blob: add05d808a4a6fe96823cf3a5899d1c9a5a3f95b [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 *
53 * The c_cpuarray may not be read with enabled local interrupts -
54 * 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>
92#include <linux/swap.h>
93#include <linux/cache.h>
94#include <linux/interrupt.h>
95#include <linux/init.h>
96#include <linux/compiler.h>
97#include <linux/seq_file.h>
98#include <linux/notifier.h>
99#include <linux/kallsyms.h>
100#include <linux/cpu.h>
101#include <linux/sysctl.h>
102#include <linux/module.h>
103#include <linux/rcupdate.h>
Paulo Marques543537b2005-06-23 00:09:02 -0700104#include <linux/string.h>
Christoph Lametere498be72005-09-09 13:03:32 -0700105#include <linux/nodemask.h>
Christoph Lameterdc85da12006-01-18 17:42:36 -0800106#include <linux/mempolicy.h>
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800107#include <linux/mutex.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700108
109#include <asm/uaccess.h>
110#include <asm/cacheflush.h>
111#include <asm/tlbflush.h>
112#include <asm/page.h>
113
114/*
115 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL,
116 * SLAB_RED_ZONE & SLAB_POISON.
117 * 0 for faster, smaller code (especially in the critical paths).
118 *
119 * STATS - 1 to collect stats for /proc/slabinfo.
120 * 0 for faster, smaller code (especially in the critical paths).
121 *
122 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
123 */
124
125#ifdef CONFIG_DEBUG_SLAB
126#define DEBUG 1
127#define STATS 1
128#define FORCED_DEBUG 1
129#else
130#define DEBUG 0
131#define STATS 0
132#define FORCED_DEBUG 0
133#endif
134
Linus Torvalds1da177e2005-04-16 15:20:36 -0700135/* Shouldn't this be in a header file somewhere? */
136#define BYTES_PER_WORD sizeof(void *)
137
138#ifndef cache_line_size
139#define cache_line_size() L1_CACHE_BYTES
140#endif
141
142#ifndef ARCH_KMALLOC_MINALIGN
143/*
144 * Enforce a minimum alignment for the kmalloc caches.
145 * Usually, the kmalloc caches are cache_line_size() aligned, except when
146 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
147 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
148 * alignment larger than BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that.
149 * Note that this flag disables some debug features.
150 */
151#define ARCH_KMALLOC_MINALIGN 0
152#endif
153
154#ifndef ARCH_SLAB_MINALIGN
155/*
156 * Enforce a minimum alignment for all caches.
157 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
158 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
159 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
160 * some debug features.
161 */
162#define ARCH_SLAB_MINALIGN 0
163#endif
164
165#ifndef ARCH_KMALLOC_FLAGS
166#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
167#endif
168
169/* Legal flag mask for kmem_cache_create(). */
170#if DEBUG
171# define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \
172 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
173 SLAB_NO_REAP | SLAB_CACHE_DMA | \
174 SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \
175 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
176 SLAB_DESTROY_BY_RCU)
177#else
178# define CREATE_MASK (SLAB_HWCACHE_ALIGN | SLAB_NO_REAP | \
179 SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \
180 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
181 SLAB_DESTROY_BY_RCU)
182#endif
183
184/*
185 * kmem_bufctl_t:
186 *
187 * Bufctl's are used for linking objs within a slab
188 * linked offsets.
189 *
190 * This implementation relies on "struct page" for locating the cache &
191 * slab an object belongs to.
192 * This allows the bufctl structure to be small (one int), but limits
193 * the number of objects a slab (not a cache) can contain when off-slab
194 * bufctls are used. The limit is the size of the largest general cache
195 * that does not use off-slab slabs.
196 * For 32bit archs with 4 kB pages, is this 56.
197 * This is not serious, as it is only for large objects, when it is unwise
198 * to have too many per slab.
199 * Note: This limit can be raised by introducing a general cache whose size
200 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
201 */
202
Kyle Moffettfa5b08d2005-09-03 15:55:03 -0700203typedef unsigned int kmem_bufctl_t;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700204#define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
205#define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
206#define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-2)
207
208/* Max number of objs-per-slab for caches which use off-slab slabs.
209 * Needed to avoid a possible looping condition in cache_grow().
210 */
211static unsigned long offslab_limit;
212
213/*
214 * struct slab
215 *
216 * Manages the objs in a slab. Placed either at the beginning of mem allocated
217 * for a slab, or allocated from an general cache.
218 * Slabs are chained into three list: fully used, partial, fully free slabs.
219 */
220struct slab {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800221 struct list_head list;
222 unsigned long colouroff;
223 void *s_mem; /* including colour offset */
224 unsigned int inuse; /* num of objs active in slab */
225 kmem_bufctl_t free;
226 unsigned short nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700227};
228
229/*
230 * struct slab_rcu
231 *
232 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
233 * arrange for kmem_freepages to be called via RCU. This is useful if
234 * we need to approach a kernel structure obliquely, from its address
235 * obtained without the usual locking. We can lock the structure to
236 * stabilize it and check it's still at the given address, only if we
237 * can be sure that the memory has not been meanwhile reused for some
238 * other kind of object (which our subsystem's lock might corrupt).
239 *
240 * rcu_read_lock before reading the address, then rcu_read_unlock after
241 * taking the spinlock within the structure expected at that address.
242 *
243 * We assume struct slab_rcu can overlay struct slab when destroying.
244 */
245struct slab_rcu {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800246 struct rcu_head head;
Pekka Enberg343e0d72006-02-01 03:05:50 -0800247 struct kmem_cache *cachep;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800248 void *addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700249};
250
251/*
252 * struct array_cache
253 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700254 * Purpose:
255 * - LIFO ordering, to hand out cache-warm objects from _alloc
256 * - reduce the number of linked list operations
257 * - reduce spinlock operations
258 *
259 * The limit is stored in the per-cpu structure to reduce the data cache
260 * footprint.
261 *
262 */
263struct array_cache {
264 unsigned int avail;
265 unsigned int limit;
266 unsigned int batchcount;
267 unsigned int touched;
Christoph Lametere498be72005-09-09 13:03:32 -0700268 spinlock_t lock;
269 void *entry[0]; /*
270 * Must have this definition in here for the proper
271 * alignment of array_cache. Also simplifies accessing
272 * the entries.
273 * [0] is for gcc 2.95. It should really be [].
274 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700275};
276
277/* bootstrap: The caches do not work without cpuarrays anymore,
278 * but the cpuarrays are allocated from the generic caches...
279 */
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;
294 unsigned long next_reap;
295 int free_touched;
296 unsigned int free_limit;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800297 unsigned int colour_next; /* Per-node cache coloring */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800298 spinlock_t list_lock;
299 struct array_cache *shared; /* shared per node */
300 struct array_cache **alien; /* on other nodes */
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 Lametere498be72005-09-09 13:03:32 -0700312/*
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700313 * This function must be completely optimized away if
Christoph Lametere498be72005-09-09 13:03:32 -0700314 * a constant is passed to it. Mostly the same as
315 * what is in linux/slab.h except it returns an
316 * index.
