blob: 5e59ce7a46c86922c09fb811c9c9e06bd5b99b35 [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
29 * slabs and you must pass objects with the same intializations to
30 * 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
89#include <linux/config.h>
90#include <linux/slab.h>
91#include <linux/mm.h>
Randy Dunlapc9cf5522006-06-27 02:53:52 -070092#include <linux/poison.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070093#include <linux/swap.h>
94#include <linux/cache.h>
95#include <linux/interrupt.h>
96#include <linux/init.h>
97#include <linux/compiler.h>
Paul Jackson101a5002006-03-24 03:16:07 -080098#include <linux/cpuset.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>
105#include <linux/rcupdate.h>
Paulo Marques543537b2005-06-23 00:09:02 -0700106#include <linux/string.h>
Christoph Lametere498be72005-09-09 13:03:32 -0700107#include <linux/nodemask.h>
Christoph Lameterdc85da12006-01-18 17:42:36 -0800108#include <linux/mempolicy.h>
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800109#include <linux/mutex.h>
Ingo Molnare7eebaf2006-06-27 02:54:55 -0700110#include <linux/rtmutex.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700111
112#include <asm/uaccess.h>
113#include <asm/cacheflush.h>
114#include <asm/tlbflush.h>
115#include <asm/page.h>
116
117/*
118 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL,
119 * SLAB_RED_ZONE & SLAB_POISON.
120 * 0 for faster, smaller code (especially in the critical paths).
121 *
122 * STATS - 1 to collect stats for /proc/slabinfo.
123 * 0 for faster, smaller code (especially in the critical paths).
124 *
125 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
126 */
127
128#ifdef CONFIG_DEBUG_SLAB
129#define DEBUG 1
130#define STATS 1
131#define FORCED_DEBUG 1
132#else
133#define DEBUG 0
134#define STATS 0
135#define FORCED_DEBUG 0
136#endif
137
Linus Torvalds1da177e2005-04-16 15:20:36 -0700138/* Shouldn't this be in a header file somewhere? */
139#define BYTES_PER_WORD sizeof(void *)
140
141#ifndef cache_line_size
142#define cache_line_size() L1_CACHE_BYTES
143#endif
144
145#ifndef ARCH_KMALLOC_MINALIGN
146/*
147 * Enforce a minimum alignment for the kmalloc caches.
148 * Usually, the kmalloc caches are cache_line_size() aligned, except when
149 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
150 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
151 * alignment larger than BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that.
152 * Note that this flag disables some debug features.
153 */
154#define ARCH_KMALLOC_MINALIGN 0
155#endif
156
157#ifndef ARCH_SLAB_MINALIGN
158/*
159 * Enforce a minimum alignment for all caches.
160 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
161 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
162 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
163 * some debug features.
164 */
165#define ARCH_SLAB_MINALIGN 0
166#endif
167
168#ifndef ARCH_KMALLOC_FLAGS
169#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
170#endif
171
172/* Legal flag mask for kmem_cache_create(). */
173#if DEBUG
174# define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \
175 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
Christoph Lameterac2b8982006-03-22 00:08:15 -0800176 SLAB_CACHE_DMA | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700177 SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \
178 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
Paul Jackson101a5002006-03-24 03:16:07 -0800179 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700180#else
Christoph Lameterac2b8982006-03-22 00:08:15 -0800181# define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
Linus Torvalds1da177e2005-04-16 15:20:36 -0700182 SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \
183 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
Paul Jackson101a5002006-03-24 03:16:07 -0800184 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700185#endif
186
187/*
188 * kmem_bufctl_t:
189 *
190 * Bufctl's are used for linking objs within a slab
191 * linked offsets.
192 *
193 * This implementation relies on "struct page" for locating the cache &
194 * slab an object belongs to.
195 * This allows the bufctl structure to be small (one int), but limits
196 * the number of objects a slab (not a cache) can contain when off-slab
197 * bufctls are used. The limit is the size of the largest general cache
198 * that does not use off-slab slabs.
199 * For 32bit archs with 4 kB pages, is this 56.
200 * This is not serious, as it is only for large objects, when it is unwise
201 * to have too many per slab.
202 * Note: This limit can be raised by introducing a general cache whose size
203 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
204 */
205
Kyle Moffettfa5b08d2005-09-03 15:55:03 -0700206typedef unsigned int kmem_bufctl_t;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700207#define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
208#define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
Al Viro871751e2006-03-25 03:06:39 -0800209#define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
210#define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700211
Linus Torvalds1da177e2005-04-16 15:20:36 -0700212/*
213 * struct slab
214 *
215 * Manages the objs in a slab. Placed either at the beginning of mem allocated
216 * for a slab, or allocated from an general cache.
217 * Slabs are chained into three list: fully used, partial, fully free slabs.
218 */
219struct slab {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800220 struct list_head list;
221 unsigned long colouroff;
222 void *s_mem; /* including colour offset */
223 unsigned int inuse; /* num of objs active in slab */
224 kmem_bufctl_t free;
225 unsigned short nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700226};
227
228/*
229 * struct slab_rcu
230 *
231 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
232 * arrange for kmem_freepages to be called via RCU. This is useful if
233 * we need to approach a kernel structure obliquely, from its address
234 * obtained without the usual locking. We can lock the structure to
235 * stabilize it and check it's still at the given address, only if we
236 * can be sure that the memory has not been meanwhile reused for some
237 * other kind of object (which our subsystem's lock might corrupt).
238 *
239 * rcu_read_lock before reading the address, then rcu_read_unlock after
240 * taking the spinlock within the structure expected at that address.
241 *
242 * We assume struct slab_rcu can overlay struct slab when destroying.
243 */
244struct slab_rcu {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800245 struct rcu_head head;
Pekka Enberg343e0d72006-02-01 03:05:50 -0800246 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800247 void *addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700248};
249
250/*
251 * struct array_cache
252 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700253 * Purpose:
254 * - LIFO ordering, to hand out cache-warm objects from _alloc
255 * - reduce the number of linked list operations
256 * - reduce spinlock operations
257 *
258 * The limit is stored in the per-cpu structure to reduce the data cache
259 * footprint.
260 *
261 */
262struct array_cache {
263 unsigned int avail;
264 unsigned int limit;
265 unsigned int batchcount;
266 unsigned int touched;
Christoph Lametere498be72005-09-09 13:03:32 -0700267 spinlock_t lock;
Andrew Mortona737b3e2006-03-22 00:08:11 -0800268 void *entry[0]; /*
269 * Must have this definition in here for the proper
270 * alignment of array_cache. Also simplifies accessing
271 * the entries.
272 * [0] is for gcc 2.95. It should really be [].
273 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700274};
275
Andrew Mortona737b3e2006-03-22 00:08:11 -0800276/*
277 * bootstrap: The caches do not work without cpuarrays anymore, but the
278 * cpuarrays are allocated from the generic caches...
Linus Torvalds1da177e2005-04-16 15:20:36 -0700279 */
280#define BOOT_CPUCACHE_ENTRIES 1
281struct arraycache_init {
282 struct array_cache cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800283 void *entries[BOOT_CPUCACHE_ENTRIES];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700284};
285
286/*
Christoph Lametere498be72005-09-09 13:03:32 -0700287 * The slab lists for all objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700288 */
289struct kmem_list3 {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800290 struct list_head slabs_partial; /* partial list first, better asm code */
291 struct list_head slabs_full;
292 struct list_head slabs_free;
293 unsigned long free_objects;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800294 unsigned int free_limit;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800295 unsigned int colour_next; /* Per-node cache coloring */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800296 spinlock_t list_lock;
297 struct array_cache *shared; /* shared per node */
298 struct array_cache **alien; /* on other nodes */
Christoph Lameter35386e32006-03-22 00:09:05 -0800299 unsigned long next_reap; /* updated without locking */
300 int free_touched; /* updated without locking */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700301};
302
Christoph Lametere498be72005-09-09 13:03:32 -0700303/*
304 * Need this for bootstrapping a per node allocator.
305 */
306#define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
307struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
308#define CACHE_CACHE 0
309#define SIZE_AC 1
310#define SIZE_L3 (1 + MAX_NUMNODES)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700311
Christoph Lametered11d9e2006-06-30 01:55:45 -0700312static int drain_freelist(struct kmem_cache *cache,
313 struct kmem_list3 *l3, int tofree);
314static void free_block(struct kmem_cache *cachep, void **objpp, int len,
315 int node);
316static void enable_cpucache(struct kmem_cache *cachep);
317static void cache_reap(void *unused);
318
Christoph Lametere498be72005-09-09 13:03:32 -0700319/*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800320 * This function must be completely optimized away if a constant is passed to
321 * it. Mostly the same as what is in linux/slab.h except it returns an index.
Christoph Lametere498be72005-09-09 13:03:32 -0700322 */
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700323static __always_inline int index_of(const size_t size)
Christoph Lametere498be72005-09-09 13:03:32 -0700324{
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800325 extern void __bad_size(void);
326
Christoph Lametere498be72005-09-09 13:03:32 -0700327 if (__builtin_constant_p(size)) {
328 int i = 0;
329
330#define CACHE(x) \
331 if (size <=x) \
332 return i; \
333 else \
334 i++;
335#include "linux/kmalloc_sizes.h"
336#undef CACHE
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800337 __bad_size();
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700338 } else
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800339 __bad_size();
Christoph Lametere498be72005-09-09 13:03:32 -0700340 return 0;
341}
342
Ingo Molnare0a42722006-06-23 02:03:46 -0700343static int slab_early_init = 1;
344
Christoph Lametere498be72005-09-09 13:03:32 -0700345#define INDEX_AC index_of(sizeof(struct arraycache_init))
346#define INDEX_L3 index_of(sizeof(struct kmem_list3))
347
Pekka Enberg5295a742006-02-01 03:05:48 -0800348static void kmem_list3_init(struct kmem_list3 *parent)
Christoph Lametere498be72005-09-09 13:03:32 -0700349{
350 INIT_LIST_HEAD(&parent->slabs_full);
351 INIT_LIST_HEAD(&parent->slabs_partial);
352 INIT_LIST_HEAD(&parent->slabs_free);
353 parent->shared = NULL;
354 parent->alien = NULL;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800355 parent->colour_next = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700356 spin_lock_init(&parent->list_lock);
357 parent->free_objects = 0;
358 parent->free_touched = 0;
359}
360
Andrew Mortona737b3e2006-03-22 00:08:11 -0800361#define MAKE_LIST(cachep, listp, slab, nodeid) \
362 do { \
363 INIT_LIST_HEAD(listp); \
364 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
Christoph Lametere498be72005-09-09 13:03:32 -0700365 } while (0)
366
Andrew Mortona737b3e2006-03-22 00:08:11 -0800367#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
368 do { \
Christoph Lametere498be72005-09-09 13:03:32 -0700369 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
370 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
371 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
372 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700373
374/*
Pekka Enberg343e0d72006-02-01 03:05:50 -0800375 * struct kmem_cache
Linus Torvalds1da177e2005-04-16 15:20:36 -0700376 *
377 * manages a cache.
378 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800379
Pekka J Enberg2109a2d2005-11-07 00:58:01 -0800380struct kmem_cache {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700381/* 1) per-cpu data, touched during every alloc/free */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800382 struct array_cache *array[NR_CPUS];
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800383/* 2) Cache tunables. Protected by cache_chain_mutex */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800384 unsigned int batchcount;
385 unsigned int limit;
386 unsigned int shared;
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800387
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800388 unsigned int buffer_size;
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800389/* 3) touched by every alloc & free from the backend */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800390 struct kmem_list3 *nodelists[MAX_NUMNODES];
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800391
Andrew Mortona737b3e2006-03-22 00:08:11 -0800392 unsigned int flags; /* constant flags */
393 unsigned int num; /* # of objs per slab */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700394
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -0800395/* 4) cache_grow/shrink */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700396 /* order of pgs per slab (2^n) */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800397 unsigned int gfporder;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700398
399 /* force GFP flags, e.g. GFP_DMA */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800400 gfp_t gfpflags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700401
Andrew Mortona737b3e2006-03-22 00:08:11 -0800402 size_t colour; /* cache colouring range */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800403 unsigned int colour_off; /* colour offset */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800404 struct kmem_cache *slabp_cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800405 unsigned int slab_size;
Andrew Mortona737b3e2006-03-22 00:08:11 -0800406 unsigned int dflags; /* dynamic flags */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700407
408 /* constructor func */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800409 void (*ctor) (void *, struct kmem_cache *, unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700410
411 /* de-constructor func */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800412 void (*dtor) (void *, struct kmem_cache *, unsigned long);
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
445};
446
447#define CFLGS_OFF_SLAB (0x80000000UL)
448#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
449
450#define BATCHREFILL_LIMIT 16
Andrew Mortona737b3e2006-03-22 00:08:11 -0800451/*
452 * Optimization question: fewer reaps means less probability for unnessary
453 * cpucache drain/refill cycles.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700454 *
Adrian Bunkdc6f3f22005-11-08 16:44:08 +0100455 * OTOH the cpuarrays can contain lots of objects,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700456 * which could lock up otherwise freeable slabs.
457 */
458#define REAPTIMEOUT_CPUC (2*HZ)
459#define REAPTIMEOUT_LIST3 (4*HZ)
460
461#if STATS
462#define STATS_INC_ACTIVE(x) ((x)->num_active++)
463#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
464#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
465#define STATS_INC_GROWN(x) ((x)->grown++)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700466#define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
Andrew Mortona737b3e2006-03-22 00:08:11 -0800467#define STATS_SET_HIGH(x) \
468 do { \
469 if ((x)->num_active > (x)->high_mark) \
470 (x)->high_mark = (x)->num_active; \
471 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700472#define STATS_INC_ERR(x) ((x)->errors++)
473#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
Christoph Lametere498be72005-09-09 13:03:32 -0700474#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700475#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800476#define STATS_SET_FREEABLE(x, i) \
477 do { \
478 if ((x)->max_freeable < i) \
479 (x)->max_freeable = i; \
480 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700481#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
482#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
483#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
484#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
485#else
486#define STATS_INC_ACTIVE(x) do { } while (0)
487#define STATS_DEC_ACTIVE(x) do { } while (0)
488#define STATS_INC_ALLOCED(x) do { } while (0)
489#define STATS_INC_GROWN(x) do { } while (0)
Christoph Lametered11d9e2006-06-30 01:55:45 -0700490#define STATS_ADD_REAPED(x,y) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700491#define STATS_SET_HIGH(x) do { } while (0)
492#define STATS_INC_ERR(x) do { } while (0)
493#define STATS_INC_NODEALLOCS(x) do { } while (0)
Christoph Lametere498be72005-09-09 13:03:32 -0700494#define STATS_INC_NODEFREES(x) do { } while (0)
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -0700495#define STATS_INC_ACOVERFLOW(x) do { } while (0)
Andrew Mortona737b3e2006-03-22 00:08:11 -0800496#define STATS_SET_FREEABLE(x, i) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700497#define STATS_INC_ALLOCHIT(x) do { } while (0)
498#define STATS_INC_ALLOCMISS(x) do { } while (0)
499#define STATS_INC_FREEHIT(x) do { } while (0)
500#define STATS_INC_FREEMISS(x) do { } while (0)
501#endif
502
503#if DEBUG
Linus Torvalds1da177e2005-04-16 15:20:36 -0700504
Andrew Mortona737b3e2006-03-22 00:08:11 -0800505/*
506 * memory layout of objects:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700507 * 0 : objp
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800508 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
Linus Torvalds1da177e2005-04-16 15:20:36 -0700509 * the end of an object is aligned with the end of the real
510 * allocation. Catches writes behind the end of the allocation.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800511 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700512 * redzone word.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800513 * cachep->obj_offset: The real object.
514 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
Andrew Mortona737b3e2006-03-22 00:08:11 -0800515 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
516 * [BYTES_PER_WORD long]
Linus Torvalds1da177e2005-04-16 15:20:36 -0700517 */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800518static int obj_offset(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700519{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800520 return cachep->obj_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700521}
522
Pekka Enberg343e0d72006-02-01 03:05:50 -0800523static int obj_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700524{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800525 return cachep->obj_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700526}
527
Pekka Enberg343e0d72006-02-01 03:05:50 -0800528static unsigned long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700529{
530 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800531 return (unsigned long*) (objp+obj_offset(cachep)-BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700532}
533
Pekka Enberg343e0d72006-02-01 03:05:50 -0800534static unsigned long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700535{
536 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
537 if (cachep->flags & SLAB_STORE_USER)
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800538 return (unsigned long *)(objp + cachep->buffer_size -
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800539 2 * BYTES_PER_WORD);
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800540 return (unsigned long *)(objp + cachep->buffer_size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700541}
542
Pekka Enberg343e0d72006-02-01 03:05:50 -0800543static void **dbg_userword(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700544{
545 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800546 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700547}
548
549#else
550
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800551#define obj_offset(x) 0
552#define obj_size(cachep) (cachep->buffer_size)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700553#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;})
554#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;})
555#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
556
557#endif
558
559/*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800560 * Maximum size of an obj (in 2^order pages) and absolute limit for the gfp
561 * order.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700562 */
563#if defined(CONFIG_LARGE_ALLOCS)
564#define MAX_OBJ_ORDER 13 /* up to 32Mb */
565#define MAX_GFP_ORDER 13 /* up to 32Mb */
566#elif defined(CONFIG_MMU)
567#define MAX_OBJ_ORDER 5 /* 32 pages */
568#define MAX_GFP_ORDER 5 /* 32 pages */
569#else
570#define MAX_OBJ_ORDER 8 /* up to 1Mb */
571#define MAX_GFP_ORDER 8 /* up to 1Mb */
572#endif
573
574/*
575 * Do not go above this order unless 0 objects fit into the slab.
576 */
577#define BREAK_GFP_ORDER_HI 1
578#define BREAK_GFP_ORDER_LO 0
579static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
580
Andrew Mortona737b3e2006-03-22 00:08:11 -0800581/*
582 * Functions for storing/retrieving the cachep and or slab from the page
583 * allocator. These are used to find the slab an obj belongs to. With kfree(),
584 * these are used to find the cache which an obj belongs to.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700585 */
Pekka Enberg065d41c2005-11-13 16:06:46 -0800586static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
587{
588 page->lru.next = (struct list_head *)cache;
589}
590
591static inline struct kmem_cache *page_get_cache(struct page *page)
592{
Nick Piggin84097512006-03-22 00:08:34 -0800593 if (unlikely(PageCompound(page)))
594 page = (struct page *)page_private(page);
Pekka Enbergddc2e812006-06-23 02:03:40 -0700595 BUG_ON(!PageSlab(page));
Pekka Enberg065d41c2005-11-13 16:06:46 -0800596 return (struct kmem_cache *)page->lru.next;
597}
598
599static inline void page_set_slab(struct page *page, struct slab *slab)
600{
601 page->lru.prev = (struct list_head *)slab;
602}
603
604static inline struct slab *page_get_slab(struct page *page)
605{
Nick Piggin84097512006-03-22 00:08:34 -0800606 if (unlikely(PageCompound(page)))
607 page = (struct page *)page_private(page);
Pekka Enbergddc2e812006-06-23 02:03:40 -0700608 BUG_ON(!PageSlab(page));
Pekka Enberg065d41c2005-11-13 16:06:46 -0800609 return (struct slab *)page->lru.prev;
610}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700611
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -0800612static inline struct kmem_cache *virt_to_cache(const void *obj)
613{
614 struct page *page = virt_to_page(obj);
615 return page_get_cache(page);
616}
617
618static inline struct slab *virt_to_slab(const void *obj)
619{
620 struct page *page = virt_to_page(obj);
621 return page_get_slab(page);
622}
623
Pekka Enberg8fea4e92006-03-22 00:08:10 -0800624static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
625 unsigned int idx)
626{
627 return slab->s_mem + cache->buffer_size * idx;
628}
629
630static inline unsigned int obj_to_index(struct kmem_cache *cache,
631 struct slab *slab, void *obj)
632{
633 return (unsigned)(obj - slab->s_mem) / cache->buffer_size;
634}
635
Andrew Mortona737b3e2006-03-22 00:08:11 -0800636/*
637 * These are the default caches for kmalloc. Custom caches can have other sizes.
