0800619 return page_get_cache(page);
620}
621
622static inline struct slab *virt_to_slab(const void *obj)
623{
Christoph Lameterb49af682007-05-06 14:49:41 -0700624 struct page *page = virt_to_head_page(obj);
Pekka Enberg6ed5eb22006-02-01 03:05:49 -0800625 return page_get_slab(page);
626}
627
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800628static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
629 unsigned int idx)
630{
631 return slab->s_mem + cache->buffer_size * idx;
632}
633
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800634/*
635 * We want to avoid an expensive divide : (offset / cache->buffer_size)
636 * Using the fact that buffer_size is a constant for a particular cache,
637 * we can replace (offset / cache->buffer_size) by
638 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
639 */
640static inline unsigned int obj_to_index(const struct kmem_cache *cache,
641 const struct slab *slab, void *obj)
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800642{
Eric Dumazet6a2d7a92006-12-13 00:34:27 -0800643 u32 offset = (obj - slab->s_mem);
644 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800645}
646
Andrew Mortona737b3e2006-03-22 00:08:11 -0800647/*
648 * These are the default caches for kmalloc. Custom caches can have other sizes.
649 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700650struct cache_sizes malloc_sizes[] = {
651#define CACHE(x) { .cs_size = (x) },
652#include <linux/kmalloc_sizes.h>
653 CACHE(ULONG_MAX)
654#undef CACHE
655};
656EXPORT_SYMBOL(malloc_sizes);
657
658/* Must match cache_sizes above. Out of line to keep cache footprint low. */
659struct cache_names {
660 char *name;
661 char *name_dma;
662};
663
664static struct cache_names __initdata cache_names[] = {
665#define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
666#include <linux/kmalloc_sizes.h>
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800667 {NULL,}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700668#undef CACHE
669};
670
671static struct arraycache_init initarray_cache __initdata =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800672 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700673static struct arraycache_init initarray_generic =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800674 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700675
676/* internal cache of cache description objs */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800677static struct kmem_cache cache_cache = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800678 .batchcount = 1,
679 .limit = BOOT_CPUCACHE_ENTRIES,
680 .shared = 1,
Pekka Enberg343e0d72006-02-01 03:05:50 -0800681 .buffer_size = sizeof(struct kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800682 .name = "kmem_cache",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700683};
684
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700685#define BAD_ALIEN_MAGIC 0x01020304ul
686
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200687#ifdef CONFIG_LOCKDEP
688
689/*
690 * Slab sometimes uses the kmalloc slabs to store the slab headers
691 * for other slabs "off slab".
692 * The locking for this is tricky in that it nests within the locks
693 * of all other slabs in a few places; to deal with this special
694 * locking we put on-slab caches into a separate lock-class.
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700695 *
696 * We set lock class for alien array caches which are up during init.
697 * The lock annotation will be lost if all cpus of a node goes down and
698 * then comes back up during hotplug
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200699 */
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700700static struct lock_class_key on_slab_l3_key;
701static struct lock_class_key on_slab_alc_key;
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200702
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700703static inline void init_lock_keys(void)
704
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200705{
706 int q;
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700707 struct cache_sizes *s = malloc_sizes;
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200708
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700709 while (s->cs_size != ULONG_MAX) {
710 for_each_node(q) {
711 struct array_cache **alc;
712 int r;
713 struct kmem_list3 *l3 = s->cs_cachep->nodelists[q];
714 if (!l3 || OFF_SLAB(s->cs_cachep))
715 continue;
716 lockdep_set_class(&l3->list_lock, &on_slab_l3_key);
717 alc = l3->alien;
718 /*
719 * FIXME: This check for BAD_ALIEN_MAGIC
720 * should go away when common slab code is taught to
721 * work even without alien caches.
722 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
723 * for alloc_alien_cache,
724 */
725 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
726 continue;
727 for_each_node(r) {
728 if (alc[r])
729 lockdep_set_class(&alc[r]->lock,
730 &on_slab_alc_key);
731 }
732 }
733 s++;
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200734 }
735}
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200736#else
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -0700737static inline void init_lock_keys(void)
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200738{
739}
740#endif
741
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -0800742/*
Gautham R Shenoy95402b32008-01-25 21:08:02 +0100743 * Guard access to the cache-chain.
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -0800744 */
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800745static DEFINE_MUTEX(cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700746static struct list_head cache_chain;
747
748/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700749 * chicken and egg problem: delay the per-cpu array allocation
750 * until the general caches are up.
751 */
752static enum {
753 NONE,
Christoph Lametere498be72005-09-09 13:03:32 -0700754 PARTIAL_AC,
755 PARTIAL_L3,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700756 FULL
757} g_cpucache_up;
758
Mike Kravetz39d24e62006-05-15 09:44:13 -0700759/*
760 * used by boot code to determine if it can use slab based allocator
761 */
762int slab_is_available(void)
763{
764 return g_cpucache_up == FULL;
765}
766
David Howells52bad642006-11-22 14:54:01 +0000767static DEFINE_PER_CPU(struct delayed_work, reap_work);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700768
Pekka Enberg343e0d72006-02-01 03:05:50 -0800769static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700770{
771 return cachep->array[smp_processor_id()];
772}
773
Andrew Mortona737b3e2006-03-22 00:08:11 -0800774static inline struct kmem_cache *__find_general_cachep(size_t size,
775 gfp_t gfpflags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700776{
777 struct cache_sizes *csizep = malloc_sizes;
778
779#if DEBUG
780 /* This happens if someone tries to call
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800781 * kmem_cache_create(), or __kmalloc(), before
782 * the generic caches are initialized.
783 */
Alok Katariac7e43c72005-09-14 12:17:53 -0700784 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700785#endif
Christoph Lameter6cb8f912007-07-17 04:03:22 -0700786 if (!size)
787 return ZERO_SIZE_PTR;
788
Linus Torvalds1da177e2005-04-16 15:20:36 -0700789 while (size > csizep->cs_size)
790 csizep++;
791
792 /*
Martin Hicks0abf40c2005-09-03 15:54:54 -0700793 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
Linus Torvalds1da177e2005-04-16 15:20:36 -0700794 * has cs_{dma,}cachep==NULL. Thus no special case
795 * for large kmalloc calls required.
796 */
Christoph Lameter4b51d662007-02-10 01:43:10 -0800797#ifdef CONFIG_ZONE_DMA
Linus Torvalds1da177e2005-04-16 15:20:36 -0700798 if (unlikely(gfpflags & GFP_DMA))
799 return csizep->cs_dmacachep;
Christoph Lameter4b51d662007-02-10 01:43:10 -0800800#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700801 return csizep->cs_cachep;
802}
803
Adrian Bunkb2213852006-09-25 23:31:02 -0700804static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700805{
806 return __find_general_cachep(size, gfpflags);
807}
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700808
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800809static size_t slab_mgmt_size(size_t nr_objs, size_t align)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700810{
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800811 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
812}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700813
Andrew Mortona737b3e2006-03-22 00:08:11 -0800814/*
815 * Calculate the number of objects and left-over bytes for a given buffer size.
816 */
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800817static void cache_estimate(unsigned long gfporder, size_t buffer_size,
818 size_t align, int flags, size_t *left_over,
819 unsigned int *num)
820{
821 int nr_objs;
822 size_t mgmt_size;
823 size_t slab_size = PAGE_SIZE << gfporder;
824
825 /*
826 * The slab management structure can be either off the slab or
827 * on it. For the latter case, the memory allocated for a
828 * slab is used for:
829 *
830 * - The struct slab
831 * - One kmem_bufctl_t for each object
832 * - Padding to respect alignment of @align
833 * - @buffer_size bytes for each object
834 *
835 * If the slab management structure is off the slab, then the
836 * alignment will already be calculated into the size. Because
837 * the slabs are all pages aligned, the objects will be at the
838 * correct alignment when allocated.
839 */
840 if (flags & CFLGS_OFF_SLAB) {
841 mgmt_size = 0;
842 nr_objs = slab_size / buffer_size;
843
844 if (nr_objs > SLAB_LIMIT)
845 nr_objs = SLAB_LIMIT;
846 } else {
847 /*
848 * Ignore padding for the initial guess. The padding
849 * is at most @align-1 bytes, and @buffer_size is at
850 * least @align. In the worst case, this result will
851 * be one greater than the number of objects that fit
852 * into the memory allocation when taking the padding
853 * into account.
854 */
855 nr_objs = (slab_size - sizeof(struct slab)) /
856 (buffer_size + sizeof(kmem_bufctl_t));
857
858 /*
859 * This calculated number will be either the right
860 * amount, or one greater than what we want.
861 */
862 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
863 > slab_size)
864 nr_objs--;
865
866 if (nr_objs > SLAB_LIMIT)
867 nr_objs = SLAB_LIMIT;
868
869 mgmt_size = slab_mgmt_size(nr_objs, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700870 }
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800871 *num = nr_objs;
872 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700873}
874
Harvey Harrisond40cee22008-04-30 00:55:07 -0700875#define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700876
Andrew Mortona737b3e2006-03-22 00:08:11 -0800877static void __slab_error(const char *function, struct kmem_cache *cachep,
878 char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700879{
880 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800881 function, cachep->name, msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700882 dump_stack();
883}
884
Paul Menage3395ee02006-12-06 20:32:16 -0800885/*
886 * By default on NUMA we use alien caches to stage the freeing of
887 * objects allocated from other nodes. This causes massive memory
888 * inefficiencies when using fake NUMA setup to split memory into a
889 * large number of small nodes, so it can be disabled on the command
890 * line
891 */
892
893static int use_alien_caches __read_mostly = 1;
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -0700894static int numa_platform __read_mostly = 1;
Paul Menage3395ee02006-12-06 20:32:16 -0800895static int __init noaliencache_setup(char *s)
896{
897 use_alien_caches = 0;
898 return 1;
899}
900__setup("noaliencache", noaliencache_setup);
901
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800902#ifdef CONFIG_NUMA
903/*
904 * Special reaping functions for NUMA systems called from cache_reap().
905 * These take care of doing round robin flushing of alien caches (containing
906 * objects freed on different nodes from which they were allocated) and the
907 * flushing of remote pcps by calling drain_node_pages.
908 */
909static DEFINE_PER_CPU(unsigned long, reap_node);
910
911static void init_reap_node(int cpu)
912{
913 int node;
914
915 node = next_node(cpu_to_node(cpu), node_online_map);
916 if (node == MAX_NUMNODES)
Paul Jackson442295c2006-03-22 00:09:11 -0800917 node = first_node(node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800918
Daniel Yeisley7f6b8872006-11-02 22:07:14 -0800919 per_cpu(reap_node, cpu) = node;
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800920}
921
922static void next_reap_node(void)
923{
924 int node = __get_cpu_var(reap_node);
925
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800926 node = next_node(node, node_online_map);
927 if (unlikely(node >= MAX_NUMNODES))
928 node = first_node(node_online_map);
929 __get_cpu_var(reap_node) = node;
930}
931
932#else
933#define init_reap_node(cpu) do { } while (0)
934#define next_reap_node(void) do { } while (0)
935#endif
936
Linus Torvalds1da177e2005-04-16 15:20:36 -0700937/*
938 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
939 * via the workqueue/eventd.
940 * Add the CPU number into the expiration time to minimize the possibility of
941 * the CPUs getting into lockstep and contending for the global cache chain
942 * lock.
943 */
Adrian Bunk897e6792007-07-15 23:38:20 -0700944static void __cpuinit start_cpu_timer(int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700945{
David Howells52bad642006-11-22 14:54:01 +0000946 struct delayed_work *reap_work = &per_cpu(reap_work, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700947
948 /*
949 * When this gets called from do_initcalls via cpucache_init(),
950 * init_workqueues() has already run, so keventd will be setup
951 * at that time.
952 */
David Howells52bad642006-11-22 14:54:01 +0000953 if (keventd_up() && reap_work->work.func == NULL) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800954 init_reap_node(cpu);
David Howells65f27f32006-11-22 14:55:48 +0000955 INIT_DELAYED_WORK(reap_work, cache_reap);
Arjan van de Ven2b284212006-12-10 02:21:28 -0800956 schedule_delayed_work_on(cpu, reap_work,
957 __round_jiffies_relative(HZ, cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700958 }
959}
960
Christoph Lametere498be72005-09-09 13:03:32 -0700961static struct array_cache *alloc_arraycache(int node, int entries,
Pekka Enberg83b519e2009-06-10 19:40:04 +0300962 int batchcount, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700963{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800964 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700965 struct array_cache *nc = NULL;
966
Pekka Enberg83b519e2009-06-10 19:40:04 +0300967 nc = kmalloc_node(memsize, gfp, node);
Catalin Marinasd5cff632009-06-11 13:22:40 +0100968 /*
969 * The array_cache structures contain pointers to free object.
970 * However, when such objects are allocated or transfered to another
971 * cache the pointers are not cleared and they could be counted as
972 * valid references during a kmemleak scan. Therefore, kmemleak must
973 * not scan such objects.
974 */
975 kmemleak_no_scan(nc);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700976 if (nc) {
977 nc->avail = 0;
978 nc->limit = entries;
979 nc->batchcount = batchcount;
980 nc->touched = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700981 spin_lock_init(&nc->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700982 }
983 return nc;
984}
985
Christoph Lameter3ded1752006-03-25 03:06:44 -0800986/*
987 * Transfer objects in one arraycache to another.
988 * Locking must be handled by the caller.
989 *
990 * Return the number of entries transferred.
991 */
992static int transfer_objects(struct array_cache *to,
993 struct array_cache *from, unsigned int max)
994{
995 /* Figure out how many entries to transfer */
996 int nr = min(min(from->avail, max), to->limit - to->avail);
997
998 if (!nr)
999 return 0;
1000
1001 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
1002 sizeof(void *) *nr);
1003
1004 from->avail -= nr;
1005 to->avail += nr;
1006 to->touched = 1;
1007 return nr;
1008}
1009
Christoph Lameter765c4502006-09-27 01:50:08 -07001010#ifndef CONFIG_NUMA
1011
1012#define drain_alien_cache(cachep, alien) do { } while (0)
1013#define reap_alien(cachep, l3) do { } while (0)
1014
Pekka Enberg83b519e2009-06-10 19:40:04 +03001015static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
Christoph Lameter765c4502006-09-27 01:50:08 -07001016{
1017 return (struct array_cache **)BAD_ALIEN_MAGIC;
1018}
1019
1020static inline void free_alien_cache(struct array_cache **ac_ptr)
1021{
1022}
1023
1024static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1025{
1026 return 0;
1027}
1028
1029static inline void *alternate_node_alloc(struct kmem_cache *cachep,
1030 gfp_t flags)
1031{
1032 return NULL;
1033}
1034
Christoph Hellwig8b98c162006-12-06 20:32:30 -08001035static inline void *____cache_alloc_node(struct kmem_cache *cachep,
Christoph Lameter765c4502006-09-27 01:50:08 -07001036 gfp_t flags, int nodeid)
1037{
1038 return NULL;
1039}
1040
1041#else /* CONFIG_NUMA */
1042
Christoph Hellwig8b98c162006-12-06 20:32:30 -08001043static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
Paul Jacksonc61afb12006-03-24 03:16:08 -08001044static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
Christoph Lameterdc85da12006-01-18 17:42:36 -08001045
Pekka Enberg83b519e2009-06-10 19:40:04 +03001046static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -07001047{
1048 struct array_cache **ac_ptr;
Christoph Lameter8ef82862007-02-20 13:57:52 -08001049 int memsize = sizeof(void *) * nr_node_ids;
Christoph Lametere498be72005-09-09 13:03:32 -07001050 int i;
1051
1052 if (limit > 1)
1053 limit = 12;
Pekka Enberg83b519e2009-06-10 19:40:04 +03001054 ac_ptr = kmalloc_node(memsize, gfp, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001055 if (ac_ptr) {
1056 for_each_node(i) {
1057 if (i == node || !node_online(i)) {
1058 ac_ptr[i] = NULL;
1059 continue;
1060 }
Pekka Enberg83b519e2009-06-10 19:40:04 +03001061 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
Christoph Lametere498be72005-09-09 13:03:32 -07001062 if (!ac_ptr[i]) {
Akinobu Mitacc550de2007-11-14 16:58:35 -08001063 for (i--; i >= 0; i--)
Christoph Lametere498be72005-09-09 13:03:32 -07001064 kfree(ac_ptr[i]);
1065 kfree(ac_ptr);
1066 return NULL;
1067 }
1068 }
1069 }
1070 return ac_ptr;
1071}
1072
Pekka Enberg5295a742006-02-01 03:05:48 -08001073static void free_alien_cache(struct array_cache **ac_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -07001074{
1075 int i;
1076
1077 if (!ac_ptr)
1078 return;
Christoph Lametere498be72005-09-09 13:03:32 -07001079 for_each_node(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001080 kfree(ac_ptr[i]);
Christoph Lametere498be72005-09-09 13:03:32 -07001081 kfree(ac_ptr);
1082}
1083
Pekka Enberg343e0d72006-02-01 03:05:50 -08001084static void __drain_alien_cache(struct kmem_cache *cachep,
Pekka Enberg5295a742006-02-01 03:05:48 -08001085 struct array_cache *ac, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07001086{
1087 struct kmem_list3 *rl3 = cachep->nodelists[node];
1088
1089 if (ac->avail) {
1090 spin_lock(&rl3->list_lock);
Christoph Lametere00946f2006-03-25 03:06:45 -08001091 /*
1092 * Stuff objects into the remote nodes shared array first.
1093 * That way we could avoid the overhead of putting the objects
1094 * into the free lists and getting them back later.
1095 */
shin, jacob693f7d32006-04-28 10:54:37 -05001096 if (rl3->shared)
1097 transfer_objects(rl3->shared, ac, ac->limit);
Christoph Lametere00946f2006-03-25 03:06:45 -08001098
Christoph Lameterff694162005-09-22 21:44:02 -07001099 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001100 ac->avail = 0;
1101 spin_unlock(&rl3->list_lock);
1102 }
1103}
1104
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001105/*
1106 * Called from cache_reap() to regularly drain alien caches round robin.