317 */
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700318static __always_inline int index_of(const size_t size)
Christoph Lametere498be72005-09-09 13:03:32 -0700319{
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800320 extern void __bad_size(void);
321
Christoph Lametere498be72005-09-09 13:03:32 -0700322 if (__builtin_constant_p(size)) {
323 int i = 0;
324
325#define CACHE(x) \
326 if (size <=x) \
327 return i; \
328 else \
329 i++;
330#include "linux/kmalloc_sizes.h"
331#undef CACHE
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800332 __bad_size();
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700333 } else
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800334 __bad_size();
Christoph Lametere498be72005-09-09 13:03:32 -0700335 return 0;
336}
337
338#define INDEX_AC index_of(sizeof(struct arraycache_init))
339#define INDEX_L3 index_of(sizeof(struct kmem_list3))
340
Pekka Enberg5295a742006-02-01 03:05:48 -0800341static void kmem_list3_init(struct kmem_list3 *parent)
Christoph Lametere498be72005-09-09 13:03:32 -0700342{
343 INIT_LIST_HEAD(&parent->slabs_full);
344 INIT_LIST_HEAD(&parent->slabs_partial);
345 INIT_LIST_HEAD(&parent->slabs_free);
346 parent->shared = NULL;
347 parent->alien = NULL;
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -0800348 parent->colour_next = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700349 spin_lock_init(&parent->list_lock);
350 parent->free_objects = 0;
351 parent->free_touched = 0;
352}
353
354#define MAKE_LIST(cachep, listp, slab, nodeid) \
355 do { \
356 INIT_LIST_HEAD(listp); \
357 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
358 } while (0)
359
360#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
361 do { \
362 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
363 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
364 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
365 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700366
367/*
Pekka Enberg343e0d72006-02-01 03:05:50 -0800368 * struct kmem_cache
Linus Torvalds1da177e2005-04-16 15:20:36 -0700369 *
370 * manages a cache.
371 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800372
Pekka J Enberg2109a2d2005-11-07 00:58:01 -0800373struct kmem_cache {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700374/* 1) per-cpu data, touched during every alloc/free */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800375 struct array_cache *array[NR_CPUS];
376 unsigned int batchcount;
377 unsigned int limit;
378 unsigned int shared;
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800379 unsigned int buffer_size;
Christoph Lametere498be72005-09-09 13:03:32 -0700380/* 2) touched by every alloc & free from the backend */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800381 struct kmem_list3 *nodelists[MAX_NUMNODES];
382 unsigned int flags; /* constant flags */
383 unsigned int num; /* # of objs per slab */
384 spinlock_t spinlock;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700385
386/* 3) cache_grow/shrink */
387 /* order of pgs per slab (2^n) */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800388 unsigned int gfporder;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700389
390 /* force GFP flags, e.g. GFP_DMA */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800391 gfp_t gfpflags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700392
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800393 size_t colour; /* cache colouring range */
394 unsigned int colour_off; /* colour offset */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800395 struct kmem_cache *slabp_cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800396 unsigned int slab_size;
397 unsigned int dflags; /* dynamic flags */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700398
399 /* constructor func */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800400 void (*ctor) (void *, struct kmem_cache *, unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700401
402 /* de-constructor func */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800403 void (*dtor) (void *, struct kmem_cache *, unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700404
405/* 4) cache creation/removal */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800406 const char *name;
407 struct list_head next;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700408
409/* 5) statistics */
410#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800411 unsigned long num_active;
412 unsigned long num_allocations;
413 unsigned long high_mark;
414 unsigned long grown;
415 unsigned long reaped;
416 unsigned long errors;
417 unsigned long max_freeable;
418 unsigned long node_allocs;
419 unsigned long node_frees;
420 atomic_t allochit;
421 atomic_t allocmiss;
422 atomic_t freehit;
423 atomic_t freemiss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700424#endif
425#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800426 /*
427 * If debugging is enabled, then the allocator can add additional
428 * fields and/or padding to every object. buffer_size contains the total
429 * object size including these internal fields, the following two
430 * variables contain the offset to the user object and its size.
431 */
432 int obj_offset;
433 int obj_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700434#endif
435};
436
437#define CFLGS_OFF_SLAB (0x80000000UL)
438#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
439
440#define BATCHREFILL_LIMIT 16
441/* Optimization question: fewer reaps means less
442 * probability for unnessary cpucache drain/refill cycles.
443 *
Adrian Bunkdc6f3f22005-11-08 16:44:08 +0100444 * OTOH the cpuarrays can contain lots of objects,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700445 * which could lock up otherwise freeable slabs.
446 */
447#define REAPTIMEOUT_CPUC (2*HZ)
448#define REAPTIMEOUT_LIST3 (4*HZ)
449
450#if STATS
451#define STATS_INC_ACTIVE(x) ((x)->num_active++)
452#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
453#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
454#define STATS_INC_GROWN(x) ((x)->grown++)
455#define STATS_INC_REAPED(x) ((x)->reaped++)
456#define STATS_SET_HIGH(x) do { if ((x)->num_active > (x)->high_mark) \
457 (x)->high_mark = (x)->num_active; \
458 } while (0)
459#define STATS_INC_ERR(x) ((x)->errors++)
460#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
Christoph Lametere498be72005-09-09 13:03:32 -0700461#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700462#define STATS_SET_FREEABLE(x, i) \
463 do { if ((x)->max_freeable < i) \
464 (x)->max_freeable = i; \
465 } while (0)
466
467#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
468#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
469#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
470#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
471#else
472#define STATS_INC_ACTIVE(x) do { } while (0)
473#define STATS_DEC_ACTIVE(x) do { } while (0)
474#define STATS_INC_ALLOCED(x) do { } while (0)
475#define STATS_INC_GROWN(x) do { } while (0)
476#define STATS_INC_REAPED(x) do { } while (0)
477#define STATS_SET_HIGH(x) do { } while (0)
478#define STATS_INC_ERR(x) do { } while (0)
479#define STATS_INC_NODEALLOCS(x) do { } while (0)
Christoph Lametere498be72005-09-09 13:03:32 -0700480#define STATS_INC_NODEFREES(x) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700481#define STATS_SET_FREEABLE(x, i) \
482 do { } while (0)
483
484#define STATS_INC_ALLOCHIT(x) do { } while (0)
485#define STATS_INC_ALLOCMISS(x) do { } while (0)
486#define STATS_INC_FREEHIT(x) do { } while (0)
487#define STATS_INC_FREEMISS(x) do { } while (0)
488#endif
489
490#if DEBUG
491/* Magic nums for obj red zoning.
492 * Placed in the first word before and the first word after an obj.
493 */
494#define RED_INACTIVE 0x5A2CF071UL /* when obj is inactive */
495#define RED_ACTIVE 0x170FC2A5UL /* when obj is active */
496
497/* ...and for poisoning */
498#define POISON_INUSE 0x5a /* for use-uninitialised poisoning */
499#define POISON_FREE 0x6b /* for use-after-free poisoning */
500#define POISON_END 0xa5 /* end-byte of poisoning */
501
502/* memory layout of objects:
503 * 0 : objp
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800504 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
Linus Torvalds1da177e2005-04-16 15:20:36 -0700505 * the end of an object is aligned with the end of the real
506 * allocation. Catches writes behind the end of the allocation.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800507 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700508 * redzone word.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800509 * cachep->obj_offset: The real object.