638 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700639struct cache_sizes malloc_sizes[] = {
640#define CACHE(x) { .cs_size = (x) },
641#include <linux/kmalloc_sizes.h>
642 CACHE(ULONG_MAX)
643#undef CACHE
644};
645EXPORT_SYMBOL(malloc_sizes);
646
647/* Must match cache_sizes above. Out of line to keep cache footprint low. */
648struct cache_names {
649 char *name;
650 char *name_dma;
651};
652
653static struct cache_names __initdata cache_names[] = {
654#define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
655#include <linux/kmalloc_sizes.h>
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800656 {NULL,}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700657#undef CACHE
658};
659
660static struct arraycache_init initarray_cache __initdata =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800661 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700662static struct arraycache_init initarray_generic =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800663 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700664
665/* internal cache of cache description objs */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800666static struct kmem_cache cache_cache = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800667 .batchcount = 1,
668 .limit = BOOT_CPUCACHE_ENTRIES,
669 .shared = 1,
Pekka Enberg343e0d72006-02-01 03:05:50 -0800670 .buffer_size = sizeof(struct kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800671 .name = "kmem_cache",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700672#if DEBUG
Pekka Enberg343e0d72006-02-01 03:05:50 -0800673 .obj_size = sizeof(struct kmem_cache),
Linus Torvalds1da177e2005-04-16 15:20:36 -0700674#endif
675};
676
Arjan van de Venf1aaee52006-07-13 14:46:03 +0200677#ifdef CONFIG_LOCKDEP
678
679/*
680 * Slab sometimes uses the kmalloc slabs to store the slab headers
681 * for other slabs "off slab".
682 * The locking for this is tricky in that it nests within the locks
683 * of all other slabs in a few places; to deal with this special
684 * locking we put on-slab caches into a separate lock-class.
685 */
686static struct lock_class_key on_slab_key;
687
688static inline void init_lock_keys(struct cache_sizes *s)
689{
690 int q;
691
692 for (q = 0; q < MAX_NUMNODES; q++) {
693 if (!s->cs_cachep->nodelists[q] || OFF_SLAB(s->cs_cachep))
694 continue;
695 lockdep_set_class(&s->cs_cachep->nodelists[q]->list_lock,
696 &on_slab_key);
697 }
698}
699
700#else
701static inline void init_lock_keys(struct cache_sizes *s)
702{
703}
704#endif
705
706
707
Linus Torvalds1da177e2005-04-16 15:20:36 -0700708/* Guard access to the cache-chain. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800709static DEFINE_MUTEX(cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700710static struct list_head cache_chain;
711
712/*
Andrew Mortona737b3e2006-03-22 00:08:11 -0800713 * vm_enough_memory() looks at this to determine how many slab-allocated pages
714 * are possibly freeable under pressure
Linus Torvalds1da177e2005-04-16 15:20:36 -0700715 *
716 * SLAB_RECLAIM_ACCOUNT turns this on per-slab
717 */
718atomic_t slab_reclaim_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700719
720/*
721 * chicken and egg problem: delay the per-cpu array allocation
722 * until the general caches are up.
723 */
724static enum {
725 NONE,
Christoph Lametere498be72005-09-09 13:03:32 -0700726 PARTIAL_AC,
727 PARTIAL_L3,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700728 FULL
729} g_cpucache_up;
730
Mike Kravetz39d24e62006-05-15 09:44:13 -0700731/*
732 * used by boot code to determine if it can use slab based allocator
733 */
734int slab_is_available(void)
735{
736 return g_cpucache_up == FULL;
737}
738
Linus Torvalds1da177e2005-04-16 15:20:36 -0700739static DEFINE_PER_CPU(struct work_struct, reap_work);
740
Pekka Enberg343e0d72006-02-01 03:05:50 -0800741static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700742{
743 return cachep->array[smp_processor_id()];
744}
745
Andrew Mortona737b3e2006-03-22 00:08:11 -0800746static inline struct kmem_cache *__find_general_cachep(size_t size,
747 gfp_t gfpflags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700748{
749 struct cache_sizes *csizep = malloc_sizes;
750
751#if DEBUG
752 /* This happens if someone tries to call
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800753 * kmem_cache_create(), or __kmalloc(), before
754 * the generic caches are initialized.
755 */
Alok Katariac7e43c72005-09-14 12:17:53 -0700756 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700757#endif
758 while (size > csizep->cs_size)
759 csizep++;
760
761 /*
Martin Hicks0abf40c2005-09-03 15:54:54 -0700762 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
Linus Torvalds1da177e2005-04-16 15:20:36 -0700763 * has cs_{dma,}cachep==NULL. Thus no special case
764 * for large kmalloc calls required.
765 */
766 if (unlikely(gfpflags & GFP_DMA))
767 return csizep->cs_dmacachep;
768 return csizep->cs_cachep;
769}
770
Adrian Bunkb2213852006-09-25 23:31:02 -0700771static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700772{
773 return __find_general_cachep(size, gfpflags);
774}
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700775
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800776static size_t slab_mgmt_size(size_t nr_objs, size_t align)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700777{
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800778 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
779}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700780
Andrew Mortona737b3e2006-03-22 00:08:11 -0800781/*
782 * Calculate the number of objects and left-over bytes for a given buffer size.
783 */
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800784static void cache_estimate(unsigned long gfporder, size_t buffer_size,
785 size_t align, int flags, size_t *left_over,
786 unsigned int *num)
787{
788 int nr_objs;
789 size_t mgmt_size;
790 size_t slab_size = PAGE_SIZE << gfporder;
791
792 /*
793 * The slab management structure can be either off the slab or
794 * on it. For the latter case, the memory allocated for a
795 * slab is used for:
796 *
797 * - The struct slab
798 * - One kmem_bufctl_t for each object
799 * - Padding to respect alignment of @align
800 * - @buffer_size bytes for each object
801 *
802 * If the slab management structure is off the slab, then the
803 * alignment will already be calculated into the size. Because
804 * the slabs are all pages aligned, the objects will be at the
805 * correct alignment when allocated.
806 */
807 if (flags & CFLGS_OFF_SLAB) {
808 mgmt_size = 0;
809 nr_objs = slab_size / buffer_size;
810
811 if (nr_objs > SLAB_LIMIT)
812 nr_objs = SLAB_LIMIT;
813 } else {
814 /*
815 * Ignore padding for the initial guess. The padding
816 * is at most @align-1 bytes, and @buffer_size is at
817 * least @align. In the worst case, this result will
818 * be one greater than the number of objects that fit
819 * into the memory allocation when taking the padding
820 * into account.
821 */
822 nr_objs = (slab_size - sizeof(struct slab)) /
823 (buffer_size + sizeof(kmem_bufctl_t));
824
825 /*
826 * This calculated number will be either the right
827 * amount, or one greater than what we want.
828 */
829 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
830 > slab_size)
831 nr_objs--;
832
833 if (nr_objs > SLAB_LIMIT)
834 nr_objs = SLAB_LIMIT;
835
836 mgmt_size = slab_mgmt_size(nr_objs, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700837 }
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800838 *num = nr_objs;
839 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700840}
841
842#define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
843
Andrew Mortona737b3e2006-03-22 00:08:11 -0800844static void __slab_error(const char *function, struct kmem_cache *cachep,
845 char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700846{
847 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800848 function, cachep->name, msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700849 dump_stack();
850}
851
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800852#ifdef CONFIG_NUMA
853/*
854 * Special reaping functions for NUMA systems called from cache_reap().
855 * These take care of doing round robin flushing of alien caches (containing
856 * objects freed on different nodes from which they were allocated) and the
857 * flushing of remote pcps by calling drain_node_pages.
858 */
859static DEFINE_PER_CPU(unsigned long, reap_node);
860
861static void init_reap_node(int cpu)
862{
863 int node;
864
865 node = next_node(cpu_to_node(cpu), node_online_map);
866 if (node == MAX_NUMNODES)
Paul Jackson442295c2006-03-22 00:09:11 -0800867 node = first_node(node_online_map);
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800868
869 __get_cpu_var(reap_node) = node;
870}
871
872static void next_reap_node(void)
873{
874 int node = __get_cpu_var(reap_node);
875
876 /*
877 * Also drain per cpu pages on remote zones
878 */
879 if (node != numa_node_id())
880 drain_node_pages(node);
881
882 node = next_node(node, node_online_map);
883 if (unlikely(node >= MAX_NUMNODES))
884 node = first_node(node_online_map);
885 __get_cpu_var(reap_node) = node;
886}
887
888#else
889#define init_reap_node(cpu) do { } while (0)
890#define next_reap_node(void) do { } while (0)
891#endif
892
Linus Torvalds1da177e2005-04-16 15:20:36 -0700893/*
894 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
895 * via the workqueue/eventd.
896 * Add the CPU number into the expiration time to minimize the possibility of
897 * the CPUs getting into lockstep and contending for the global cache chain
898 * lock.
899 */
900static void __devinit start_cpu_timer(int cpu)
901{
902 struct work_struct *reap_work = &per_cpu(reap_work, cpu);
903
904 /*
905 * When this gets called from do_initcalls via cpucache_init(),
906 * init_workqueues() has already run, so keventd will be setup
907 * at that time.
908 */
909 if (keventd_up() && reap_work->func == NULL) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -0800910 init_reap_node(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700911 INIT_WORK(reap_work, cache_reap, NULL);
912 schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu);
913 }
914}
915
Christoph Lametere498be72005-09-09 13:03:32 -0700916static struct array_cache *alloc_arraycache(int node, int entries,
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800917 int batchcount)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700918{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800919 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700920 struct array_cache *nc = NULL;
921
Christoph Lametere498be72005-09-09 13:03:32 -0700922 nc = kmalloc_node(memsize, GFP_KERNEL, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700923 if (nc) {
924 nc->avail = 0;
925 nc->limit = entries;
926 nc->batchcount = batchcount;
927 nc->touched = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700928 spin_lock_init(&nc->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700929 }
930 return nc;
931}
932
Christoph Lameter3ded1752006-03-25 03:06:44 -0800933/*
934 * Transfer objects in one arraycache to another.
935 * Locking must be handled by the caller.
936 *
937 * Return the number of entries transferred.
938 */
939static int transfer_objects(struct array_cache *to,
940 struct array_cache *from, unsigned int max)
941{
942 /* Figure out how many entries to transfer */
943 int nr = min(min(from->avail, max), to->limit - to->avail);
944
945 if (!nr)
946 return 0;
947
948 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
949 sizeof(void *) *nr);
950
951 from->avail -= nr;
952 to->avail += nr;
953 to->touched = 1;
954 return nr;
955}
956
Christoph Lametere498be72005-09-09 13:03:32 -0700957#ifdef CONFIG_NUMA
Pekka Enberg343e0d72006-02-01 03:05:50 -0800958static void *__cache_alloc_node(struct kmem_cache *, gfp_t, int);
Paul Jacksonc61afb12006-03-24 03:16:08 -0800959static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
Christoph Lameterdc85da12006-01-18 17:42:36 -0800960
Pekka Enberg5295a742006-02-01 03:05:48 -0800961static struct array_cache **alloc_alien_cache(int node, int limit)
Christoph Lametere498be72005-09-09 13:03:32 -0700962{
963 struct array_cache **ac_ptr;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800964 int memsize = sizeof(void *) * MAX_NUMNODES;
Christoph Lametere498be72005-09-09 13:03:32 -0700965 int i;
966
967 if (limit > 1)
968 limit = 12;
969 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
970 if (ac_ptr) {
971 for_each_node(i) {
972 if (i == node || !node_online(i)) {
973 ac_ptr[i] = NULL;
974 continue;
975 }
976 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
977 if (!ac_ptr[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800978 for (i--; i <= 0; i--)
Christoph Lametere498be72005-09-09 13:03:32 -0700979 kfree(ac_ptr[i]);
980 kfree(ac_ptr);
981 return NULL;
982 }
983 }
984 }
985 return ac_ptr;
986}
987
Pekka Enberg5295a742006-02-01 03:05:48 -0800988static void free_alien_cache(struct array_cache **ac_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -0700989{
990 int i;
991
992 if (!ac_ptr)
993 return;
Christoph Lametere498be72005-09-09 13:03:32 -0700994 for_each_node(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800995 kfree(ac_ptr[i]);
Christoph Lametere498be72005-09-09 13:03:32 -0700996 kfree(ac_ptr);
997}
998
Pekka Enberg343e0d72006-02-01 03:05:50 -0800999static void __drain_alien_cache(struct kmem_cache *cachep,
Pekka Enberg5295a742006-02-01 03:05:48 -08001000 struct array_cache *ac, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07001001{
1002 struct kmem_list3 *rl3 = cachep->nodelists[node];
1003
1004 if (ac->avail) {
1005 spin_lock(&rl3->list_lock);
Christoph Lametere00946f2006-03-25 03:06:45 -08001006 /*
1007 * Stuff objects into the remote nodes shared array first.
1008 * That way we could avoid the overhead of putting the objects
1009 * into the free lists and getting them back later.
1010 */
shin, jacob693f7d32006-04-28 10:54:37 -05001011 if (rl3->shared)
1012 transfer_objects(rl3->shared, ac, ac->limit);
Christoph Lametere00946f2006-03-25 03:06:45 -08001013
Christoph Lameterff694162005-09-22 21:44:02 -07001014 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001015 ac->avail = 0;
1016 spin_unlock(&rl3->list_lock);
1017 }
1018}
1019
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001020/*
1021 * Called from cache_reap() to regularly drain alien caches round robin.
1022 */
1023static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1024{
1025 int node = __get_cpu_var(reap_node);
1026
1027 if (l3->alien) {
1028 struct array_cache *ac = l3->alien[node];
Christoph Lametere00946f2006-03-25 03:06:45 -08001029
1030 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001031 __drain_alien_cache(cachep, ac, node);
1032 spin_unlock_irq(&ac->lock);
1033 }
1034 }
1035}
1036
Andrew Mortona737b3e2006-03-22 00:08:11 -08001037static void drain_alien_cache(struct kmem_cache *cachep,
1038 struct array_cache **alien)
Christoph Lametere498be72005-09-09 13:03:32 -07001039{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001040 int i = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07001041 struct array_cache *ac;
1042 unsigned long flags;
1043
1044 for_each_online_node(i) {
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001045 ac = alien[i];
Christoph Lametere498be72005-09-09 13:03:32 -07001046 if (ac) {
1047 spin_lock_irqsave(&ac->lock, flags);
1048 __drain_alien_cache(cachep, ac, i);
1049 spin_unlock_irqrestore(&ac->lock, flags);
1050 }
1051 }
1052}
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001053
Ingo Molnar873623d2006-07-13 14:44:38 +02001054static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001055{
1056 struct slab *slabp = virt_to_slab(objp);
1057 int nodeid = slabp->nodeid;
1058 struct kmem_list3 *l3;
1059 struct array_cache *alien = NULL;
1060
1061 /*
1062 * Make sure we are not freeing a object from another node to the array
1063 * cache on this cpu.
1064 */
1065 if (likely(slabp->nodeid == numa_node_id()))
1066 return 0;
1067
1068 l3 = cachep->nodelists[numa_node_id()];
1069 STATS_INC_NODEFREES(cachep);
1070 if (l3->alien && l3->alien[nodeid]) {
1071 alien = l3->alien[nodeid];
Ingo Molnar873623d2006-07-13 14:44:38 +02001072 spin_lock(&alien->lock);
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001073 if (unlikely(alien->avail == alien->limit)) {
1074 STATS_INC_ACOVERFLOW(cachep);
1075 __drain_alien_cache(cachep, alien, nodeid);
1076 }
1077 alien->entry[alien->avail++] = objp;
1078 spin_unlock(&alien->lock);
1079 } else {
1080 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1081 free_block(cachep, &objp, 1, nodeid);
1082 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1083 }
1084 return 1;
1085}
1086
Christoph Lametere498be72005-09-09 13:03:32 -07001087#else
Linus Torvalds7a21ef62006-02-05 11:26:38 -08001088
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001089#define drain_alien_cache(cachep, alien) do { } while (0)
Christoph Lameter8fce4d82006-03-09 17:33:54 -08001090#define reap_alien(cachep, l3) do { } while (0)
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001091
Linus Torvalds7a21ef62006-02-05 11:26:38 -08001092static inline struct array_cache **alloc_alien_cache(int node, int limit)
1093{
1094 return (struct array_cache **) 0x01020304ul;
1095}
1096
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001097static inline void free_alien_cache(struct array_cache **ac_ptr)
1098{
1099}
Linus Torvalds7a21ef62006-02-05 11:26:38 -08001100
Ingo Molnar873623d2006-07-13 14:44:38 +02001101static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
Pekka Enberg729bd0b2006-06-23 02:03:05 -07001102{
1103 return 0;
1104}
1105
Christoph Lametere498be72005-09-09 13:03:32 -07001106#endif
1107
Chandra Seetharaman8c78f302006-07-30 03:03:35 -07001108static int __cpuinit cpuup_callback(struct notifier_block *nfb,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001109 unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001110{
1111 long cpu = (long)hcpu;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001112 struct kmem_cache *cachep;
Christoph Lametere498be72005-09-09 13:03:32 -07001113 struct kmem_list3 *l3 = NULL;
1114 int node = cpu_to_node(cpu);
1115 int memsize = sizeof(struct kmem_list3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001116
1117 switch (action) {
1118 case CPU_UP_PREPARE:
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001119 mutex_lock(&cache_chain_mutex);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001120 /*
1121 * We need to do this right in the beginning since
Christoph Lametere498be72005-09-09 13:03:32 -07001122 * alloc_arraycache's are going to use this list.
1123 * kmalloc_node allows us to add the slab to the right
1124 * kmem_list3 and not this cpu's kmem_list3
1125 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001126
Christoph Lametere498be72005-09-09 13:03:32 -07001127 list_for_each_entry(cachep, &cache_chain, next) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08001128 /*
1129 * Set up the size64 kmemlist for cpu before we can
Christoph Lametere498be72005-09-09 13:03:32 -07001130 * begin anything. Make sure some other cpu on this
1131 * node has not already allocated this
1132 */
1133 if (!cachep->nodelists[node]) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08001134 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1135 if (!l3)
Christoph Lametere498be72005-09-09 13:03:32 -07001136 goto bad;
1137 kmem_list3_init(l3);
1138 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001139 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -07001140
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001141 /*
1142 * The l3s don't come and go as CPUs come and
1143 * go. cache_chain_mutex is sufficient
1144 * protection here.