1107 */
1108static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1109{
1110 int node = __get_cpu_var(reap_node);
1111
1112 if (l3->alien) {
1113 struct array_cache *ac = l3->alien[node];
Christoph Lametere00946f2006-03-25 03:06:45 -08001114
1115 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001116 __drain_alien_cache(cachep, ac, node);
1117 spin_unlock_irq(&ac->lock);
1118 }
1119 }
1120}
1121
Andrew Mortona737b3e2006-03-22 00:08:11 -08001122static void drain_alien_cache(struct kmem_cache *cachep,
1123 struct array_cache **alien)
Christoph Lametere498be72005-09-09 13:03:32 -07001124{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001125 int i = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07001126 struct array_cache *ac;
1127 unsigned long flags;
1128
1129 for_each_online_node(i) {
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001130 ac = alien[i];
Christoph Lametere498be72005-09-09 13:03:32 -07001131 if (ac) {
1132 spin_lock_irqsave(&ac->lock, flags);
1133 __drain_alien_cache(cachep, ac, i);
1134 spin_unlock_irqrestore(&ac->lock, flags);
1135 }
1136 }
1137}
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001138
Ingo Molnar873623d2006-07-13 14:44:38 +02001139static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001140{
1141 struct slab *slabp = virt_to_slab(objp);
1142 int nodeid = slabp->nodeid;
1143 struct kmem_list3 *l3;
1144 struct array_cache *alien = NULL;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001145 int node;
1146
1147 node = numa_node_id();
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001148
1149 /*
1150 * Make sure we are not freeing a object from another node to the array
1151 * cache on this cpu.
1152 */
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001153 if (likely(slabp->nodeid == node))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001154 return 0;
1155
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001156 l3 = cachep->nodelists[node];
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001157 STATS_INC_NODEFREES(cachep);
1158 if (l3->alien && l3->alien[nodeid]) {
1159 alien = l3->alien[nodeid];
Ingo Molnar873623d2006-07-13 14:44:38 +02001160 spin_lock(&alien->lock);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001161 if (unlikely(alien->avail == alien->limit)) {
1162 STATS_INC_ACOVERFLOW(cachep);
1163 __drain_alien_cache(cachep, alien, nodeid);
1164 }
1165 alien->entry[alien->avail++] = objp;
1166 spin_unlock(&alien->lock);
1167 } else {
1168 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1169 free_block(cachep, &objp, 1, nodeid);
1170 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1171 }
1172 return 1;
1173}
Christoph Lametere498be72005-09-09 13:03:32 -07001174#endif
1175
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001176static void __cpuinit cpuup_canceled(long cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001177{
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001178 struct kmem_cache *cachep;
1179 struct kmem_list3 *l3 = NULL;
1180 int node = cpu_to_node(cpu);
Rusty Russella70f7302009-03-13 14:49:46 +10301181 const struct cpumask *mask = cpumask_of_node(node);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001182
1183 list_for_each_entry(cachep, &cache_chain, next) {
1184 struct array_cache *nc;
1185 struct array_cache *shared;
1186 struct array_cache **alien;
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001187
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001188 /* cpu is dead; no one can alloc from it. */
1189 nc = cachep->array[cpu];
1190 cachep->array[cpu] = NULL;
1191 l3 = cachep->nodelists[node];
1192
1193 if (!l3)
1194 goto free_array_cache;
1195
1196 spin_lock_irq(&l3->list_lock);
1197
1198 /* Free limit for this kmem_list3 */
1199 l3->free_limit -= cachep->batchcount;
1200 if (nc)
1201 free_block(cachep, nc->entry, nc->avail, node);
1202
Mike Travisc5f59f02008-04-04 18:11:10 -07001203 if (!cpus_empty(*mask)) {
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001204 spin_unlock_irq(&l3->list_lock);
1205 goto free_array_cache;
1206 }
1207
1208 shared = l3->shared;
1209 if (shared) {
1210 free_block(cachep, shared->entry,
1211 shared->avail, node);
1212 l3->shared = NULL;
1213 }
1214
1215 alien = l3->alien;
1216 l3->alien = NULL;
1217
1218 spin_unlock_irq(&l3->list_lock);
1219
1220 kfree(shared);
1221 if (alien) {
1222 drain_alien_cache(cachep, alien);
1223 free_alien_cache(alien);
1224 }
1225free_array_cache:
1226 kfree(nc);
1227 }
1228 /*
1229 * In the previous loop, all the objects were freed to
1230 * the respective cache's slabs, now we can go ahead and
1231 * shrink each nodelist to its limit.
1232 */
1233 list_for_each_entry(cachep, &cache_chain, next) {
1234 l3 = cachep->nodelists[node];
1235 if (!l3)
1236 continue;
1237 drain_freelist(cachep, l3, l3->free_objects);
1238 }
1239}
1240
1241static int __cpuinit cpuup_prepare(long cpu)
1242{
Pekka Enberg343e0d72006-02-01 03:05:50 -08001243 struct kmem_cache *cachep;
Christoph Lametere498be72005-09-09 13:03:32 -07001244 struct kmem_list3 *l3 = NULL;
1245 int node = cpu_to_node(cpu);
David Howellsea02e3d2007-07-19 01:49:09 -07001246 const int memsize = sizeof(struct kmem_list3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001247
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001248 /*
1249 * We need to do this right in the beginning since
1250 * alloc_arraycache's are going to use this list.
1251 * kmalloc_node allows us to add the slab to the right
1252 * kmem_list3 and not this cpu's kmem_list3
1253 */
1254
1255 list_for_each_entry(cachep, &cache_chain, next) {
1256 /*
1257 * Set up the size64 kmemlist for cpu before we can
1258 * begin anything. Make sure some other cpu on this
1259 * node has not already allocated this
1260 */
1261 if (!cachep->nodelists[node]) {
1262 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1263 if (!l3)
1264 goto bad;
1265 kmem_list3_init(l3);
1266 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1267 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1268
1269 /*
1270 * The l3s don't come and go as CPUs come and
1271 * go. cache_chain_mutex is sufficient
1272 * protection here.
1273 */
1274 cachep->nodelists[node] = l3;
1275 }
1276
1277 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1278 cachep->nodelists[node]->free_limit =
1279 (1 + nr_cpus_node(node)) *
1280 cachep->batchcount + cachep->num;
1281 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1282 }
1283
1284 /*
1285 * Now we can go ahead with allocating the shared arrays and
1286 * array caches
1287 */
1288 list_for_each_entry(cachep, &cache_chain, next) {
1289 struct array_cache *nc;
1290 struct array_cache *shared = NULL;
1291 struct array_cache **alien = NULL;
1292
1293 nc = alloc_arraycache(node, cachep->limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03001294 cachep->batchcount, GFP_KERNEL);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001295 if (!nc)
1296 goto bad;
1297 if (cachep->shared) {
1298 shared = alloc_arraycache(node,
1299 cachep->shared * cachep->batchcount,
Pekka Enberg83b519e2009-06-10 19:40:04 +03001300 0xbaadf00d, GFP_KERNEL);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001301 if (!shared) {
1302 kfree(nc);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001303 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001304 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001305 }
1306 if (use_alien_caches) {
Pekka Enberg83b519e2009-06-10 19:40:04 +03001307 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
Akinobu Mita12d00f62007-10-18 03:05:11 -07001308 if (!alien) {
1309 kfree(shared);
1310 kfree(nc);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001311 goto bad;
Akinobu Mita12d00f62007-10-18 03:05:11 -07001312 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001313 }
1314 cachep->array[cpu] = nc;
1315 l3 = cachep->nodelists[node];
1316 BUG_ON(!l3);
1317
1318 spin_lock_irq(&l3->list_lock);
1319 if (!l3->shared) {
1320 /*
1321 * We are serialised from CPU_DEAD or
1322 * CPU_UP_CANCELLED by the cpucontrol lock
1323 */
1324 l3->shared = shared;
1325 shared = NULL;
1326 }
1327#ifdef CONFIG_NUMA
1328 if (!l3->alien) {
1329 l3->alien = alien;
1330 alien = NULL;
1331 }
1332#endif
1333 spin_unlock_irq(&l3->list_lock);
1334 kfree(shared);
1335 free_alien_cache(alien);
1336 }
1337 return 0;
1338bad:
Akinobu Mita12d00f62007-10-18 03:05:11 -07001339 cpuup_canceled(cpu);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001340 return -ENOMEM;
1341}
1342
1343static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1344 unsigned long action, void *hcpu)
1345{
1346 long cpu = (long)hcpu;
1347 int err = 0;
1348
Linus Torvalds1da177e2005-04-16 15:20:36 -07001349 switch (action) {
Heiko Carstens38c3bd92007-05-09 02:34:05 -07001350 case CPU_UP_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001351 case CPU_UP_PREPARE_FROZEN:
Gautham R Shenoy95402b32008-01-25 21:08:02 +01001352 mutex_lock(&cache_chain_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001353 err = cpuup_prepare(cpu);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01001354 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001355 break;
1356 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001357 case CPU_ONLINE_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001358 start_cpu_timer(cpu);
1359 break;
1360#ifdef CONFIG_HOTPLUG_CPU
Christoph Lameter5830c592007-05-09 02:34:22 -07001361 case CPU_DOWN_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001362 case CPU_DOWN_PREPARE_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001363 /*
1364 * Shutdown cache reaper. Note that the cache_chain_mutex is
1365 * held so that if cache_reap() is invoked it cannot do
1366 * anything expensive but will only modify reap_work
1367 * and reschedule the timer.
1368 */
1369 cancel_rearming_delayed_work(&per_cpu(reap_work, cpu));
1370 /* Now the cache_reaper is guaranteed to be not running. */
1371 per_cpu(reap_work, cpu).work.func = NULL;
1372 break;
1373 case CPU_DOWN_FAILED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001374 case CPU_DOWN_FAILED_FROZEN:
Christoph Lameter5830c592007-05-09 02:34:22 -07001375 start_cpu_timer(cpu);
1376 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001377 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001378 case CPU_DEAD_FROZEN:
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001379 /*
1380 * Even if all the cpus of a node are down, we don't free the
1381 * kmem_list3 of any cache. This to avoid a race between
1382 * cpu_down, and a kmalloc allocation from another cpu for
1383 * memory from the node of the cpu going down. The list3
1384 * structure is usually allocated from kmem_cache_create() and
1385 * gets destroyed at kmem_cache_destroy().
1386 */
Simon Arlott183ff222007-10-20 01:27:18 +02001387 /* fall through */
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08001388#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001389 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001390 case CPU_UP_CANCELED_FROZEN:
Gautham R Shenoy95402b32008-01-25 21:08:02 +01001391 mutex_lock(&cache_chain_mutex);
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001392 cpuup_canceled(cpu);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001393 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001394 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001395 }
Akinobu Mitafbf1e472007-10-18 03:05:09 -07001396 return err ? NOTIFY_BAD : NOTIFY_OK;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001397}
1398
Chandra Seetharaman74b85f32006-06-27 02:54:09 -07001399static struct notifier_block __cpuinitdata cpucache_notifier = {
1400 &cpuup_callback, NULL, 0
1401};
Linus Torvalds1da177e2005-04-16 15:20:36 -07001402
Christoph Lametere498be72005-09-09 13:03:32 -07001403/*
1404 * swap the static kmem_list3 with kmalloced memory
1405 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001406static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1407 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001408{
1409 struct kmem_list3 *ptr;
1410
Pekka Enberg83b519e2009-06-10 19:40:04 +03001411 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07001412 BUG_ON(!ptr);
1413
Christoph Lametere498be72005-09-09 13:03:32 -07001414 memcpy(ptr, list, sizeof(struct kmem_list3));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001415 /*
1416 * Do not assume that spinlocks can be initialized via memcpy:
1417 */
1418 spin_lock_init(&ptr->list_lock);
1419
Christoph Lametere498be72005-09-09 13:03:32 -07001420 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1421 cachep->nodelists[nodeid] = ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001422}
1423
Andrew Mortona737b3e2006-03-22 00:08:11 -08001424/*
Pekka Enberg556a1692008-01-25 08:20:51 +02001425 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1426 * size of kmem_list3.
1427 */
1428static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1429{
1430 int node;
1431
1432 for_each_online_node(node) {
1433 cachep->nodelists[node] = &initkmem_list3[index + node];
1434 cachep->nodelists[node]->next_reap = jiffies +
1435 REAPTIMEOUT_LIST3 +
1436 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1437 }
1438}
1439
1440/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08001441 * Initialisation. Called after the page allocator have been initialised and
1442 * before smp_init().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001443 */
1444void __init kmem_cache_init(void)
1445{
1446 size_t left_over;
1447 struct cache_sizes *sizes;
1448 struct cache_names *names;
Christoph Lametere498be72005-09-09 13:03:32 -07001449 int i;
Jack Steiner07ed76b2006-03-07 21:55:46 -08001450 int order;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001451 int node;
Christoph Lametere498be72005-09-09 13:03:32 -07001452
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -07001453 if (num_possible_nodes() == 1) {
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001454 use_alien_caches = 0;
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -07001455 numa_platform = 0;
1456 }
Siddha, Suresh B62918a02007-05-02 19:27:18 +02001457
Christoph Lametere498be72005-09-09 13:03:32 -07001458 for (i = 0; i < NUM_INIT_LISTS; i++) {
1459 kmem_list3_init(&initkmem_list3[i]);
1460 if (i < MAX_NUMNODES)
1461 cache_cache.nodelists[i] = NULL;
1462 }
Pekka Enberg556a1692008-01-25 08:20:51 +02001463 set_up_list3s(&cache_cache, CACHE_CACHE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001464
1465 /*
1466 * Fragmentation resistance on low memory - only use bigger
1467 * page orders on machines with more than 32MB of memory.
1468 */
1469 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1470 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1471
Linus Torvalds1da177e2005-04-16 15:20:36 -07001472 /* Bootstrap is tricky, because several objects are allocated
1473 * from caches that do not exist yet:
Andrew Mortona737b3e2006-03-22 00:08:11 -08001474 * 1) initialize the cache_cache cache: it contains the struct
1475 * kmem_cache structures of all caches, except cache_cache itself:
1476 * cache_cache is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001477 * Initially an __init data area is used for the head array and the
1478 * kmem_list3 structures, it's replaced with a kmalloc allocated
1479 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001480 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001481 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001482 * An __init data area is used for the head array.
1483 * 3) Create the remaining kmalloc caches, with minimally sized
1484 * head arrays.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001485 * 4) Replace the __init data head arrays for cache_cache and the first
1486 * kmalloc cache with kmalloc allocated arrays.
Christoph Lametere498be72005-09-09 13:03:32 -07001487 * 5) Replace the __init data for kmem_list3 for cache_cache and
1488 * the other cache's with kmalloc allocated memory.
1489 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001490 */
1491
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001492 node = numa_node_id();
1493
Linus Torvalds1da177e2005-04-16 15:20:36 -07001494 /* 1) create the cache_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001495 INIT_LIST_HEAD(&cache_chain);
1496 list_add(&cache_cache.next, &cache_chain);
1497 cache_cache.colour_off = cache_line_size();
1498 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
Daniel Yeisleyec1f5ee2008-03-25 23:59:08 +02001499 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001500
Eric Dumazet8da34302007-05-06 14:49:29 -07001501 /*
1502 * struct kmem_cache size depends on nr_node_ids, which
1503 * can be less than MAX_NUMNODES.
1504 */
1505 cache_cache.buffer_size = offsetof(struct kmem_cache, nodelists) +
1506 nr_node_ids * sizeof(struct kmem_list3 *);
1507#if DEBUG
1508 cache_cache.obj_size = cache_cache.buffer_size;
1509#endif
Andrew Mortona737b3e2006-03-22 00:08:11 -08001510 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1511 cache_line_size());
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08001512 cache_cache.reciprocal_buffer_size =
1513 reciprocal_value(cache_cache.buffer_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001514
Jack Steiner07ed76b2006-03-07 21:55:46 -08001515 for (order = 0; order < MAX_ORDER; order++) {
1516 cache_estimate(order, cache_cache.buffer_size,
1517 cache_line_size(), 0, &left_over, &cache_cache.num);
1518 if (cache_cache.num)
1519 break;
1520 }
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02001521 BUG_ON(!cache_cache.num);
Jack Steiner07ed76b2006-03-07 21:55:46 -08001522 cache_cache.gfporder = order;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001523 cache_cache.colour = left_over / cache_cache.colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001524 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1525 sizeof(struct slab), cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001526
1527 /* 2+3) create the kmalloc caches */
1528 sizes = malloc_sizes;
1529 names = cache_names;
1530
Andrew Mortona737b3e2006-03-22 00:08:11 -08001531 /*
1532 * Initialize the caches that provide memory for the array cache and the
1533 * kmem_list3 structures first. Without this, further allocations will
1534 * bug.
Christoph Lametere498be72005-09-09 13:03:32 -07001535 */
1536
1537 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001538 sizes[INDEX_AC].cs_size,
1539 ARCH_KMALLOC_MINALIGN,
1540 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001541 NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07001542
Andrew Mortona737b3e2006-03-22 00:08:11 -08001543 if (INDEX_AC != INDEX_L3) {
Christoph Lametere498be72005-09-09 13:03:32 -07001544 sizes[INDEX_L3].cs_cachep =
Andrew Mortona737b3e2006-03-22 00:08:11 -08001545 kmem_cache_create(names[INDEX_L3].name,
1546 sizes[INDEX_L3].cs_size,
1547 ARCH_KMALLOC_MINALIGN,
1548 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001549 NULL);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001550 }
Christoph Lametere498be72005-09-09 13:03:32 -07001551
Ingo Molnare0a42722006-06-23 02:03:46 -07001552 slab_early_init = 0;
1553
Linus Torvalds1da177e2005-04-16 15:20:36 -07001554 while (sizes->cs_size != ULONG_MAX) {
Christoph Lametere498be72005-09-09 13:03:32 -07001555 /*
1556 * For performance, all the general caches are L1 aligned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001557 * This should be particularly beneficial on SMP boxes, as it
1558 * eliminates "false sharing".
1559 * Note for systems short on memory removing the alignment will
Christoph Lametere498be72005-09-09 13:03:32 -07001560 * allow tighter packing of the smaller caches.