510 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
511 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address [BYTES_PER_WORD long]
Linus Torvalds1da177e2005-04-16 15:20:36 -0700512 */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800513static int obj_offset(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700514{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800515 return cachep->obj_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700516}
517
Pekka Enberg343e0d72006-02-01 03:05:50 -0800518static int obj_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700519{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800520 return cachep->obj_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700521}
522
Pekka Enberg343e0d72006-02-01 03:05:50 -0800523static unsigned long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700524{
525 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800526 return (unsigned long*) (objp+obj_offset(cachep)-BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700527}
528
Pekka Enberg343e0d72006-02-01 03:05:50 -0800529static unsigned long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700530{
531 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
532 if (cachep->flags & SLAB_STORE_USER)
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800533 return (unsigned long *)(objp + cachep->buffer_size -
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800534 2 * BYTES_PER_WORD);
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800535 return (unsigned long *)(objp + cachep->buffer_size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700536}
537
Pekka Enberg343e0d72006-02-01 03:05:50 -0800538static void **dbg_userword(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700539{
540 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800541 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700542}
543
544#else
545
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800546#define obj_offset(x) 0
547#define obj_size(cachep) (cachep->buffer_size)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700548#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;})
549#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;})
550#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
551
552#endif
553
554/*
555 * Maximum size of an obj (in 2^order pages)
556 * and absolute limit for the gfp order.
557 */
558#if defined(CONFIG_LARGE_ALLOCS)
559#define MAX_OBJ_ORDER 13 /* up to 32Mb */
560#define MAX_GFP_ORDER 13 /* up to 32Mb */
561#elif defined(CONFIG_MMU)
562#define MAX_OBJ_ORDER 5 /* 32 pages */
563#define MAX_GFP_ORDER 5 /* 32 pages */
564#else
565#define MAX_OBJ_ORDER 8 /* up to 1Mb */
566#define MAX_GFP_ORDER 8 /* up to 1Mb */
567#endif
568
569/*
570 * Do not go above this order unless 0 objects fit into the slab.
571 */
572#define BREAK_GFP_ORDER_HI 1
573#define BREAK_GFP_ORDER_LO 0
574static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
575
Pekka Enberg065d41c2005-11-13 16:06:46 -0800576/* Functions for storing/retrieving the cachep and or slab from the
Linus Torvalds1da177e2005-04-16 15:20:36 -0700577 * global 'mem_map'. These are used to find the slab an obj belongs to.
578 * With kfree(), these are used to find the cache which an obj belongs to.
579 */
Pekka Enberg065d41c2005-11-13 16:06:46 -0800580static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
581{
582 page->lru.next = (struct list_head *)cache;
583}
584
585static inline struct kmem_cache *page_get_cache(struct page *page)
586{
587 return (struct kmem_cache *)page->lru.next;
588}
589
590static inline void page_set_slab(struct page *page, struct slab *slab)
591{
592 page->lru.prev = (struct list_head *)slab;
593}
594
595static inline struct slab *page_get_slab(struct page *page)
596{
597 return (struct slab *)page->lru.prev;
598}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700599
Pekka Enberg6ed5eb22006-02-01 03:05:49 -0800600static inline struct kmem_cache *virt_to_cache(const void *obj)
601{
602 struct page *page = virt_to_page(obj);
603 return page_get_cache(page);
604}
605
606static inline struct slab *virt_to_slab(const void *obj)
607{
608 struct page *page = virt_to_page(obj);
609 return page_get_slab(page);
610}
611
Linus Torvalds1da177e2005-04-16 15:20:36 -0700612/* These are the default caches for kmalloc. Custom caches can have other sizes. */
613struct cache_sizes malloc_sizes[] = {
614#define CACHE(x) { .cs_size = (x) },
615#include <linux/kmalloc_sizes.h>
616 CACHE(ULONG_MAX)
617#undef CACHE
618};
619EXPORT_SYMBOL(malloc_sizes);
620
621/* Must match cache_sizes above. Out of line to keep cache footprint low. */
622struct cache_names {
623 char *name;
624 char *name_dma;
625};
626
627static struct cache_names __initdata cache_names[] = {
628#define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
629#include <linux/kmalloc_sizes.h>
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800630 {NULL,}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700631#undef CACHE
632};
633
634static struct arraycache_init initarray_cache __initdata =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800635 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700636static struct arraycache_init initarray_generic =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800637 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700638
639/* internal cache of cache description objs */
Pekka Enberg343e0d72006-02-01 03:05:50 -0800640static struct kmem_cache cache_cache = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800641 .batchcount = 1,
642 .limit = BOOT_CPUCACHE_ENTRIES,
643 .shared = 1,
Pekka Enberg343e0d72006-02-01 03:05:50 -0800644 .buffer_size = sizeof(struct kmem_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800645 .flags = SLAB_NO_REAP,
646 .spinlock = SPIN_LOCK_UNLOCKED,
647 .name = "kmem_cache",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700648#if DEBUG
Pekka Enberg343e0d72006-02-01 03:05:50 -0800649 .obj_size = sizeof(struct kmem_cache),
Linus Torvalds1da177e2005-04-16 15:20:36 -0700650#endif
651};
652
653/* Guard access to the cache-chain. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800654static DEFINE_MUTEX(cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700655static struct list_head cache_chain;
656
657/*
658 * vm_enough_memory() looks at this to determine how many
659 * slab-allocated pages are possibly freeable under pressure
660 *
661 * SLAB_RECLAIM_ACCOUNT turns this on per-slab
662 */
663atomic_t slab_reclaim_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700664
665/*
666 * chicken and egg problem: delay the per-cpu array allocation
667 * until the general caches are up.
668 */
669static enum {
670 NONE,
Christoph Lametere498be72005-09-09 13:03:32 -0700671 PARTIAL_AC,
672 PARTIAL_L3,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700673 FULL
674} g_cpucache_up;
675
676static DEFINE_PER_CPU(struct work_struct, reap_work);
677
Pekka Enberg343e0d72006-02-01 03:05:50 -0800678static void free_block(struct kmem_cache *cachep, void **objpp, int len, int node);
679static void enable_cpucache(struct kmem_cache *cachep);
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800680static void cache_reap(void *unused);
Pekka Enberg343e0d72006-02-01 03:05:50 -0800681static int __node_shrink(struct kmem_cache *cachep, int node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700682
Pekka Enberg343e0d72006-02-01 03:05:50 -0800683static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700684{
685 return cachep->array[smp_processor_id()];
686}
687
Pekka Enberg343e0d72006-02-01 03:05:50 -0800688static inline struct kmem_cache *__find_general_cachep(size_t size, gfp_t gfpflags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700689{
690 struct cache_sizes *csizep = malloc_sizes;
691
692#if DEBUG
693 /* This happens if someone tries to call
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800694 * kmem_cache_create(), or __kmalloc(), before
695 * the generic caches are initialized.
696 */
Alok Katariac7e43c72005-09-14 12:17:53 -0700697 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700698#endif
699 while (size > csizep->cs_size)
700 csizep++;
701
702 /*
Martin Hicks0abf40c2005-09-03 15:54:54 -0700703 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
Linus Torvalds1da177e2005-04-16 15:20:36 -0700704 * has cs_{dma,}cachep==NULL. Thus no special case
705 * for large kmalloc calls required.
706 */
707 if (unlikely(gfpflags & GFP_DMA))
708 return csizep->cs_dmacachep;
709 return csizep->cs_cachep;
710}
711
Pekka Enberg343e0d72006-02-01 03:05:50 -0800712struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700713{
714 return __find_general_cachep(size, gfpflags);
715}
716EXPORT_SYMBOL(kmem_find_general_cachep);
717
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800718static size_t slab_mgmt_size(size_t nr_objs, size_t align)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700719{
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800720 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
721}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700722
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800723/* Calculate the number of objects and left-over bytes for a given
724 buffer size. */
725static void cache_estimate(unsigned long gfporder, size_t buffer_size,
726 size_t align, int flags, size_t *left_over,
727 unsigned int *num)
728{
729 int nr_objs;
730 size_t mgmt_size;
731 size_t slab_size = PAGE_SIZE << gfporder;
732
733 /*
734 * The slab management structure can be either off the slab or
735 * on it. For the latter case, the memory allocated for a
736 * slab is used for:
737 *
738 * - The struct slab
739 * - One kmem_bufctl_t for each object
740 * - Padding to respect alignment of @align
741 * - @buffer_size bytes for each object
742 *
743 * If the slab management structure is off the slab, then the
744 * alignment will already be calculated into the size. Because
745 * the slabs are all pages aligned, the objects will be at the
746 * correct alignment when allocated.