1145 */
Christoph Lametere498be72005-09-09 13:03:32 -07001146 cachep->nodelists[node] = l3;
1147 }
1148
1149 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1150 cachep->nodelists[node]->free_limit =
Andrew Mortona737b3e2006-03-22 00:08:11 -08001151 (1 + nr_cpus_node(node)) *
1152 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07001153 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1154 }
1155
Andrew Mortona737b3e2006-03-22 00:08:11 -08001156 /*
1157 * Now we can go ahead with allocating the shared arrays and
1158 * array caches
1159 */
Christoph Lametere498be72005-09-09 13:03:32 -07001160 list_for_each_entry(cachep, &cache_chain, next) {
Tobias Klausercd105df2006-01-08 01:00:59 -08001161 struct array_cache *nc;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001162 struct array_cache *shared;
1163 struct array_cache **alien;
Tobias Klausercd105df2006-01-08 01:00:59 -08001164
Christoph Lametere498be72005-09-09 13:03:32 -07001165 nc = alloc_arraycache(node, cachep->limit,
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001166 cachep->batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001167 if (!nc)
1168 goto bad;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001169 shared = alloc_arraycache(node,
1170 cachep->shared * cachep->batchcount,
1171 0xbaadf00d);
1172 if (!shared)
1173 goto bad;
Linus Torvalds7a21ef62006-02-05 11:26:38 -08001174
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001175 alien = alloc_alien_cache(node, cachep->limit);
1176 if (!alien)
1177 goto bad;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001178 cachep->array[cpu] = nc;
Christoph Lametere498be72005-09-09 13:03:32 -07001179 l3 = cachep->nodelists[node];
1180 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07001181
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001182 spin_lock_irq(&l3->list_lock);
1183 if (!l3->shared) {
1184 /*
1185 * We are serialised from CPU_DEAD or
1186 * CPU_UP_CANCELLED by the cpucontrol lock
1187 */
1188 l3->shared = shared;
1189 shared = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07001190 }
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001191#ifdef CONFIG_NUMA
1192 if (!l3->alien) {
1193 l3->alien = alien;
1194 alien = NULL;
1195 }
1196#endif
1197 spin_unlock_irq(&l3->list_lock);
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001198 kfree(shared);
1199 free_alien_cache(alien);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001200 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001201 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001202 break;
1203 case CPU_ONLINE:
1204 start_cpu_timer(cpu);
1205 break;
1206#ifdef CONFIG_HOTPLUG_CPU
1207 case CPU_DEAD:
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001208 /*
1209 * Even if all the cpus of a node are down, we don't free the
1210 * kmem_list3 of any cache. This to avoid a race between
1211 * cpu_down, and a kmalloc allocation from another cpu for
1212 * memory from the node of the cpu going down. The list3
1213 * structure is usually allocated from kmem_cache_create() and
1214 * gets destroyed at kmem_cache_destroy().
1215 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001216 /* fall thru */
1217 case CPU_UP_CANCELED:
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001218 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001219 list_for_each_entry(cachep, &cache_chain, next) {
1220 struct array_cache *nc;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001221 struct array_cache *shared;
1222 struct array_cache **alien;
Christoph Lametere498be72005-09-09 13:03:32 -07001223 cpumask_t mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001224
Christoph Lametere498be72005-09-09 13:03:32 -07001225 mask = node_to_cpumask(node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001226 /* cpu is dead; no one can alloc from it. */
1227 nc = cachep->array[cpu];
1228 cachep->array[cpu] = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07001229 l3 = cachep->nodelists[node];
1230
1231 if (!l3)
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001232 goto free_array_cache;
Christoph Lametere498be72005-09-09 13:03:32 -07001233
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08001234 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07001235
1236 /* Free limit for this kmem_list3 */
1237 l3->free_limit -= cachep->batchcount;
1238 if (nc)
Christoph Lameterff694162005-09-22 21:44:02 -07001239 free_block(cachep, nc->entry, nc->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001240
1241 if (!cpus_empty(mask)) {
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08001242 spin_unlock_irq(&l3->list_lock);
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001243 goto free_array_cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001244 }
Christoph Lametere498be72005-09-09 13:03:32 -07001245
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001246 shared = l3->shared;
1247 if (shared) {
Christoph Lametere498be72005-09-09 13:03:32 -07001248 free_block(cachep, l3->shared->entry,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001249 l3->shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001250 l3->shared = NULL;
1251 }
Christoph Lametere498be72005-09-09 13:03:32 -07001252
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001253 alien = l3->alien;
1254 l3->alien = NULL;
1255
1256 spin_unlock_irq(&l3->list_lock);
1257
1258 kfree(shared);
1259 if (alien) {
1260 drain_alien_cache(cachep, alien);
1261 free_alien_cache(alien);
Christoph Lametere498be72005-09-09 13:03:32 -07001262 }
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001263free_array_cache:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001264 kfree(nc);
1265 }
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001266 /*
1267 * In the previous loop, all the objects were freed to
1268 * the respective cache's slabs, now we can go ahead and
1269 * shrink each nodelist to its limit.
1270 */
1271 list_for_each_entry(cachep, &cache_chain, next) {
1272 l3 = cachep->nodelists[node];
1273 if (!l3)
1274 continue;
Christoph Lametered11d9e2006-06-30 01:55:45 -07001275 drain_freelist(cachep, l3, l3->free_objects);
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001276 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001277 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001278 break;
1279#endif
1280 }
1281 return NOTIFY_OK;
Andrew Mortona737b3e2006-03-22 00:08:11 -08001282bad:
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001283 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001284 return NOTIFY_BAD;
1285}
1286
Chandra Seetharaman74b85f32006-06-27 02:54:09 -07001287static struct notifier_block __cpuinitdata cpucache_notifier = {
1288 &cpuup_callback, NULL, 0
1289};
Linus Torvalds1da177e2005-04-16 15:20:36 -07001290
Christoph Lametere498be72005-09-09 13:03:32 -07001291/*
1292 * swap the static kmem_list3 with kmalloced memory
1293 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001294static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1295 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001296{
1297 struct kmem_list3 *ptr;
1298
1299 BUG_ON(cachep->nodelists[nodeid] != list);
1300 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1301 BUG_ON(!ptr);
1302
1303 local_irq_disable();
1304 memcpy(ptr, list, sizeof(struct kmem_list3));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001305 /*
1306 * Do not assume that spinlocks can be initialized via memcpy:
1307 */
1308 spin_lock_init(&ptr->list_lock);
1309
Christoph Lametere498be72005-09-09 13:03:32 -07001310 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1311 cachep->nodelists[nodeid] = ptr;
1312 local_irq_enable();
1313}
1314
Andrew Mortona737b3e2006-03-22 00:08:11 -08001315/*
1316 * Initialisation. Called after the page allocator have been initialised and
1317 * before smp_init().
Linus Torvalds1da177e2005-04-16 15:20:36 -07001318 */
1319void __init kmem_cache_init(void)
1320{
1321 size_t left_over;
1322 struct cache_sizes *sizes;
1323 struct cache_names *names;
Christoph Lametere498be72005-09-09 13:03:32 -07001324 int i;
Jack Steiner07ed76b2006-03-07 21:55:46 -08001325 int order;
Christoph Lametere498be72005-09-09 13:03:32 -07001326
1327 for (i = 0; i < NUM_INIT_LISTS; i++) {
1328 kmem_list3_init(&initkmem_list3[i]);
1329 if (i < MAX_NUMNODES)
1330 cache_cache.nodelists[i] = NULL;
1331 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001332
1333 /*
1334 * Fragmentation resistance on low memory - only use bigger
1335 * page orders on machines with more than 32MB of memory.
1336 */
1337 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1338 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1339
Linus Torvalds1da177e2005-04-16 15:20:36 -07001340 /* Bootstrap is tricky, because several objects are allocated
1341 * from caches that do not exist yet:
Andrew Mortona737b3e2006-03-22 00:08:11 -08001342 * 1) initialize the cache_cache cache: it contains the struct
1343 * kmem_cache structures of all caches, except cache_cache itself:
1344 * cache_cache is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001345 * Initially an __init data area is used for the head array and the
1346 * kmem_list3 structures, it's replaced with a kmalloc allocated
1347 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001348 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001349 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001350 * An __init data area is used for the head array.
1351 * 3) Create the remaining kmalloc caches, with minimally sized
1352 * head arrays.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001353 * 4) Replace the __init data head arrays for cache_cache and the first
1354 * kmalloc cache with kmalloc allocated arrays.
Christoph Lametere498be72005-09-09 13:03:32 -07001355 * 5) Replace the __init data for kmem_list3 for cache_cache and
1356 * the other cache's with kmalloc allocated memory.
1357 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001358 */
1359
1360 /* 1) create the cache_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001361 INIT_LIST_HEAD(&cache_chain);
1362 list_add(&cache_cache.next, &cache_chain);
1363 cache_cache.colour_off = cache_line_size();
1364 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
Christoph Lametere498be72005-09-09 13:03:32 -07001365 cache_cache.nodelists[numa_node_id()] = &initkmem_list3[CACHE_CACHE];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001366
Andrew Mortona737b3e2006-03-22 00:08:11 -08001367 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1368 cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001369
Jack Steiner07ed76b2006-03-07 21:55:46 -08001370 for (order = 0; order < MAX_ORDER; order++) {
1371 cache_estimate(order, cache_cache.buffer_size,
1372 cache_line_size(), 0, &left_over, &cache_cache.num);
1373 if (cache_cache.num)
1374 break;
1375 }
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02001376 BUG_ON(!cache_cache.num);
Jack Steiner07ed76b2006-03-07 21:55:46 -08001377 cache_cache.gfporder = order;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001378 cache_cache.colour = left_over / cache_cache.colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001379 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1380 sizeof(struct slab), cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001381
1382 /* 2+3) create the kmalloc caches */
1383 sizes = malloc_sizes;
1384 names = cache_names;
1385
Andrew Mortona737b3e2006-03-22 00:08:11 -08001386 /*
1387 * Initialize the caches that provide memory for the array cache and the
1388 * kmem_list3 structures first. Without this, further allocations will
1389 * bug.
Christoph Lametere498be72005-09-09 13:03:32 -07001390 */
1391
1392 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001393 sizes[INDEX_AC].cs_size,
1394 ARCH_KMALLOC_MINALIGN,
1395 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1396 NULL, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07001397
Andrew Mortona737b3e2006-03-22 00:08:11 -08001398 if (INDEX_AC != INDEX_L3) {
Christoph Lametere498be72005-09-09 13:03:32 -07001399 sizes[INDEX_L3].cs_cachep =
Andrew Mortona737b3e2006-03-22 00:08:11 -08001400 kmem_cache_create(names[INDEX_L3].name,
1401 sizes[INDEX_L3].cs_size,
1402 ARCH_KMALLOC_MINALIGN,
1403 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1404 NULL, NULL);
1405 }
Christoph Lametere498be72005-09-09 13:03:32 -07001406
Ingo Molnare0a42722006-06-23 02:03:46 -07001407 slab_early_init = 0;
1408
Linus Torvalds1da177e2005-04-16 15:20:36 -07001409 while (sizes->cs_size != ULONG_MAX) {
Christoph Lametere498be72005-09-09 13:03:32 -07001410 /*
1411 * For performance, all the general caches are L1 aligned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001412 * This should be particularly beneficial on SMP boxes, as it
1413 * eliminates "false sharing".
1414 * Note for systems short on memory removing the alignment will
Christoph Lametere498be72005-09-09 13:03:32 -07001415 * allow tighter packing of the smaller caches.
1416 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001417 if (!sizes->cs_cachep) {
Christoph Lametere498be72005-09-09 13:03:32 -07001418 sizes->cs_cachep = kmem_cache_create(names->name,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001419 sizes->cs_size,
1420 ARCH_KMALLOC_MINALIGN,
1421 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1422 NULL, NULL);
1423 }
Arjan van de Venf1aaee52006-07-13 14:46:03 +02001424 init_lock_keys(sizes);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001425
Linus Torvalds1da177e2005-04-16 15:20:36 -07001426 sizes->cs_dmacachep = kmem_cache_create(names->name_dma,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001427 sizes->cs_size,
1428 ARCH_KMALLOC_MINALIGN,
1429 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1430 SLAB_PANIC,
1431 NULL, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001432 sizes++;
1433 names++;
1434 }
1435 /* 4) Replace the bootstrap head arrays */
1436 {
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001437 struct array_cache *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001438
Linus Torvalds1da177e2005-04-16 15:20:36 -07001439 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001440
Linus Torvalds1da177e2005-04-16 15:20:36 -07001441 local_irq_disable();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001442 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1443 memcpy(ptr, cpu_cache_get(&cache_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001444 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001445 /*
1446 * Do not assume that spinlocks can be initialized via memcpy:
1447 */
1448 spin_lock_init(&ptr->lock);
1449
Linus Torvalds1da177e2005-04-16 15:20:36 -07001450 cache_cache.array[smp_processor_id()] = ptr;
1451 local_irq_enable();
Christoph Lametere498be72005-09-09 13:03:32 -07001452
Linus Torvalds1da177e2005-04-16 15:20:36 -07001453 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001454
Linus Torvalds1da177e2005-04-16 15:20:36 -07001455 local_irq_disable();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001456 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001457 != &initarray_generic.cache);
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001458 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001459 sizeof(struct arraycache_init));
Ingo Molnar2b2d5492006-07-03 00:25:28 -07001460 /*
1461 * Do not assume that spinlocks can be initialized via memcpy:
1462 */
1463 spin_lock_init(&ptr->lock);
1464
Christoph Lametere498be72005-09-09 13:03:32 -07001465 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001466 ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001467 local_irq_enable();
1468 }
Christoph Lametere498be72005-09-09 13:03:32 -07001469 /* 5) Replace the bootstrap kmem_list3's */
1470 {
1471 int node;
1472 /* Replace the static kmem_list3 structures for the boot cpu */
1473 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE],
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001474 numa_node_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001475
Christoph Lametere498be72005-09-09 13:03:32 -07001476 for_each_online_node(node) {
1477 init_list(malloc_sizes[INDEX_AC].cs_cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001478 &initkmem_list3[SIZE_AC + node], node);
Christoph Lametere498be72005-09-09 13:03:32 -07001479
1480 if (INDEX_AC != INDEX_L3) {
1481 init_list(malloc_sizes[INDEX_L3].cs_cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001482 &initkmem_list3[SIZE_L3 + node],
1483 node);
Christoph Lametere498be72005-09-09 13:03:32 -07001484 }
1485 }
1486 }
1487
1488 /* 6) resize the head arrays to their final sizes */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001489 {
Pekka Enberg343e0d72006-02-01 03:05:50 -08001490 struct kmem_cache *cachep;
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001491 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001492 list_for_each_entry(cachep, &cache_chain, next)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001493 enable_cpucache(cachep);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001494 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001495 }
1496
1497 /* Done! */
1498 g_cpucache_up = FULL;
1499
Andrew Mortona737b3e2006-03-22 00:08:11 -08001500 /*
1501 * Register a cpu startup notifier callback that initializes
1502 * cpu_cache_get for all new cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07001503 */
1504 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001505
Andrew Mortona737b3e2006-03-22 00:08:11 -08001506 /*
1507 * The reap timers are started later, with a module init call: That part
1508 * of the kernel is not yet operational.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001509 */
1510}
1511
1512static int __init cpucache_init(void)
1513{
1514 int cpu;
1515
Andrew Mortona737b3e2006-03-22 00:08:11 -08001516 /*
1517 * Register the timers that return unneeded pages to the page allocator
Linus Torvalds1da177e2005-04-16 15:20:36 -07001518 */
Christoph Lametere498be72005-09-09 13:03:32 -07001519 for_each_online_cpu(cpu)
Andrew Mortona737b3e2006-03-22 00:08:11 -08001520 start_cpu_timer(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001521 return 0;
1522}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001523__initcall(cpucache_init);
1524
1525/*
1526 * Interface to system's page allocator. No need to hold the cache-lock.
1527 *
1528 * If we requested dmaable memory, we will get it. Even if we
1529 * did not request dmaable memory, we might get it, but that
1530 * would be relatively rare and ignorable.
1531 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001532static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001533{
1534 struct page *page;
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001535 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001536 int i;
1537
Luke Yangd6fef9d2006-04-10 22:52:56 -07001538#ifndef CONFIG_MMU
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001539 /*
1540 * Nommu uses slab's for process anonymous memory allocations, and thus
1541 * requires __GFP_COMP to properly refcount higher order allocations
Luke Yangd6fef9d2006-04-10 22:52:56 -07001542 */
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001543 flags |= __GFP_COMP;
Luke Yangd6fef9d2006-04-10 22:52:56 -07001544#endif
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001545 flags |= cachep->gfpflags;
1546
1547 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001548 if (!page)
1549 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001550
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001551 nr_pages = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001552 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001553 atomic_add(nr_pages, &slab_reclaim_pages);
Christoph Lameter9a865ff2006-06-30 01:55:38 -07001554 add_zone_page_state(page_zone(page), NR_SLAB, nr_pages);
Christoph Hellwige1b6aa62006-06-23 02:03:17 -07001555 for (i = 0; i < nr_pages; i++)
1556 __SetPageSlab(page + i);
1557 return page_address(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001558}
1559
1560/*
1561 * Interface to system's page release.