1561 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001562 if (!sizes->cs_cachep) {
Christoph Lametere498be72005-09-09 13:03:32 -07001563 sizes->cs_cachep = kmem_cache_create(names->name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001564 sizes->cs_size,
1565 ARCH_KMALLOC_MINALIGN,
1566 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001567 NULL);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001568 }
Christoph Lameter4b51d662007-02-10 01:43:10 -08001569#ifdef CONFIG_ZONE_DMA
1570 sizes->cs_dmacachep = kmem_cache_create(
1571 names->name_dma,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001572 sizes->cs_size,
1573 ARCH_KMALLOC_MINALIGN,
1574 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1575 SLAB_PANIC,
Paul Mundt20c2df82007-07-20 10:11:58 +09001576 NULL);
Christoph Lameter4b51d662007-02-10 01:43:10 -08001577#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07001578 sizes++;
1579 names++;
1580 }
1581 /* 4) Replace the bootstrap head arrays */
1582 {
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001583 struct array_cache *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001584
Pekka Enberg83b519e2009-06-10 19:40:04 +03001585 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
Christoph Lametere498be72005-09-09 13:03:32 -07001586
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001587 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1588 memcpy(ptr, cpu_cache_get(&cache_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001589 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001590 /*
1591 * Do not assume that spinlocks can be initialized via memcpy:
1592 */
1593 spin_lock_init(&ptr->lock);
1594
Linus Torvalds1da177e2005-04-16 15:20:36 -07001595 cache_cache.array[smp_processor_id()] = ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001596
Pekka Enberg83b519e2009-06-10 19:40:04 +03001597 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
Christoph Lametere498be72005-09-09 13:03:32 -07001598
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001599 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001600 != &initarray_generic.cache);
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001601 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001602 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001603 /*
1604 * Do not assume that spinlocks can be initialized via memcpy:
1605 */
1606 spin_lock_init(&ptr->lock);
1607
Christoph Lametere498be72005-09-09 13:03:32 -07001608 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001609 ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001610 }
Christoph Lametere498be72005-09-09 13:03:32 -07001611 /* 5) Replace the bootstrap kmem_list3's */
1612 {
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001613 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001614
Mel Gorman9c09a952008-01-24 05:49:54 -08001615 for_each_online_node(nid) {
Daniel Yeisleyec1f5ee2008-03-25 23:59:08 +02001616 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid);
Pekka Enberg556a1692008-01-25 08:20:51 +02001617
Christoph Lametere498be72005-09-09 13:03:32 -07001618 init_list(malloc_sizes[INDEX_AC].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001619 &initkmem_list3[SIZE_AC + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001620
1621 if (INDEX_AC != INDEX_L3) {
1622 init_list(malloc_sizes[INDEX_L3].cs_cachep,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07001623 &initkmem_list3[SIZE_L3 + nid], nid);
Christoph Lametere498be72005-09-09 13:03:32 -07001624 }
1625 }
1626 }
1627
1628 /* 6) resize the head arrays to their final sizes */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001629 {
Pekka Enberg343e0d72006-02-01 03:05:50 -08001630 struct kmem_cache *cachep;
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001631 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001632 list_for_each_entry(cachep, &cache_chain, next)
Pekka Enberg83b519e2009-06-10 19:40:04 +03001633 if (enable_cpucache(cachep, GFP_NOWAIT))
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07001634 BUG();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001635 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001636 }
1637
Ravikiran G Thirumalai056c6242006-09-25 23:31:38 -07001638 /* Annotate slab for lockdep -- annotate the malloc caches */
1639 init_lock_keys();
1640
1641
Linus Torvalds1da177e2005-04-16 15:20:36 -07001642 /* Done! */
1643 g_cpucache_up = FULL;
1644
Andrew Mortona737b3e2006-03-22 00:08:11 -08001645 /*
1646 * Register a cpu startup notifier callback that initializes
1647 * cpu_cache_get for all new cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07001648 */
1649 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001650
Andrew Mortona737b3e2006-03-22 00:08:11 -08001651 /*
1652 * The reap timers are started later, with a module init call: That part
1653 * of the kernel is not yet operational.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001654 */
1655}
1656
1657static int __init cpucache_init(void)
1658{
1659 int cpu;
1660
Andrew Mortona737b3e2006-03-22 00:08:11 -08001661 /*
1662 * Register the timers that return unneeded pages to the page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -07001663 */
Christoph Lametere498be72005-09-09 13:03:32 -07001664 for_each_online_cpu(cpu)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001665 start_cpu_timer(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001666 return 0;
1667}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001668__initcall(cpucache_init);
1669
1670/*
1671 * Interface to system's page allocator. No need to hold the cache-lock.
1672 *
1673 * If we requested dmaable memory, we will get it. Even if we
1674 * did not request dmaable memory, we might get it, but that
1675 * would be relatively rare and ignorable.
1676 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001677static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001678{
1679 struct page *page;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001680 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001681 int i;
1682
Luke Yangd6fef9d2006-04-10 22:52:56 -07001683#ifndef CONFIG_MMU
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001684 /*
1685 * Nommu uses slab's for process anonymous memory allocations, and thus
1686 * requires __GFP_COMP to properly refcount higher order allocations
Luke Yangd6fef9d2006-04-10 22:52:56 -07001687 */
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001688 flags |= __GFP_COMP;
Luke Yangd6fef9d2006-04-10 22:52:56 -07001689#endif
Christoph Lameter765c4502006-09-27 01:50:08 -07001690
Christoph Lameter3c517a62006-12-06 20:33:29 -08001691 flags |= cachep->gfpflags;
Mel Gormane12ba742007-10-16 01:25:52 -07001692 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1693 flags |= __GFP_RECLAIMABLE;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001694
1695 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001696 if (!page)
1697 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001698
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001699 nr_pages = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001700 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
Christoph Lameter972d1a72006-09-25 23:31:51 -07001701 add_zone_page_state(page_zone(page),
1702 NR_SLAB_RECLAIMABLE, nr_pages);
1703 else
1704 add_zone_page_state(page_zone(page),
1705 NR_SLAB_UNRECLAIMABLE, nr_pages);
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001706 for (i = 0; i < nr_pages; i++)
1707 __SetPageSlab(page + i);
1708 return page_address(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001709}
1710
1711/*
1712 * Interface to system's page release.
1713 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001714static void kmem_freepages(struct kmem_cache *cachep, void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001715{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001716 unsigned long i = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001717 struct page *page = virt_to_page(addr);
1718 const unsigned long nr_freed = i;
1719
Christoph Lameter972d1a72006-09-25 23:31:51 -07001720 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1721 sub_zone_page_state(page_zone(page),
1722 NR_SLAB_RECLAIMABLE, nr_freed);
1723 else
1724 sub_zone_page_state(page_zone(page),
1725 NR_SLAB_UNRECLAIMABLE, nr_freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001726 while (i--) {
Nick Pigginf205b2f2006-03-22 00:08:02 -08001727 BUG_ON(!PageSlab(page));
1728 __ClearPageSlab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001729 page++;
1730 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001731 if (current->reclaim_state)
1732 current->reclaim_state->reclaimed_slab += nr_freed;
1733 free_pages((unsigned long)addr, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001734}
1735
1736static void kmem_rcu_free(struct rcu_head *head)
1737{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001738 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001739 struct kmem_cache *cachep = slab_rcu->cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001740
1741 kmem_freepages(cachep, slab_rcu->addr);
1742 if (OFF_SLAB(cachep))
1743 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1744}
1745
1746#if DEBUG
1747
1748#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001749static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001750 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001751{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001752 int size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001753
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001754 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001755
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001756 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001757 return;
1758
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001759 *addr++ = 0x12345678;
1760 *addr++ = caller;
1761 *addr++ = smp_processor_id();
1762 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001763 {
1764 unsigned long *sptr = &caller;
1765 unsigned long svalue;
1766
1767 while (!kstack_end(sptr)) {
1768 svalue = *sptr++;
1769 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001770 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001771 size -= sizeof(unsigned long);
1772 if (size <= sizeof(unsigned long))
1773 break;
1774 }
1775 }
1776
1777 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001778 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001779}
1780#endif
1781
Pekka Enberg343e0d72006-02-01 03:05:50 -08001782static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001783{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001784 int size = obj_size(cachep);
1785 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001786
1787 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001788 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001789}
1790
1791static void dump_line(char *data, int offset, int limit)
1792{
1793 int i;
Dave Jonesaa83aa42006-09-29 01:59:51 -07001794 unsigned char error = 0;
1795 int bad_count = 0;
1796
Linus Torvalds1da177e2005-04-16 15:20:36 -07001797 printk(KERN_ERR "%03x:", offset);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001798 for (i = 0; i < limit; i++) {
1799 if (data[offset + i] != POISON_FREE) {
1800 error = data[offset + i];
1801 bad_count++;
1802 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001803 printk(" %02x", (unsigned char)data[offset + i]);
Dave Jonesaa83aa42006-09-29 01:59:51 -07001804 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001805 printk("\n");
Dave Jonesaa83aa42006-09-29 01:59:51 -07001806
1807 if (bad_count == 1) {
1808 error ^= POISON_FREE;
1809 if (!(error & (error - 1))) {
1810 printk(KERN_ERR "Single bit error detected. Probably "
1811 "bad RAM.\n");
1812#ifdef CONFIG_X86
1813 printk(KERN_ERR "Run memtest86+ or a similar memory "
1814 "test tool.\n");
1815#else
1816 printk(KERN_ERR "Run a memory test tool.\n");
1817#endif
1818 }
1819 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001820}
1821#endif
1822
1823#if DEBUG
1824
Pekka Enberg343e0d72006-02-01 03:05:50 -08001825static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001826{
1827 int i, size;
1828 char *realobj;
1829
1830 if (cachep->flags & SLAB_RED_ZONE) {
David Woodhouseb46b8f12007-05-08 00:22:59 -07001831 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001832 *dbg_redzone1(cachep, objp),
1833 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001834 }
1835
1836 if (cachep->flags & SLAB_STORE_USER) {
1837 printk(KERN_ERR "Last user: [<%p>]",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001838 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001839 print_symbol("(%s)",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001840 (unsigned long)*dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001841 printk("\n");
1842 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001843 realobj = (char *)objp + obj_offset(cachep);
1844 size = obj_size(cachep);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001845 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001846 int limit;
1847 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001848 if (i + limit > size)
1849 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001850 dump_line(realobj, i, limit);
1851 }
1852}
1853
Pekka Enberg343e0d72006-02-01 03:05:50 -08001854static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001855{
1856 char *realobj;
1857 int size, i;
1858 int lines = 0;
1859
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001860 realobj = (char *)objp + obj_offset(cachep);
1861 size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001862
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001863 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001864 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001865 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001866 exp = POISON_END;
1867 if (realobj[i] != exp) {
1868 int limit;
1869 /* Mismatch ! */
1870 /* Print header */
1871 if (lines == 0) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001872 printk(KERN_ERR
David Howellse94a40c2007-04-02 23:46:28 +01001873 "Slab corruption: %s start=%p, len=%d\n",
1874 cachep->name, realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001875 print_objinfo(cachep, objp, 0);
1876 }
1877 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001878 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001879 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001880 if (i + limit > size)
1881 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001882 dump_line(realobj, i, limit);
1883 i += 16;
1884 lines++;
1885 /* Limit to 5 lines */
1886 if (lines > 5)
1887 break;
1888 }
1889 }
1890 if (lines != 0) {
1891 /* Print some data about the neighboring objects, if they
1892 * exist:
1893 */
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08001894 struct slab *slabp = virt_to_slab(objp);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001895 unsigned int objnr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001896
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001897 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001898 if (objnr) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001899 objp = index_to_obj(cachep, slabp, objnr - 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001900 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001901 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001902 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001903 print_objinfo(cachep, objp, 2);
1904 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001905 if (objnr + 1 < cachep->num) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001906 objp = index_to_obj(cachep, slabp, objnr + 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001907 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001908 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001909 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001910 print_objinfo(cachep, objp, 2);
1911 }
1912 }
1913}
1914#endif
1915
Linus Torvalds1da177e2005-04-16 15:20:36 -07001916#if DEBUG
Rabin Vincente79aec22008-07-04 00:40:32 +05301917static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001918{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001919 int i;
1920 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001921 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001922
1923 if (cachep->flags & SLAB_POISON) {
1924#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08001925 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1926 OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001927 kernel_map_pages(virt_to_page(objp),
Andrew Mortona737b3e2006-03-22 00:08:11 -08001928 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001929 else
1930 check_poison_obj(cachep, objp);
1931#else
1932 check_poison_obj(cachep, objp);
1933#endif
1934 }
1935 if (cachep->flags & SLAB_RED_ZONE) {
1936 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1937 slab_error(cachep, "start of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001938 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001939 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1940 slab_error(cachep, "end of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001941 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001942 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001943 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001944}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001945#else
Rabin Vincente79aec22008-07-04 00:40:32 +05301946static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001947{
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001948}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001949#endif
1950
Randy Dunlap911851e2006-03-22 00:08:14 -08001951/**
1952 * slab_destroy - destroy and release all objects in a slab
1953 * @cachep: cache pointer being destroyed
1954 * @slabp: slab pointer being destroyed
1955 *
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001956 * Destroy all the objs in a slab, and release the mem back to the system.
Andrew Mortona737b3e2006-03-22 00:08:11 -08001957 * Before calling the slab must have been unlinked from the cache. The
1958 * cache-lock is not held/needed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001959 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001960static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001961{
1962 void *addr = slabp->s_mem - slabp->colouroff;
1963
Rabin Vincente79aec22008-07-04 00:40:32 +05301964 slab_destroy_debugcheck(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001965 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1966 struct slab_rcu *slab_rcu;
1967
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001968 slab_rcu = (struct slab_rcu *)slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001969 slab_rcu->cachep = cachep;
1970 slab_rcu->addr = addr;
1971 call_rcu(&slab_rcu->head, kmem_rcu_free);
1972 } else {
1973 kmem_freepages(cachep, addr);
Ingo Molnar873623d2006-07-13 14:44:38 +02001974 if (OFF_SLAB(cachep))
1975 kmem_cache_free(cachep->slabp_cache, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001976 }
1977}
1978
Christoph Lameter117f6eb2006-09-25 23:31:37 -07001979static void __kmem_cache_destroy(struct kmem_cache *cachep)
1980{
1981 int i;
1982 struct kmem_list3 *l3;
1983
1984 for_each_online_cpu(i)
1985 kfree(cachep->array[i]);
1986
1987 /* NUMA: free the list3 structures */
1988 for_each_online_node(i) {
1989 l3 = cachep->nodelists[i];
1990 if (l3) {
1991 kfree(l3->shared);
1992 free_alien_cache(l3->alien);
1993 kfree(l3);
1994 }
1995 }
1996 kmem_cache_free(&cache_cache, cachep);
1997}
1998
1999
Linus Torvalds1da177e2005-04-16 15:20:36 -07002000/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08002001 * calculate_slab_order - calculate size (page order) of slabs
2002 * @cachep: pointer to the cache that is being created
2003 * @size: size of objects to be created in this cache.
2004 * @align: required alignment for the objects.
2005 * @flags: slab allocation flags
2006 *
2007 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002008 *
2009 * This could be made much more intelligent. For now, try to avoid using
2010 * high order pages for slabs. When the gfp() functions are more friendly
2011 * towards high-order requests, this should be changed.
2012 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002013static size_t calculate_slab_order(struct kmem_cache *cachep,
Randy Dunlapee13d782006-02-01 03:05:53 -08002014 size_t size, size_t align, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002015{
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002016 unsigned long offslab_limit;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002017 size_t left_over = 0;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002018 int gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002019
Christoph Lameter0aa817f2007-05-16 22:11:01 -07002020 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002021 unsigned int num;
2022 size_t remainder;
2023
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002024 cache_estimate(gfporder, size, align, flags, &remainder, &num);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002025 if (!num)
2026 continue;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002027
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02002028 if (flags & CFLGS_OFF_SLAB) {
2029 /*
2030 * Max number of objs-per-slab for caches which
2031 * use off-slab slabs. Needed to avoid a possible
2032 * looping condition in cache_grow().
2033 */
2034 offslab_limit = size - sizeof(struct slab);
2035 offslab_limit /= sizeof(kmem_bufctl_t);
2036
2037 if (num > offslab_limit)
2038 break;
2039 }
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002040
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002041 /* Found something acceptable - save it away */
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002042 cachep->num = num;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002043 cachep->gfporder = gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002044 left_over = remainder;
2045
2046 /*
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002047 * A VFS-reclaimable slab tends to have most allocations
2048 * as GFP_NOFS and we really don't want to have to be allocating
2049 * higher-order pages when we are unable to shrink dcache.
2050 */
2051 if (flags & SLAB_RECLAIM_ACCOUNT)
2052 break;
2053
2054 /*
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002055 * Large number of objects is good, but very large slabs are
2056 * currently bad for the gfp()s.
2057 */
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002058 if (gfporder >= slab_break_gfp_order)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002059 break;
2060
Linus Torvalds9888e6f2006-03-06 17:44:43 -08002061 /*
2062 * Acceptable internal fragmentation?
2063 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002064 if (left_over * 8 <= (PAGE_SIZE << gfporder))
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002065 break;
2066 }
2067 return left_over;
2068}
2069
Pekka Enberg83b519e2009-06-10 19:40:04 +03002070static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002071{
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002072 if (g_cpucache_up == FULL)
Pekka Enberg83b519e2009-06-10 19:40:04 +03002073 return enable_cpucache(cachep, gfp);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002074
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002075 if (g_cpucache_up == NONE) {
2076 /*
2077 * Note: the first kmem_cache_create must create the cache
2078 * that's used by kmalloc(24), otherwise the creation of
2079 * further caches will BUG().
2080 */
2081 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2082
2083 /*
2084 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2085 * the first cache, then we need to set up all its list3s,
2086 * otherwise the creation of further caches will BUG().
2087 */
2088 set_up_list3s(cachep, SIZE_AC);
2089 if (INDEX_AC == INDEX_L3)
2090 g_cpucache_up = PARTIAL_L3;
2091 else
2092 g_cpucache_up = PARTIAL_AC;
2093 } else {
2094 cachep->array[smp_processor_id()] =
Pekka Enberg83b519e2009-06-10 19:40:04 +03002095 kmalloc(sizeof(struct arraycache_init), gfp);
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002096
2097 if (g_cpucache_up == PARTIAL_AC) {
2098 set_up_list3s(cachep, SIZE_L3);
2099 g_cpucache_up = PARTIAL_L3;
2100 } else {
2101 int node;
Pekka Enberg556a1692008-01-25 08:20:51 +02002102 for_each_online_node(node) {
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002103 cachep->nodelists[node] =
2104 kmalloc_node(sizeof(struct kmem_list3),
2105 GFP_KERNEL, node);
2106 BUG_ON(!cachep->nodelists[node]);
2107 kmem_list3_init(cachep->nodelists[node]);
2108 }
2109 }
2110 }
2111 cachep->nodelists[numa_node_id()]->next_reap =
2112 jiffies + REAPTIMEOUT_LIST3 +
2113 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2114
2115 cpu_cache_get(cachep)->avail = 0;
2116 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2117 cpu_cache_get(cachep)->batchcount = 1;
2118 cpu_cache_get(cachep)->touched = 0;
2119 cachep->batchcount = 1;
2120 cachep->limit = BOOT_CPUCACHE_ENTRIES;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002121 return 0;
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002122}
2123
Pekka Enberg4d268eb2006-01-08 01:00:36 -08002124/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07002125 * kmem_cache_create - Create a cache.
2126 * @name: A string which is used in /proc/slabinfo to identify this cache.
2127 * @size: The size of objects to be created in this cache.
2128 * @align: The required alignment for the objects.