747 */
748 if (flags & CFLGS_OFF_SLAB) {
749 mgmt_size = 0;
750 nr_objs = slab_size / buffer_size;
751
752 if (nr_objs > SLAB_LIMIT)
753 nr_objs = SLAB_LIMIT;
754 } else {
755 /*
756 * Ignore padding for the initial guess. The padding
757 * is at most @align-1 bytes, and @buffer_size is at
758 * least @align. In the worst case, this result will
759 * be one greater than the number of objects that fit
760 * into the memory allocation when taking the padding
761 * into account.
762 */
763 nr_objs = (slab_size - sizeof(struct slab)) /
764 (buffer_size + sizeof(kmem_bufctl_t));
765
766 /*
767 * This calculated number will be either the right
768 * amount, or one greater than what we want.
769 */
770 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
771 > slab_size)
772 nr_objs--;
773
774 if (nr_objs > SLAB_LIMIT)
775 nr_objs = SLAB_LIMIT;
776
777 mgmt_size = slab_mgmt_size(nr_objs, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700778 }
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800779 *num = nr_objs;
780 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700781}
782
783#define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
784
Pekka Enberg343e0d72006-02-01 03:05:50 -0800785static void __slab_error(const char *function, struct kmem_cache *cachep, char *msg)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700786{
787 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800788 function, cachep->name, msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700789 dump_stack();
790}
791
792/*
793 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
794 * via the workqueue/eventd.
795 * Add the CPU number into the expiration time to minimize the possibility of
796 * the CPUs getting into lockstep and contending for the global cache chain
797 * lock.
798 */
799static void __devinit start_cpu_timer(int cpu)
800{
801 struct work_struct *reap_work = &per_cpu(reap_work, cpu);
802
803 /*
804 * When this gets called from do_initcalls via cpucache_init(),
805 * init_workqueues() has already run, so keventd will be setup
806 * at that time.
807 */
808 if (keventd_up() && reap_work->func == NULL) {
809 INIT_WORK(reap_work, cache_reap, NULL);
810 schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu);
811 }
812}
813
Christoph Lametere498be72005-09-09 13:03:32 -0700814static struct array_cache *alloc_arraycache(int node, int entries,
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800815 int batchcount)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700816{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800817 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700818 struct array_cache *nc = NULL;
819
Christoph Lametere498be72005-09-09 13:03:32 -0700820 nc = kmalloc_node(memsize, GFP_KERNEL, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700821 if (nc) {
822 nc->avail = 0;
823 nc->limit = entries;
824 nc->batchcount = batchcount;
825 nc->touched = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700826 spin_lock_init(&nc->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700827 }
828 return nc;
829}
830
Christoph Lametere498be72005-09-09 13:03:32 -0700831#ifdef CONFIG_NUMA
Pekka Enberg343e0d72006-02-01 03:05:50 -0800832static void *__cache_alloc_node(struct kmem_cache *, gfp_t, int);
Christoph Lameterdc85da12006-01-18 17:42:36 -0800833
Pekka Enberg5295a742006-02-01 03:05:48 -0800834static struct array_cache **alloc_alien_cache(int node, int limit)
Christoph Lametere498be72005-09-09 13:03:32 -0700835{
836 struct array_cache **ac_ptr;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800837 int memsize = sizeof(void *) * MAX_NUMNODES;
Christoph Lametere498be72005-09-09 13:03:32 -0700838 int i;
839
840 if (limit > 1)
841 limit = 12;
842 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
843 if (ac_ptr) {
844 for_each_node(i) {
845 if (i == node || !node_online(i)) {
846 ac_ptr[i] = NULL;
847 continue;
848 }
849 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
850 if (!ac_ptr[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800851 for (i--; i <= 0; i--)
Christoph Lametere498be72005-09-09 13:03:32 -0700852 kfree(ac_ptr[i]);
853 kfree(ac_ptr);
854 return NULL;
855 }
856 }
857 }
858 return ac_ptr;
859}
860
Pekka Enberg5295a742006-02-01 03:05:48 -0800861static void free_alien_cache(struct array_cache **ac_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -0700862{
863 int i;
864
865 if (!ac_ptr)
866 return;
867
868 for_each_node(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800869 kfree(ac_ptr[i]);
Christoph Lametere498be72005-09-09 13:03:32 -0700870
871 kfree(ac_ptr);
872}
873
Pekka Enberg343e0d72006-02-01 03:05:50 -0800874static void __drain_alien_cache(struct kmem_cache *cachep,
Pekka Enberg5295a742006-02-01 03:05:48 -0800875 struct array_cache *ac, int node)
Christoph Lametere498be72005-09-09 13:03:32 -0700876{
877 struct kmem_list3 *rl3 = cachep->nodelists[node];
878
879 if (ac->avail) {
880 spin_lock(&rl3->list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -0700881 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -0700882 ac->avail = 0;
883 spin_unlock(&rl3->list_lock);
884 }
885}
886
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -0800887static void drain_alien_cache(struct kmem_cache *cachep, struct array_cache **alien)
Christoph Lametere498be72005-09-09 13:03:32 -0700888{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800889 int i = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700890 struct array_cache *ac;
891 unsigned long flags;
892
893 for_each_online_node(i) {
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -0800894 ac = alien[i];
Christoph Lametere498be72005-09-09 13:03:32 -0700895 if (ac) {
896 spin_lock_irqsave(&ac->lock, flags);
897 __drain_alien_cache(cachep, ac, i);
898 spin_unlock_irqrestore(&ac->lock, flags);
899 }
900 }
901}
902#else
Linus Torvalds7a21ef62006-02-05 11:26:38 -0800903
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -0800904#define drain_alien_cache(cachep, alien) do { } while (0)
905
Linus Torvalds7a21ef62006-02-05 11:26:38 -0800906static inline struct array_cache **alloc_alien_cache(int node, int limit)
907{
908 return (struct array_cache **) 0x01020304ul;
909}
910
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -0800911static inline void free_alien_cache(struct array_cache **ac_ptr)
912{
913}
Linus Torvalds7a21ef62006-02-05 11:26:38 -0800914
Christoph Lametere498be72005-09-09 13:03:32 -0700915#endif
916
Linus Torvalds1da177e2005-04-16 15:20:36 -0700917static int __devinit cpuup_callback(struct notifier_block *nfb,
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800918 unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700919{
920 long cpu = (long)hcpu;
Pekka Enberg343e0d72006-02-01 03:05:50 -0800921 struct kmem_cache *cachep;
Christoph Lametere498be72005-09-09 13:03:32 -0700922 struct kmem_list3 *l3 = NULL;
923 int node = cpu_to_node(cpu);
924 int memsize = sizeof(struct kmem_list3);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700925
926 switch (action) {
927 case CPU_UP_PREPARE:
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800928 mutex_lock(&cache_chain_mutex);
Christoph Lametere498be72005-09-09 13:03:32 -0700929 /* we need to do this right in the beginning since
930 * alloc_arraycache's are going to use this list.