1562 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001563static void kmem_freepages(struct kmem_cache *cachep, void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001564{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001565 unsigned long i = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001566 struct page *page = virt_to_page(addr);
1567 const unsigned long nr_freed = i;
1568
Christoph Lameter9a865ff2006-06-30 01:55:38 -07001569 sub_zone_page_state(page_zone(page), NR_SLAB, nr_freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001570 while (i--) {
Nick Pigginf205b2f2006-03-22 00:08:02 -08001571 BUG_ON(!PageSlab(page));
1572 __ClearPageSlab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001573 page++;
1574 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001575 if (current->reclaim_state)
1576 current->reclaim_state->reclaimed_slab += nr_freed;
1577 free_pages((unsigned long)addr, cachep->gfporder);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001578 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1579 atomic_sub(1 << cachep->gfporder, &slab_reclaim_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001580}
1581
1582static void kmem_rcu_free(struct rcu_head *head)
1583{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001584 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001585 struct kmem_cache *cachep = slab_rcu->cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001586
1587 kmem_freepages(cachep, slab_rcu->addr);
1588 if (OFF_SLAB(cachep))
1589 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1590}
1591
1592#if DEBUG
1593
1594#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001595static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001596 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001597{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001598 int size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001599
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001600 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001601
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001602 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001603 return;
1604
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001605 *addr++ = 0x12345678;
1606 *addr++ = caller;
1607 *addr++ = smp_processor_id();
1608 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001609 {
1610 unsigned long *sptr = &caller;
1611 unsigned long svalue;
1612
1613 while (!kstack_end(sptr)) {
1614 svalue = *sptr++;
1615 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001616 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001617 size -= sizeof(unsigned long);
1618 if (size <= sizeof(unsigned long))
1619 break;
1620 }
1621 }
1622
1623 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001624 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001625}
1626#endif
1627
Pekka Enberg343e0d72006-02-01 03:05:50 -08001628static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001629{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001630 int size = obj_size(cachep);
1631 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001632
1633 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001634 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001635}
1636
1637static void dump_line(char *data, int offset, int limit)
1638{
1639 int i;
1640 printk(KERN_ERR "%03x:", offset);
Andrew Mortona737b3e2006-03-22 00:08:11 -08001641 for (i = 0; i < limit; i++)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001642 printk(" %02x", (unsigned char)data[offset + i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001643 printk("\n");
1644}
1645#endif
1646
1647#if DEBUG
1648
Pekka Enberg343e0d72006-02-01 03:05:50 -08001649static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001650{
1651 int i, size;
1652 char *realobj;
1653
1654 if (cachep->flags & SLAB_RED_ZONE) {
1655 printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001656 *dbg_redzone1(cachep, objp),
1657 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001658 }
1659
1660 if (cachep->flags & SLAB_STORE_USER) {
1661 printk(KERN_ERR "Last user: [<%p>]",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001662 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001663 print_symbol("(%s)",
Andrew Mortona737b3e2006-03-22 00:08:11 -08001664 (unsigned long)*dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001665 printk("\n");
1666 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001667 realobj = (char *)objp + obj_offset(cachep);
1668 size = obj_size(cachep);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001669 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001670 int limit;
1671 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001672 if (i + limit > size)
1673 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001674 dump_line(realobj, i, limit);
1675 }
1676}
1677
Pekka Enberg343e0d72006-02-01 03:05:50 -08001678static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001679{
1680 char *realobj;
1681 int size, i;
1682 int lines = 0;
1683
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001684 realobj = (char *)objp + obj_offset(cachep);
1685 size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001686
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001687 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001688 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001689 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001690 exp = POISON_END;
1691 if (realobj[i] != exp) {
1692 int limit;
1693 /* Mismatch ! */
1694 /* Print header */
1695 if (lines == 0) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001696 printk(KERN_ERR
Andrew Mortona737b3e2006-03-22 00:08:11 -08001697 "Slab corruption: start=%p, len=%d\n",
1698 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001699 print_objinfo(cachep, objp, 0);
1700 }
1701 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001702 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001703 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001704 if (i + limit > size)
1705 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001706 dump_line(realobj, i, limit);
1707 i += 16;
1708 lines++;
1709 /* Limit to 5 lines */
1710 if (lines > 5)
1711 break;
1712 }
1713 }
1714 if (lines != 0) {
1715 /* Print some data about the neighboring objects, if they
1716 * exist:
1717 */
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08001718 struct slab *slabp = virt_to_slab(objp);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001719 unsigned int objnr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001720
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001721 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001722 if (objnr) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001723 objp = index_to_obj(cachep, slabp, objnr - 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001724 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001725 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001726 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001727 print_objinfo(cachep, objp, 2);
1728 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001729 if (objnr + 1 < cachep->num) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001730 objp = index_to_obj(cachep, slabp, objnr + 1);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001731 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001732 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001733 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001734 print_objinfo(cachep, objp, 2);
1735 }
1736 }
1737}
1738#endif
1739
Linus Torvalds1da177e2005-04-16 15:20:36 -07001740#if DEBUG
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001741/**
Randy Dunlap911851e2006-03-22 00:08:14 -08001742 * slab_destroy_objs - destroy a slab and its objects
1743 * @cachep: cache pointer being destroyed
1744 * @slabp: slab pointer being destroyed
1745 *
1746 * Call the registered destructor for each object in a slab that is being
1747 * destroyed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001748 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001749static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001750{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001751 int i;
1752 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001753 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001754
1755 if (cachep->flags & SLAB_POISON) {
1756#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08001757 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1758 OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001759 kernel_map_pages(virt_to_page(objp),
Andrew Mortona737b3e2006-03-22 00:08:11 -08001760 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001761 else
1762 check_poison_obj(cachep, objp);
1763#else
1764 check_poison_obj(cachep, objp);
1765#endif
1766 }
1767 if (cachep->flags & SLAB_RED_ZONE) {
1768 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1769 slab_error(cachep, "start of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001770 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001771 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1772 slab_error(cachep, "end of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001773 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001774 }
1775 if (cachep->dtor && !(cachep->flags & SLAB_POISON))
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001776 (cachep->dtor) (objp + obj_offset(cachep), cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001777 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001778}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001779#else
Pekka Enberg343e0d72006-02-01 03:05:50 -08001780static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001781{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001782 if (cachep->dtor) {
1783 int i;
1784 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08001785 void *objp = index_to_obj(cachep, slabp, i);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001786 (cachep->dtor) (objp, cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001787 }
1788 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001789}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001790#endif
1791
Randy Dunlap911851e2006-03-22 00:08:14 -08001792/**
1793 * slab_destroy - destroy and release all objects in a slab
1794 * @cachep: cache pointer being destroyed
1795 * @slabp: slab pointer being destroyed
1796 *
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001797 * Destroy all the objs in a slab, and release the mem back to the system.
Andrew Mortona737b3e2006-03-22 00:08:11 -08001798 * Before calling the slab must have been unlinked from the cache. The
1799 * cache-lock is not held/needed.
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001800 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001801static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001802{
1803 void *addr = slabp->s_mem - slabp->colouroff;
1804
1805 slab_destroy_objs(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001806 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1807 struct slab_rcu *slab_rcu;
1808
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001809 slab_rcu = (struct slab_rcu *)slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001810 slab_rcu->cachep = cachep;
1811 slab_rcu->addr = addr;
1812 call_rcu(&slab_rcu->head, kmem_rcu_free);
1813 } else {
1814 kmem_freepages(cachep, addr);
Ingo Molnar873623d2006-07-13 14:44:38 +02001815 if (OFF_SLAB(cachep))
1816 kmem_cache_free(cachep->slabp_cache, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001817 }
1818}
1819
Andrew Mortona737b3e2006-03-22 00:08:11 -08001820/*
1821 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1822 * size of kmem_list3.
1823 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001824static void set_up_list3s(struct kmem_cache *cachep, int index)
Christoph Lametere498be72005-09-09 13:03:32 -07001825{
1826 int node;
1827
1828 for_each_online_node(node) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001829 cachep->nodelists[node] = &initkmem_list3[index + node];
Christoph Lametere498be72005-09-09 13:03:32 -07001830 cachep->nodelists[node]->next_reap = jiffies +
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001831 REAPTIMEOUT_LIST3 +
1832 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -07001833 }
1834}
1835
Linus Torvalds1da177e2005-04-16 15:20:36 -07001836/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08001837 * calculate_slab_order - calculate size (page order) of slabs
1838 * @cachep: pointer to the cache that is being created
1839 * @size: size of objects to be created in this cache.
1840 * @align: required alignment for the objects.
1841 * @flags: slab allocation flags
1842 *
1843 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001844 *
1845 * This could be made much more intelligent. For now, try to avoid using
1846 * high order pages for slabs. When the gfp() functions are more friendly
1847 * towards high-order requests, this should be changed.
1848 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001849static size_t calculate_slab_order(struct kmem_cache *cachep,
Randy Dunlapee13d782006-02-01 03:05:53 -08001850 size_t size, size_t align, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001851{
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001852 unsigned long offslab_limit;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001853 size_t left_over = 0;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001854 int gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001855
Andrew Mortona737b3e2006-03-22 00:08:11 -08001856 for (gfporder = 0; gfporder <= MAX_GFP_ORDER; gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001857 unsigned int num;
1858 size_t remainder;
1859
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001860 cache_estimate(gfporder, size, align, flags, &remainder, &num);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001861 if (!num)
1862 continue;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001863
Ingo Molnarb1ab41c2006-06-02 15:44:58 +02001864 if (flags & CFLGS_OFF_SLAB) {
1865 /*
1866 * Max number of objs-per-slab for caches which
1867 * use off-slab slabs. Needed to avoid a possible
1868 * looping condition in cache_grow().
1869 */
1870 offslab_limit = size - sizeof(struct slab);
1871 offslab_limit /= sizeof(kmem_bufctl_t);
1872
1873 if (num > offslab_limit)
1874 break;
1875 }
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001876
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001877 /* Found something acceptable - save it away */
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001878 cachep->num = num;
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001879 cachep->gfporder = gfporder;
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001880 left_over = remainder;
1881
1882 /*
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08001883 * A VFS-reclaimable slab tends to have most allocations
1884 * as GFP_NOFS and we really don't want to have to be allocating
1885 * higher-order pages when we are unable to shrink dcache.
1886 */
1887 if (flags & SLAB_RECLAIM_ACCOUNT)
1888 break;
1889
1890 /*
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001891 * Large number of objects is good, but very large slabs are
1892 * currently bad for the gfp()s.
1893 */
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001894 if (gfporder >= slab_break_gfp_order)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001895 break;
1896
Linus Torvalds9888e6f2006-03-06 17:44:43 -08001897 /*
1898 * Acceptable internal fragmentation?
1899 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08001900 if (left_over * 8 <= (PAGE_SIZE << gfporder))
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001901 break;
1902 }
1903 return left_over;
1904}
1905
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08001906static void setup_cpu_cache(struct kmem_cache *cachep)
1907{
1908 if (g_cpucache_up == FULL) {
1909 enable_cpucache(cachep);
1910 return;
1911 }
1912 if (g_cpucache_up == NONE) {
1913 /*
1914 * Note: the first kmem_cache_create must create the cache
1915 * that's used by kmalloc(24), otherwise the creation of
1916 * further caches will BUG().
1917 */
1918 cachep->array[smp_processor_id()] = &initarray_generic.cache;
1919
1920 /*
1921 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
1922 * the first cache, then we need to set up all its list3s,
1923 * otherwise the creation of further caches will BUG().
1924 */
1925 set_up_list3s(cachep, SIZE_AC);
1926 if (INDEX_AC == INDEX_L3)
1927 g_cpucache_up = PARTIAL_L3;
1928 else
1929 g_cpucache_up = PARTIAL_AC;
1930 } else {
1931 cachep->array[smp_processor_id()] =
1932 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1933
1934 if (g_cpucache_up == PARTIAL_AC) {
1935 set_up_list3s(cachep, SIZE_L3);
1936 g_cpucache_up = PARTIAL_L3;
1937 } else {
1938 int node;
1939 for_each_online_node(node) {
1940 cachep->nodelists[node] =
1941 kmalloc_node(sizeof(struct kmem_list3),
1942 GFP_KERNEL, node);
1943 BUG_ON(!cachep->nodelists[node]);
1944 kmem_list3_init(cachep->nodelists[node]);
1945 }
1946 }
1947 }
1948 cachep->nodelists[numa_node_id()]->next_reap =
1949 jiffies + REAPTIMEOUT_LIST3 +
1950 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1951
1952 cpu_cache_get(cachep)->avail = 0;
1953 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1954 cpu_cache_get(cachep)->batchcount = 1;
1955 cpu_cache_get(cachep)->touched = 0;
1956 cachep->batchcount = 1;
1957 cachep->limit = BOOT_CPUCACHE_ENTRIES;
1958}
1959
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001960/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07001961 * kmem_cache_create - Create a cache.
1962 * @name: A string which is used in /proc/slabinfo to identify this cache.
1963 * @size: The size of objects to be created in this cache.
1964 * @align: The required alignment for the objects.
1965 * @flags: SLAB flags
1966 * @ctor: A constructor for the objects.
1967 * @dtor: A destructor for the objects.
1968 *
1969 * Returns a ptr to the cache on success, NULL on failure.
1970 * Cannot be called within a int, but can be interrupted.
1971 * The @ctor is run when new pages are allocated by the cache
1972 * and the @dtor is run before the pages are handed back.
1973 *
1974 * @name must be valid until the cache is destroyed. This implies that
Andrew Mortona737b3e2006-03-22 00:08:11 -08001975 * the module calling this has to destroy the cache before getting unloaded.
1976 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001977 * The flags are
1978 *
1979 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1980 * to catch references to uninitialised memory.
1981 *
1982 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1983 * for buffer overruns.
1984 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001985 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1986 * cacheline. This can be beneficial if you're counting cycles as closely
1987 * as davem.
1988 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001989struct kmem_cache *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001990kmem_cache_create (const char *name, size_t size, size_t align,
Andrew Mortona737b3e2006-03-22 00:08:11 -08001991 unsigned long flags,
1992 void (*ctor)(void*, struct kmem_cache *, unsigned long),
Pekka Enberg343e0d72006-02-01 03:05:50 -08001993 void (*dtor)(void*, struct kmem_cache *, unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001994{
1995 size_t left_over, slab_size, ralign;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07001996 struct kmem_cache *cachep = NULL, *pc;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001997
1998 /*
1999 * Sanity checks... these are all serious usage bugs.
2000 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002001 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002002 (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002003 printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__,
2004 name);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002005 BUG();
2006 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002007
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002008 /*
2009 * Prevent CPUs from coming and going.
2010 * lock_cpu_hotplug() nests outside cache_chain_mutex
2011 */
2012 lock_cpu_hotplug();
2013
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002014 mutex_lock(&cache_chain_mutex);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002015
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07002016 list_for_each_entry(pc, &cache_chain, next) {
Andrew Morton4f12bb42005-11-07 00:58:00 -08002017 mm_segment_t old_fs = get_fs();
2018 char tmp;
2019 int res;
2020
2021 /*
2022 * This happens when the module gets unloaded and doesn't
2023 * destroy its slab cache and no-one else reuses the vmalloc
2024 * area of the module. Print a warning.
2025 */
2026 set_fs(KERNEL_DS);
2027 res = __get_user(tmp, pc->name);
2028 set_fs(old_fs);
2029 if (res) {
2030 printk("SLAB: cache with size %d has lost its name\n",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002031 pc->buffer_size);
Andrew Morton4f12bb42005-11-07 00:58:00 -08002032 continue;
2033 }
2034
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002035 if (!strcmp(pc->name, name)) {
Andrew Morton4f12bb42005-11-07 00:58:00 -08002036 printk("kmem_cache_create: duplicate cache %s\n", name);
2037 dump_stack();
2038 goto oops;
2039 }
2040 }
2041
Linus Torvalds1da177e2005-04-16 15:20:36 -07002042#if DEBUG
2043 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
2044 if ((flags & SLAB_DEBUG_INITIAL) && !ctor) {
2045 /* No constructor, but inital state check requested */
2046 printk(KERN_ERR "%s: No con, but init state check "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002047 "requested - %s\n", __FUNCTION__, name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002048 flags &= ~SLAB_DEBUG_INITIAL;
2049 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002050#if FORCED_DEBUG
2051 /*
2052 * Enable redzoning and last user accounting, except for caches with
2053 * large objects, if the increased size would increase the object size
2054 * above the next power of two: caches with object sizes just above a
2055 * power of two have a significant amount of internal fragmentation.
2056 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002057 if (size < 4096 || fls(size - 1) == fls(size-1 + 3 * BYTES_PER_WORD))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002058 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002059 if (!(flags & SLAB_DESTROY_BY_RCU))
2060 flags |= SLAB_POISON;
2061#endif
2062 if (flags & SLAB_DESTROY_BY_RCU)
2063 BUG_ON(flags & SLAB_POISON);
2064#endif
2065 if (flags & SLAB_DESTROY_BY_RCU)
2066 BUG_ON(dtor);
2067
2068 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002069 * Always checks flags, a caller might be expecting debug support which
2070 * isn't available.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002071 */
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002072 BUG_ON(flags & ~CREATE_MASK);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002073
Andrew Mortona737b3e2006-03-22 00:08:11 -08002074 /*
2075 * Check that size is in terms of words. This is needed to avoid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002076 * unaligned accesses for some archs when redzoning is used, and makes
2077 * sure any on-slab bufctl's are also correctly aligned.
2078 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002079 if (size & (BYTES_PER_WORD - 1)) {
2080 size += (BYTES_PER_WORD - 1);
2081 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002082 }
2083
Andrew Mortona737b3e2006-03-22 00:08:11 -08002084 /* calculate the final buffer alignment: */
2085
Linus Torvalds1da177e2005-04-16 15:20:36 -07002086 /* 1) arch recommendation: can be overridden for debug */
2087 if (flags & SLAB_HWCACHE_ALIGN) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002088 /*
2089 * Default alignment: as specified by the arch code. Except if
2090 * an object is really small, then squeeze multiple objects into
2091 * one cacheline.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002092 */
2093 ralign = cache_line_size();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002094 while (size <= ralign / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002095 ralign /= 2;
2096 } else {
2097 ralign = BYTES_PER_WORD;
2098 }
Pekka Enbergca5f9702006-09-25 23:31:25 -07002099
2100 /*
2101 * Redzoning and user store require word alignment. Note this will be
2102 * overridden by architecture or caller mandated alignment if either
2103 * is greater than BYTES_PER_WORD.
2104 */
2105 if (flags & SLAB_RED_ZONE || flags & SLAB_STORE_USER)
2106 ralign = BYTES_PER_WORD;
2107
Linus Torvalds1da177e2005-04-16 15:20:36 -07002108 /* 2) arch mandated alignment: disables debug if necessary */
2109 if (ralign < ARCH_SLAB_MINALIGN) {
2110 ralign = ARCH_SLAB_MINALIGN;
2111 if (ralign > BYTES_PER_WORD)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002112 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002113 }
2114 /* 3) caller mandated alignment: disables debug if necessary */
2115 if (ralign < align) {
2116 ralign = align;
2117 if (ralign > BYTES_PER_WORD)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002118 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002119 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08002120 /*
Pekka Enbergca5f9702006-09-25 23:31:25 -07002121 * 4) Store it.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002122 */
2123 align = ralign;
2124
2125 /* Get cache's description obj. */
Pekka Enbergc5e3b832006-03-25 03:06:43 -08002126 cachep = kmem_cache_zalloc(&cache_cache, SLAB_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002127 if (!cachep)
Andrew Morton4f12bb42005-11-07 00:58:00 -08002128 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002129
2130#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002131 cachep->obj_size = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002132
Pekka Enbergca5f9702006-09-25 23:31:25 -07002133 /*
2134 * Both debugging options require word-alignment which is calculated
2135 * into align above.
2136 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002137 if (flags & SLAB_RED_ZONE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002138 /* add space for red zone words */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002139 cachep->obj_offset += BYTES_PER_WORD;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002140 size += 2 * BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002141 }
2142 if (flags & SLAB_STORE_USER) {
Pekka Enbergca5f9702006-09-25 23:31:25 -07002143 /* user store requires one word storage behind the end of
2144 * the real object.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002145 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002146 size += BYTES_PER_WORD;
2147 }
2148#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002149 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002150 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2151 cachep->obj_offset += PAGE_SIZE - size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002152 size = PAGE_SIZE;
2153 }
2154#endif
2155#endif
2156
Ingo Molnare0a42722006-06-23 02:03:46 -07002157 /*
2158 * Determine if the slab management is 'on' or 'off' slab.
2159 * (bootstrapping cannot cope with offslab caches so don't do
2160 * it too early on.)
2161 */
2162 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002163 /*
2164 * Size is large, assume best to place the slab management obj
2165 * off-slab (should allow better packing of objs).
2166 */
2167 flags |= CFLGS_OFF_SLAB;
2168
2169 size = ALIGN(size, align);
2170
Linus Torvaldsf78bb8a2006-03-08 10:33:05 -08002171 left_over = calculate_slab_order(cachep, size, align, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002172
2173 if (!cachep->num) {
2174 printk("kmem_cache_create: couldn't create cache %s.\n", name);
2175 kmem_cache_free(&cache_cache, cachep);
2176 cachep = NULL;
Andrew Morton4f12bb42005-11-07 00:58:00 -08002177 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002178 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002179 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2180 + sizeof(struct slab), align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002181
2182 /*
2183 * If the slab has been placed off-slab, and we have enough space then
2184 * move it on-slab. This is at the expense of any extra colouring.
2185 */
2186 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2187 flags &= ~CFLGS_OFF_SLAB;
2188 left_over -= slab_size;
2189 }
2190
2191 if (flags & CFLGS_OFF_SLAB) {
2192 /* really off slab. No need for manual alignment */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002193 slab_size =
2194 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002195 }
2196
2197 cachep->colour_off = cache_line_size();
2198 /* Offset must be a multiple of the alignment. */
2199 if (cachep->colour_off < align)
2200 cachep->colour_off = align;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002201 cachep->colour = left_over / cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002202 cachep->slab_size = slab_size;
2203 cachep->flags = flags;
2204 cachep->gfpflags = 0;
2205 if (flags & SLAB_CACHE_DMA)
2206 cachep->gfpflags |= GFP_DMA;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002207 cachep->buffer_size = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002208
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002209 if (flags & CFLGS_OFF_SLAB) {
Victor Fuscob2d55072005-09-10 00:26:36 -07002210 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002211 /*
2212 * This is a possibility for one of the malloc_sizes caches.