2129 * @flags: SLAB flags
2130 * @ctor: A constructor for the objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002131 *
2132 * Returns a ptr to the cache on success, NULL on failure.
2133 * Cannot be called within a int, but can be interrupted.
Paul Mundt20c2df82007-07-20 10:11:58 +09002134 * The @ctor is run when new pages are allocated by the cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002135 *
2136 * @name must be valid until the cache is destroyed. This implies that
Andrew Mortona737b3e2006-03-22 00:08:11 -08002137 * the module calling this has to destroy the cache before getting unloaded.
Catalin Marinas249da162008-11-21 12:56:22 +00002138 * Note that kmem_cache_name() is not guaranteed to return the same pointer,
2139 * therefore applications must manage it themselves.
Andrew Mortona737b3e2006-03-22 00:08:11 -08002140 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002141 * The flags are
2142 *
2143 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2144 * to catch references to uninitialised memory.
2145 *
2146 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2147 * for buffer overruns.
2148 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002149 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2150 * cacheline. This can be beneficial if you're counting cycles as closely
2151 * as davem.
2152 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002153struct kmem_cache *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002154kmem_cache_create (const char *name, size_t size, size_t align,
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002155 unsigned long flags, void (*ctor)(void *))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002156{
2157 size_t left_over, slab_size, ralign;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07002158 struct kmem_cache *cachep = NULL, *pc;
Pekka Enberg83b519e2009-06-10 19:40:04 +03002159 gfp_t gfp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002160
2161 /*
2162 * Sanity checks... these are all serious usage bugs.
2163 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002164 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
Paul Mundt20c2df82007-07-20 10:11:58 +09002165 size > KMALLOC_MAX_SIZE) {
Harvey Harrisond40cee22008-04-30 00:55:07 -07002166 printk(KERN_ERR "%s: Early error in slab %s\n", __func__,
Andrew Mortona737b3e2006-03-22 00:08:11 -08002167 name);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002168 BUG();
2169 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002170
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002171 /*
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002172 * We use cache_chain_mutex to ensure a consistent view of
Rusty Russell174596a2009-01-01 10:12:29 +10302173 * cpu_online_mask as well. Please see cpuup_callback
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002174 */
Pekka Enberg83b519e2009-06-10 19:40:04 +03002175 if (slab_is_available()) {
2176 get_online_cpus();
2177 mutex_lock(&cache_chain_mutex);
2178 }
Andrew Morton4f12bb42005-11-07 00:58:00 -08002179
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07002180 list_for_each_entry(pc, &cache_chain, next) {
Andrew Morton4f12bb42005-11-07 00:58:00 -08002181 char tmp;
2182 int res;
2183
2184 /*
2185 * This happens when the module gets unloaded and doesn't
2186 * destroy its slab cache and no-one else reuses the vmalloc
2187 * area of the module. Print a warning.
2188 */
Andrew Morton138ae662006-12-06 20:36:41 -08002189 res = probe_kernel_address(pc->name, tmp);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002190 if (res) {
matzeb4169522007-05-06 14:49:52 -07002191 printk(KERN_ERR
2192 "SLAB: cache with size %d has lost its name\n",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002193 pc->buffer_size);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002194 continue;
2195 }
2196
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002197 if (!strcmp(pc->name, name)) {
matzeb4169522007-05-06 14:49:52 -07002198 printk(KERN_ERR
2199 "kmem_cache_create: duplicate cache %s\n", name);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002200 dump_stack();
2201 goto oops;
2202 }
2203 }
2204
Linus Torvalds1da177e2005-04-16 15:20:36 -07002205#if DEBUG
2206 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002207#if FORCED_DEBUG
2208 /*
2209 * Enable redzoning and last user accounting, except for caches with
2210 * large objects, if the increased size would increase the object size
2211 * above the next power of two: caches with object sizes just above a
2212 * power of two have a significant amount of internal fragmentation.
2213 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002214 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2215 2 * sizeof(unsigned long long)))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002216 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002217 if (!(flags & SLAB_DESTROY_BY_RCU))
2218 flags |= SLAB_POISON;
2219#endif
2220 if (flags & SLAB_DESTROY_BY_RCU)
2221 BUG_ON(flags & SLAB_POISON);
2222#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002223 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002224 * Always checks flags, a caller might be expecting debug support which
2225 * isn't available.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002226 */
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002227 BUG_ON(flags & ~CREATE_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002228
Andrew Mortona737b3e2006-03-22 00:08:11 -08002229 /*
2230 * Check that size is in terms of words. This is needed to avoid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002231 * unaligned accesses for some archs when redzoning is used, and makes
2232 * sure any on-slab bufctl's are also correctly aligned.
2233 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002234 if (size & (BYTES_PER_WORD - 1)) {
2235 size += (BYTES_PER_WORD - 1);
2236 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002237 }
2238
Andrew Mortona737b3e2006-03-22 00:08:11 -08002239 /* calculate the final buffer alignment: */
2240
Linus Torvalds1da177e2005-04-16 15:20:36 -07002241 /* 1) arch recommendation: can be overridden for debug */
2242 if (flags & SLAB_HWCACHE_ALIGN) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002243 /*
2244 * Default alignment: as specified by the arch code. Except if
2245 * an object is really small, then squeeze multiple objects into
2246 * one cacheline.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002247 */
2248 ralign = cache_line_size();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002249 while (size <= ralign / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002250 ralign /= 2;
2251 } else {
2252 ralign = BYTES_PER_WORD;
2253 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002254
2255 /*
David Woodhouse87a927c2007-07-04 21:26:44 -04002256 * Redzoning and user store require word alignment or possibly larger.
2257 * Note this will be overridden by architecture or caller mandated
2258 * alignment if either is greater than BYTES_PER_WORD.
Pekka Enbergca5f9702006-09-25 23:31:25 -07002259 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002260 if (flags & SLAB_STORE_USER)
2261 ralign = BYTES_PER_WORD;
2262
2263 if (flags & SLAB_RED_ZONE) {
2264 ralign = REDZONE_ALIGN;
2265 /* If redzoning, ensure that the second redzone is suitably
2266 * aligned, by adjusting the object size accordingly. */
2267 size += REDZONE_ALIGN - 1;
2268 size &= ~(REDZONE_ALIGN - 1);
2269 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002270
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002271 /* 2) arch mandated alignment */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002272 if (ralign < ARCH_SLAB_MINALIGN) {
2273 ralign = ARCH_SLAB_MINALIGN;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002274 }
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002275 /* 3) caller mandated alignment */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002276 if (ralign < align) {
2277 ralign = align;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002278 }
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002279 /* disable debug if necessary */
David Woodhouseb46b8f12007-05-08 00:22:59 -07002280 if (ralign > __alignof__(unsigned long long))
Kevin Hilmana44b56d2006-12-06 20:32:11 -08002281 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002282 /*
Pekka Enbergca5f9702006-09-25 23:31:25 -07002283 * 4) Store it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002284 */
2285 align = ralign;
2286
Pekka Enberg83b519e2009-06-10 19:40:04 +03002287 if (slab_is_available())
2288 gfp = GFP_KERNEL;
2289 else
2290 gfp = GFP_NOWAIT;
2291
Linus Torvalds1da177e2005-04-16 15:20:36 -07002292 /* Get cache's description obj. */
Pekka Enberg83b519e2009-06-10 19:40:04 +03002293 cachep = kmem_cache_zalloc(&cache_cache, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002294 if (!cachep)
Andrew Morton4f12bb42005-11-07 00:58:00 -08002295 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002296
2297#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002298 cachep->obj_size = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002299
Pekka Enbergca5f9702006-09-25 23:31:25 -07002300 /*
2301 * Both debugging options require word-alignment which is calculated
2302 * into align above.
2303 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002304 if (flags & SLAB_RED_ZONE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002305 /* add space for red zone words */
David Woodhouseb46b8f12007-05-08 00:22:59 -07002306 cachep->obj_offset += sizeof(unsigned long long);
2307 size += 2 * sizeof(unsigned long long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002308 }
2309 if (flags & SLAB_STORE_USER) {
Pekka Enbergca5f9702006-09-25 23:31:25 -07002310 /* user store requires one word storage behind the end of
David Woodhouse87a927c2007-07-04 21:26:44 -04002311 * the real object. But if the second red zone needs to be
2312 * aligned to 64 bits, we must allow that much space.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002313 */
David Woodhouse87a927c2007-07-04 21:26:44 -04002314 if (flags & SLAB_RED_ZONE)
2315 size += REDZONE_ALIGN;
2316 else
2317 size += BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002318 }
2319#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002320 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002321 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2322 cachep->obj_offset += PAGE_SIZE - size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002323 size = PAGE_SIZE;
2324 }
2325#endif
2326#endif
2327
Ingo Molnare0a42722006-06-23 02:03:46 -07002328 /*
2329 * Determine if the slab management is 'on' or 'off' slab.
2330 * (bootstrapping cannot cope with offslab caches so don't do
2331 * it too early on.)
2332 */
2333 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002334 /*
2335 * Size is large, assume best to place the slab management obj
2336 * off-slab (should allow better packing of objs).
2337 */
2338 flags |= CFLGS_OFF_SLAB;
2339
2340 size = ALIGN(size, align);
2341
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002342 left_over = calculate_slab_order(cachep, size, align, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002343
2344 if (!cachep->num) {
matzeb4169522007-05-06 14:49:52 -07002345 printk(KERN_ERR
2346 "kmem_cache_create: couldn't create cache %s.\n", name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002347 kmem_cache_free(&cache_cache, cachep);
2348 cachep = NULL;
Andrew Morton4f12bb42005-11-07 00:58:00 -08002349 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002350 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002351 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2352 + sizeof(struct slab), align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002353
2354 /*
2355 * If the slab has been placed off-slab, and we have enough space then
2356 * move it on-slab. This is at the expense of any extra colouring.
2357 */
2358 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2359 flags &= ~CFLGS_OFF_SLAB;
2360 left_over -= slab_size;
2361 }
2362
2363 if (flags & CFLGS_OFF_SLAB) {
2364 /* really off slab. No need for manual alignment */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002365 slab_size =
2366 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002367 }
2368
2369 cachep->colour_off = cache_line_size();
2370 /* Offset must be a multiple of the alignment. */
2371 if (cachep->colour_off < align)
2372 cachep->colour_off = align;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002373 cachep->colour = left_over / cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002374 cachep->slab_size = slab_size;
2375 cachep->flags = flags;
2376 cachep->gfpflags = 0;
Christoph Lameter4b51d662007-02-10 01:43:10 -08002377 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002378 cachep->gfpflags |= GFP_DMA;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002379 cachep->buffer_size = size;
Eric Dumazet6a2d7a92006-12-13 00:34:27 -08002380 cachep->reciprocal_buffer_size = reciprocal_value(size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002381
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002382 if (flags & CFLGS_OFF_SLAB) {
Victor Fuscob2d55072005-09-10 00:26:36 -07002383 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002384 /*
2385 * This is a possibility for one of the malloc_sizes caches.
2386 * But since we go off slab only for object size greater than
2387 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2388 * this should not happen at all.
2389 * But leave a BUG_ON for some lucky dude.
2390 */
Christoph Lameter6cb8f912007-07-17 04:03:22 -07002391 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002392 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002393 cachep->ctor = ctor;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002394 cachep->name = name;
2395
Pekka Enberg83b519e2009-06-10 19:40:04 +03002396 if (setup_cpu_cache(cachep, gfp)) {
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07002397 __kmem_cache_destroy(cachep);
2398 cachep = NULL;
2399 goto oops;
2400 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002401
Linus Torvalds1da177e2005-04-16 15:20:36 -07002402 /* cache setup completed, link it into the list */
2403 list_add(&cachep->next, &cache_chain);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002404oops:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002405 if (!cachep && (flags & SLAB_PANIC))
2406 panic("kmem_cache_create(): failed to create slab `%s'\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002407 name);
Pekka Enberg83b519e2009-06-10 19:40:04 +03002408 if (slab_is_available()) {
2409 mutex_unlock(&cache_chain_mutex);
2410 put_online_cpus();
2411 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002412 return cachep;
2413}
2414EXPORT_SYMBOL(kmem_cache_create);
2415
2416#if DEBUG
2417static void check_irq_off(void)
2418{
2419 BUG_ON(!irqs_disabled());
2420}
2421
2422static void check_irq_on(void)
2423{
2424 BUG_ON(irqs_disabled());
2425}
2426
Pekka Enberg343e0d72006-02-01 03:05:50 -08002427static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002428{
2429#ifdef CONFIG_SMP
2430 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002431 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002432#endif
2433}
Christoph Lametere498be72005-09-09 13:03:32 -07002434
Pekka Enberg343e0d72006-02-01 03:05:50 -08002435static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002436{
2437#ifdef CONFIG_SMP
2438 check_irq_off();
2439 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2440#endif
2441}
2442
Linus Torvalds1da177e2005-04-16 15:20:36 -07002443#else
2444#define check_irq_off() do { } while(0)
2445#define check_irq_on() do { } while(0)
2446#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002447#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002448#endif
2449
Christoph Lameteraab22072006-03-22 00:09:06 -08002450static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2451 struct array_cache *ac,
2452 int force, int node);
2453
Linus Torvalds1da177e2005-04-16 15:20:36 -07002454static void do_drain(void *arg)
2455{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002456 struct kmem_cache *cachep = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002457 struct array_cache *ac;
Christoph Lameterff694162005-09-22 21:44:02 -07002458 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002459
2460 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002461 ac = cpu_cache_get(cachep);
Christoph Lameterff694162005-09-22 21:44:02 -07002462 spin_lock(&cachep->nodelists[node]->list_lock);
2463 free_block(cachep, ac->entry, ac->avail, node);
2464 spin_unlock(&cachep->nodelists[node]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002465 ac->avail = 0;
2466}
2467
Pekka Enberg343e0d72006-02-01 03:05:50 -08002468static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002469{
Christoph Lametere498be72005-09-09 13:03:32 -07002470 struct kmem_list3 *l3;
2471 int node;
2472
Jens Axboe15c8b6c2008-05-09 09:39:44 +02002473 on_each_cpu(do_drain, cachep, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002474 check_irq_on();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002475 for_each_online_node(node) {
Christoph Lametere498be72005-09-09 13:03:32 -07002476 l3 = cachep->nodelists[node];
Roland Dreiera4523a82006-05-15 11:41:00 -07002477 if (l3 && l3->alien)
2478 drain_alien_cache(cachep, l3->alien);
2479 }
2480
2481 for_each_online_node(node) {
2482 l3 = cachep->nodelists[node];
2483 if (l3)
Christoph Lameteraab22072006-03-22 00:09:06 -08002484 drain_array(cachep, l3, l3->shared, 1, node);
Christoph Lametere498be72005-09-09 13:03:32 -07002485 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002486}
2487
Christoph Lametered11d9e2006-06-30 01:55:45 -07002488/*
2489 * Remove slabs from the list of free slabs.
2490 * Specify the number of slabs to drain in tofree.
2491 *
2492 * Returns the actual number of slabs released.
2493 */
2494static int drain_freelist(struct kmem_cache *cache,
2495 struct kmem_list3 *l3, int tofree)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002496{
Christoph Lametered11d9e2006-06-30 01:55:45 -07002497 struct list_head *p;
2498 int nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002499 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002500
Christoph Lametered11d9e2006-06-30 01:55:45 -07002501 nr_freed = 0;
2502 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002503
Christoph Lametered11d9e2006-06-30 01:55:45 -07002504 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002505 p = l3->slabs_free.prev;
Christoph Lametered11d9e2006-06-30 01:55:45 -07002506 if (p == &l3->slabs_free) {
2507 spin_unlock_irq(&l3->list_lock);
2508 goto out;
2509 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002510
Christoph Lametered11d9e2006-06-30 01:55:45 -07002511 slabp = list_entry(p, struct slab, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002512#if DEBUG
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002513 BUG_ON(slabp->inuse);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002514#endif
2515 list_del(&slabp->list);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002516 /*
2517 * Safe to drop the lock. The slab is no longer linked
2518 * to the cache.
2519 */
2520 l3->free_objects -= cache->num;
Christoph Lametere498be72005-09-09 13:03:32 -07002521 spin_unlock_irq(&l3->list_lock);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002522 slab_destroy(cache, slabp);
2523 nr_freed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002524 }
Christoph Lametered11d9e2006-06-30 01:55:45 -07002525out:
2526 return nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002527}
2528
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002529/* Called with cache_chain_mutex held to protect against cpu hotplug */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002530static int __cache_shrink(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07002531{
2532 int ret = 0, i = 0;
2533 struct kmem_list3 *l3;
2534
2535 drain_cpu_caches(cachep);
2536
2537 check_irq_on();
2538 for_each_online_node(i) {
2539 l3 = cachep->nodelists[i];
Christoph Lametered11d9e2006-06-30 01:55:45 -07002540 if (!l3)
2541 continue;
2542
2543 drain_freelist(cachep, l3, l3->free_objects);
2544
2545 ret += !list_empty(&l3->slabs_full) ||
2546 !list_empty(&l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07002547 }
2548 return (ret ? 1 : 0);
2549}
2550
Linus Torvalds1da177e2005-04-16 15:20:36 -07002551/**
2552 * kmem_cache_shrink - Shrink a cache.
2553 * @cachep: The cache to shrink.
2554 *
2555 * Releases as many slabs as possible for a cache.
2556 * To help debugging, a zero exit status indicates all slabs were released.
2557 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002558int kmem_cache_shrink(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002559{
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002560 int ret;
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002561 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002562
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002563 get_online_cpus();
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002564 mutex_lock(&cache_chain_mutex);
2565 ret = __cache_shrink(cachep);
2566 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002567 put_online_cpus();
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002568 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002569}
2570EXPORT_SYMBOL(kmem_cache_shrink);
2571
2572/**
2573 * kmem_cache_destroy - delete a cache
2574 * @cachep: the cache to destroy
2575 *
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08002576 * Remove a &struct kmem_cache object from the slab cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002577 *
2578 * It is expected this function will be called by a module when it is
2579 * unloaded. This will remove the cache completely, and avoid a duplicate
2580 * cache being allocated each time a module is loaded and unloaded, if the
2581 * module doesn't have persistent in-kernel storage across loads and unloads.
2582 *
2583 * The cache must be empty before calling this function.
2584 *
2585 * The caller must guarantee that noone will allocate memory from the cache
2586 * during the kmem_cache_destroy().