931 * kmalloc_node allows us to add the slab to the right
932 * kmem_list3 and not this cpu's kmem_list3
933 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700934
Christoph Lametere498be72005-09-09 13:03:32 -0700935 list_for_each_entry(cachep, &cache_chain, next) {
936 /* setup the size64 kmemlist for cpu before we can
937 * begin anything. Make sure some other cpu on this
938 * node has not already allocated this
939 */
940 if (!cachep->nodelists[node]) {
941 if (!(l3 = kmalloc_node(memsize,
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800942 GFP_KERNEL, node)))
Christoph Lametere498be72005-09-09 13:03:32 -0700943 goto bad;
944 kmem_list3_init(l3);
945 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800946 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -0700947
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -0800948 /*
949 * The l3s don't come and go as CPUs come and
950 * go. cache_chain_mutex is sufficient
951 * protection here.
952 */
Christoph Lametere498be72005-09-09 13:03:32 -0700953 cachep->nodelists[node] = l3;
954 }
955
956 spin_lock_irq(&cachep->nodelists[node]->list_lock);
957 cachep->nodelists[node]->free_limit =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800958 (1 + nr_cpus_node(node)) *
959 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -0700960 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
961 }
962
963 /* Now we can go ahead with allocating the shared array's
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800964 & array cache's */
Christoph Lametere498be72005-09-09 13:03:32 -0700965 list_for_each_entry(cachep, &cache_chain, next) {
Tobias Klausercd105df2006-01-08 01:00:59 -0800966 struct array_cache *nc;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -0800967 struct array_cache *shared;
968 struct array_cache **alien;
Tobias Klausercd105df2006-01-08 01:00:59 -0800969
Christoph Lametere498be72005-09-09 13:03:32 -0700970 nc = alloc_arraycache(node, cachep->limit,
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -0800971 cachep->batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700972 if (!nc)
973 goto bad;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -0800974 shared = alloc_arraycache(node,
975 cachep->shared * cachep->batchcount,
976 0xbaadf00d);
977 if (!shared)
978 goto bad;
Linus Torvalds7a21ef62006-02-05 11:26:38 -0800979
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -0800980 alien = alloc_alien_cache(node, cachep->limit);
981 if (!alien)
982 goto bad;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700983 cachep->array[cpu] = nc;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700984
Christoph Lametere498be72005-09-09 13:03:32 -0700985 l3 = cachep->nodelists[node];
986 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -0700987
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -0800988 spin_lock_irq(&l3->list_lock);
989 if (!l3->shared) {
990 /*
991 * We are serialised from CPU_DEAD or
992 * CPU_UP_CANCELLED by the cpucontrol lock
993 */
994 l3->shared = shared;
995 shared = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -0700996 }
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -0800997#ifdef CONFIG_NUMA
998 if (!l3->alien) {
999 l3->alien = alien;
1000 alien = NULL;
1001 }
1002#endif
1003 spin_unlock_irq(&l3->list_lock);
1004
1005 kfree(shared);
1006 free_alien_cache(alien);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001007 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001008 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001009 break;
1010 case CPU_ONLINE:
1011 start_cpu_timer(cpu);
1012 break;
1013#ifdef CONFIG_HOTPLUG_CPU
1014 case CPU_DEAD:
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001015 /*
1016 * Even if all the cpus of a node are down, we don't free the
1017 * kmem_list3 of any cache. This to avoid a race between
1018 * cpu_down, and a kmalloc allocation from another cpu for
1019 * memory from the node of the cpu going down. The list3
1020 * structure is usually allocated from kmem_cache_create() and
1021 * gets destroyed at kmem_cache_destroy().
1022 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001023 /* fall thru */
1024 case CPU_UP_CANCELED:
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001025 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001026
1027 list_for_each_entry(cachep, &cache_chain, next) {
1028 struct array_cache *nc;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001029 struct array_cache *shared;
1030 struct array_cache **alien;
Christoph Lametere498be72005-09-09 13:03:32 -07001031 cpumask_t mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001032
Christoph Lametere498be72005-09-09 13:03:32 -07001033 mask = node_to_cpumask(node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001034 /* cpu is dead; no one can alloc from it. */
1035 nc = cachep->array[cpu];
1036 cachep->array[cpu] = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07001037 l3 = cachep->nodelists[node];
1038
1039 if (!l3)
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001040 goto free_array_cache;
Christoph Lametere498be72005-09-09 13:03:32 -07001041
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08001042 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07001043
1044 /* Free limit for this kmem_list3 */
1045 l3->free_limit -= cachep->batchcount;
1046 if (nc)
Christoph Lameterff694162005-09-22 21:44:02 -07001047 free_block(cachep, nc->entry, nc->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001048
1049 if (!cpus_empty(mask)) {
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08001050 spin_unlock_irq(&l3->list_lock);
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001051 goto free_array_cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001052 }
Christoph Lametere498be72005-09-09 13:03:32 -07001053
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001054 shared = l3->shared;
1055 if (shared) {
Christoph Lametere498be72005-09-09 13:03:32 -07001056 free_block(cachep, l3->shared->entry,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001057 l3->shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001058 l3->shared = NULL;
1059 }
Christoph Lametere498be72005-09-09 13:03:32 -07001060
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001061 alien = l3->alien;
1062 l3->alien = NULL;
1063
1064 spin_unlock_irq(&l3->list_lock);
1065
1066 kfree(shared);
1067 if (alien) {
1068 drain_alien_cache(cachep, alien);
1069 free_alien_cache(alien);
Christoph Lametere498be72005-09-09 13:03:32 -07001070 }
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001071free_array_cache:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001072 kfree(nc);
1073 }
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08001074 /*
1075 * In the previous loop, all the objects were freed to
1076 * the respective cache's slabs, now we can go ahead and
1077 * shrink each nodelist to its limit.
1078 */
1079 list_for_each_entry(cachep, &cache_chain, next) {
1080 l3 = cachep->nodelists[node];
1081 if (!l3)
1082 continue;
1083 spin_lock_irq(&l3->list_lock);
1084 /* free slabs belonging to this node */
1085 __node_shrink(cachep, node);
1086 spin_unlock_irq(&l3->list_lock);
1087 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001088 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001089 break;
1090#endif
1091 }
1092 return NOTIFY_OK;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001093 bad:
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001094 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001095 return NOTIFY_BAD;
1096}
1097
1098static struct notifier_block cpucache_notifier = { &cpuup_callback, NULL, 0 };
1099
Christoph Lametere498be72005-09-09 13:03:32 -07001100/*
1101 * swap the static kmem_list3 with kmalloced memory
1102 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001103static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list, int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001104{
1105 struct kmem_list3 *ptr;
1106
1107 BUG_ON(cachep->nodelists[nodeid] != list);
1108 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1109 BUG_ON(!ptr);
1110
1111 local_irq_disable();
1112 memcpy(ptr, list, sizeof(struct kmem_list3));
1113 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1114 cachep->nodelists[nodeid] = ptr;
1115 local_irq_enable();
1116}
1117
Linus Torvalds1da177e2005-04-16 15:20:36 -07001118/* Initialisation.
1119 * Called after the gfp() functions have been enabled, and before smp_init().