2213 * But since we go off slab only for object size greater than
2214 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2215 * this should not happen at all.
2216 * But leave a BUG_ON for some lucky dude.
2217 */
2218 BUG_ON(!cachep->slabp_cache);
2219 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002220 cachep->ctor = ctor;
2221 cachep->dtor = dtor;
2222 cachep->name = name;
2223
Linus Torvalds1da177e2005-04-16 15:20:36 -07002224
Pekka Enbergf30cf7d2006-03-22 00:08:11 -08002225 setup_cpu_cache(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002226
Linus Torvalds1da177e2005-04-16 15:20:36 -07002227 /* cache setup completed, link it into the list */
2228 list_add(&cachep->next, &cache_chain);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002229oops:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002230 if (!cachep && (flags & SLAB_PANIC))
2231 panic("kmem_cache_create(): failed to create slab `%s'\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002232 name);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002233 mutex_unlock(&cache_chain_mutex);
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08002234 unlock_cpu_hotplug();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002235 return cachep;
2236}
2237EXPORT_SYMBOL(kmem_cache_create);
2238
2239#if DEBUG
2240static void check_irq_off(void)
2241{
2242 BUG_ON(!irqs_disabled());
2243}
2244
2245static void check_irq_on(void)
2246{
2247 BUG_ON(irqs_disabled());
2248}
2249
Pekka Enberg343e0d72006-02-01 03:05:50 -08002250static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002251{
2252#ifdef CONFIG_SMP
2253 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002254 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002255#endif
2256}
Christoph Lametere498be72005-09-09 13:03:32 -07002257
Pekka Enberg343e0d72006-02-01 03:05:50 -08002258static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002259{
2260#ifdef CONFIG_SMP
2261 check_irq_off();
2262 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2263#endif
2264}
2265
Linus Torvalds1da177e2005-04-16 15:20:36 -07002266#else
2267#define check_irq_off() do { } while(0)
2268#define check_irq_on() do { } while(0)
2269#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002270#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002271#endif
2272
Christoph Lameteraab22072006-03-22 00:09:06 -08002273static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2274 struct array_cache *ac,
2275 int force, int node);
2276
Linus Torvalds1da177e2005-04-16 15:20:36 -07002277static void do_drain(void *arg)
2278{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002279 struct kmem_cache *cachep = arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002280 struct array_cache *ac;
Christoph Lameterff694162005-09-22 21:44:02 -07002281 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002282
2283 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002284 ac = cpu_cache_get(cachep);
Christoph Lameterff694162005-09-22 21:44:02 -07002285 spin_lock(&cachep->nodelists[node]->list_lock);
2286 free_block(cachep, ac->entry, ac->avail, node);
2287 spin_unlock(&cachep->nodelists[node]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002288 ac->avail = 0;
2289}
2290
Pekka Enberg343e0d72006-02-01 03:05:50 -08002291static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002292{
Christoph Lametere498be72005-09-09 13:03:32 -07002293 struct kmem_list3 *l3;
2294 int node;
2295
Andrew Mortona07fa392006-03-22 00:08:17 -08002296 on_each_cpu(do_drain, cachep, 1, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002297 check_irq_on();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002298 for_each_online_node(node) {
Christoph Lametere498be72005-09-09 13:03:32 -07002299 l3 = cachep->nodelists[node];
Roland Dreiera4523a82006-05-15 11:41:00 -07002300 if (l3 && l3->alien)
2301 drain_alien_cache(cachep, l3->alien);
2302 }
2303
2304 for_each_online_node(node) {
2305 l3 = cachep->nodelists[node];
2306 if (l3)
Christoph Lameteraab22072006-03-22 00:09:06 -08002307 drain_array(cachep, l3, l3->shared, 1, node);
Christoph Lametere498be72005-09-09 13:03:32 -07002308 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002309}
2310
Christoph Lametered11d9e2006-06-30 01:55:45 -07002311/*
2312 * Remove slabs from the list of free slabs.
2313 * Specify the number of slabs to drain in tofree.
2314 *
2315 * Returns the actual number of slabs released.
2316 */
2317static int drain_freelist(struct kmem_cache *cache,
2318 struct kmem_list3 *l3, int tofree)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002319{
Christoph Lametered11d9e2006-06-30 01:55:45 -07002320 struct list_head *p;
2321 int nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002322 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002323
Christoph Lametered11d9e2006-06-30 01:55:45 -07002324 nr_freed = 0;
2325 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002326
Christoph Lametered11d9e2006-06-30 01:55:45 -07002327 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002328 p = l3->slabs_free.prev;
Christoph Lametered11d9e2006-06-30 01:55:45 -07002329 if (p == &l3->slabs_free) {
2330 spin_unlock_irq(&l3->list_lock);
2331 goto out;
2332 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002333
Christoph Lametered11d9e2006-06-30 01:55:45 -07002334 slabp = list_entry(p, struct slab, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002335#if DEBUG
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002336 BUG_ON(slabp->inuse);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002337#endif
2338 list_del(&slabp->list);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002339 /*
2340 * Safe to drop the lock. The slab is no longer linked
2341 * to the cache.
2342 */
2343 l3->free_objects -= cache->num;
Christoph Lametere498be72005-09-09 13:03:32 -07002344 spin_unlock_irq(&l3->list_lock);
Christoph Lametered11d9e2006-06-30 01:55:45 -07002345 slab_destroy(cache, slabp);
2346 nr_freed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002347 }
Christoph Lametered11d9e2006-06-30 01:55:45 -07002348out:
2349 return nr_freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002350}
2351
Pekka Enberg343e0d72006-02-01 03:05:50 -08002352static int __cache_shrink(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07002353{
2354 int ret = 0, i = 0;
2355 struct kmem_list3 *l3;
2356
2357 drain_cpu_caches(cachep);
2358
2359 check_irq_on();
2360 for_each_online_node(i) {
2361 l3 = cachep->nodelists[i];
Christoph Lametered11d9e2006-06-30 01:55:45 -07002362 if (!l3)
2363 continue;
2364
2365 drain_freelist(cachep, l3, l3->free_objects);
2366
2367 ret += !list_empty(&l3->slabs_full) ||
2368 !list_empty(&l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07002369 }
2370 return (ret ? 1 : 0);
2371}
2372
Linus Torvalds1da177e2005-04-16 15:20:36 -07002373/**
2374 * kmem_cache_shrink - Shrink a cache.
2375 * @cachep: The cache to shrink.
2376 *
2377 * Releases as many slabs as possible for a cache.
2378 * To help debugging, a zero exit status indicates all slabs were released.
2379 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002380int kmem_cache_shrink(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002381{
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002382 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002383
2384 return __cache_shrink(cachep);
2385}
2386EXPORT_SYMBOL(kmem_cache_shrink);
2387
2388/**
2389 * kmem_cache_destroy - delete a cache
2390 * @cachep: the cache to destroy
2391 *
Pekka Enberg343e0d72006-02-01 03:05:50 -08002392 * Remove a struct kmem_cache object from the slab cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002393 * Returns 0 on success.
2394 *
2395 * It is expected this function will be called by a module when it is
2396 * unloaded. This will remove the cache completely, and avoid a duplicate
2397 * cache being allocated each time a module is loaded and unloaded, if the
2398 * module doesn't have persistent in-kernel storage across loads and unloads.
2399 *
2400 * The cache must be empty before calling this function.
2401 *
2402 * The caller must guarantee that noone will allocate memory from the cache
2403 * during the kmem_cache_destroy().
2404 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002405int kmem_cache_destroy(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002406{
2407 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07002408 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002409
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002410 BUG_ON(!cachep || in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002411
2412 /* Don't let CPUs to come and go */
2413 lock_cpu_hotplug();
2414
2415 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002416 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002417 /*
2418 * the chain is never empty, cache_cache is never destroyed
2419 */
2420 list_del(&cachep->next);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002421 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002422
2423 if (__cache_shrink(cachep)) {
2424 slab_error(cachep, "Can't free all objects");
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002425 mutex_lock(&cache_chain_mutex);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002426 list_add(&cachep->next, &cache_chain);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002427 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002428 unlock_cpu_hotplug();
2429 return 1;
2430 }
2431
2432 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
Paul E. McKenneyfbd568a3e2005-05-01 08:59:04 -07002433 synchronize_rcu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002434
Christoph Lametere498be72005-09-09 13:03:32 -07002435 for_each_online_cpu(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002436 kfree(cachep->array[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002437
2438 /* NUMA: free the list3 structures */
Christoph Lametere498be72005-09-09 13:03:32 -07002439 for_each_online_node(i) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002440 l3 = cachep->nodelists[i];
2441 if (l3) {
Christoph Lametere498be72005-09-09 13:03:32 -07002442 kfree(l3->shared);
2443 free_alien_cache(l3->alien);
2444 kfree(l3);
2445 }
2446 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002447 kmem_cache_free(&cache_cache, cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002448 unlock_cpu_hotplug();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002449 return 0;
2450}
2451EXPORT_SYMBOL(kmem_cache_destroy);
2452
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07002453/*
2454 * Get the memory for a slab management obj.
2455 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2456 * always come from malloc_sizes caches. The slab descriptor cannot
2457 * come from the same cache which is getting created because,
2458 * when we are searching for an appropriate cache for these
2459 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2460 * If we are creating a malloc_sizes cache here it would not be visible to
2461 * kmem_find_general_cachep till the initialization is complete.
2462 * Hence we cannot have slabp_cache same as the original cache.
2463 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002464static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002465 int colour_off, gfp_t local_flags,
2466 int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002467{
2468 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002469
Linus Torvalds1da177e2005-04-16 15:20:36 -07002470 if (OFF_SLAB(cachep)) {
2471 /* Slab management obj is off-slab. */
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002472 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2473 local_flags, nodeid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002474 if (!slabp)
2475 return NULL;
2476 } else {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002477 slabp = objp + colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002478 colour_off += cachep->slab_size;
2479 }
2480 slabp->inuse = 0;
2481 slabp->colouroff = colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002482 slabp->s_mem = objp + colour_off;
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002483 slabp->nodeid = nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002484 return slabp;
2485}
2486
2487static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2488{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002489 return (kmem_bufctl_t *) (slabp + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002490}
2491
Pekka Enberg343e0d72006-02-01 03:05:50 -08002492static void cache_init_objs(struct kmem_cache *cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002493 struct slab *slabp, unsigned long ctor_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002494{
2495 int i;
2496
2497 for (i = 0; i < cachep->num; i++) {
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002498 void *objp = index_to_obj(cachep, slabp, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002499#if DEBUG
2500 /* need to poison the objs? */
2501 if (cachep->flags & SLAB_POISON)
2502 poison_obj(cachep, objp, POISON_FREE);
2503 if (cachep->flags & SLAB_STORE_USER)
2504 *dbg_userword(cachep, objp) = NULL;
2505
2506 if (cachep->flags & SLAB_RED_ZONE) {
2507 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2508 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2509 }
2510 /*
Andrew Mortona737b3e2006-03-22 00:08:11 -08002511 * Constructors are not allowed to allocate memory from the same
2512 * cache which they are a constructor for. Otherwise, deadlock.
2513 * They must also be threaded.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002514 */
2515 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002516 cachep->ctor(objp + obj_offset(cachep), cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002517 ctor_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002518
2519 if (cachep->flags & SLAB_RED_ZONE) {
2520 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2521 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002522 " end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002523 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2524 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002525 " start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002526 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08002527 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2528 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002529 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002530 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002531#else
2532 if (cachep->ctor)
2533 cachep->ctor(objp, cachep, ctor_flags);
2534#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002535 slab_bufctl(slabp)[i] = i + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002536 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002537 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002538 slabp->free = 0;
2539}
2540
Pekka Enberg343e0d72006-02-01 03:05:50 -08002541static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002542{
Andrew Mortona737b3e2006-03-22 00:08:11 -08002543 if (flags & SLAB_DMA)
2544 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2545 else
2546 BUG_ON(cachep->gfpflags & GFP_DMA);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002547}
2548
Andrew Mortona737b3e2006-03-22 00:08:11 -08002549static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2550 int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002551{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002552 void *objp = index_to_obj(cachep, slabp, slabp->free);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002553 kmem_bufctl_t next;
2554
2555 slabp->inuse++;
2556 next = slab_bufctl(slabp)[slabp->free];
2557#if DEBUG
2558 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2559 WARN_ON(slabp->nodeid != nodeid);
2560#endif
2561 slabp->free = next;
2562
2563 return objp;
2564}
2565
Andrew Mortona737b3e2006-03-22 00:08:11 -08002566static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2567 void *objp, int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002568{
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002569 unsigned int objnr = obj_to_index(cachep, slabp, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002570
2571#if DEBUG
2572 /* Verify that the slab belongs to the intended node */
2573 WARN_ON(slabp->nodeid != nodeid);
2574
Al Viro871751e2006-03-25 03:06:39 -08002575 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
Matthew Dobson78d382d2006-02-01 03:05:47 -08002576 printk(KERN_ERR "slab: double free detected in cache "
Andrew Mortona737b3e2006-03-22 00:08:11 -08002577 "'%s', objp %p\n", cachep->name, objp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002578 BUG();
2579 }
2580#endif
2581 slab_bufctl(slabp)[objnr] = slabp->free;
2582 slabp->free = objnr;
2583 slabp->inuse--;
2584}
2585
Pekka Enberg47768742006-06-23 02:03:07 -07002586/*
2587 * Map pages beginning at addr to the given cache and slab. This is required
2588 * for the slab allocator to be able to lookup the cache and slab of a
2589 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2590 */
2591static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2592 void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002593{
Pekka Enberg47768742006-06-23 02:03:07 -07002594 int nr_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002595 struct page *page;
2596
Pekka Enberg47768742006-06-23 02:03:07 -07002597 page = virt_to_page(addr);
Nick Piggin84097512006-03-22 00:08:34 -08002598
Pekka Enberg47768742006-06-23 02:03:07 -07002599 nr_pages = 1;
Nick Piggin84097512006-03-22 00:08:34 -08002600 if (likely(!PageCompound(page)))
Pekka Enberg47768742006-06-23 02:03:07 -07002601 nr_pages <<= cache->gfporder;
2602
Linus Torvalds1da177e2005-04-16 15:20:36 -07002603 do {
Pekka Enberg47768742006-06-23 02:03:07 -07002604 page_set_cache(page, cache);
2605 page_set_slab(page, slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002606 page++;
Pekka Enberg47768742006-06-23 02:03:07 -07002607 } while (--nr_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002608}
2609
2610/*
2611 * Grow (by 1) the number of slabs within a cache. This is called by
2612 * kmem_cache_alloc() when there are no active objs left in a cache.
2613 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002614static int cache_grow(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002615{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002616 struct slab *slabp;
2617 void *objp;
2618 size_t offset;
2619 gfp_t local_flags;
2620 unsigned long ctor_flags;
Christoph Lametere498be72005-09-09 13:03:32 -07002621 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002622
Andrew Mortona737b3e2006-03-22 00:08:11 -08002623 /*
2624 * Be lazy and only check for valid flags here, keeping it out of the
2625 * critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002626 */
Eric Sesterhenn40094fa2006-04-02 13:49:25 +02002627 BUG_ON(flags & ~(SLAB_DMA | SLAB_LEVEL_MASK | SLAB_NO_GROW));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002628 if (flags & SLAB_NO_GROW)
2629 return 0;
2630
2631 ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2632 local_flags = (flags & SLAB_LEVEL_MASK);
2633 if (!(local_flags & __GFP_WAIT))
2634 /*
2635 * Not allowed to sleep. Need to tell a constructor about
2636 * this - it might need to know...
2637 */
2638 ctor_flags |= SLAB_CTOR_ATOMIC;
2639
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002640 /* Take the l3 list lock to change the colour_next on this node */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002641 check_irq_off();
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002642 l3 = cachep->nodelists[nodeid];
2643 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002644
2645 /* Get colour for the slab, and cal the next value. */
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002646 offset = l3->colour_next;
2647 l3->colour_next++;
2648 if (l3->colour_next >= cachep->colour)
2649 l3->colour_next = 0;
2650 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002651
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002652 offset *= cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002653
2654 if (local_flags & __GFP_WAIT)
2655 local_irq_enable();
2656
2657 /*
2658 * The test for missing atomic flag is performed here, rather than
2659 * the more obvious place, simply to reduce the critical path length
2660 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2661 * will eventually be caught here (where it matters).
2662 */
2663 kmem_flagcheck(cachep, flags);
2664
Andrew Mortona737b3e2006-03-22 00:08:11 -08002665 /*
2666 * Get mem for the objs. Attempt to allocate a physical page from
2667 * 'nodeid'.
Christoph Lametere498be72005-09-09 13:03:32 -07002668 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08002669 objp = kmem_getpages(cachep, flags, nodeid);
2670 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002671 goto failed;
2672
2673 /* Get slab management. */
Ravikiran G Thirumalai5b74ada2006-04-10 22:52:53 -07002674 slabp = alloc_slabmgmt(cachep, objp, offset, local_flags, nodeid);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002675 if (!slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002676 goto opps1;
2677
Christoph Lametere498be72005-09-09 13:03:32 -07002678 slabp->nodeid = nodeid;
Pekka Enberg47768742006-06-23 02:03:07 -07002679 slab_map_pages(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002680
2681 cache_init_objs(cachep, slabp, ctor_flags);
2682
2683 if (local_flags & __GFP_WAIT)
2684 local_irq_disable();
2685 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002686 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002687
2688 /* Make slab active. */
Christoph Lametere498be72005-09-09 13:03:32 -07002689 list_add_tail(&slabp->list, &(l3->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002690 STATS_INC_GROWN(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002691 l3->free_objects += cachep->num;
2692 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002693 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002694opps1:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002695 kmem_freepages(cachep, objp);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002696failed:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002697 if (local_flags & __GFP_WAIT)
2698 local_irq_disable();
2699 return 0;
2700}
2701
2702#if DEBUG
2703
2704/*
2705 * Perform extra freeing checks:
2706 * - detect bad pointers.
2707 * - POISON/RED_ZONE checking
2708 * - destructor calls, for caches with POISON+dtor
2709 */
2710static void kfree_debugcheck(const void *objp)
2711{
2712 struct page *page;
2713
2714 if (!virt_addr_valid(objp)) {
2715 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002716 (unsigned long)objp);
2717 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002718 }
2719 page = virt_to_page(objp);
2720 if (!PageSlab(page)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002721 printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n",
2722 (unsigned long)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002723 BUG();
2724 }
2725}
2726
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002727static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2728{
2729 unsigned long redzone1, redzone2;
2730
2731 redzone1 = *dbg_redzone1(cache, obj);
2732 redzone2 = *dbg_redzone2(cache, obj);
2733
2734 /*
2735 * Redzone is ok.