2587 */
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002588void kmem_cache_destroy(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002589{
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002590 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002591
Linus Torvalds1da177e2005-04-16 15:20:36 -07002592 /* Find the cache in the chain of caches. */
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002593 get_online_cpus();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002594 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002595 /*
2596 * the chain is never empty, cache_cache is never destroyed
2597 */
2598 list_del(&cachep->next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002599 if (__cache_shrink(cachep)) {
2600 slab_error(cachep, "Can't free all objects");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002601 list_add(&cachep->next, &cache_chain);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002602 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002603 put_online_cpus();
Alexey Dobriyan133d2052006-09-27 01:49:41 -07002604 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002605 }
2606
2607 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
Paul E. McKenneyfbd568a3e2005-05-01 08:59:04 -07002608 synchronize_rcu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002609
Christoph Lameter117f6eb2006-09-25 23:31:37 -07002610 __kmem_cache_destroy(cachep);
Ravikiran G Thirumalai8f5be202006-12-06 20:32:14 -08002611 mutex_unlock(&cache_chain_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01002612 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002613}
2614EXPORT_SYMBOL(kmem_cache_destroy);
2615
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002616/*
2617 * Get the memory for a slab management obj.
2618 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2619 * always come from malloc_sizes caches. The slab descriptor cannot
2620 * come from the same cache which is getting created because,
2621 * when we are searching for an appropriate cache for these
2622 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2623 * If we are creating a malloc_sizes cache here it would not be visible to
2624 * kmem_find_general_cachep till the initialization is complete.
2625 * Hence we cannot have slabp_cache same as the original cache.
2626 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002627static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002628 int colour_off, gfp_t local_flags,
2629 int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002630{
2631 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002632
Linus Torvalds1da177e2005-04-16 15:20:36 -07002633 if (OFF_SLAB(cachep)) {
2634 /* Slab management obj is off-slab. */
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002635 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
Pekka Enberg8759ec52008-11-26 10:01:31 +02002636 local_flags, nodeid);
Catalin Marinasd5cff632009-06-11 13:22:40 +01002637 /*
2638 * If the first object in the slab is leaked (it's allocated
2639 * but no one has a reference to it), we want to make sure
2640 * kmemleak does not treat the ->s_mem pointer as a reference
2641 * to the object. Otherwise we will not report the leak.
2642 */
2643 kmemleak_scan_area(slabp, offsetof(struct slab, list),
2644 sizeof(struct list_head), local_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002645 if (!slabp)
2646 return NULL;
2647 } else {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002648 slabp = objp + colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002649 colour_off += cachep->slab_size;
2650 }
2651 slabp->inuse = 0;
2652 slabp->colouroff = colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002653 slabp->s_mem = objp + colour_off;
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002654 slabp->nodeid = nodeid;
Marcin Slusarze51bfd02008-02-10 11:21:54 +01002655 slabp->free = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002656 return slabp;
2657}
2658
2659static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2660{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002661 return (kmem_bufctl_t *) (slabp + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002662}
2663
Pekka Enberg343e0d72006-02-01 03:05:50 -08002664static void cache_init_objs(struct kmem_cache *cachep,
Christoph Lametera35afb82007-05-16 22:10:57 -07002665 struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002666{
2667 int i;
2668
2669 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002670 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002671#if DEBUG
2672 /* need to poison the objs? */
2673 if (cachep->flags & SLAB_POISON)
2674 poison_obj(cachep, objp, POISON_FREE);
2675 if (cachep->flags & SLAB_STORE_USER)
2676 *dbg_userword(cachep, objp) = NULL;
2677
2678 if (cachep->flags & SLAB_RED_ZONE) {
2679 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2680 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2681 }
2682 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002683 * Constructors are not allowed to allocate memory from the same
2684 * cache which they are a constructor for. Otherwise, deadlock.
2685 * They must also be threaded.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002686 */
2687 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002688 cachep->ctor(objp + obj_offset(cachep));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002689
2690 if (cachep->flags & SLAB_RED_ZONE) {
2691 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2692 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002693 " end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002694 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2695 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002696 " start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002697 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08002698 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2699 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002700 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002701 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002702#else
2703 if (cachep->ctor)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07002704 cachep->ctor(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002705#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002706 slab_bufctl(slabp)[i] = i + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002707 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002708 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002709}
2710
Pekka Enberg343e0d72006-02-01 03:05:50 -08002711static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002712{
Christoph Lameter4b51d662007-02-10 01:43:10 -08002713 if (CONFIG_ZONE_DMA_FLAG) {
2714 if (flags & GFP_DMA)
2715 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2716 else
2717 BUG_ON(cachep->gfpflags & GFP_DMA);
2718 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002719}
2720
Andrew Mortona737b3e2006-03-22 00:08:11 -08002721static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2722 int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002723{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002724 void *objp = index_to_obj(cachep, slabp, slabp->free);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002725 kmem_bufctl_t next;
2726
2727 slabp->inuse++;
2728 next = slab_bufctl(slabp)[slabp->free];
2729#if DEBUG
2730 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2731 WARN_ON(slabp->nodeid != nodeid);
2732#endif
2733 slabp->free = next;
2734
2735 return objp;
2736}
2737
Andrew Mortona737b3e2006-03-22 00:08:11 -08002738static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2739 void *objp, int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002740{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002741 unsigned int objnr = obj_to_index(cachep, slabp, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002742
2743#if DEBUG
2744 /* Verify that the slab belongs to the intended node */
2745 WARN_ON(slabp->nodeid != nodeid);
2746
Al Viro871751e2006-03-25 03:06:39 -08002747 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
Matthew Dobson78d382d2006-02-01 03:05:47 -08002748 printk(KERN_ERR "slab: double free detected in cache "
Andrew Mortona737b3e2006-03-22 00:08:11 -08002749 "'%s', objp %p\n", cachep->name, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002750 BUG();
2751 }
2752#endif
2753 slab_bufctl(slabp)[objnr] = slabp->free;
2754 slabp->free = objnr;
2755 slabp->inuse--;
2756}
2757
Pekka Enberg47768742006-06-23 02:03:07 -07002758/*
2759 * Map pages beginning at addr to the given cache and slab. This is required
2760 * for the slab allocator to be able to lookup the cache and slab of a
2761 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2762 */
2763static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2764 void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002765{
Pekka Enberg47768742006-06-23 02:03:07 -07002766 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002767 struct page *page;
2768
Pekka Enberg47768742006-06-23 02:03:07 -07002769 page = virt_to_page(addr);
Nick Piggin84097512006-03-22 00:08:34 -08002770
Pekka Enberg47768742006-06-23 02:03:07 -07002771 nr_pages = 1;
Nick Piggin84097512006-03-22 00:08:34 -08002772 if (likely(!PageCompound(page)))
Pekka Enberg47768742006-06-23 02:03:07 -07002773 nr_pages <<= cache->gfporder;
2774
Linus Torvalds1da177e2005-04-16 15:20:36 -07002775 do {
Pekka Enberg47768742006-06-23 02:03:07 -07002776 page_set_cache(page, cache);
2777 page_set_slab(page, slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002778 page++;
Pekka Enberg47768742006-06-23 02:03:07 -07002779 } while (--nr_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002780}
2781
2782/*
2783 * Grow (by 1) the number of slabs within a cache. This is called by
2784 * kmem_cache_alloc() when there are no active objs left in a cache.
2785 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002786static int cache_grow(struct kmem_cache *cachep,
2787 gfp_t flags, int nodeid, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002788{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002789 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002790 size_t offset;
2791 gfp_t local_flags;
Christoph Lametere498be72005-09-09 13:03:32 -07002792 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002793
Andrew Mortona737b3e2006-03-22 00:08:11 -08002794 /*
2795 * Be lazy and only check for valid flags here, keeping it out of the
2796 * critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002797 */
Christoph Lameter6cb06222007-10-16 01:25:41 -07002798 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2799 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002800
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002801 /* Take the l3 list lock to change the colour_next on this node */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002802 check_irq_off();
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002803 l3 = cachep->nodelists[nodeid];
2804 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002805
2806 /* Get colour for the slab, and cal the next value. */
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002807 offset = l3->colour_next;
2808 l3->colour_next++;
2809 if (l3->colour_next >= cachep->colour)
2810 l3->colour_next = 0;
2811 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002812
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002813 offset *= cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002814
2815 if (local_flags & __GFP_WAIT)
2816 local_irq_enable();
2817
2818 /*
2819 * The test for missing atomic flag is performed here, rather than
2820 * the more obvious place, simply to reduce the critical path length
2821 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2822 * will eventually be caught here (where it matters).
2823 */
2824 kmem_flagcheck(cachep, flags);
2825
Andrew Mortona737b3e2006-03-22 00:08:11 -08002826 /*
2827 * Get mem for the objs. Attempt to allocate a physical page from
2828 * 'nodeid'.
Christoph Lametere498be72005-09-09 13:03:32 -07002829 */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002830 if (!objp)
Andrew Mortonb8c1c5d2007-07-24 12:02:40 -07002831 objp = kmem_getpages(cachep, local_flags, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002832 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002833 goto failed;
2834
2835 /* Get slab management. */
Christoph Lameter3c517a62006-12-06 20:33:29 -08002836 slabp = alloc_slabmgmt(cachep, objp, offset,
Christoph Lameter6cb06222007-10-16 01:25:41 -07002837 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002838 if (!slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002839 goto opps1;
2840
Pekka Enberg47768742006-06-23 02:03:07 -07002841 slab_map_pages(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002842
Christoph Lametera35afb82007-05-16 22:10:57 -07002843 cache_init_objs(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002844
2845 if (local_flags & __GFP_WAIT)
2846 local_irq_disable();
2847 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002848 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002849
2850 /* Make slab active. */
Christoph Lametere498be72005-09-09 13:03:32 -07002851 list_add_tail(&slabp->list, &(l3->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002852 STATS_INC_GROWN(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002853 l3->free_objects += cachep->num;
2854 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002855 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002856opps1:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002857 kmem_freepages(cachep, objp);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002858failed:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002859 if (local_flags & __GFP_WAIT)
2860 local_irq_disable();
2861 return 0;
2862}
2863
2864#if DEBUG
2865
2866/*
2867 * Perform extra freeing checks:
2868 * - detect bad pointers.
2869 * - POISON/RED_ZONE checking
Linus Torvalds1da177e2005-04-16 15:20:36 -07002870 */
2871static void kfree_debugcheck(const void *objp)
2872{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002873 if (!virt_addr_valid(objp)) {
2874 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002875 (unsigned long)objp);
2876 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002877 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002878}
2879
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002880static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2881{
David Woodhouseb46b8f12007-05-08 00:22:59 -07002882 unsigned long long redzone1, redzone2;
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002883
2884 redzone1 = *dbg_redzone1(cache, obj);
2885 redzone2 = *dbg_redzone2(cache, obj);
2886
2887 /*
2888 * Redzone is ok.
2889 */
2890 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2891 return;
2892
2893 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2894 slab_error(cache, "double free detected");
2895 else
2896 slab_error(cache, "memory outside object was overwritten");
2897
David Woodhouseb46b8f12007-05-08 00:22:59 -07002898 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002899 obj, redzone1, redzone2);
2900}
2901
Pekka Enberg343e0d72006-02-01 03:05:50 -08002902static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002903 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002904{
2905 struct page *page;
2906 unsigned int objnr;
2907 struct slab *slabp;
2908
Matthew Wilcox80cbd912007-11-29 12:05:13 -07002909 BUG_ON(virt_to_cache(objp) != cachep);
2910
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002911 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002912 kfree_debugcheck(objp);
Christoph Lameterb49af682007-05-06 14:49:41 -07002913 page = virt_to_head_page(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002914
Pekka Enberg065d41c2005-11-13 16:06:46 -08002915 slabp = page_get_slab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002916
2917 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002918 verify_redzone_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002919 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2920 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2921 }
2922 if (cachep->flags & SLAB_STORE_USER)
2923 *dbg_userword(cachep, objp) = caller;
2924
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002925 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002926
2927 BUG_ON(objnr >= cachep->num);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002928 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002929
Al Viro871751e2006-03-25 03:06:39 -08002930#ifdef CONFIG_DEBUG_SLAB_LEAK
2931 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2932#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002933 if (cachep->flags & SLAB_POISON) {
2934#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08002935 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002936 store_stackinfo(cachep, objp, (unsigned long)caller);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002937 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002938 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002939 } else {
2940 poison_obj(cachep, objp, POISON_FREE);
2941 }
2942#else
2943 poison_obj(cachep, objp, POISON_FREE);
2944#endif
2945 }
2946 return objp;
2947}
2948
Pekka Enberg343e0d72006-02-01 03:05:50 -08002949static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002950{
2951 kmem_bufctl_t i;
2952 int entries = 0;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002953
Linus Torvalds1da177e2005-04-16 15:20:36 -07002954 /* Check slab's freelist to see if this obj is there. */
2955 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2956 entries++;
2957 if (entries > cachep->num || i >= cachep->num)
2958 goto bad;
2959 }
2960 if (entries != cachep->num - slabp->inuse) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002961bad:
2962 printk(KERN_ERR "slab: Internal list corruption detected in "
2963 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2964 cachep->name, cachep->num, slabp, slabp->inuse);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002965 for (i = 0;
Linus Torvalds264132b2006-03-06 12:10:07 -08002966 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002967 i++) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002968 if (i % 16 == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002969 printk("\n%03x:", i);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002970 printk(" %02x", ((unsigned char *)slabp)[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002971 }
2972 printk("\n");
2973 BUG();
2974 }
2975}
2976#else
2977#define kfree_debugcheck(x) do { } while(0)
2978#define cache_free_debugcheck(x,objp,z) (objp)
2979#define check_slabp(x,y) do { } while(0)
2980#endif
2981
Pekka Enberg343e0d72006-02-01 03:05:50 -08002982static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002983{
2984 int batchcount;
2985 struct kmem_list3 *l3;
2986 struct array_cache *ac;
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07002987 int node;
2988
Andrew Mortona737b3e2006-03-22 00:08:11 -08002989retry:
Joe Korty6d2144d2008-03-05 15:04:59 -08002990 check_irq_off();
2991 node = numa_node_id();
2992 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002993 batchcount = ac->batchcount;
2994 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002995 /*
2996 * If there was little recent activity on this cache, then
2997 * perform only a partial refill. Otherwise we could generate
2998 * refill bouncing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002999 */
3000 batchcount = BATCHREFILL_LIMIT;
3001 }
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07003002 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003003
Christoph Lametere498be72005-09-09 13:03:32 -07003004 BUG_ON(ac->avail > 0 || !l3);
3005 spin_lock(&l3->list_lock);
3006
Christoph Lameter3ded1752006-03-25 03:06:44 -08003007 /* See if we can refill from the shared array */
3008 if (l3->shared && transfer_objects(ac, l3->shared, batchcount))
3009 goto alloc_done;
3010
Linus Torvalds1da177e2005-04-16 15:20:36 -07003011 while (batchcount > 0) {
3012 struct list_head *entry;
3013 struct slab *slabp;
3014 /* Get slab alloc is to come from. */
3015 entry = l3->slabs_partial.next;
3016 if (entry == &l3->slabs_partial) {
3017 l3->free_touched = 1;
3018 entry = l3->slabs_free.next;
3019 if (entry == &l3->slabs_free)
3020 goto must_grow;
3021 }
3022
3023 slabp = list_entry(entry, struct slab, list);
3024 check_slabp(cachep, slabp);
3025 check_spinlock_acquired(cachep);
Pekka Enberg714b8172007-05-06 14:49:03 -07003026
3027 /*
3028 * The slab was either on partial or free list so
3029 * there must be at least one object available for
3030 * allocation.
3031 */
roel kluin249b9f32008-10-29 17:18:07 -04003032 BUG_ON(slabp->inuse >= cachep->num);
Pekka Enberg714b8172007-05-06 14:49:03 -07003033
Linus Torvalds1da177e2005-04-16 15:20:36 -07003034 while (slabp->inuse < cachep->num && batchcount--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003035 STATS_INC_ALLOCED(cachep);
3036 STATS_INC_ACTIVE(cachep);
3037 STATS_SET_HIGH(cachep);
3038
Matthew Dobson78d382d2006-02-01 03:05:47 -08003039 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
Pekka Enberg1ca4cb22006-10-06 00:43:52 -07003040 node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003041 }
3042 check_slabp(cachep, slabp);
3043
3044 /* move slabp to correct slabp list: */
3045 list_del(&slabp->list);
3046 if (slabp->free == BUFCTL_END)
3047 list_add(&slabp->list, &l3->slabs_full);
3048 else
3049 list_add(&slabp->list, &l3->slabs_partial);
3050 }
3051
Andrew Mortona737b3e2006-03-22 00:08:11 -08003052must_grow:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003053 l3->free_objects -= ac->avail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003054alloc_done:
Christoph Lametere498be72005-09-09 13:03:32 -07003055 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003056
3057 if (unlikely(!ac->avail)) {
3058 int x;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003059 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07003060
Andrew Mortona737b3e2006-03-22 00:08:11 -08003061 /* cache_grow can reenable interrupts, then ac could change. */
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003062 ac = cpu_cache_get(cachep);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003063 if (!x && ac->avail == 0) /* no objects in sight? abort */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003064 return NULL;
3065
Andrew Mortona737b3e2006-03-22 00:08:11 -08003066 if (!ac->avail) /* objects refilled by interrupt? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003067 goto retry;
3068 }
3069 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07003070 return ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003071}
3072
Andrew Mortona737b3e2006-03-22 00:08:11 -08003073static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3074 gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003075{
3076 might_sleep_if(flags & __GFP_WAIT);
3077#if DEBUG
3078 kmem_flagcheck(cachep, flags);
3079#endif
3080}
3081
3082#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003083static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3084 gfp_t flags, void *objp, void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003085{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003086 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003087 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003088 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003089#ifdef CONFIG_DEBUG_PAGEALLOC
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003090 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003091 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003092 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003093 else
3094 check_poison_obj(cachep, objp);
3095#else
3096 check_poison_obj(cachep, objp);
3097#endif
3098 poison_obj(cachep, objp, POISON_INUSE);
3099 }
3100 if (cachep->flags & SLAB_STORE_USER)
3101 *dbg_userword(cachep, objp) = caller;
3102
3103 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003104 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3105 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3106 slab_error(cachep, "double free, or memory outside"
3107 " object was overwritten");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003108 printk(KERN_ERR
David Woodhouseb46b8f12007-05-08 00:22:59 -07003109 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08003110 objp, *dbg_redzone1(cachep, objp),
3111 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003112 }
3113 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3114 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3115 }
Al Viro871751e2006-03-25 03:06:39 -08003116#ifdef CONFIG_DEBUG_SLAB_LEAK
3117 {
3118 struct slab *slabp;
3119 unsigned objnr;
3120
Christoph Lameterb49af682007-05-06 14:49:41 -07003121 slabp = page_get_slab(virt_to_head_page(objp));
Al Viro871751e2006-03-25 03:06:39 -08003122 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3123 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3124 }
3125#endif
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003126 objp += obj_offset(cachep);
Christoph Lameter4f104932007-05-06 14:50:17 -07003127 if (cachep->ctor && cachep->flags & SLAB_POISON)
Alexey Dobriyan51cc5062008-07-25 19:45:34 -07003128 cachep->ctor(objp);
Kevin Hilmana44b56d2006-12-06 20:32:11 -08003129#if ARCH_SLAB_MINALIGN
3130 if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) {
3131 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3132 objp, ARCH_SLAB_MINALIGN);
3133 }
3134#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003135 return objp;
3136}
3137#else
3138#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3139#endif
3140
Akinobu Mita773ff602008-12-23 19:37:01 +09003141static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003142{
3143 if (cachep == &cache_cache)
Akinobu Mita773ff602008-12-23 19:37:01 +09003144 return false;
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003145
Akinobu Mita773ff602008-12-23 19:37:01 +09003146 return should_failslab(obj_size(cachep), flags);
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003147}
3148
Pekka Enberg343e0d72006-02-01 03:05:50 -08003149static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003150{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003151 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003152 struct array_cache *ac;
3153
Alok N Kataria5c382302005-09-27 21:45:46 -07003154 check_irq_off();
Akinobu Mita8a8b6502006-12-08 02:39:44 -08003155
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003156 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003157 if (likely(ac->avail)) {
3158 STATS_INC_ALLOCHIT(cachep);
3159 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07003160 objp = ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003161 } else {
3162 STATS_INC_ALLOCMISS(cachep);
3163 objp = cache_alloc_refill(cachep, flags);
3164 }
Catalin Marinasd5cff632009-06-11 13:22:40 +01003165 /*
3166 * To avoid a false negative, if an object that is in one of the
3167 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3168 * treat the array pointers as a reference to the object.