1120 */
1121void __init kmem_cache_init(void)
1122{
1123 size_t left_over;
1124 struct cache_sizes *sizes;
1125 struct cache_names *names;
Christoph Lametere498be72005-09-09 13:03:32 -07001126 int i;
1127
1128 for (i = 0; i < NUM_INIT_LISTS; i++) {
1129 kmem_list3_init(&initkmem_list3[i]);
1130 if (i < MAX_NUMNODES)
1131 cache_cache.nodelists[i] = NULL;
1132 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001133
1134 /*
1135 * Fragmentation resistance on low memory - only use bigger
1136 * page orders on machines with more than 32MB of memory.
1137 */
1138 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1139 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1140
Linus Torvalds1da177e2005-04-16 15:20:36 -07001141 /* Bootstrap is tricky, because several objects are allocated
1142 * from caches that do not exist yet:
Pekka Enberg343e0d72006-02-01 03:05:50 -08001143 * 1) initialize the cache_cache cache: it contains the struct kmem_cache
Linus Torvalds1da177e2005-04-16 15:20:36 -07001144 * structures of all caches, except cache_cache itself: cache_cache
1145 * is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001146 * Initially an __init data area is used for the head array and the
1147 * kmem_list3 structures, it's replaced with a kmalloc allocated
1148 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001149 * 2) Create the first kmalloc cache.
Pekka Enberg343e0d72006-02-01 03:05:50 -08001150 * The struct kmem_cache for the new cache is allocated normally.
Christoph Lametere498be72005-09-09 13:03:32 -07001151 * An __init data area is used for the head array.
1152 * 3) Create the remaining kmalloc caches, with minimally sized
1153 * head arrays.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001154 * 4) Replace the __init data head arrays for cache_cache and the first
1155 * kmalloc cache with kmalloc allocated arrays.
Christoph Lametere498be72005-09-09 13:03:32 -07001156 * 5) Replace the __init data for kmem_list3 for cache_cache and
1157 * the other cache's with kmalloc allocated memory.
1158 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001159 */
1160
1161 /* 1) create the cache_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001162 INIT_LIST_HEAD(&cache_chain);
1163 list_add(&cache_cache.next, &cache_chain);
1164 cache_cache.colour_off = cache_line_size();
1165 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
Christoph Lametere498be72005-09-09 13:03:32 -07001166 cache_cache.nodelists[numa_node_id()] = &initkmem_list3[CACHE_CACHE];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001167
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001168 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size, cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001169
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001170 cache_estimate(0, cache_cache.buffer_size, cache_line_size(), 0,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001171 &left_over, &cache_cache.num);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001172 if (!cache_cache.num)
1173 BUG();
1174
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001175 cache_cache.colour = left_over / cache_cache.colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001176 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1177 sizeof(struct slab), cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001178
1179 /* 2+3) create the kmalloc caches */
1180 sizes = malloc_sizes;
1181 names = cache_names;
1182
Christoph Lametere498be72005-09-09 13:03:32 -07001183 /* Initialize the caches that provide memory for the array cache
1184 * and the kmem_list3 structures first.
1185 * Without this, further allocations will bug
1186 */
1187
1188 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001189 sizes[INDEX_AC].cs_size,
1190 ARCH_KMALLOC_MINALIGN,
1191 (ARCH_KMALLOC_FLAGS |
1192 SLAB_PANIC), NULL, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07001193
1194 if (INDEX_AC != INDEX_L3)
1195 sizes[INDEX_L3].cs_cachep =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001196 kmem_cache_create(names[INDEX_L3].name,
1197 sizes[INDEX_L3].cs_size,
1198 ARCH_KMALLOC_MINALIGN,
1199 (ARCH_KMALLOC_FLAGS | SLAB_PANIC), NULL,
1200 NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07001201
Linus Torvalds1da177e2005-04-16 15:20:36 -07001202 while (sizes->cs_size != ULONG_MAX) {
Christoph Lametere498be72005-09-09 13:03:32 -07001203 /*
1204 * For performance, all the general caches are L1 aligned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001205 * This should be particularly beneficial on SMP boxes, as it
1206 * eliminates "false sharing".
1207 * Note for systems short on memory removing the alignment will
Christoph Lametere498be72005-09-09 13:03:32 -07001208 * allow tighter packing of the smaller caches.
1209 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001210 if (!sizes->cs_cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07001211 sizes->cs_cachep = kmem_cache_create(names->name,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001212 sizes->cs_size,
1213 ARCH_KMALLOC_MINALIGN,
1214 (ARCH_KMALLOC_FLAGS
1215 | SLAB_PANIC),
1216 NULL, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001217
1218 /* Inc off-slab bufctl limit until the ceiling is hit. */
1219 if (!(OFF_SLAB(sizes->cs_cachep))) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001220 offslab_limit = sizes->cs_size - sizeof(struct slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001221 offslab_limit /= sizeof(kmem_bufctl_t);
1222 }
1223
1224 sizes->cs_dmacachep = kmem_cache_create(names->name_dma,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001225 sizes->cs_size,
1226 ARCH_KMALLOC_MINALIGN,
1227 (ARCH_KMALLOC_FLAGS |
1228 SLAB_CACHE_DMA |
1229 SLAB_PANIC), NULL,
1230 NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001231
1232 sizes++;
1233 names++;
1234 }
1235 /* 4) Replace the bootstrap head arrays */
1236 {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001237 void *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001238
Linus Torvalds1da177e2005-04-16 15:20:36 -07001239 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001240
Linus Torvalds1da177e2005-04-16 15:20:36 -07001241 local_irq_disable();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001242 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1243 memcpy(ptr, cpu_cache_get(&cache_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001244 sizeof(struct arraycache_init));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001245 cache_cache.array[smp_processor_id()] = ptr;
1246 local_irq_enable();
Christoph Lametere498be72005-09-09 13:03:32 -07001247
Linus Torvalds1da177e2005-04-16 15:20:36 -07001248 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001249
Linus Torvalds1da177e2005-04-16 15:20:36 -07001250 local_irq_disable();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001251 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001252 != &initarray_generic.cache);
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001253 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001254 sizeof(struct arraycache_init));
Christoph Lametere498be72005-09-09 13:03:32 -07001255 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001256 ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001257 local_irq_enable();
1258 }
Christoph Lametere498be72005-09-09 13:03:32 -07001259 /* 5) Replace the bootstrap kmem_list3's */
1260 {
1261 int node;
1262 /* Replace the static kmem_list3 structures for the boot cpu */
1263 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE],
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001264 numa_node_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001265
Christoph Lametere498be72005-09-09 13:03:32 -07001266 for_each_online_node(node) {
1267 init_list(malloc_sizes[INDEX_AC].cs_cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001268 &initkmem_list3[SIZE_AC + node], node);
Christoph Lametere498be72005-09-09 13:03:32 -07001269
1270 if (INDEX_AC != INDEX_L3) {
1271 init_list(malloc_sizes[INDEX_L3].cs_cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001272 &initkmem_list3[SIZE_L3 + node],
1273 node);
Christoph Lametere498be72005-09-09 13:03:32 -07001274 }
1275 }
1276 }
1277
1278 /* 6) resize the head arrays to their final sizes */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001279 {
Pekka Enberg343e0d72006-02-01 03:05:50 -08001280 struct kmem_cache *cachep;
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001281 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001282 list_for_each_entry(cachep, &cache_chain, next)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001283 enable_cpucache(cachep);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001284 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001285 }
1286
1287 /* Done! */
1288 g_cpucache_up = FULL;
1289
1290 /* Register a cpu startup notifier callback
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001291 * that initializes cpu_cache_get for all new cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07001292 */
1293 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001294
1295 /* The reap timers are started later, with a module init call:
1296 * That part of the kernel is not yet operational.