2736 */
2737 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2738 return;
2739
2740 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2741 slab_error(cache, "double free detected");
2742 else
2743 slab_error(cache, "memory outside object was overwritten");
2744
2745 printk(KERN_ERR "%p: redzone 1:0x%lx, redzone 2:0x%lx.\n",
2746 obj, redzone1, redzone2);
2747}
2748
Pekka Enberg343e0d72006-02-01 03:05:50 -08002749static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002750 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002751{
2752 struct page *page;
2753 unsigned int objnr;
2754 struct slab *slabp;
2755
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002756 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002757 kfree_debugcheck(objp);
2758 page = virt_to_page(objp);
2759
Pekka Enberg065d41c2005-11-13 16:06:46 -08002760 slabp = page_get_slab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002761
2762 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enberg58ce1fd2006-06-23 02:03:24 -07002763 verify_redzone_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002764 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2765 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2766 }
2767 if (cachep->flags & SLAB_STORE_USER)
2768 *dbg_userword(cachep, objp) = caller;
2769
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002770 objnr = obj_to_index(cachep, slabp, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002771
2772 BUG_ON(objnr >= cachep->num);
Pekka Enberg8fea4e92006-03-22 00:08:10 -08002773 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002774
2775 if (cachep->flags & SLAB_DEBUG_INITIAL) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002776 /*
2777 * Need to call the slab's constructor so the caller can
2778 * perform a verify of its state (debugging). Called without
2779 * the cache-lock held.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002780 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002781 cachep->ctor(objp + obj_offset(cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002782 cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002783 }
2784 if (cachep->flags & SLAB_POISON && cachep->dtor) {
2785 /* we want to cache poison the object,
2786 * call the destruction callback
2787 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002788 cachep->dtor(objp + obj_offset(cachep), cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002789 }
Al Viro871751e2006-03-25 03:06:39 -08002790#ifdef CONFIG_DEBUG_SLAB_LEAK
2791 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2792#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002793 if (cachep->flags & SLAB_POISON) {
2794#ifdef CONFIG_DEBUG_PAGEALLOC
Andrew Mortona737b3e2006-03-22 00:08:11 -08002795 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002796 store_stackinfo(cachep, objp, (unsigned long)caller);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002797 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002798 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002799 } else {
2800 poison_obj(cachep, objp, POISON_FREE);
2801 }
2802#else
2803 poison_obj(cachep, objp, POISON_FREE);
2804#endif
2805 }
2806 return objp;
2807}
2808
Pekka Enberg343e0d72006-02-01 03:05:50 -08002809static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002810{
2811 kmem_bufctl_t i;
2812 int entries = 0;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002813
Linus Torvalds1da177e2005-04-16 15:20:36 -07002814 /* Check slab's freelist to see if this obj is there. */
2815 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2816 entries++;
2817 if (entries > cachep->num || i >= cachep->num)
2818 goto bad;
2819 }
2820 if (entries != cachep->num - slabp->inuse) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002821bad:
2822 printk(KERN_ERR "slab: Internal list corruption detected in "
2823 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2824 cachep->name, cachep->num, slabp, slabp->inuse);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002825 for (i = 0;
Linus Torvalds264132b2006-03-06 12:10:07 -08002826 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002827 i++) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002828 if (i % 16 == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002829 printk("\n%03x:", i);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002830 printk(" %02x", ((unsigned char *)slabp)[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002831 }
2832 printk("\n");
2833 BUG();
2834 }
2835}
2836#else
2837#define kfree_debugcheck(x) do { } while(0)
2838#define cache_free_debugcheck(x,objp,z) (objp)
2839#define check_slabp(x,y) do { } while(0)
2840#endif
2841
Pekka Enberg343e0d72006-02-01 03:05:50 -08002842static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002843{
2844 int batchcount;
2845 struct kmem_list3 *l3;
2846 struct array_cache *ac;
2847
2848 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002849 ac = cpu_cache_get(cachep);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002850retry:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002851 batchcount = ac->batchcount;
2852 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002853 /*
2854 * If there was little recent activity on this cache, then
2855 * perform only a partial refill. Otherwise we could generate
2856 * refill bouncing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002857 */
2858 batchcount = BATCHREFILL_LIMIT;
2859 }
Christoph Lametere498be72005-09-09 13:03:32 -07002860 l3 = cachep->nodelists[numa_node_id()];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002861
Christoph Lametere498be72005-09-09 13:03:32 -07002862 BUG_ON(ac->avail > 0 || !l3);
2863 spin_lock(&l3->list_lock);
2864
Christoph Lameter3ded1752006-03-25 03:06:44 -08002865 /* See if we can refill from the shared array */
2866 if (l3->shared && transfer_objects(ac, l3->shared, batchcount))
2867 goto alloc_done;
2868
Linus Torvalds1da177e2005-04-16 15:20:36 -07002869 while (batchcount > 0) {
2870 struct list_head *entry;
2871 struct slab *slabp;
2872 /* Get slab alloc is to come from. */
2873 entry = l3->slabs_partial.next;
2874 if (entry == &l3->slabs_partial) {
2875 l3->free_touched = 1;
2876 entry = l3->slabs_free.next;
2877 if (entry == &l3->slabs_free)
2878 goto must_grow;
2879 }
2880
2881 slabp = list_entry(entry, struct slab, list);
2882 check_slabp(cachep, slabp);
2883 check_spinlock_acquired(cachep);
2884 while (slabp->inuse < cachep->num && batchcount--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002885 STATS_INC_ALLOCED(cachep);
2886 STATS_INC_ACTIVE(cachep);
2887 STATS_SET_HIGH(cachep);
2888
Matthew Dobson78d382d2006-02-01 03:05:47 -08002889 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
2890 numa_node_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002891 }
2892 check_slabp(cachep, slabp);
2893
2894 /* move slabp to correct slabp list: */
2895 list_del(&slabp->list);
2896 if (slabp->free == BUFCTL_END)
2897 list_add(&slabp->list, &l3->slabs_full);
2898 else
2899 list_add(&slabp->list, &l3->slabs_partial);
2900 }
2901
Andrew Mortona737b3e2006-03-22 00:08:11 -08002902must_grow:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002903 l3->free_objects -= ac->avail;
Andrew Mortona737b3e2006-03-22 00:08:11 -08002904alloc_done:
Christoph Lametere498be72005-09-09 13:03:32 -07002905 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002906
2907 if (unlikely(!ac->avail)) {
2908 int x;
Christoph Lametere498be72005-09-09 13:03:32 -07002909 x = cache_grow(cachep, flags, numa_node_id());
2910
Andrew Mortona737b3e2006-03-22 00:08:11 -08002911 /* cache_grow can reenable interrupts, then ac could change. */
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002912 ac = cpu_cache_get(cachep);
Andrew Mortona737b3e2006-03-22 00:08:11 -08002913 if (!x && ac->avail == 0) /* no objects in sight? abort */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002914 return NULL;
2915
Andrew Mortona737b3e2006-03-22 00:08:11 -08002916 if (!ac->avail) /* objects refilled by interrupt? */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002917 goto retry;
2918 }
2919 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07002920 return ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002921}
2922
Andrew Mortona737b3e2006-03-22 00:08:11 -08002923static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
2924 gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002925{
2926 might_sleep_if(flags & __GFP_WAIT);
2927#if DEBUG
2928 kmem_flagcheck(cachep, flags);
2929#endif
2930}
2931
2932#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08002933static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
2934 gfp_t flags, void *objp, void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002935{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002936 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002937 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002938 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002939#ifdef CONFIG_DEBUG_PAGEALLOC
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002940 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002941 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002942 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002943 else
2944 check_poison_obj(cachep, objp);
2945#else
2946 check_poison_obj(cachep, objp);
2947#endif
2948 poison_obj(cachep, objp, POISON_INUSE);
2949 }
2950 if (cachep->flags & SLAB_STORE_USER)
2951 *dbg_userword(cachep, objp) = caller;
2952
2953 if (cachep->flags & SLAB_RED_ZONE) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08002954 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
2955 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2956 slab_error(cachep, "double free, or memory outside"
2957 " object was overwritten");
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002958 printk(KERN_ERR
Andrew Mortona737b3e2006-03-22 00:08:11 -08002959 "%p: redzone 1:0x%lx, redzone 2:0x%lx\n",
2960 objp, *dbg_redzone1(cachep, objp),
2961 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002962 }
2963 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2964 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2965 }
Al Viro871751e2006-03-25 03:06:39 -08002966#ifdef CONFIG_DEBUG_SLAB_LEAK
2967 {
2968 struct slab *slabp;
2969 unsigned objnr;
2970
2971 slabp = page_get_slab(virt_to_page(objp));
2972 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
2973 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
2974 }
2975#endif
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002976 objp += obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002977 if (cachep->ctor && cachep->flags & SLAB_POISON) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002978 unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002979
2980 if (!(flags & __GFP_WAIT))
2981 ctor_flags |= SLAB_CTOR_ATOMIC;
2982
2983 cachep->ctor(objp, cachep, ctor_flags);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002984 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002985 return objp;
2986}
2987#else
2988#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2989#endif
2990
Pekka Enberg343e0d72006-02-01 03:05:50 -08002991static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002992{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002993 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002994 struct array_cache *ac;
2995
Christoph Lameterdc85da12006-01-18 17:42:36 -08002996#ifdef CONFIG_NUMA
Paul Jacksonb2455392006-03-24 03:16:12 -08002997 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
Paul Jacksonc61afb12006-03-24 03:16:08 -08002998 objp = alternate_node_alloc(cachep, flags);
2999 if (objp != NULL)
3000 return objp;
Paul Jackson101a5002006-03-24 03:16:07 -08003001 }
Christoph Lameterdc85da12006-01-18 17:42:36 -08003002#endif
3003
Alok N Kataria5c382302005-09-27 21:45:46 -07003004 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003005 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003006 if (likely(ac->avail)) {
3007 STATS_INC_ALLOCHIT(cachep);
3008 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07003009 objp = ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003010 } else {
3011 STATS_INC_ALLOCMISS(cachep);
3012 objp = cache_alloc_refill(cachep, flags);
3013 }
Alok N Kataria5c382302005-09-27 21:45:46 -07003014 return objp;
3015}
3016
Andrew Mortona737b3e2006-03-22 00:08:11 -08003017static __always_inline void *__cache_alloc(struct kmem_cache *cachep,
3018 gfp_t flags, void *caller)
Alok N Kataria5c382302005-09-27 21:45:46 -07003019{
3020 unsigned long save_flags;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003021 void *objp;
Alok N Kataria5c382302005-09-27 21:45:46 -07003022
3023 cache_alloc_debugcheck_before(cachep, flags);
3024
3025 local_irq_save(save_flags);
3026 objp = ____cache_alloc(cachep, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003027 local_irq_restore(save_flags);
Eric Dumazet34342e82005-09-03 15:55:06 -07003028 objp = cache_alloc_debugcheck_after(cachep, flags, objp,
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003029 caller);
Eric Dumazet34342e82005-09-03 15:55:06 -07003030 prefetchw(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003031 return objp;
3032}
3033
Christoph Lametere498be72005-09-09 13:03:32 -07003034#ifdef CONFIG_NUMA
3035/*
Paul Jacksonb2455392006-03-24 03:16:12 -08003036 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
Paul Jacksonc61afb12006-03-24 03:16:08 -08003037 *
3038 * If we are in_interrupt, then process context, including cpusets and
3039 * mempolicy, may not apply and should not be used for allocation policy.
3040 */
3041static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3042{
3043 int nid_alloc, nid_here;
3044
3045 if (in_interrupt())
3046 return NULL;
3047 nid_alloc = nid_here = numa_node_id();
3048 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3049 nid_alloc = cpuset_mem_spread_node();
3050 else if (current->mempolicy)
3051 nid_alloc = slab_node(current->mempolicy);
3052 if (nid_alloc != nid_here)
3053 return __cache_alloc_node(cachep, flags, nid_alloc);
3054 return NULL;
3055}
3056
3057/*
Christoph Lametere498be72005-09-09 13:03:32 -07003058 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07003059 */
Andrew Mortona737b3e2006-03-22 00:08:11 -08003060static void *__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3061 int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07003062{
3063 struct list_head *entry;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003064 struct slab *slabp;
3065 struct kmem_list3 *l3;
3066 void *obj;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003067 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003068
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003069 l3 = cachep->nodelists[nodeid];
3070 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07003071
Andrew Mortona737b3e2006-03-22 00:08:11 -08003072retry:
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003073 check_irq_off();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003074 spin_lock(&l3->list_lock);
3075 entry = l3->slabs_partial.next;
3076 if (entry == &l3->slabs_partial) {
3077 l3->free_touched = 1;
3078 entry = l3->slabs_free.next;
3079 if (entry == &l3->slabs_free)
3080 goto must_grow;
3081 }
Christoph Lametere498be72005-09-09 13:03:32 -07003082
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003083 slabp = list_entry(entry, struct slab, list);
3084 check_spinlock_acquired_node(cachep, nodeid);
3085 check_slabp(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07003086
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003087 STATS_INC_NODEALLOCS(cachep);
3088 STATS_INC_ACTIVE(cachep);
3089 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003090
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003091 BUG_ON(slabp->inuse == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07003092
Matthew Dobson78d382d2006-02-01 03:05:47 -08003093 obj = slab_get_obj(cachep, slabp, nodeid);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003094 check_slabp(cachep, slabp);
3095 l3->free_objects--;
3096 /* move slabp to correct slabp list: */
3097 list_del(&slabp->list);
Christoph Lametere498be72005-09-09 13:03:32 -07003098
Andrew Mortona737b3e2006-03-22 00:08:11 -08003099 if (slabp->free == BUFCTL_END)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003100 list_add(&slabp->list, &l3->slabs_full);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003101 else
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003102 list_add(&slabp->list, &l3->slabs_partial);
Christoph Lametere498be72005-09-09 13:03:32 -07003103
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003104 spin_unlock(&l3->list_lock);
3105 goto done;
Christoph Lametere498be72005-09-09 13:03:32 -07003106
Andrew Mortona737b3e2006-03-22 00:08:11 -08003107must_grow:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003108 spin_unlock(&l3->list_lock);
3109 x = cache_grow(cachep, flags, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07003110
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003111 if (!x)
3112 return NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003113
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003114 goto retry;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003115done:
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003116 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07003117}
3118#endif
3119
3120/*
3121 * Caller needs to acquire correct kmem_list's list_lock
3122 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003123static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003124 int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003125{
3126 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07003127 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003128
3129 for (i = 0; i < nr_objects; i++) {
3130 void *objp = objpp[i];
3131 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003132
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003133 slabp = virt_to_slab(objp);
Christoph Lameterff694162005-09-22 21:44:02 -07003134 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07003135 list_del(&slabp->list);
Christoph Lameterff694162005-09-22 21:44:02 -07003136 check_spinlock_acquired_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003137 check_slabp(cachep, slabp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08003138 slab_put_obj(cachep, slabp, objp, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003139 STATS_DEC_ACTIVE(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003140 l3->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003141 check_slabp(cachep, slabp);
3142
3143 /* fixup slab chains */
3144 if (slabp->inuse == 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07003145 if (l3->free_objects > l3->free_limit) {
3146 l3->free_objects -= cachep->num;
Ravikiran G Thirumalaie5ac9c52006-09-25 23:31:34 -07003147 /* No need to drop any previously held
3148 * lock here, even if we have a off-slab slab
3149 * descriptor it is guaranteed to come from
3150 * a different cache, refer to comments before
3151 * alloc_slabmgmt.
3152 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003153 slab_destroy(cachep, slabp);
3154 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07003155 list_add(&slabp->list, &l3->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003156 }
3157 } else {
3158 /* Unconditionally move a slab to the end of the
3159 * partial list on free - maximum time for the
3160 * other objects to be freed, too.
3161 */
Christoph Lametere498be72005-09-09 13:03:32 -07003162 list_add_tail(&slabp->list, &l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003163 }
3164 }
3165}
3166
Pekka Enberg343e0d72006-02-01 03:05:50 -08003167static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003168{
3169 int batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07003170 struct kmem_list3 *l3;
Christoph Lameterff694162005-09-22 21:44:02 -07003171 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003172
3173 batchcount = ac->batchcount;
3174#if DEBUG
3175 BUG_ON(!batchcount || batchcount > ac->avail);
3176#endif
3177 check_irq_off();
Christoph Lameterff694162005-09-22 21:44:02 -07003178 l3 = cachep->nodelists[node];
Ingo Molnar873623d2006-07-13 14:44:38 +02003179 spin_lock(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003180 if (l3->shared) {
3181 struct array_cache *shared_array = l3->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003182 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003183 if (max) {
3184 if (batchcount > max)
3185 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07003186 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003187 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003188 shared_array->avail += batchcount;
3189 goto free_done;
3190 }
3191 }
3192
Christoph Lameterff694162005-09-22 21:44:02 -07003193 free_block(cachep, ac->entry, batchcount, node);
Andrew Mortona737b3e2006-03-22 00:08:11 -08003194free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003195#if STATS
3196 {
3197 int i = 0;
3198 struct list_head *p;
3199
Christoph Lametere498be72005-09-09 13:03:32 -07003200 p = l3->slabs_free.next;
3201 while (p != &(l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003202 struct slab *slabp;
3203
3204 slabp = list_entry(p, struct slab, list);
3205 BUG_ON(slabp->inuse);
3206
3207 i++;
3208 p = p->next;
3209 }
3210 STATS_SET_FREEABLE(cachep, i);
3211 }
3212#endif
Christoph Lametere498be72005-09-09 13:03:32 -07003213 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003214 ac->avail -= batchcount;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003215 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003216}
3217
3218/*
Andrew Mortona737b3e2006-03-22 00:08:11 -08003219 * Release an obj back to its cache. If the obj has a constructed state, it must
3220 * be in this state _before_ it is released. Called with disabled ints.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003221 */
Ingo Molnar873623d2006-07-13 14:44:38 +02003222static inline void __cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003223{
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003224 struct array_cache *ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003225
3226 check_irq_off();
3227 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3228
Ingo Molnar873623d2006-07-13 14:44:38 +02003229 if (cache_free_alien(cachep, objp))
Pekka Enberg729bd0b2006-06-23 02:03:05 -07003230 return;
Christoph Lametere498be72005-09-09 13:03:32 -07003231
Linus Torvalds1da177e2005-04-16 15:20:36 -07003232 if (likely(ac->avail < ac->limit)) {
3233 STATS_INC_FREEHIT(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003234 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003235 return;
3236 } else {
3237 STATS_INC_FREEMISS(cachep);
3238 cache_flusharray(cachep, ac);
Christoph Lametere498be72005-09-09 13:03:32 -07003239 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003240 }
3241}
3242
3243/**
3244 * kmem_cache_alloc - Allocate an object
3245 * @cachep: The cache to allocate from.
3246 * @flags: See kmalloc().
3247 *
3248 * Allocate an object from this cache. The flags are only relevant
3249 * if the cache has no available objects.
3250 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003251void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003252{
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003253 return __cache_alloc(cachep, flags, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003254}
3255EXPORT_SYMBOL(kmem_cache_alloc);
3256
3257/**
Rolf Eike Beerb8008b22006-07-30 03:04:04 -07003258 * kmem_cache_zalloc - Allocate an object. The memory is set to zero.
Pekka Enberga8c0f9a2006-03-25 03:06:42 -08003259 * @cache: The cache to allocate from.
3260 * @flags: See kmalloc().
3261 *
3262 * Allocate an object from this cache and set the allocated memory to zero.
3263 * The flags are only relevant if the cache has no available objects.
3264 */
3265void *kmem_cache_zalloc(struct kmem_cache *cache, gfp_t flags)
3266{
3267 void *ret = __cache_alloc(cache, flags, __builtin_return_address(0));
3268 if (ret)
3269 memset(ret, 0, obj_size(cache));
3270 return ret;
3271}
3272EXPORT_SYMBOL(kmem_cache_zalloc);
3273
3274/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003275 * kmem_ptr_validate - check if an untrusted pointer might
3276 * be a slab entry.
3277 * @cachep: the cache we're checking against
3278 * @ptr: pointer to validate
3279 *
3280 * This verifies that the untrusted pointer looks sane:
3281 * it is _not_ a guarantee that the pointer is actually
3282 * part of the slab cache in question, but it at least
3283 * validates that the pointer can be dereferenced and
3284 * looks half-way sane.
3285 *
3286 * Currently only used for dentry validation.