3169 */
3170 kmemleak_erase(&ac->entry[ac->avail]);
Alok N Kataria5c382302005-09-27 21:45:46 -07003171 return objp;
3172}
3173
Christoph Lametere498be72005-09-09 13:03:32 -07003174#ifdef CONFIG_NUMA
3175/*
Paul Jacksonb2455392006-03-24 03:16:12 -08003176 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
Paul Jacksonc61afb12006-03-24 03:16:08 -08003177 *
3178 * If we are in_interrupt, then process context, including cpusets and
3179 * mempolicy, may not apply and should not be used for allocation policy.
3180 */
3181static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3182{
3183 int nid_alloc, nid_here;
3184
Christoph Lameter765c4502006-09-27 01:50:08 -07003185 if (in_interrupt() || (flags & __GFP_THISNODE))
Paul Jacksonc61afb12006-03-24 03:16:08 -08003186 return NULL;
3187 nid_alloc = nid_here = numa_node_id();
3188 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3189 nid_alloc = cpuset_mem_spread_node();
3190 else if (current->mempolicy)
3191 nid_alloc = slab_node(current->mempolicy);
3192 if (nid_alloc != nid_here)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003193 return ____cache_alloc_node(cachep, flags, nid_alloc);
Paul Jacksonc61afb12006-03-24 03:16:08 -08003194 return NULL;
3195}
3196
3197/*
Christoph Lameter765c4502006-09-27 01:50:08 -07003198 * Fallback function if there was no memory available and no objects on a
Christoph Lameter3c517a62006-12-06 20:33:29 -08003199 * certain node and fall back is permitted. First we scan all the
3200 * available nodelists for available objects. If that fails then we
3201 * perform an allocation without specifying a node. This allows the page
3202 * allocator to do its reclaim / fallback magic. We then insert the
3203 * slab into the proper nodelist and then allocate from it.
Christoph Lameter765c4502006-09-27 01:50:08 -07003204 */
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003205static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
Christoph Lameter765c4502006-09-27 01:50:08 -07003206{
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003207 struct zonelist *zonelist;
3208 gfp_t local_flags;
Mel Gormandd1a2392008-04-28 02:12:17 -07003209 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07003210 struct zone *zone;
3211 enum zone_type high_zoneidx = gfp_zone(flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003212 void *obj = NULL;
Christoph Lameter3c517a62006-12-06 20:33:29 -08003213 int nid;
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003214
3215 if (flags & __GFP_THISNODE)
3216 return NULL;
3217
Mel Gorman0e884602008-04-28 02:12:14 -07003218 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
Christoph Lameter6cb06222007-10-16 01:25:41 -07003219 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
Christoph Lameter765c4502006-09-27 01:50:08 -07003220
Christoph Lameter3c517a62006-12-06 20:33:29 -08003221retry:
3222 /*
3223 * Look through allowed nodes for objects available
3224 * from existing per node queues.
3225 */
Mel Gorman54a6eb52008-04-28 02:12:16 -07003226 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3227 nid = zone_to_nid(zone);
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003228
Mel Gorman54a6eb52008-04-28 02:12:16 -07003229 if (cpuset_zone_allowed_hardwall(zone, flags) &&
Christoph Lameter3c517a62006-12-06 20:33:29 -08003230 cache->nodelists[nid] &&
Christoph Lameter481c5342008-06-21 16:46:35 -07003231 cache->nodelists[nid]->free_objects) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003232 obj = ____cache_alloc_node(cache,
3233 flags | GFP_THISNODE, nid);
Christoph Lameter481c5342008-06-21 16:46:35 -07003234 if (obj)
3235 break;
3236 }
Christoph Lameter3c517a62006-12-06 20:33:29 -08003237 }
3238
Christoph Lametercfce6602007-05-06 14:50:17 -07003239 if (!obj) {
Christoph Lameter3c517a62006-12-06 20:33:29 -08003240 /*
3241 * This allocation will be performed within the constraints
3242 * of the current cpuset / memory policy requirements.
3243 * We may trigger various forms of reclaim on the allowed
3244 * set and go into memory reserves if necessary.
3245 */
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003246 if (local_flags & __GFP_WAIT)
3247 local_irq_enable();
3248 kmem_flagcheck(cache, flags);
Christoph Lameter9ac33b22008-03-04 12:24:22 -08003249 obj = kmem_getpages(cache, local_flags, -1);
Christoph Lameterdd47ea72006-12-13 00:34:11 -08003250 if (local_flags & __GFP_WAIT)
3251 local_irq_disable();
Christoph Lameter3c517a62006-12-06 20:33:29 -08003252 if (obj) {
3253 /*
3254 * Insert into the appropriate per node queues
3255 */
3256 nid = page_to_nid(virt_to_page(obj));
3257 if (cache_grow(cache, flags, nid, obj)) {
3258 obj = ____cache_alloc_node(cache,
3259 flags | GFP_THISNODE, nid);
3260 if (!obj)
3261 /*
3262 * Another processor may allocate the
3263 * objects in the slab since we are
3264 * not holding any locks.
3265 */
3266 goto retry;
3267 } else {
Hugh Dickinsb6a60452007-01-05 16:36:36 -08003268 /* cache_grow already freed obj */
Christoph Lameter3c517a62006-12-06 20:33:29 -08003269 obj = NULL;
3270 }
3271 }
Christoph Lameteraedb0eb2006-10-21 10:24:16 -07003272 }
Christoph Lameter765c4502006-09-27 01:50:08 -07003273 return obj;
3274}
3275
3276/*
Christoph Lametere498be72005-09-09 13:03:32 -07003277 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07003278 */
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003279static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003280 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07003281{
3282 struct list_head *entry;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003283 struct slab *slabp;
3284 struct kmem_list3 *l3;
3285 void *obj;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003286 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003287
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003288 l3 = cachep->nodelists[nodeid];
3289 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07003290
Andrew Mortona737b3e2006-03-22 00:08:11 -08003291retry:
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003292 check_irq_off();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003293 spin_lock(&l3->list_lock);
3294 entry = l3->slabs_partial.next;
3295 if (entry == &l3->slabs_partial) {
3296 l3->free_touched = 1;
3297 entry = l3->slabs_free.next;
3298 if (entry == &l3->slabs_free)
3299 goto must_grow;
3300 }
Christoph Lametere498be72005-09-09 13:03:32 -07003301
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003302 slabp = list_entry(entry, struct slab, list);
3303 check_spinlock_acquired_node(cachep, nodeid);
3304 check_slabp(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07003305
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003306 STATS_INC_NODEALLOCS(cachep);
3307 STATS_INC_ACTIVE(cachep);
3308 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003309
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003310 BUG_ON(slabp->inuse == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07003311
Matthew Dobson78d382d2006-02-01 03:05:47 -08003312 obj = slab_get_obj(cachep, slabp, nodeid);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003313 check_slabp(cachep, slabp);
3314 l3->free_objects--;
3315 /* move slabp to correct slabp list: */
3316 list_del(&slabp->list);
Christoph Lametere498be72005-09-09 13:03:32 -07003317
Andrew Mortona737b3e2006-03-22 00:08:11 -08003318 if (slabp->free == BUFCTL_END)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003319 list_add(&slabp->list, &l3->slabs_full);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003320 else
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003321 list_add(&slabp->list, &l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07003322
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003323 spin_unlock(&l3->list_lock);
3324 goto done;
Christoph Lametere498be72005-09-09 13:03:32 -07003325
Andrew Mortona737b3e2006-03-22 00:08:11 -08003326must_grow:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003327 spin_unlock(&l3->list_lock);
Christoph Lameter3c517a62006-12-06 20:33:29 -08003328 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
Christoph Lameter765c4502006-09-27 01:50:08 -07003329 if (x)
3330 goto retry;
Christoph Lametere498be72005-09-09 13:03:32 -07003331
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003332 return fallback_alloc(cachep, flags);
Christoph Lameter765c4502006-09-27 01:50:08 -07003333
Andrew Mortona737b3e2006-03-22 00:08:11 -08003334done:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003335 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07003336}
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003337
3338/**
3339 * kmem_cache_alloc_node - Allocate an object on the specified node
3340 * @cachep: The cache to allocate from.
3341 * @flags: See kmalloc().
3342 * @nodeid: node number of the target node.
3343 * @caller: return address of caller, used for debug information
3344 *
3345 * Identical to kmem_cache_alloc but it will allocate memory on the given
3346 * node, which can improve the performance for cpu bound structures.
3347 *
3348 * Fallback to other node is possible if __GFP_THISNODE is not set.
3349 */
3350static __always_inline void *
3351__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3352 void *caller)
3353{
3354 unsigned long save_flags;
3355 void *ptr;
3356
Nick Piggincf40bd12009-01-21 08:12:39 +01003357 lockdep_trace_alloc(flags);
3358
Akinobu Mita773ff602008-12-23 19:37:01 +09003359 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003360 return NULL;
3361
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003362 cache_alloc_debugcheck_before(cachep, flags);
3363 local_irq_save(save_flags);
3364
3365 if (unlikely(nodeid == -1))
3366 nodeid = numa_node_id();
3367
3368 if (unlikely(!cachep->nodelists[nodeid])) {
3369 /* Node not bootstrapped yet */
3370 ptr = fallback_alloc(cachep, flags);
3371 goto out;
3372 }
3373
3374 if (nodeid == numa_node_id()) {
3375 /*
3376 * Use the locally cached objects if possible.
3377 * However ____cache_alloc does not allow fallback
3378 * to other nodes. It may fail while we still have
3379 * objects on other nodes available.
3380 */
3381 ptr = ____cache_alloc(cachep, flags);
3382 if (ptr)
3383 goto out;
3384 }
3385 /* ___cache_alloc_node can fall back to other nodes */
3386 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3387 out:
3388 local_irq_restore(save_flags);
3389 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
Catalin Marinasd5cff632009-06-11 13:22:40 +01003390 kmemleak_alloc_recursive(ptr, obj_size(cachep), 1, cachep->flags,
3391 flags);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003392
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003393 if (unlikely((flags & __GFP_ZERO) && ptr))
3394 memset(ptr, 0, obj_size(cachep));
3395
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003396 return ptr;
3397}
3398
3399static __always_inline void *
3400__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3401{
3402 void *objp;
3403
3404 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3405 objp = alternate_node_alloc(cache, flags);
3406 if (objp)
3407 goto out;
3408 }
3409 objp = ____cache_alloc(cache, flags);
3410
3411 /*
3412 * We may just have run out of memory on the local node.
3413 * ____cache_alloc_node() knows how to locate memory on other nodes
3414 */
3415 if (!objp)
3416 objp = ____cache_alloc_node(cache, flags, numa_node_id());
3417
3418 out:
3419 return objp;
3420}
3421#else
3422
3423static __always_inline void *
3424__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3425{
3426 return ____cache_alloc(cachep, flags);
3427}
3428
3429#endif /* CONFIG_NUMA */
3430
3431static __always_inline void *
3432__cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3433{
3434 unsigned long save_flags;
3435 void *objp;
3436
Nick Piggincf40bd12009-01-21 08:12:39 +01003437 lockdep_trace_alloc(flags);
3438
Akinobu Mita773ff602008-12-23 19:37:01 +09003439 if (slab_should_failslab(cachep, flags))
Akinobu Mita824ebef2007-05-06 14:49:58 -07003440 return NULL;
3441
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003442 cache_alloc_debugcheck_before(cachep, flags);
3443 local_irq_save(save_flags);
3444 objp = __do_cache_alloc(cachep, flags);
3445 local_irq_restore(save_flags);
3446 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
Catalin Marinasd5cff632009-06-11 13:22:40 +01003447 kmemleak_alloc_recursive(objp, obj_size(cachep), 1, cachep->flags,
3448 flags);
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003449 prefetchw(objp);
3450
Christoph Lameterd07dbea2007-07-17 04:03:23 -07003451 if (unlikely((flags & __GFP_ZERO) && objp))
3452 memset(objp, 0, obj_size(cachep));
3453
Pekka Enberg8c8cc2c2007-02-10 01:42:53 -08003454 return objp;
3455}
Christoph Lametere498be72005-09-09 13:03:32 -07003456
3457/*
3458 * Caller needs to acquire correct kmem_list's list_lock
3459 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003460static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003461 int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003462{
3463 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07003464 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003465
3466 for (i = 0; i < nr_objects; i++) {
3467 void *objp = objpp[i];
3468 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003469
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08003470 slabp = virt_to_slab(objp);
Christoph Lameterff694162005-09-22 21:44:02 -07003471 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003472 list_del(&slabp->list);
Christoph Lameterff694162005-09-22 21:44:02 -07003473 check_spinlock_acquired_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003474 check_slabp(cachep, slabp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08003475 slab_put_obj(cachep, slabp, objp, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003476 STATS_DEC_ACTIVE(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003477 l3->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003478 check_slabp(cachep, slabp);
3479
3480 /* fixup slab chains */
3481 if (slabp->inuse == 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07003482 if (l3->free_objects > l3->free_limit) {
3483 l3->free_objects -= cachep->num;
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07003484 /* No need to drop any previously held
3485 * lock here, even if we have a off-slab slab
3486 * descriptor it is guaranteed to come from
3487 * a different cache, refer to comments before
3488 * alloc_slabmgmt.
3489 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003490 slab_destroy(cachep, slabp);
3491 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07003492 list_add(&slabp->list, &l3->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003493 }
3494 } else {
3495 /* Unconditionally move a slab to the end of the
3496 * partial list on free - maximum time for the
3497 * other objects to be freed, too.
3498 */
Christoph Lametere498be72005-09-09 13:03:32 -07003499 list_add_tail(&slabp->list, &l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003500 }
3501 }
3502}
3503
Pekka Enberg343e0d72006-02-01 03:05:50 -08003504static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003505{
3506 int batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07003507 struct kmem_list3 *l3;
Christoph Lameterff694162005-09-22 21:44:02 -07003508 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003509
3510 batchcount = ac->batchcount;
3511#if DEBUG
3512 BUG_ON(!batchcount || batchcount > ac->avail);
3513#endif
3514 check_irq_off();
Christoph Lameterff694162005-09-22 21:44:02 -07003515 l3 = cachep->nodelists[node];
Ingo Molnar873623d2006-07-13 14:44:38 +02003516 spin_lock(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003517 if (l3->shared) {
3518 struct array_cache *shared_array = l3->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003519 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003520 if (max) {
3521 if (batchcount > max)
3522 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07003523 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003524 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003525 shared_array->avail += batchcount;
3526 goto free_done;
3527 }
3528 }
3529
Christoph Lameterff694162005-09-22 21:44:02 -07003530 free_block(cachep, ac->entry, batchcount, node);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003531free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003532#if STATS
3533 {
3534 int i = 0;
3535 struct list_head *p;
3536
Christoph Lametere498be72005-09-09 13:03:32 -07003537 p = l3->slabs_free.next;
3538 while (p != &(l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003539 struct slab *slabp;
3540
3541 slabp = list_entry(p, struct slab, list);
3542 BUG_ON(slabp->inuse);
3543
3544 i++;
3545 p = p->next;
3546 }
3547 STATS_SET_FREEABLE(cachep, i);
3548 }
3549#endif
Christoph Lametere498be72005-09-09 13:03:32 -07003550 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003551 ac->avail -= batchcount;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003552 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003553}
3554
3555/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08003556 * Release an obj back to its cache. If the obj has a constructed state, it must
3557 * be in this state _before_ it is released. Called with disabled ints.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003558 */
Ingo Molnar873623d2006-07-13 14:44:38 +02003559static inline void __cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003560{
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003561 struct array_cache *ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003562
3563 check_irq_off();
Catalin Marinasd5cff632009-06-11 13:22:40 +01003564 kmemleak_free_recursive(objp, cachep->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003565 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3566
Siddha, Suresh B1807a1a2007-08-22 14:01:49 -07003567 /*
3568 * Skip calling cache_free_alien() when the platform is not numa.
3569 * This will avoid cache misses that happen while accessing slabp (which
3570 * is per page memory reference) to get nodeid. Instead use a global
3571 * variable to skip the call, which is mostly likely to be present in
3572 * the cache.
3573 */
3574 if (numa_platform && cache_free_alien(cachep, objp))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07003575 return;
Christoph Lametere498be72005-09-09 13:03:32 -07003576
Linus Torvalds1da177e2005-04-16 15:20:36 -07003577 if (likely(ac->avail < ac->limit)) {
3578 STATS_INC_FREEHIT(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003579 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003580 return;
3581 } else {
3582 STATS_INC_FREEMISS(cachep);
3583 cache_flusharray(cachep, ac);
Christoph Lametere498be72005-09-09 13:03:32 -07003584 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003585 }
3586}
3587
3588/**
3589 * kmem_cache_alloc - Allocate an object
3590 * @cachep: The cache to allocate from.
3591 * @flags: See kmalloc().
3592 *
3593 * Allocate an object from this cache. The flags are only relevant
3594 * if the cache has no available objects.