1297 */
1298}
1299
1300static int __init cpucache_init(void)
1301{
1302 int cpu;
1303
1304 /*
1305 * Register the timers that return unneeded
1306 * pages to gfp.
1307 */
Christoph Lametere498be72005-09-09 13:03:32 -07001308 for_each_online_cpu(cpu)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001309 start_cpu_timer(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001310
1311 return 0;
1312}
1313
1314__initcall(cpucache_init);
1315
1316/*
1317 * Interface to system's page allocator. No need to hold the cache-lock.
1318 *
1319 * If we requested dmaable memory, we will get it. Even if we
1320 * did not request dmaable memory, we might get it, but that
1321 * would be relatively rare and ignorable.
1322 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001323static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001324{
1325 struct page *page;
1326 void *addr;
1327 int i;
1328
1329 flags |= cachep->gfpflags;
Christoph Lameter50c85a12005-11-13 16:06:47 -08001330 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001331 if (!page)
1332 return NULL;
1333 addr = page_address(page);
1334
1335 i = (1 << cachep->gfporder);
1336 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1337 atomic_add(i, &slab_reclaim_pages);
1338 add_page_state(nr_slab, i);
1339 while (i--) {
1340 SetPageSlab(page);
1341 page++;
1342 }
1343 return addr;
1344}
1345
1346/*
1347 * Interface to system's page release.
1348 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001349static void kmem_freepages(struct kmem_cache *cachep, void *addr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001350{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001351 unsigned long i = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001352 struct page *page = virt_to_page(addr);
1353 const unsigned long nr_freed = i;
1354
1355 while (i--) {
1356 if (!TestClearPageSlab(page))
1357 BUG();
1358 page++;
1359 }
1360 sub_page_state(nr_slab, nr_freed);
1361 if (current->reclaim_state)
1362 current->reclaim_state->reclaimed_slab += nr_freed;
1363 free_pages((unsigned long)addr, cachep->gfporder);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001364 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1365 atomic_sub(1 << cachep->gfporder, &slab_reclaim_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001366}
1367
1368static void kmem_rcu_free(struct rcu_head *head)
1369{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001370 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001371 struct kmem_cache *cachep = slab_rcu->cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001372
1373 kmem_freepages(cachep, slab_rcu->addr);
1374 if (OFF_SLAB(cachep))
1375 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1376}
1377
1378#if DEBUG
1379
1380#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enberg343e0d72006-02-01 03:05:50 -08001381static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001382 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001383{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001384 int size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001385
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001386 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001387
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001388 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001389 return;
1390
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001391 *addr++ = 0x12345678;
1392 *addr++ = caller;
1393 *addr++ = smp_processor_id();
1394 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001395 {
1396 unsigned long *sptr = &caller;
1397 unsigned long svalue;
1398
1399 while (!kstack_end(sptr)) {
1400 svalue = *sptr++;
1401 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001402 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001403 size -= sizeof(unsigned long);
1404 if (size <= sizeof(unsigned long))
1405 break;
1406 }
1407 }
1408
1409 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001410 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001411}
1412#endif
1413
Pekka Enberg343e0d72006-02-01 03:05:50 -08001414static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001415{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001416 int size = obj_size(cachep);
1417 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001418
1419 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001420 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001421}
1422
1423static void dump_line(char *data, int offset, int limit)
1424{
1425 int i;
1426 printk(KERN_ERR "%03x:", offset);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001427 for (i = 0; i < limit; i++) {
1428 printk(" %02x", (unsigned char)data[offset + i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001429 }
1430 printk("\n");
1431}
1432#endif
1433
1434#if DEBUG
1435
Pekka Enberg343e0d72006-02-01 03:05:50 -08001436static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001437{
1438 int i, size;
1439 char *realobj;
1440
1441 if (cachep->flags & SLAB_RED_ZONE) {
1442 printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001443 *dbg_redzone1(cachep, objp),
1444 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001445 }
1446
1447 if (cachep->flags & SLAB_STORE_USER) {
1448 printk(KERN_ERR "Last user: [<%p>]",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001449 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001450 print_symbol("(%s)",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001451 (unsigned long)*dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001452 printk("\n");
1453 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001454 realobj = (char *)objp + obj_offset(cachep);
1455 size = obj_size(cachep);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001456 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001457 int limit;
1458 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001459 if (i + limit > size)
1460 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001461 dump_line(realobj, i, limit);
1462 }
1463}
1464
Pekka Enberg343e0d72006-02-01 03:05:50 -08001465static void check_poison_obj(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001466{
1467 char *realobj;
1468 int size, i;
1469 int lines = 0;
1470
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001471 realobj = (char *)objp + obj_offset(cachep);
1472 size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001473
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001474 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001475 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001476 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001477 exp = POISON_END;
1478 if (realobj[i] != exp) {
1479 int limit;
1480 /* Mismatch ! */
1481 /* Print header */
1482 if (lines == 0) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001483 printk(KERN_ERR
1484 "Slab corruption: start=%p, len=%d\n",
1485 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001486 print_objinfo(cachep, objp, 0);
1487 }
1488 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001489 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001490 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001491 if (i + limit > size)
1492 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001493 dump_line(realobj, i, limit);
1494 i += 16;
1495 lines++;
1496 /* Limit to 5 lines */
1497 if (lines > 5)
1498 break;
1499 }
1500 }
1501 if (lines != 0) {
1502 /* Print some data about the neighboring objects, if they
1503 * exist:
1504 */
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08001505 struct slab *slabp = virt_to_slab(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001506 int objnr;
1507
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001508 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001509 if (objnr) {
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001510 objp = slabp->s_mem + (objnr - 1) * cachep->buffer_size;
1511 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001512 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001513 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001514 print_objinfo(cachep, objp, 2);
1515 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001516 if (objnr + 1 < cachep->num) {
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001517 objp = slabp->s_mem + (objnr + 1) * cachep->buffer_size;
1518 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001519 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001520 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001521 print_objinfo(cachep, objp, 2);
1522 }
1523 }
1524}
1525#endif
1526
Linus Torvalds1da177e2005-04-16 15:20:36 -07001527#if DEBUG
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001528/**
1529 * slab_destroy_objs - call the registered destructor for each object in
1530 * a slab that is to be destroyed.
1531 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001532static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001533{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001534 int i;
1535 for (i = 0; i < cachep->num; i++) {
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001536 void *objp = slabp->s_mem + cachep->buffer_size * i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001537
1538 if (cachep->flags & SLAB_POISON) {
1539#ifdef CONFIG_DEBUG_PAGEALLOC
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001540 if ((cachep->buffer_size % PAGE_SIZE) == 0
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001541 && OFF_SLAB(cachep))
1542 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001543 cachep->buffer_size / PAGE_SIZE,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001544 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001545 else
1546 check_poison_obj(cachep, objp);
1547#else
1548 check_poison_obj(cachep, objp);
1549#endif
1550 }
1551 if (cachep->flags & SLAB_RED_ZONE) {
1552 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1553 slab_error(cachep, "start of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001554 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001555 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1556 slab_error(cachep, "end of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001557 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001558 }
1559 if (cachep->dtor && !(cachep->flags & SLAB_POISON))
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001560 (cachep->dtor) (objp + obj_offset(cachep), cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001561 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001562}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001563#else
Pekka Enberg343e0d72006-02-01 03:05:50 -08001564static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001565{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001566 if (cachep->dtor) {
1567 int i;
1568 for (i = 0; i < cachep->num; i++) {
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001569 void *objp = slabp->s_mem + cachep->buffer_size * i;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001570 (cachep->dtor) (objp, cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001571 }
1572 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001573}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001574#endif
1575
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001576/**
1577 * Destroy all the objs in a slab, and release the mem back to the system.