3287 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003288int fastcall kmem_ptr_validate(struct kmem_cache *cachep, void *ptr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003289{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003290 unsigned long addr = (unsigned long)ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003291 unsigned long min_addr = PAGE_OFFSET;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003292 unsigned long align_mask = BYTES_PER_WORD - 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003293 unsigned long size = cachep->buffer_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003294 struct page *page;
3295
3296 if (unlikely(addr < min_addr))
3297 goto out;
3298 if (unlikely(addr > (unsigned long)high_memory - size))
3299 goto out;
3300 if (unlikely(addr & align_mask))
3301 goto out;
3302 if (unlikely(!kern_addr_valid(addr)))
3303 goto out;
3304 if (unlikely(!kern_addr_valid(addr + size - 1)))
3305 goto out;
3306 page = virt_to_page(ptr);
3307 if (unlikely(!PageSlab(page)))
3308 goto out;
Pekka Enberg065d41c2005-11-13 16:06:46 -08003309 if (unlikely(page_get_cache(page) != cachep))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003310 goto out;
3311 return 1;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003312out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003313 return 0;
3314}
3315
3316#ifdef CONFIG_NUMA
3317/**
3318 * kmem_cache_alloc_node - Allocate an object on the specified node
3319 * @cachep: The cache to allocate from.
3320 * @flags: See kmalloc().
3321 * @nodeid: node number of the target node.
3322 *
3323 * Identical to kmem_cache_alloc, except that this function is slow
3324 * and can sleep. And it will allocate memory on the given node, which
3325 * can improve the performance for cpu bound structures.
Christoph Lametere498be72005-09-09 13:03:32 -07003326 * New and improved: it will now make sure that the object gets
3327 * put on the correct node list so that there is no false sharing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003328 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003329void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003330{
Christoph Lametere498be72005-09-09 13:03:32 -07003331 unsigned long save_flags;
3332 void *ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003333
Christoph Lametere498be72005-09-09 13:03:32 -07003334 cache_alloc_debugcheck_before(cachep, flags);
3335 local_irq_save(save_flags);
Christoph Lameter18f820f2006-02-01 03:05:43 -08003336
3337 if (nodeid == -1 || nodeid == numa_node_id() ||
Andrew Mortona737b3e2006-03-22 00:08:11 -08003338 !cachep->nodelists[nodeid])
Alok N Kataria5c382302005-09-27 21:45:46 -07003339 ptr = ____cache_alloc(cachep, flags);
3340 else
3341 ptr = __cache_alloc_node(cachep, flags, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07003342 local_irq_restore(save_flags);
Christoph Lameter18f820f2006-02-01 03:05:43 -08003343
3344 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr,
3345 __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003346
Christoph Lametere498be72005-09-09 13:03:32 -07003347 return ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003348}
3349EXPORT_SYMBOL(kmem_cache_alloc_node);
3350
Christoph Hellwigdbe5e692006-09-25 23:31:36 -07003351void *__kmalloc_node(size_t size, gfp_t flags, int node)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003352{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003353 struct kmem_cache *cachep;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003354
3355 cachep = kmem_find_general_cachep(size, flags);
3356 if (unlikely(cachep == NULL))
3357 return NULL;
3358 return kmem_cache_alloc_node(cachep, flags, node);
3359}
Christoph Hellwigdbe5e692006-09-25 23:31:36 -07003360EXPORT_SYMBOL(__kmalloc_node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003361#endif
3362
3363/**
Paul Drynoff800590f2006-06-23 02:03:48 -07003364 * __do_kmalloc - allocate memory
Linus Torvalds1da177e2005-04-16 15:20:36 -07003365 * @size: how many bytes of memory are required.
Paul Drynoff800590f2006-06-23 02:03:48 -07003366 * @flags: the type of memory to allocate (see kmalloc).
Randy Dunlap911851e2006-03-22 00:08:14 -08003367 * @caller: function caller for debug tracking of the caller
Linus Torvalds1da177e2005-04-16 15:20:36 -07003368 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003369static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3370 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003371{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003372 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003373
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003374 /* If you want to save a few bytes .text space: replace
3375 * __ with kmem_.
3376 * Then kmalloc uses the uninlined functions instead of the inline
3377 * functions.
3378 */
3379 cachep = __find_general_cachep(size, flags);
Andrew Mortondbdb9042005-09-23 13:24:10 -07003380 if (unlikely(cachep == NULL))
3381 return NULL;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003382 return __cache_alloc(cachep, flags, caller);
3383}
3384
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003385
3386void *__kmalloc(size_t size, gfp_t flags)
3387{
Al Viro871751e2006-03-25 03:06:39 -08003388#ifndef CONFIG_DEBUG_SLAB
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003389 return __do_kmalloc(size, flags, NULL);
Al Viro871751e2006-03-25 03:06:39 -08003390#else
3391 return __do_kmalloc(size, flags, __builtin_return_address(0));
3392#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003393}
3394EXPORT_SYMBOL(__kmalloc);
3395
Al Viro871751e2006-03-25 03:06:39 -08003396#ifdef CONFIG_DEBUG_SLAB
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003397void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller)
3398{
3399 return __do_kmalloc(size, flags, caller);
3400}
3401EXPORT_SYMBOL(__kmalloc_track_caller);
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003402#endif
3403
Linus Torvalds1da177e2005-04-16 15:20:36 -07003404#ifdef CONFIG_SMP
3405/**
Martin Peschke7ff6f082006-09-25 23:31:21 -07003406 * percpu_depopulate - depopulate per-cpu data for given cpu
3407 * @__pdata: per-cpu data to depopulate
3408 * @cpu: depopulate per-cpu data for this cpu
Linus Torvalds1da177e2005-04-16 15:20:36 -07003409 *
Martin Peschke7ff6f082006-09-25 23:31:21 -07003410 * Depopulating per-cpu data for a cpu going offline would be a typical
3411 * use case. You need to register a cpu hotplug handler for that purpose.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003412 */
Martin Peschke7ff6f082006-09-25 23:31:21 -07003413void percpu_depopulate(void *__pdata, int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003414{
Martin Peschke7ff6f082006-09-25 23:31:21 -07003415 struct percpu_data *pdata = __percpu_disguise(__pdata);
3416 if (pdata->ptrs[cpu]) {
3417 kfree(pdata->ptrs[cpu]);
3418 pdata->ptrs[cpu] = NULL;
3419 }
3420}
3421EXPORT_SYMBOL_GPL(percpu_depopulate);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003422
Martin Peschke7ff6f082006-09-25 23:31:21 -07003423/**
3424 * percpu_depopulate_mask - depopulate per-cpu data for some cpu's
3425 * @__pdata: per-cpu data to depopulate
3426 * @mask: depopulate per-cpu data for cpu's selected through mask bits
3427 */
3428void __percpu_depopulate_mask(void *__pdata, cpumask_t *mask)
3429{
3430 int cpu;
3431 for_each_cpu_mask(cpu, *mask)
3432 percpu_depopulate(__pdata, cpu);
3433}
3434EXPORT_SYMBOL_GPL(__percpu_depopulate_mask);
3435
3436/**
3437 * percpu_populate - populate per-cpu data for given cpu
3438 * @__pdata: per-cpu data to populate further
3439 * @size: size of per-cpu object
3440 * @gfp: may sleep or not etc.
3441 * @cpu: populate per-data for this cpu
3442 *
3443 * Populating per-cpu data for a cpu coming online would be a typical
3444 * use case. You need to register a cpu hotplug handler for that purpose.
3445 * Per-cpu object is populated with zeroed buffer.
3446 */
3447void *percpu_populate(void *__pdata, size_t size, gfp_t gfp, int cpu)
3448{
3449 struct percpu_data *pdata = __percpu_disguise(__pdata);
3450 int node = cpu_to_node(cpu);
3451
3452 BUG_ON(pdata->ptrs[cpu]);
3453 if (node_online(node)) {
3454 /* FIXME: kzalloc_node(size, gfp, node) */
3455 pdata->ptrs[cpu] = kmalloc_node(size, gfp, node);
3456 if (pdata->ptrs[cpu])
3457 memset(pdata->ptrs[cpu], 0, size);
3458 } else
3459 pdata->ptrs[cpu] = kzalloc(size, gfp);
3460 return pdata->ptrs[cpu];
3461}
3462EXPORT_SYMBOL_GPL(percpu_populate);
3463
3464/**
3465 * percpu_populate_mask - populate per-cpu data for more cpu's
3466 * @__pdata: per-cpu data to populate further
3467 * @size: size of per-cpu object
3468 * @gfp: may sleep or not etc.
3469 * @mask: populate per-cpu data for cpu's selected through mask bits
3470 *
3471 * Per-cpu objects are populated with zeroed buffers.
3472 */
3473int __percpu_populate_mask(void *__pdata, size_t size, gfp_t gfp,
3474 cpumask_t *mask)
3475{
3476 cpumask_t populated = CPU_MASK_NONE;
3477 int cpu;
3478
3479 for_each_cpu_mask(cpu, *mask)
3480 if (unlikely(!percpu_populate(__pdata, size, gfp, cpu))) {
3481 __percpu_depopulate_mask(__pdata, &populated);
3482 return -ENOMEM;
3483 } else
3484 cpu_set(cpu, populated);
3485 return 0;
3486}
3487EXPORT_SYMBOL_GPL(__percpu_populate_mask);
3488
3489/**
3490 * percpu_alloc_mask - initial setup of per-cpu data
3491 * @size: size of per-cpu object
3492 * @gfp: may sleep or not etc.
3493 * @mask: populate per-data for cpu's selected through mask bits
3494 *
3495 * Populating per-cpu data for all online cpu's would be a typical use case,
3496 * which is simplified by the percpu_alloc() wrapper.
3497 * Per-cpu objects are populated with zeroed buffers.
3498 */
3499void *__percpu_alloc_mask(size_t size, gfp_t gfp, cpumask_t *mask)
3500{
3501 void *pdata = kzalloc(sizeof(struct percpu_data), gfp);
3502 void *__pdata = __percpu_disguise(pdata);
3503
3504 if (unlikely(!pdata))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003505 return NULL;
Martin Peschke7ff6f082006-09-25 23:31:21 -07003506 if (likely(!__percpu_populate_mask(__pdata, size, gfp, mask)))
3507 return __pdata;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003508 kfree(pdata);
3509 return NULL;
3510}
Martin Peschke7ff6f082006-09-25 23:31:21 -07003511EXPORT_SYMBOL_GPL(__percpu_alloc_mask);
3512
3513/**
3514 * percpu_free - final cleanup of per-cpu data
3515 * @__pdata: object to clean up
3516 *
3517 * We simply clean up any per-cpu object left. No need for the client to
3518 * track and specify through a bis mask which per-cpu objects are to free.
3519 */
3520void percpu_free(void *__pdata)
3521{
3522 __percpu_depopulate_mask(__pdata, &cpu_possible_map);
3523 kfree(__percpu_disguise(__pdata));
3524}
3525EXPORT_SYMBOL_GPL(percpu_free);
3526#endif /* CONFIG_SMP */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003527
3528/**
3529 * kmem_cache_free - Deallocate an object
3530 * @cachep: The cache the allocation was from.
3531 * @objp: The previously allocated object.
3532 *
3533 * Free an object which was previously allocated from this
3534 * cache.
3535 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003536void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003537{
3538 unsigned long flags;
3539
Pekka Enbergddc2e812006-06-23 02:03:40 -07003540 BUG_ON(virt_to_cache(objp) != cachep);
3541
Linus Torvalds1da177e2005-04-16 15:20:36 -07003542 local_irq_save(flags);
Ingo Molnar873623d2006-07-13 14:44:38 +02003543 __cache_free(cachep, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003544 local_irq_restore(flags);
3545}
3546EXPORT_SYMBOL(kmem_cache_free);
3547
3548/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003549 * kfree - free previously allocated memory
3550 * @objp: pointer returned by kmalloc.
3551 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003552 * If @objp is NULL, no operation is performed.
3553 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003554 * Don't free memory not originally allocated by kmalloc()
3555 * or you will run into trouble.
3556 */
3557void kfree(const void *objp)
3558{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003559 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003560 unsigned long flags;
3561
3562 if (unlikely(!objp))
3563 return;
3564 local_irq_save(flags);
3565 kfree_debugcheck(objp);
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08003566 c = virt_to_cache(objp);
Ingo Molnarf9b84042006-06-27 02:54:49 -07003567 debug_check_no_locks_freed(objp, obj_size(c));
Ingo Molnar873623d2006-07-13 14:44:38 +02003568 __cache_free(c, (void *)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003569 local_irq_restore(flags);
3570}
3571EXPORT_SYMBOL(kfree);
3572
Pekka Enberg343e0d72006-02-01 03:05:50 -08003573unsigned int kmem_cache_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003574{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003575 return obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003576}
3577EXPORT_SYMBOL(kmem_cache_size);
3578
Pekka Enberg343e0d72006-02-01 03:05:50 -08003579const char *kmem_cache_name(struct kmem_cache *cachep)
Arnaldo Carvalho de Melo19449722005-06-18 22:46:19 -07003580{
3581 return cachep->name;
3582}
3583EXPORT_SYMBOL_GPL(kmem_cache_name);
3584
Christoph Lametere498be72005-09-09 13:03:32 -07003585/*
Christoph Lameter0718dc22006-03-25 03:06:47 -08003586 * This initializes kmem_list3 or resizes varioius caches for all nodes.
Christoph Lametere498be72005-09-09 13:03:32 -07003587 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003588static int alloc_kmemlist(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07003589{
3590 int node;
3591 struct kmem_list3 *l3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003592 struct array_cache *new_shared;
3593 struct array_cache **new_alien;
Christoph Lametere498be72005-09-09 13:03:32 -07003594
3595 for_each_online_node(node) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003596
Andrew Mortona737b3e2006-03-22 00:08:11 -08003597 new_alien = alloc_alien_cache(node, cachep->limit);
3598 if (!new_alien)
Christoph Lametere498be72005-09-09 13:03:32 -07003599 goto fail;
Christoph Lametercafeb022006-03-25 03:06:46 -08003600
Christoph Lameter0718dc22006-03-25 03:06:47 -08003601 new_shared = alloc_arraycache(node,
3602 cachep->shared*cachep->batchcount,
Andrew Mortona737b3e2006-03-22 00:08:11 -08003603 0xbaadf00d);
Christoph Lameter0718dc22006-03-25 03:06:47 -08003604 if (!new_shared) {
3605 free_alien_cache(new_alien);
Christoph Lametere498be72005-09-09 13:03:32 -07003606 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003607 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003608
Andrew Mortona737b3e2006-03-22 00:08:11 -08003609 l3 = cachep->nodelists[node];
3610 if (l3) {
Christoph Lametercafeb022006-03-25 03:06:46 -08003611 struct array_cache *shared = l3->shared;
3612
Christoph Lametere498be72005-09-09 13:03:32 -07003613 spin_lock_irq(&l3->list_lock);
3614
Christoph Lametercafeb022006-03-25 03:06:46 -08003615 if (shared)
Christoph Lameter0718dc22006-03-25 03:06:47 -08003616 free_block(cachep, shared->entry,
3617 shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07003618
Christoph Lametercafeb022006-03-25 03:06:46 -08003619 l3->shared = new_shared;
3620 if (!l3->alien) {
Christoph Lametere498be72005-09-09 13:03:32 -07003621 l3->alien = new_alien;
3622 new_alien = NULL;
3623 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003624 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003625 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003626 spin_unlock_irq(&l3->list_lock);
Christoph Lametercafeb022006-03-25 03:06:46 -08003627 kfree(shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003628 free_alien_cache(new_alien);
3629 continue;
3630 }
Andrew Mortona737b3e2006-03-22 00:08:11 -08003631 l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node);
Christoph Lameter0718dc22006-03-25 03:06:47 -08003632 if (!l3) {
3633 free_alien_cache(new_alien);
3634 kfree(new_shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003635 goto fail;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003636 }
Christoph Lametere498be72005-09-09 13:03:32 -07003637
3638 kmem_list3_init(l3);
3639 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Andrew Mortona737b3e2006-03-22 00:08:11 -08003640 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametercafeb022006-03-25 03:06:46 -08003641 l3->shared = new_shared;
Christoph Lametere498be72005-09-09 13:03:32 -07003642 l3->alien = new_alien;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003643 l3->free_limit = (1 + nr_cpus_node(node)) *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003644 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003645 cachep->nodelists[node] = l3;
3646 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003647 return 0;
Christoph Lameter0718dc22006-03-25 03:06:47 -08003648
Andrew Mortona737b3e2006-03-22 00:08:11 -08003649fail:
Christoph Lameter0718dc22006-03-25 03:06:47 -08003650 if (!cachep->next.next) {
3651 /* Cache is not active yet. Roll back what we did */
3652 node--;
3653 while (node >= 0) {
3654 if (cachep->nodelists[node]) {
3655 l3 = cachep->nodelists[node];
3656
3657 kfree(l3->shared);
3658 free_alien_cache(l3->alien);
3659 kfree(l3);
3660 cachep->nodelists[node] = NULL;
3661 }
3662 node--;
3663 }
3664 }
Christoph Lametercafeb022006-03-25 03:06:46 -08003665 return -ENOMEM;
Christoph Lametere498be72005-09-09 13:03:32 -07003666}
3667
Linus Torvalds1da177e2005-04-16 15:20:36 -07003668struct ccupdate_struct {
Pekka Enberg343e0d72006-02-01 03:05:50 -08003669 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003670 struct array_cache *new[NR_CPUS];
3671};
3672
3673static void do_ccupdate_local(void *info)
3674{
Andrew Mortona737b3e2006-03-22 00:08:11 -08003675 struct ccupdate_struct *new = info;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003676 struct array_cache *old;
3677
3678 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003679 old = cpu_cache_get(new->cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003680
Linus Torvalds1da177e2005-04-16 15:20:36 -07003681 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3682 new->new[smp_processor_id()] = old;
3683}
3684
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003685/* Always called with the cache_chain_mutex held */
Andrew Mortona737b3e2006-03-22 00:08:11 -08003686static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3687 int batchcount, int shared)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003688{
3689 struct ccupdate_struct new;
Christoph Lametere498be72005-09-09 13:03:32 -07003690 int i, err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003691
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003692 memset(&new.new, 0, sizeof(new.new));
Christoph Lametere498be72005-09-09 13:03:32 -07003693 for_each_online_cpu(i) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08003694 new.new[i] = alloc_arraycache(cpu_to_node(i), limit,
3695 batchcount);
Christoph Lametere498be72005-09-09 13:03:32 -07003696 if (!new.new[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003697 for (i--; i >= 0; i--)
3698 kfree(new.new[i]);
Christoph Lametere498be72005-09-09 13:03:32 -07003699 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003700 }
3701 }
3702 new.cachep = cachep;
3703
Andrew Mortona07fa392006-03-22 00:08:17 -08003704 on_each_cpu(do_ccupdate_local, (void *)&new, 1, 1);
Christoph Lametere498be72005-09-09 13:03:32 -07003705
Linus Torvalds1da177e2005-04-16 15:20:36 -07003706 check_irq_on();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003707 cachep->batchcount = batchcount;
3708 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07003709 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003710
Christoph Lametere498be72005-09-09 13:03:32 -07003711 for_each_online_cpu(i) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003712 struct array_cache *ccold = new.new[i];
3713 if (!ccold)
3714 continue;
Christoph Lametere498be72005-09-09 13:03:32 -07003715 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -07003716 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
Christoph Lametere498be72005-09-09 13:03:32 -07003717 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003718 kfree(ccold);
3719 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003720
Christoph Lametere498be72005-09-09 13:03:32 -07003721 err = alloc_kmemlist(cachep);
3722 if (err) {
3723 printk(KERN_ERR "alloc_kmemlist failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003724 cachep->name, -err);
Christoph Lametere498be72005-09-09 13:03:32 -07003725 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003726 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003727 return 0;
3728}
3729
Ravikiran G Thirumalaib5d8ca72006-03-22 00:08:12 -08003730/* Called with cache_chain_mutex held always */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003731static void enable_cpucache(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003732{
3733 int err;
3734 int limit, shared;
3735
Andrew Mortona737b3e2006-03-22 00:08:11 -08003736 /*
3737 * The head array serves three purposes:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003738 * - create a LIFO ordering, i.e. return objects that are cache-warm
3739 * - reduce the number of spinlock operations.