3595 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003596void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003597{
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003598 void *ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3599
Eduard - Gabriel Munteanuca2b84c2009-03-23 15:12:24 +02003600 trace_kmem_cache_alloc(_RET_IP_, ret,
3601 obj_size(cachep), cachep->buffer_size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003602
3603 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003604}
3605EXPORT_SYMBOL(kmem_cache_alloc);
3606
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003607#ifdef CONFIG_KMEMTRACE
3608void *kmem_cache_alloc_notrace(struct kmem_cache *cachep, gfp_t flags)
3609{
3610 return __cache_alloc(cachep, flags, __builtin_return_address(0));
3611}
3612EXPORT_SYMBOL(kmem_cache_alloc_notrace);
3613#endif
3614
Linus Torvalds1da177e2005-04-16 15:20:36 -07003615/**
Randy Dunlap76824862008-03-19 17:00:40 -07003616 * kmem_ptr_validate - check if an untrusted pointer might be a slab entry.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003617 * @cachep: the cache we're checking against
3618 * @ptr: pointer to validate
3619 *
Randy Dunlap76824862008-03-19 17:00:40 -07003620 * This verifies that the untrusted pointer looks sane;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003621 * it is _not_ a guarantee that the pointer is actually
3622 * part of the slab cache in question, but it at least
3623 * validates that the pointer can be dereferenced and
3624 * looks half-way sane.
3625 *
3626 * Currently only used for dentry validation.
3627 */
Christoph Lameterb7f869a2006-12-22 01:06:44 -08003628int kmem_ptr_validate(struct kmem_cache *cachep, const void *ptr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003629{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003630 unsigned long addr = (unsigned long)ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003631 unsigned long min_addr = PAGE_OFFSET;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003632 unsigned long align_mask = BYTES_PER_WORD - 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003633 unsigned long size = cachep->buffer_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003634 struct page *page;
3635
3636 if (unlikely(addr < min_addr))
3637 goto out;
3638 if (unlikely(addr > (unsigned long)high_memory - size))
3639 goto out;
3640 if (unlikely(addr & align_mask))
3641 goto out;
3642 if (unlikely(!kern_addr_valid(addr)))
3643 goto out;
3644 if (unlikely(!kern_addr_valid(addr + size - 1)))
3645 goto out;
3646 page = virt_to_page(ptr);
3647 if (unlikely(!PageSlab(page)))
3648 goto out;
Pekka Enberg065d41c2005-11-13 16:06:46 -08003649 if (unlikely(page_get_cache(page) != cachep))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003650 goto out;
3651 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003652out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003653 return 0;
3654}
3655
3656#ifdef CONFIG_NUMA
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003657void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3658{
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003659 void *ret = __cache_alloc_node(cachep, flags, nodeid,
3660 __builtin_return_address(0));
3661
Eduard - Gabriel Munteanuca2b84c2009-03-23 15:12:24 +02003662 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3663 obj_size(cachep), cachep->buffer_size,
3664 flags, nodeid);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003665
3666 return ret;
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003667}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003668EXPORT_SYMBOL(kmem_cache_alloc_node);
3669
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003670#ifdef CONFIG_KMEMTRACE
3671void *kmem_cache_alloc_node_notrace(struct kmem_cache *cachep,
3672 gfp_t flags,
3673 int nodeid)
3674{
3675 return __cache_alloc_node(cachep, flags, nodeid,
3676 __builtin_return_address(0));
3677}
3678EXPORT_SYMBOL(kmem_cache_alloc_node_notrace);
3679#endif
3680
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003681static __always_inline void *
3682__do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003683{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003684 struct kmem_cache *cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003685 void *ret;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003686
3687 cachep = kmem_find_general_cachep(size, flags);
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003688 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3689 return cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003690 ret = kmem_cache_alloc_node_notrace(cachep, flags, node);
3691
Eduard - Gabriel Munteanuca2b84c2009-03-23 15:12:24 +02003692 trace_kmalloc_node((unsigned long) caller, ret,
3693 size, cachep->buffer_size, flags, node);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003694
3695 return ret;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003696}
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003697
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003698#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_KMEMTRACE)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003699void *__kmalloc_node(size_t size, gfp_t flags, int node)
3700{
3701 return __do_kmalloc_node(size, flags, node,
3702 __builtin_return_address(0));
3703}
Christoph Hellwigdbe5e692006-09-25 23:31:36 -07003704EXPORT_SYMBOL(__kmalloc_node);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003705
3706void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003707 int node, unsigned long caller)
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003708{
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003709 return __do_kmalloc_node(size, flags, node, (void *)caller);
Christoph Hellwig8b98c162006-12-06 20:32:30 -08003710}
3711EXPORT_SYMBOL(__kmalloc_node_track_caller);
3712#else
3713void *__kmalloc_node(size_t size, gfp_t flags, int node)
3714{
3715 return __do_kmalloc_node(size, flags, node, NULL);
3716}
3717EXPORT_SYMBOL(__kmalloc_node);
3718#endif /* CONFIG_DEBUG_SLAB */
3719#endif /* CONFIG_NUMA */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003720
3721/**
Paul Drynoff800590f2006-06-23 02:03:48 -07003722 * __do_kmalloc - allocate memory
Linus Torvalds1da177e2005-04-16 15:20:36 -07003723 * @size: how many bytes of memory are required.
Paul Drynoff800590f2006-06-23 02:03:48 -07003724 * @flags: the type of memory to allocate (see kmalloc).
Randy Dunlap911851e2006-03-22 00:08:14 -08003725 * @caller: function caller for debug tracking of the caller
Linus Torvalds1da177e2005-04-16 15:20:36 -07003726 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003727static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3728 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003729{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003730 struct kmem_cache *cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003731 void *ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003732
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003733 /* If you want to save a few bytes .text space: replace
3734 * __ with kmem_.
3735 * Then kmalloc uses the uninlined functions instead of the inline
3736 * functions.
3737 */
3738 cachep = __find_general_cachep(size, flags);
Linus Torvaldsa5c96d82007-07-19 13:17:15 -07003739 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3740 return cachep;
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003741 ret = __cache_alloc(cachep, flags, caller);
3742
Eduard - Gabriel Munteanuca2b84c2009-03-23 15:12:24 +02003743 trace_kmalloc((unsigned long) caller, ret,
3744 size, cachep->buffer_size, flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003745
3746 return ret;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003747}
3748
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003749
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003750#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_KMEMTRACE)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003751void *__kmalloc(size_t size, gfp_t flags)
3752{
Al Viro871751e2006-03-25 03:06:39 -08003753 return __do_kmalloc(size, flags, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003754}
3755EXPORT_SYMBOL(__kmalloc);
3756
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003757void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003758{
Eduard - Gabriel Munteanuce71e272008-08-19 20:43:25 +03003759 return __do_kmalloc(size, flags, (void *)caller);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003760}
3761EXPORT_SYMBOL(__kmalloc_track_caller);
Christoph Hellwig1d2c8ee2006-10-04 02:15:25 -07003762
3763#else
3764void *__kmalloc(size_t size, gfp_t flags)
3765{
3766 return __do_kmalloc(size, flags, NULL);
3767}
3768EXPORT_SYMBOL(__kmalloc);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003769#endif
3770
Linus Torvalds1da177e2005-04-16 15:20:36 -07003771/**
3772 * kmem_cache_free - Deallocate an object
3773 * @cachep: The cache the allocation was from.
3774 * @objp: The previously allocated object.
3775 *
3776 * Free an object which was previously allocated from this
3777 * cache.
3778 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003779void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003780{
3781 unsigned long flags;
3782
3783 local_irq_save(flags);
Ingo Molnar898552c2007-02-10 01:44:57 -08003784 debug_check_no_locks_freed(objp, obj_size(cachep));
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07003785 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3786 debug_check_no_obj_freed(objp, obj_size(cachep));
Ingo Molnar873623d2006-07-13 14:44:38 +02003787 __cache_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003788 local_irq_restore(flags);
Eduard - Gabriel Munteanu36555752008-08-10 20:14:05 +03003789
Eduard - Gabriel Munteanuca2b84c2009-03-23 15:12:24 +02003790 trace_kmem_cache_free(_RET_IP_, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003791}
3792EXPORT_SYMBOL(kmem_cache_free);
3793
3794/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003795 * kfree - free previously allocated memory
3796 * @objp: pointer returned by kmalloc.
3797 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003798 * If @objp is NULL, no operation is performed.
3799 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003800 * Don't free memory not originally allocated by kmalloc()
3801 * or you will run into trouble.
3802 */
3803void kfree(const void *objp)
3804{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003805 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003806 unsigned long flags;
3807
Pekka Enberg2121db72009-03-25 11:05:57 +02003808 trace_kfree(_RET_IP_, objp);
3809
Christoph Lameter6cb8f912007-07-17 04:03:22 -07003810 if (unlikely(ZERO_OR_NULL_PTR(objp)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003811 return;
3812 local_irq_save(flags);
3813 kfree_debugcheck(objp);
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08003814 c = virt_to_cache(objp);
Ingo Molnarf9b84042006-06-27 02:54:49 -07003815 debug_check_no_locks_freed(objp, obj_size(c));
Thomas Gleixner3ac7fe52008-04-30 00:55:01 -07003816 debug_check_no_obj_freed(objp, obj_size(c));
Ingo Molnar873623d2006-07-13 14:44:38 +02003817 __cache_free(c, (void *)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003818 local_irq_restore(flags);
3819}
3820EXPORT_SYMBOL(kfree);
3821
Pekka Enberg343e0d72006-02-01 03:05:50 -08003822unsigned int kmem_cache_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003823{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003824 return obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003825}
3826EXPORT_SYMBOL(kmem_cache_size);
3827
Pekka Enberg343e0d72006-02-01 03:05:50 -08003828const char *kmem_cache_name(struct kmem_cache *cachep)
Arnaldo Carvalho de Melo19449722005-06-18 22:46:19 -07003829{
3830 return cachep->name;
3831}
3832EXPORT_SYMBOL_GPL(kmem_cache_name);
3833
Christoph Lametere498be72005-09-09 13:03:32 -07003834/*
Simon Arlott183ff222007-10-20 01:27:18 +02003835 * This initializes kmem_list3 or resizes various caches for all nodes.
Christoph Lametere498be72005-09-09 13:03:32 -07003836 */
Pekka Enberg83b519e2009-06-10 19:40:04 +03003837static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
Christoph Lametere498be72005-09-09 13:03:32 -07003838{
3839 int node;
3840 struct kmem_list3 *l3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003841 struct array_cache *new_shared;
Paul Menage3395ee02006-12-06 20:32:16 -08003842 struct array_cache **new_alien = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003843
Mel Gorman9c09a952008-01-24 05:49:54 -08003844 for_each_online_node(node) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003845
Paul Menage3395ee02006-12-06 20:32:16 -08003846 if (use_alien_caches) {
Pekka Enberg83b519e2009-06-10 19:40:04 +03003847 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
Paul Menage3395ee02006-12-06 20:32:16 -08003848 if (!new_alien)
3849 goto fail;
3850 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003851
Eric Dumazet63109842007-05-06 14:49:28 -07003852 new_shared = NULL;
3853 if (cachep->shared) {
3854 new_shared = alloc_arraycache(node,
Christoph Lameter0718dc22006-03-25 03:06:47 -08003855 cachep->shared*cachep->batchcount,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003856 0xbaadf00d, gfp);
Eric Dumazet63109842007-05-06 14:49:28 -07003857 if (!new_shared) {
3858 free_alien_cache(new_alien);
3859 goto fail;
3860 }
Christoph Lameter0718dc22006-03-25 03:06:47 -08003861 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003862
Andrew Mortona737b3e2006-03-22 00:08:11 -08003863 l3 = cachep->nodelists[node];
3864 if (l3) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003865 struct array_cache *shared = l3->shared;
3866
Christoph Lametere498be72005-09-09 13:03:32 -07003867 spin_lock_irq(&l3->list_lock);
3868
Christoph Lametercafeb022006-03-25 03:06:46 -08003869 if (shared)
Christoph Lameter0718dc22006-03-25 03:06:47 -08003870 free_block(cachep, shared->entry,
3871 shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07003872
Christoph Lametercafeb022006-03-25 03:06:46 -08003873 l3->shared = new_shared;
3874 if (!l3->alien) {
Christoph Lametere498be72005-09-09 13:03:32 -07003875 l3->alien = new_alien;
3876 new_alien = NULL;
3877 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003878 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003879 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003880 spin_unlock_irq(&l3->list_lock);
Christoph Lametercafeb022006-03-25 03:06:46 -08003881 kfree(shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003882 free_alien_cache(new_alien);
3883 continue;
3884 }
Pekka Enberg83b519e2009-06-10 19:40:04 +03003885 l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node);
Christoph Lameter0718dc22006-03-25 03:06:47 -08003886 if (!l3) {
3887 free_alien_cache(new_alien);
3888 kfree(new_shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003889 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003890 }
Christoph Lametere498be72005-09-09 13:03:32 -07003891
3892 kmem_list3_init(l3);
3893 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Andrew Mortona737b3e2006-03-22 00:08:11 -08003894 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003895 l3->shared = new_shared;
Christoph Lametere498be72005-09-09 13:03:32 -07003896 l3->alien = new_alien;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003897 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003898 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003899 cachep->nodelists[node] = l3;
3900 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003901 return 0;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003902
Andrew Mortona737b3e2006-03-22 00:08:11 -08003903fail:
Christoph Lameter0718dc22006-03-25 03:06:47 -08003904 if (!cachep->next.next) {
3905 /* Cache is not active yet. Roll back what we did */
3906 node--;
3907 while (node >= 0) {
3908 if (cachep->nodelists[node]) {
3909 l3 = cachep->nodelists[node];
3910
3911 kfree(l3->shared);
3912 free_alien_cache(l3->alien);
3913 kfree(l3);
3914 cachep->nodelists[node] = NULL;
3915 }
3916 node--;
3917 }
3918 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003919 return -ENOMEM;
Christoph Lametere498be72005-09-09 13:03:32 -07003920}
3921
Linus Torvalds1da177e2005-04-16 15:20:36 -07003922struct ccupdate_struct {
Pekka Enberg343e0d72006-02-01 03:05:50 -08003923 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003924 struct array_cache *new[NR_CPUS];
3925};
3926
3927static void do_ccupdate_local(void *info)
3928{
Andrew Mortona737b3e2006-03-22 00:08:11 -08003929 struct ccupdate_struct *new = info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003930 struct array_cache *old;
3931
3932 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003933 old = cpu_cache_get(new->cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003934
Linus Torvalds1da177e2005-04-16 15:20:36 -07003935 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3936 new->new[smp_processor_id()] = old;
3937}
3938
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003939/* Always called with the cache_chain_mutex held */
Andrew Mortona737b3e2006-03-22 00:08:11 -08003940static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003941 int batchcount, int shared, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003942{
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003943 struct ccupdate_struct *new;
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07003944 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003945
Pekka Enberg83b519e2009-06-10 19:40:04 +03003946 new = kzalloc(sizeof(*new), gfp);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003947 if (!new)
3948 return -ENOMEM;
3949
Christoph Lametere498be72005-09-09 13:03:32 -07003950 for_each_online_cpu(i) {
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003951 new->new[i] = alloc_arraycache(cpu_to_node(i), limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03003952 batchcount, gfp);
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003953 if (!new->new[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003954 for (i--; i >= 0; i--)
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003955 kfree(new->new[i]);
3956 kfree(new);
Christoph Lametere498be72005-09-09 13:03:32 -07003957 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003958 }
3959 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003960 new->cachep = cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003961
Jens Axboe15c8b6c2008-05-09 09:39:44 +02003962 on_each_cpu(do_ccupdate_local, (void *)new, 1);
Christoph Lametere498be72005-09-09 13:03:32 -07003963
Linus Torvalds1da177e2005-04-16 15:20:36 -07003964 check_irq_on();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003965 cachep->batchcount = batchcount;
3966 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07003967 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003968
Christoph Lametere498be72005-09-09 13:03:32 -07003969 for_each_online_cpu(i) {
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003970 struct array_cache *ccold = new->new[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003971 if (!ccold)
3972 continue;
Christoph Lametere498be72005-09-09 13:03:32 -07003973 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -07003974 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
Christoph Lametere498be72005-09-09 13:03:32 -07003975 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003976 kfree(ccold);
3977 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07003978 kfree(new);
Pekka Enberg83b519e2009-06-10 19:40:04 +03003979 return alloc_kmemlist(cachep, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003980}
3981
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003982/* Called with cache_chain_mutex held always */
Pekka Enberg83b519e2009-06-10 19:40:04 +03003983static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003984{
3985 int err;
3986 int limit, shared;
3987
Andrew Mortona737b3e2006-03-22 00:08:11 -08003988 /*
3989 * The head array serves three purposes:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003990 * - create a LIFO ordering, i.e. return objects that are cache-warm
3991 * - reduce the number of spinlock operations.
Andrew Mortona737b3e2006-03-22 00:08:11 -08003992 * - reduce the number of linked list operations on the slab and
Linus Torvalds1da177e2005-04-16 15:20:36 -07003993 * bufctl chains: array operations are cheaper.
3994 * The numbers are guessed, we should auto-tune as described by
3995 * Bonwick.
3996 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003997 if (cachep->buffer_size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003998 limit = 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003999 else if (cachep->buffer_size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004000 limit = 8;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004001 else if (cachep->buffer_size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004002 limit = 24;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004003 else if (cachep->buffer_size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004004 limit = 54;
4005 else
4006 limit = 120;
4007
Andrew Mortona737b3e2006-03-22 00:08:11 -08004008 /*
4009 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
Linus Torvalds1da177e2005-04-16 15:20:36 -07004010 * allocation behaviour: Most allocs on one cpu, most free operations
4011 * on another cpu. For these cases, an efficient object passing between
4012 * cpus is necessary. This is provided by a shared array. The array
4013 * replaces Bonwick's magazine layer.
4014 * On uniprocessor, it's functionally equivalent (but less efficient)
4015 * to a larger limit. Thus disabled by default.
4016 */
4017 shared = 0;
Eric Dumazet364fbb22007-05-06 14:49:27 -07004018 if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004019 shared = 8;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004020
4021#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08004022 /*
4023 * With debugging enabled, large batchcount lead to excessively long
4024 * periods with disabled local interrupts. Limit the batchcount
Linus Torvalds1da177e2005-04-16 15:20:36 -07004025 */
4026 if (limit > 32)
4027 limit = 32;
4028#endif
Pekka Enberg83b519e2009-06-10 19:40:04 +03004029 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004030 if (err)
4031 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004032 cachep->name, -err);
Christoph Lameter2ed3a4e2006-09-25 23:31:38 -07004033 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004034}
4035
Christoph Lameter1b552532006-03-22 00:09:07 -08004036/*
4037 * Drain an array if it contains any elements taking the l3 lock only if
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004038 * necessary. Note that the l3 listlock also protects the array_cache
4039 * if drain_array() is used on the shared array.