Andrew Mortona737b3e2006-03-22 00:08:11 -08003740 * - reduce the number of linked list operations on the slab and
Linus Torvalds1da177e2005-04-16 15:20:36 -07003741 * bufctl chains: array operations are cheaper.
3742 * The numbers are guessed, we should auto-tune as described by
3743 * Bonwick.
3744 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003745 if (cachep->buffer_size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003746 limit = 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003747 else if (cachep->buffer_size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003748 limit = 8;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003749 else if (cachep->buffer_size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003750 limit = 24;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003751 else if (cachep->buffer_size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003752 limit = 54;
3753 else
3754 limit = 120;
3755
Andrew Mortona737b3e2006-03-22 00:08:11 -08003756 /*
3757 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
Linus Torvalds1da177e2005-04-16 15:20:36 -07003758 * allocation behaviour: Most allocs on one cpu, most free operations
3759 * on another cpu. For these cases, an efficient object passing between
3760 * cpus is necessary. This is provided by a shared array. The array
3761 * replaces Bonwick's magazine layer.
3762 * On uniprocessor, it's functionally equivalent (but less efficient)
3763 * to a larger limit. Thus disabled by default.
3764 */
3765 shared = 0;
3766#ifdef CONFIG_SMP
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003767 if (cachep->buffer_size <= PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003768 shared = 8;
3769#endif
3770
3771#if DEBUG
Andrew Mortona737b3e2006-03-22 00:08:11 -08003772 /*
3773 * With debugging enabled, large batchcount lead to excessively long
3774 * periods with disabled local interrupts. Limit the batchcount
Linus Torvalds1da177e2005-04-16 15:20:36 -07003775 */
3776 if (limit > 32)
3777 limit = 32;
3778#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003779 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003780 if (err)
3781 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003782 cachep->name, -err);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003783}
3784
Christoph Lameter1b552532006-03-22 00:09:07 -08003785/*
3786 * Drain an array if it contains any elements taking the l3 lock only if
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003787 * necessary. Note that the l3 listlock also protects the array_cache
3788 * if drain_array() is used on the shared array.
Christoph Lameter1b552532006-03-22 00:09:07 -08003789 */
3790void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
3791 struct array_cache *ac, int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003792{
3793 int tofree;
3794
Christoph Lameter1b552532006-03-22 00:09:07 -08003795 if (!ac || !ac->avail)
3796 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003797 if (ac->touched && !force) {
3798 ac->touched = 0;
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003799 } else {
Christoph Lameter1b552532006-03-22 00:09:07 -08003800 spin_lock_irq(&l3->list_lock);
Christoph Lameterb18e7e62006-03-22 00:09:07 -08003801 if (ac->avail) {
3802 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3803 if (tofree > ac->avail)
3804 tofree = (ac->avail + 1) / 2;
3805 free_block(cachep, ac->entry, tofree, node);
3806 ac->avail -= tofree;
3807 memmove(ac->entry, &(ac->entry[tofree]),
3808 sizeof(void *) * ac->avail);
3809 }
Christoph Lameter1b552532006-03-22 00:09:07 -08003810 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003811 }
3812}
3813
3814/**
3815 * cache_reap - Reclaim memory from caches.
Randy Dunlap1e5d5332005-11-07 01:01:06 -08003816 * @unused: unused parameter
Linus Torvalds1da177e2005-04-16 15:20:36 -07003817 *
3818 * Called from workqueue/eventd every few seconds.
3819 * Purpose:
3820 * - clear the per-cpu caches for this CPU.
3821 * - return freeable pages to the main free memory pool.
3822 *
Andrew Mortona737b3e2006-03-22 00:08:11 -08003823 * If we cannot acquire the cache chain mutex then just give up - we'll try
3824 * again on the next iteration.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003825 */
3826static void cache_reap(void *unused)
3827{
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07003828 struct kmem_cache *searchp;
Christoph Lametere498be72005-09-09 13:03:32 -07003829 struct kmem_list3 *l3;
Christoph Lameteraab22072006-03-22 00:09:06 -08003830 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003831
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003832 if (!mutex_trylock(&cache_chain_mutex)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003833 /* Give up. Setup the next iteration. */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003834 schedule_delayed_work(&__get_cpu_var(reap_work),
3835 REAPTIMEOUT_CPUC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003836 return;
3837 }
3838
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07003839 list_for_each_entry(searchp, &cache_chain, next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003840 check_irq_on();
3841
Christoph Lameter35386e32006-03-22 00:09:05 -08003842 /*
3843 * We only take the l3 lock if absolutely necessary and we
3844 * have established with reasonable certainty that
3845 * we can do some work if the lock was obtained.
3846 */
Christoph Lameteraab22072006-03-22 00:09:06 -08003847 l3 = searchp->nodelists[node];
Christoph Lameter35386e32006-03-22 00:09:05 -08003848
Christoph Lameter8fce4d82006-03-09 17:33:54 -08003849 reap_alien(searchp, l3);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003850
Christoph Lameteraab22072006-03-22 00:09:06 -08003851 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003852
Christoph Lameter35386e32006-03-22 00:09:05 -08003853 /*
3854 * These are racy checks but it does not matter
3855 * if we skip one check or scan twice.
3856 */
Christoph Lametere498be72005-09-09 13:03:32 -07003857 if (time_after(l3->next_reap, jiffies))
Christoph Lameter35386e32006-03-22 00:09:05 -08003858 goto next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003859
Christoph Lametere498be72005-09-09 13:03:32 -07003860 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003861
Christoph Lameteraab22072006-03-22 00:09:06 -08003862 drain_array(searchp, l3, l3->shared, 0, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003863
Christoph Lametered11d9e2006-06-30 01:55:45 -07003864 if (l3->free_touched)
Christoph Lametere498be72005-09-09 13:03:32 -07003865 l3->free_touched = 0;
Christoph Lametered11d9e2006-06-30 01:55:45 -07003866 else {
3867 int freed;
3868
3869 freed = drain_freelist(searchp, l3, (l3->free_limit +
3870 5 * searchp->num - 1) / (5 * searchp->num));
3871 STATS_ADD_REAPED(searchp, freed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003872 }
Christoph Lameter35386e32006-03-22 00:09:05 -08003873next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003874 cond_resched();
3875 }
3876 check_irq_on();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003877 mutex_unlock(&cache_chain_mutex);
Christoph Lameter8fce4d82006-03-09 17:33:54 -08003878 next_reap_node();
Christoph Lameter2244b952006-06-30 01:55:33 -07003879 refresh_cpu_vm_stats(smp_processor_id());
Andrew Mortona737b3e2006-03-22 00:08:11 -08003880 /* Set up the next iteration */
Manfred Spraulcd61ef62005-11-07 00:58:02 -08003881 schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003882}
3883
3884#ifdef CONFIG_PROC_FS
3885
Pekka Enberg85289f92006-01-08 01:00:36 -08003886static void print_slabinfo_header(struct seq_file *m)
3887{
3888 /*
3889 * Output format version, so at least we can change it
3890 * without _too_ many complaints.
3891 */
3892#if STATS
3893 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
3894#else
3895 seq_puts(m, "slabinfo - version: 2.1\n");
3896#endif
3897 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
3898 "<objperslab> <pagesperslab>");
3899 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
3900 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
3901#if STATS
3902 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07003903 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
Pekka Enberg85289f92006-01-08 01:00:36 -08003904 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
3905#endif
3906 seq_putc(m, '\n');
3907}
3908
Linus Torvalds1da177e2005-04-16 15:20:36 -07003909static void *s_start(struct seq_file *m, loff_t *pos)
3910{
3911 loff_t n = *pos;
3912 struct list_head *p;
3913
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003914 mutex_lock(&cache_chain_mutex);
Pekka Enberg85289f92006-01-08 01:00:36 -08003915 if (!n)
3916 print_slabinfo_header(m);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003917 p = cache_chain.next;
3918 while (n--) {
3919 p = p->next;
3920 if (p == &cache_chain)
3921 return NULL;
3922 }
Pekka Enberg343e0d72006-02-01 03:05:50 -08003923 return list_entry(p, struct kmem_cache, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003924}
3925
3926static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3927{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003928 struct kmem_cache *cachep = p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003929 ++*pos;
Andrew Mortona737b3e2006-03-22 00:08:11 -08003930 return cachep->next.next == &cache_chain ?
3931 NULL : list_entry(cachep->next.next, struct kmem_cache, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003932}
3933
3934static void s_stop(struct seq_file *m, void *p)
3935{
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003936 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003937}
3938
3939static int s_show(struct seq_file *m, void *p)
3940{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003941 struct kmem_cache *cachep = p;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003942 struct slab *slabp;
3943 unsigned long active_objs;
3944 unsigned long num_objs;
3945 unsigned long active_slabs = 0;
3946 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07003947 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003948 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003949 int node;
3950 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003951
Linus Torvalds1da177e2005-04-16 15:20:36 -07003952 active_objs = 0;
3953 num_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07003954 for_each_online_node(node) {
3955 l3 = cachep->nodelists[node];
3956 if (!l3)
3957 continue;
3958
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003959 check_irq_on();
3960 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003961
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07003962 list_for_each_entry(slabp, &l3->slabs_full, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07003963 if (slabp->inuse != cachep->num && !error)
3964 error = "slabs_full accounting error";
3965 active_objs += cachep->num;
3966 active_slabs++;
3967 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07003968 list_for_each_entry(slabp, &l3->slabs_partial, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07003969 if (slabp->inuse == cachep->num && !error)
3970 error = "slabs_partial inuse accounting error";
3971 if (!slabp->inuse && !error)
3972 error = "slabs_partial/inuse accounting error";
3973 active_objs += slabp->inuse;
3974 active_slabs++;
3975 }
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07003976 list_for_each_entry(slabp, &l3->slabs_free, list) {
Christoph Lametere498be72005-09-09 13:03:32 -07003977 if (slabp->inuse && !error)
3978 error = "slabs_free/inuse accounting error";
3979 num_slabs++;
3980 }
3981 free_objects += l3->free_objects;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08003982 if (l3->shared)
3983 shared_avail += l3->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07003984
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003985 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003986 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003987 num_slabs += active_slabs;
3988 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003989 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003990 error = "free_objects accounting error";
3991
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003992 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003993 if (error)
3994 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
3995
3996 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003997 name, active_objs, num_objs, cachep->buffer_size,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003998 cachep->num, (1 << cachep->gfporder));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003999 seq_printf(m, " : tunables %4u %4u %4u",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004000 cachep->limit, cachep->batchcount, cachep->shared);
Christoph Lametere498be72005-09-09 13:03:32 -07004001 seq_printf(m, " : slabdata %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004002 active_slabs, num_slabs, shared_avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004003#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004004 { /* list3 stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004005 unsigned long high = cachep->high_mark;
4006 unsigned long allocs = cachep->num_allocations;
4007 unsigned long grown = cachep->grown;
4008 unsigned long reaped = cachep->reaped;
4009 unsigned long errors = cachep->errors;
4010 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004011 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07004012 unsigned long node_frees = cachep->node_frees;
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004013 unsigned long overflows = cachep->node_overflow;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004014
Christoph Lametere498be72005-09-09 13:03:32 -07004015 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004016 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown,
Andrew Mortona737b3e2006-03-22 00:08:11 -08004017 reaped, errors, max_freeable, node_allocs,
Ravikiran G Thirumalaifb7faf32006-04-10 22:52:54 -07004018 node_frees, overflows);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004019 }
4020 /* cpu stats */
4021 {
4022 unsigned long allochit = atomic_read(&cachep->allochit);
4023 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4024 unsigned long freehit = atomic_read(&cachep->freehit);
4025 unsigned long freemiss = atomic_read(&cachep->freemiss);
4026
4027 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004028 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004029 }
4030#endif
4031 seq_putc(m, '\n');
Linus Torvalds1da177e2005-04-16 15:20:36 -07004032 return 0;
4033}
4034
4035/*
4036 * slabinfo_op - iterator that generates /proc/slabinfo
4037 *
4038 * Output layout:
4039 * cache-name
4040 * num-active-objs
4041 * total-objs
4042 * object size
4043 * num-active-slabs
4044 * total-slabs
4045 * num-pages-per-slab
4046 * + further values on SMP and with statistics enabled
4047 */
4048
4049struct seq_operations slabinfo_op = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004050 .start = s_start,
4051 .next = s_next,
4052 .stop = s_stop,
4053 .show = s_show,
Linus Torvalds1da177e2005-04-16 15:20:36 -07004054};
4055
4056#define MAX_SLABINFO_WRITE 128
4057/**
4058 * slabinfo_write - Tuning for the slab allocator
4059 * @file: unused
4060 * @buffer: user buffer
4061 * @count: data length
4062 * @ppos: unused
4063 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004064ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4065 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004066{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004067 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004068 int limit, batchcount, shared, res;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004069 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004070
Linus Torvalds1da177e2005-04-16 15:20:36 -07004071 if (count > MAX_SLABINFO_WRITE)
4072 return -EINVAL;
4073 if (copy_from_user(&kbuf, buffer, count))
4074 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004075 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07004076
4077 tmp = strchr(kbuf, ' ');
4078 if (!tmp)
4079 return -EINVAL;
4080 *tmp = '\0';
4081 tmp++;
4082 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4083 return -EINVAL;
4084
4085 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004086 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004087 res = -EINVAL;
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004088 list_for_each_entry(cachep, &cache_chain, next) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004089 if (!strcmp(cachep->name, kbuf)) {
Andrew Mortona737b3e2006-03-22 00:08:11 -08004090 if (limit < 1 || batchcount < 1 ||
4091 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07004092 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004093 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07004094 res = do_tune_cpucache(cachep, limit,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08004095 batchcount, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004096 }
4097 break;
4098 }
4099 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08004100 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004101 if (res >= 0)
4102 res = count;
4103 return res;
4104}
Al Viro871751e2006-03-25 03:06:39 -08004105
4106#ifdef CONFIG_DEBUG_SLAB_LEAK
4107
4108static void *leaks_start(struct seq_file *m, loff_t *pos)
4109{
4110 loff_t n = *pos;
4111 struct list_head *p;
4112
4113 mutex_lock(&cache_chain_mutex);
4114 p = cache_chain.next;
4115 while (n--) {
4116 p = p->next;
4117 if (p == &cache_chain)
4118 return NULL;
4119 }
4120 return list_entry(p, struct kmem_cache, next);
4121}
4122
4123static inline int add_caller(unsigned long *n, unsigned long v)
4124{
4125 unsigned long *p;
4126 int l;
4127 if (!v)
4128 return 1;
4129 l = n[1];
4130 p = n + 2;
4131 while (l) {
4132 int i = l/2;
4133 unsigned long *q = p + 2 * i;
4134 if (*q == v) {
4135 q[1]++;
4136 return 1;
4137 }
4138 if (*q > v) {
4139 l = i;
4140 } else {
4141 p = q + 2;
4142 l -= i + 1;
4143 }
4144 }
4145 if (++n[1] == n[0])
4146 return 0;
4147 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4148 p[0] = v;
4149 p[1] = 1;
4150 return 1;
4151}
4152
4153static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4154{
4155 void *p;
4156 int i;
4157 if (n[0] == n[1])
4158 return;
4159 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4160 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4161 continue;
4162 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4163 return;
4164 }
4165}
4166
4167static void show_symbol(struct seq_file *m, unsigned long address)
4168{
4169#ifdef CONFIG_KALLSYMS
4170 char *modname;
4171 const char *name;
4172 unsigned long offset, size;
4173 char namebuf[KSYM_NAME_LEN+1];
4174
4175 name = kallsyms_lookup(address, &size, &offset, &modname, namebuf);
4176
4177 if (name) {
4178 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4179 if (modname)
4180 seq_printf(m, " [%s]", modname);
4181 return;
4182 }
4183#endif
4184 seq_printf(m, "%p", (void *)address);
4185}
4186
4187static int leaks_show(struct seq_file *m, void *p)
4188{
4189 struct kmem_cache *cachep = p;
Al Viro871751e2006-03-25 03:06:39 -08004190 struct slab *slabp;
4191 struct kmem_list3 *l3;
4192 const char *name;
4193 unsigned long *n = m->private;
4194 int node;
4195 int i;
4196
4197 if (!(cachep->flags & SLAB_STORE_USER))
4198 return 0;
4199 if (!(cachep->flags & SLAB_RED_ZONE))
4200 return 0;
4201
4202 /* OK, we can do it */
4203
4204 n[1] = 0;
4205
4206 for_each_online_node(node) {
4207 l3 = cachep->nodelists[node];
4208 if (!l3)
4209 continue;
4210
4211 check_irq_on();
4212 spin_lock_irq(&l3->list_lock);
4213
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004214 list_for_each_entry(slabp, &l3->slabs_full, list)
Al Viro871751e2006-03-25 03:06:39 -08004215 handle_slab(n, cachep, slabp);
Christoph Hellwig7a7c3812006-06-23 02:03:17 -07004216 list_for_each_entry(slabp, &l3->slabs_partial, list)
Al Viro871751e2006-03-25 03:06:39 -08004217 handle_slab(n, cachep, slabp);
Al Viro871751e2006-03-25 03:06:39 -08004218 spin_unlock_irq(&l3->list_lock);
4219 }
4220 name = cachep->name;
4221 if (n[0] == n[1]) {
4222 /* Increase the buffer size */
4223 mutex_unlock(&cache_chain_mutex);
4224 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4225 if (!m->private) {
4226 /* Too bad, we are really out */
4227 m->private = n;
4228 mutex_lock(&cache_chain_mutex);
4229 return -ENOMEM;
4230 }
4231 *(unsigned long *)m->private = n[0] * 2;
4232 kfree(n);
4233 mutex_lock(&cache_chain_mutex);
4234 /* Now make sure this entry will be retried */
4235 m->count = m->size;
4236 return 0;
4237 }
4238 for (i = 0; i < n[1]; i++) {
4239 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4240 show_symbol(m, n[2*i+2]);
4241 seq_putc(m, '\n');
4242 }
4243 return 0;
4244}
4245
4246struct seq_operations slabstats_op = {
4247 .start = leaks_start,
4248 .next = s_next,
4249 .stop = s_stop,
4250 .show = leaks_show,
4251};
4252#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004253#endif
4254
Manfred Spraul00e145b2005-09-03 15:55:07 -07004255/**
4256 * ksize - get the actual amount of memory allocated for a given object
4257 * @objp: Pointer to the object
4258 *
4259 * kmalloc may internally round up allocations and return more memory
4260 * than requested. ksize() can be used to determine the actual amount of
4261 * memory allocated. The caller may use this additional memory, even though
4262 * a smaller amount of memory was initially specified with the kmalloc call.
4263 * The caller must guarantee that objp points to a valid object previously
4264 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4265 * must not be freed during the duration of the call.
4266 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004267unsigned int ksize(const void *objp)
4268{
Manfred Spraul00e145b2005-09-03 15:55:07 -07004269 if (unlikely(objp == NULL))
4270 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004271
Pekka Enberg6ed5eb2212006-02-01 03:05:49 -08004272 return obj_size(virt_to_cache(objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004273}