Christoph Lameter1b552532006-03-22 00:09:07 -08004040 */
4041void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
4042 struct array_cache *ac, int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004043{
4044 int tofree;
4045
Christoph Lameter1b552532006-03-22 00:09:07 -08004046 if (!ac || !ac->avail)
4047 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004048 if (ac->touched && !force) {
4049 ac->touched = 0;
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004050 } else {
Christoph Lameter1b552532006-03-22 00:09:07 -08004051 spin_lock_irq(&l3->list_lock);
Christoph Lameterb18e7e62006-03-22 00:09:07 -08004052 if (ac->avail) {
4053 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4054 if (tofree > ac->avail)
4055 tofree = (ac->avail + 1) / 2;
4056 free_block(cachep, ac->entry, tofree, node);
4057 ac->avail -= tofree;
4058 memmove(ac->entry, &(ac->entry[tofree]),
4059 sizeof(void *) * ac->avail);
4060 }
Christoph Lameter1b552532006-03-22 00:09:07 -08004061 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004062 }
4063}
4064
4065/**
4066 * cache_reap - Reclaim memory from caches.
Randy Dunlap05fb6bf2007-02-28 20:12:13 -08004067 * @w: work descriptor
Linus Torvalds1da177e2005-04-16 15:20:36 -07004068 *
4069 * Called from workqueue/eventd every few seconds.
4070 * Purpose:
4071 * - clear the per-cpu caches for this CPU.
4072 * - return freeable pages to the main free memory pool.
4073 *
Andrew Mortona737b3e2006-03-22 00:08:11 -08004074 * If we cannot acquire the cache chain mutex then just give up - we'll try
4075 * again on the next iteration.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004076 */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004077static void cache_reap(struct work_struct *w)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004078{
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004079 struct kmem_cache *searchp;
Christoph Lametere498be72005-09-09 13:03:32 -07004080 struct kmem_list3 *l3;
Christoph Lameteraab22072006-03-22 00:09:06 -08004081 int node = numa_node_id();
Jean Delvarebf6aede2009-04-02 16:56:54 -07004082 struct delayed_work *work = to_delayed_work(w);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004083
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004084 if (!mutex_trylock(&cache_chain_mutex))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004085 /* Give up. Setup the next iteration. */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004086 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004087
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004088 list_for_each_entry(searchp, &cache_chain, next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004089 check_irq_on();
4090
Christoph Lameter35386e32006-03-22 00:09:05 -08004091 /*
4092 * We only take the l3 lock if absolutely necessary and we
4093 * have established with reasonable certainty that
4094 * we can do some work if the lock was obtained.
4095 */
Christoph Lameteraab22072006-03-22 00:09:06 -08004096 l3 = searchp->nodelists[node];
Christoph Lameter35386e32006-03-22 00:09:05 -08004097
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004098 reap_alien(searchp, l3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004099
Christoph Lameteraab22072006-03-22 00:09:06 -08004100 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004101
Christoph Lameter35386e32006-03-22 00:09:05 -08004102 /*
4103 * These are racy checks but it does not matter
4104 * if we skip one check or scan twice.
4105 */
Christoph Lametere498be72005-09-09 13:03:32 -07004106 if (time_after(l3->next_reap, jiffies))
Christoph Lameter35386e32006-03-22 00:09:05 -08004107 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004108
Christoph Lametere498be72005-09-09 13:03:32 -07004109 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004110
Christoph Lameteraab22072006-03-22 00:09:06 -08004111 drain_array(searchp, l3, l3->shared, 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004112
Christoph Lametered11d9e2006-06-30 01:55:45 -07004113 if (l3->free_touched)
Christoph Lametere498be72005-09-09 13:03:32 -07004114 l3->free_touched = 0;
Christoph Lametered11d9e2006-06-30 01:55:45 -07004115 else {
4116 int freed;
4117
4118 freed = drain_freelist(searchp, l3, (l3->free_limit +
4119 5 * searchp->num - 1) / (5 * searchp->num));
4120 STATS_ADD_REAPED(searchp, freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004121 }
Christoph Lameter35386e32006-03-22 00:09:05 -08004122next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004123 cond_resched();
4124 }
4125 check_irq_on();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004126 mutex_unlock(&cache_chain_mutex);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08004127 next_reap_node();
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004128out:
Andrew Mortona737b3e2006-03-22 00:08:11 -08004129 /* Set up the next iteration */
Christoph Lameter7c5cae32007-02-10 01:42:55 -08004130 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004131}
4132
Linus Torvalds158a9622008-01-02 13:04:48 -08004133#ifdef CONFIG_SLABINFO
Linus Torvalds1da177e2005-04-16 15:20:36 -07004134
Pekka Enberg85289f92006-01-08 01:00:36 -08004135static void print_slabinfo_header(struct seq_file *m)
4136{
4137 /*
4138 * Output format version, so at least we can change it
4139 * without _too_ many complaints.
4140 */
4141#if STATS
4142 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4143#else
4144 seq_puts(m, "slabinfo - version: 2.1\n");
4145#endif
4146 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4147 "<objperslab> <pagesperslab>");
4148 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4149 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4150#if STATS
4151 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004152 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
Pekka Enberg85289f92006-01-08 01:00:36 -08004153 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4154#endif
4155 seq_putc(m, '\n');
4156}
4157
Linus Torvalds1da177e2005-04-16 15:20:36 -07004158static void *s_start(struct seq_file *m, loff_t *pos)
4159{
4160 loff_t n = *pos;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004161
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004162 mutex_lock(&cache_chain_mutex);
Pekka Enberg85289f92006-01-08 01:00:36 -08004163 if (!n)
4164 print_slabinfo_header(m);
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004165
4166 return seq_list_start(&cache_chain, *pos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004167}
4168
4169static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4170{
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004171 return seq_list_next(p, &cache_chain, pos);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004172}
4173
4174static void s_stop(struct seq_file *m, void *p)
4175{
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004176 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004177}
4178
4179static int s_show(struct seq_file *m, void *p)
4180{
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004181 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004182 struct slab *slabp;
4183 unsigned long active_objs;
4184 unsigned long num_objs;
4185 unsigned long active_slabs = 0;
4186 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004187 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004188 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07004189 int node;
4190 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004191
Linus Torvalds1da177e2005-04-16 15:20:36 -07004192 active_objs = 0;
4193 num_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07004194 for_each_online_node(node) {
4195 l3 = cachep->nodelists[node];
4196 if (!l3)
4197 continue;
4198
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004199 check_irq_on();
4200 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07004201
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004202 list_for_each_entry(slabp, &l3->slabs_full, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004203 if (slabp->inuse != cachep->num && !error)
4204 error = "slabs_full accounting error";
4205 active_objs += cachep->num;
4206 active_slabs++;
4207 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004208 list_for_each_entry(slabp, &l3->slabs_partial, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004209 if (slabp->inuse == cachep->num && !error)
4210 error = "slabs_partial inuse accounting error";
4211 if (!slabp->inuse && !error)
4212 error = "slabs_partial/inuse accounting error";
4213 active_objs += slabp->inuse;
4214 active_slabs++;
4215 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004216 list_for_each_entry(slabp, &l3->slabs_free, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07004217 if (slabp->inuse && !error)
4218 error = "slabs_free/inuse accounting error";
4219 num_slabs++;
4220 }
4221 free_objects += l3->free_objects;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08004222 if (l3->shared)
4223 shared_avail += l3->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07004224
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08004225 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004226 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004227 num_slabs += active_slabs;
4228 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07004229 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004230 error = "free_objects accounting error";
4231
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004232 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004233 if (error)
4234 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4235
4236 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08004237 name, active_objs, num_objs, cachep->buffer_size,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004238 cachep->num, (1 << cachep->gfporder));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004239 seq_printf(m, " : tunables %4u %4u %4u",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004240 cachep->limit, cachep->batchcount, cachep->shared);
Christoph Lametere498be72005-09-09 13:03:32 -07004241 seq_printf(m, " : slabdata %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004242 active_slabs, num_slabs, shared_avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004243#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004244 { /* list3 stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004245 unsigned long high = cachep->high_mark;
4246 unsigned long allocs = cachep->num_allocations;
4247 unsigned long grown = cachep->grown;
4248 unsigned long reaped = cachep->reaped;
4249 unsigned long errors = cachep->errors;
4250 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004251 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07004252 unsigned long node_frees = cachep->node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004253 unsigned long overflows = cachep->node_overflow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004254
Christoph Lametere498be72005-09-09 13:03:32 -07004255 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004256 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown,
Andrew Mortona737b3e2006-03-22 00:08:11 -08004257 reaped, errors, max_freeable, node_allocs,
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004258 node_frees, overflows);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004259 }
4260 /* cpu stats */
4261 {
4262 unsigned long allochit = atomic_read(&cachep->allochit);
4263 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4264 unsigned long freehit = atomic_read(&cachep->freehit);
4265 unsigned long freemiss = atomic_read(&cachep->freemiss);
4266
4267 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004268 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004269 }
4270#endif
4271 seq_putc(m, '\n');
Linus Torvalds1da177e2005-04-16 15:20:36 -07004272 return 0;
4273}
4274
4275/*
4276 * slabinfo_op - iterator that generates /proc/slabinfo
4277 *
4278 * Output layout:
4279 * cache-name
4280 * num-active-objs
4281 * total-objs
4282 * object size
4283 * num-active-slabs
4284 * total-slabs
4285 * num-pages-per-slab
4286 * + further values on SMP and with statistics enabled
4287 */
4288
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04004289static const struct seq_operations slabinfo_op = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004290 .start = s_start,
4291 .next = s_next,
4292 .stop = s_stop,
4293 .show = s_show,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004294};
4295
4296#define MAX_SLABINFO_WRITE 128
4297/**
4298 * slabinfo_write - Tuning for the slab allocator
4299 * @file: unused
4300 * @buffer: user buffer
4301 * @count: data length
4302 * @ppos: unused
4303 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004304ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4305 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004306{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004307 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004308 int limit, batchcount, shared, res;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004309 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004310
Linus Torvalds1da177e2005-04-16 15:20:36 -07004311 if (count > MAX_SLABINFO_WRITE)
4312 return -EINVAL;
4313 if (copy_from_user(&kbuf, buffer, count))
4314 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004315 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07004316
4317 tmp = strchr(kbuf, ' ');
4318 if (!tmp)
4319 return -EINVAL;
4320 *tmp = '\0';
4321 tmp++;
4322 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4323 return -EINVAL;
4324
4325 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004326 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004327 res = -EINVAL;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004328 list_for_each_entry(cachep, &cache_chain, next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004329 if (!strcmp(cachep->name, kbuf)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08004330 if (limit < 1 || batchcount < 1 ||
4331 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07004332 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004333 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07004334 res = do_tune_cpucache(cachep, limit,
Pekka Enberg83b519e2009-06-10 19:40:04 +03004335 batchcount, shared,
4336 GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004337 }
4338 break;
4339 }
4340 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004341 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004342 if (res >= 0)
4343 res = count;
4344 return res;
4345}
Al Viro871751e2006-03-25 03:06:39 -08004346
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04004347static int slabinfo_open(struct inode *inode, struct file *file)
4348{
4349 return seq_open(file, &slabinfo_op);
4350}
4351
4352static const struct file_operations proc_slabinfo_operations = {
4353 .open = slabinfo_open,
4354 .read = seq_read,
4355 .write = slabinfo_write,
4356 .llseek = seq_lseek,
4357 .release = seq_release,
4358};
4359
Al Viro871751e2006-03-25 03:06:39 -08004360#ifdef CONFIG_DEBUG_SLAB_LEAK
4361
4362static void *leaks_start(struct seq_file *m, loff_t *pos)
4363{
Al Viro871751e2006-03-25 03:06:39 -08004364 mutex_lock(&cache_chain_mutex);
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004365 return seq_list_start(&cache_chain, *pos);
Al Viro871751e2006-03-25 03:06:39 -08004366}
4367
4368static inline int add_caller(unsigned long *n, unsigned long v)
4369{
4370 unsigned long *p;
4371 int l;
4372 if (!v)
4373 return 1;
4374 l = n[1];
4375 p = n + 2;
4376 while (l) {
4377 int i = l/2;
4378 unsigned long *q = p + 2 * i;
4379 if (*q == v) {
4380 q[1]++;
4381 return 1;
4382 }
4383 if (*q > v) {
4384 l = i;
4385 } else {
4386 p = q + 2;
4387 l -= i + 1;
4388 }
4389 }
4390 if (++n[1] == n[0])
4391 return 0;
4392 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4393 p[0] = v;
4394 p[1] = 1;
4395 return 1;
4396}
4397
4398static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4399{
4400 void *p;
4401 int i;
4402 if (n[0] == n[1])
4403 return;
4404 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4405 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4406 continue;
4407 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4408 return;
4409 }
4410}
4411
4412static void show_symbol(struct seq_file *m, unsigned long address)
4413{
4414#ifdef CONFIG_KALLSYMS
Al Viro871751e2006-03-25 03:06:39 -08004415 unsigned long offset, size;
Tejun Heo9281ace2007-07-17 04:03:51 -07004416 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
Al Viro871751e2006-03-25 03:06:39 -08004417
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004418 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
Al Viro871751e2006-03-25 03:06:39 -08004419 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
Alexey Dobriyana5c43da2007-05-08 00:28:47 -07004420 if (modname[0])
Al Viro871751e2006-03-25 03:06:39 -08004421 seq_printf(m, " [%s]", modname);
4422 return;
4423 }
4424#endif
4425 seq_printf(m, "%p", (void *)address);
4426}
4427
4428static int leaks_show(struct seq_file *m, void *p)
4429{
Pavel Emelianovb92151b2007-07-15 23:38:04 -07004430 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
Al Viro871751e2006-03-25 03:06:39 -08004431 struct slab *slabp;
4432 struct kmem_list3 *l3;
4433 const char *name;
4434 unsigned long *n = m->private;
4435 int node;
4436 int i;
4437
4438 if (!(cachep->flags & SLAB_STORE_USER))
4439 return 0;
4440 if (!(cachep->flags & SLAB_RED_ZONE))
4441 return 0;
4442
4443 /* OK, we can do it */
4444
4445 n[1] = 0;
4446
4447 for_each_online_node(node) {
4448 l3 = cachep->nodelists[node];
4449 if (!l3)
4450 continue;
4451
4452 check_irq_on();
4453 spin_lock_irq(&l3->list_lock);
4454
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004455 list_for_each_entry(slabp, &l3->slabs_full, list)
Al Viro871751e2006-03-25 03:06:39 -08004456 handle_slab(n, cachep, slabp);
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004457 list_for_each_entry(slabp, &l3->slabs_partial, list)
Al Viro871751e2006-03-25 03:06:39 -08004458 handle_slab(n, cachep, slabp);
Al Viro871751e2006-03-25 03:06:39 -08004459 spin_unlock_irq(&l3->list_lock);
4460 }
4461 name = cachep->name;
4462 if (n[0] == n[1]) {
4463 /* Increase the buffer size */
4464 mutex_unlock(&cache_chain_mutex);
4465 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4466 if (!m->private) {
4467 /* Too bad, we are really out */
4468 m->private = n;
4469 mutex_lock(&cache_chain_mutex);
4470 return -ENOMEM;
4471 }
4472 *(unsigned long *)m->private = n[0] * 2;
4473 kfree(n);
4474 mutex_lock(&cache_chain_mutex);
4475 /* Now make sure this entry will be retried */
4476 m->count = m->size;
4477 return 0;
4478 }
4479 for (i = 0; i < n[1]; i++) {
4480 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4481 show_symbol(m, n[2*i+2]);
4482 seq_putc(m, '\n');
4483 }
Siddha, Suresh Bd2e7b7d2006-09-25 23:31:47 -07004484
Al Viro871751e2006-03-25 03:06:39 -08004485 return 0;
4486}
4487
Alexey Dobriyana0ec95a82008-10-06 00:59:10 +04004488static const struct seq_operations slabstats_op = {
Al Viro871751e2006-03-25 03:06:39 -08004489 .start = leaks_start,
4490 .next = s_next,
4491 .stop = s_stop,
4492 .show = leaks_show,
4493};
Alexey Dobriyana0ec95a82008-10-06 00:59:10 +04004494
4495static int slabstats_open(struct inode *inode, struct file *file)
4496{
4497 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4498 int ret = -ENOMEM;
4499 if (n) {
4500 ret = seq_open(file, &slabstats_op);
4501 if (!ret) {
4502 struct seq_file *m = file->private_data;
4503 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4504 m->private = n;
4505 n = NULL;
4506 }
4507 kfree(n);
4508 }
4509 return ret;
4510}
4511
4512static const struct file_operations proc_slabstats_operations = {
4513 .open = slabstats_open,
4514 .read = seq_read,
4515 .llseek = seq_lseek,
4516 .release = seq_release_private,
4517};
Al Viro871751e2006-03-25 03:06:39 -08004518#endif
Alexey Dobriyana0ec95a82008-10-06 00:59:10 +04004519
4520static int __init slab_proc_init(void)
4521{
Alexey Dobriyan7b3c3a52008-10-06 02:42:17 +04004522 proc_create("slabinfo",S_IWUSR|S_IRUGO,NULL,&proc_slabinfo_operations);
Alexey Dobriyana0ec95a82008-10-06 00:59:10 +04004523#ifdef CONFIG_DEBUG_SLAB_LEAK
4524 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4525#endif
4526 return 0;
4527}
4528module_init(slab_proc_init);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004529#endif
4530
Manfred Spraul00e145b2005-09-03 15:55:07 -07004531/**
4532 * ksize - get the actual amount of memory allocated for a given object
4533 * @objp: Pointer to the object
4534 *
4535 * kmalloc may internally round up allocations and return more memory
4536 * than requested. ksize() can be used to determine the actual amount of
4537 * memory allocated. The caller may use this additional memory, even though
4538 * a smaller amount of memory was initially specified with the kmalloc call.
4539 * The caller must guarantee that objp points to a valid object previously
4540 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4541 * must not be freed during the duration of the call.
4542 */
Pekka Enbergfd76bab2007-05-06 14:48:40 -07004543size_t ksize(const void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004544{
Christoph Lameteref8b4522007-10-16 01:24:46 -07004545 BUG_ON(!objp);
4546 if (unlikely(objp == ZERO_SIZE_PTR))
Manfred Spraul00e145b2005-09-03 15:55:07 -07004547 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004548
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08004549 return obj_size(virt_to_cache(objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004550}
Kirill A. Shutemovb1aabec2009-02-10 15:21:44 +02004551EXPORT_SYMBOL(ksize);