1578 * Before calling the slab must have been unlinked from the cache.
1579 * The cache-lock is not held/needed.
1580 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001581static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001582{
1583 void *addr = slabp->s_mem - slabp->colouroff;
1584
1585 slab_destroy_objs(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001586 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1587 struct slab_rcu *slab_rcu;
1588
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001589 slab_rcu = (struct slab_rcu *)slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001590 slab_rcu->cachep = cachep;
1591 slab_rcu->addr = addr;
1592 call_rcu(&slab_rcu->head, kmem_rcu_free);
1593 } else {
1594 kmem_freepages(cachep, addr);
1595 if (OFF_SLAB(cachep))
1596 kmem_cache_free(cachep->slabp_cache, slabp);
1597 }
1598}
1599
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001600/* For setting up all the kmem_list3s for cache whose buffer_size is same
Christoph Lametere498be72005-09-09 13:03:32 -07001601 as size of kmem_list3. */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001602static void set_up_list3s(struct kmem_cache *cachep, int index)
Christoph Lametere498be72005-09-09 13:03:32 -07001603{
1604 int node;
1605
1606 for_each_online_node(node) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001607 cachep->nodelists[node] = &initkmem_list3[index + node];
Christoph Lametere498be72005-09-09 13:03:32 -07001608 cachep->nodelists[node]->next_reap = jiffies +
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001609 REAPTIMEOUT_LIST3 +
1610 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -07001611 }
1612}
1613
Linus Torvalds1da177e2005-04-16 15:20:36 -07001614/**
Randy.Dunlapa70773d2006-02-01 03:05:52 -08001615 * calculate_slab_order - calculate size (page order) of slabs
1616 * @cachep: pointer to the cache that is being created
1617 * @size: size of objects to be created in this cache.
1618 * @align: required alignment for the objects.
1619 * @flags: slab allocation flags
1620 *
1621 * Also calculates the number of objects per slab.
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001622 *
1623 * This could be made much more intelligent. For now, try to avoid using
1624 * high order pages for slabs. When the gfp() functions are more friendly
1625 * towards high-order requests, this should be changed.
1626 */
Randy Dunlapee13d782006-02-01 03:05:53 -08001627static inline size_t calculate_slab_order(struct kmem_cache *cachep,
1628 size_t size, size_t align, unsigned long flags)
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001629{
1630 size_t left_over = 0;
1631
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001632 for (;; cachep->gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001633 unsigned int num;
1634 size_t remainder;
1635
1636 if (cachep->gfporder > MAX_GFP_ORDER) {
1637 cachep->num = 0;
1638 break;
1639 }
1640
1641 cache_estimate(cachep->gfporder, size, align, flags,
1642 &remainder, &num);
1643 if (!num)
1644 continue;
1645 /* More than offslab_limit objects will cause problems */
1646 if (flags & CFLGS_OFF_SLAB && cachep->num > offslab_limit)
1647 break;
1648
1649 cachep->num = num;
1650 left_over = remainder;
1651
1652 /*
1653 * Large number of objects is good, but very large slabs are
1654 * currently bad for the gfp()s.
1655 */
1656 if (cachep->gfporder >= slab_break_gfp_order)
1657 break;
1658
1659 if ((left_over * 8) <= (PAGE_SIZE << cachep->gfporder))
1660 /* Acceptable internal fragmentation */
1661 break;
1662 }
1663 return left_over;
1664}
1665
1666/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07001667 * kmem_cache_create - Create a cache.
1668 * @name: A string which is used in /proc/slabinfo to identify this cache.
1669 * @size: The size of objects to be created in this cache.
1670 * @align: The required alignment for the objects.
1671 * @flags: SLAB flags
1672 * @ctor: A constructor for the objects.
1673 * @dtor: A destructor for the objects.
1674 *
1675 * Returns a ptr to the cache on success, NULL on failure.
1676 * Cannot be called within a int, but can be interrupted.
1677 * The @ctor is run when new pages are allocated by the cache
1678 * and the @dtor is run before the pages are handed back.
1679 *
1680 * @name must be valid until the cache is destroyed. This implies that
1681 * the module calling this has to destroy the cache before getting
1682 * unloaded.
1683 *
1684 * The flags are
1685 *
1686 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1687 * to catch references to uninitialised memory.
1688 *
1689 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1690 * for buffer overruns.
1691 *
1692 * %SLAB_NO_REAP - Don't automatically reap this cache when we're under
1693 * memory pressure.
1694 *
1695 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1696 * cacheline. This can be beneficial if you're counting cycles as closely
1697 * as davem.
1698 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001699struct kmem_cache *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001700kmem_cache_create (const char *name, size_t size, size_t align,
Pekka Enberg343e0d72006-02-01 03:05:50 -08001701 unsigned long flags, void (*ctor)(void*, struct kmem_cache *, unsigned long),
1702 void (*dtor)(void*, struct kmem_cache *, unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001703{
1704 size_t left_over, slab_size, ralign;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001705 struct kmem_cache *cachep = NULL;
Andrew Morton4f12bb42005-11-07 00:58:00 -08001706 struct list_head *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001707
1708 /*
1709 * Sanity checks... these are all serious usage bugs.
1710 */
1711 if ((!name) ||
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001712 in_interrupt() ||
1713 (size < BYTES_PER_WORD) ||
1714 (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) {
1715 printk(KERN_ERR "%s: Early error in slab %s\n",
1716 __FUNCTION__, name);
1717 BUG();
1718 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001719
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08001720 /*
1721 * Prevent CPUs from coming and going.
1722 * lock_cpu_hotplug() nests outside cache_chain_mutex
1723 */
1724 lock_cpu_hotplug();
1725
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001726 mutex_lock(&cache_chain_mutex);
Andrew Morton4f12bb42005-11-07 00:58:00 -08001727
1728 list_for_each(p, &cache_chain) {
Pekka Enberg343e0d72006-02-01 03:05:50 -08001729 struct kmem_cache *pc = list_entry(p, struct kmem_cache, next);
Andrew Morton4f12bb42005-11-07 00:58:00 -08001730 mm_segment_t old_fs = get_fs();
1731 char tmp;
1732 int res;
1733
1734 /*
1735 * This happens when the module gets unloaded and doesn't
1736 * destroy its slab cache and no-one else reuses the vmalloc
1737 * area of the module. Print a warning.
1738 */
1739 set_fs(KERNEL_DS);
1740 res = __get_user(tmp, pc->name);
1741 set_fs(old_fs);
1742 if (res) {
1743 printk("SLAB: cache with size %d has lost its name\n",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001744 pc->buffer_size);
Andrew Morton4f12bb42005-11-07 00:58:00 -08001745 continue;
1746 }
1747
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001748 if (!strcmp(pc->name, name)) {
Andrew Morton4f12bb42005-11-07 00:58:00 -08001749 printk("kmem_cache_create: duplicate cache %s\n", name);
1750 dump_stack();
1751 goto oops;
1752 }
1753 }
1754
Linus Torvalds1da177e2005-04-16 15:20:36 -07001755#if DEBUG
1756 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
1757 if ((flags & SLAB_DEBUG_INITIAL) && !ctor) {
1758 /* No constructor, but inital state check requested */
1759 printk(KERN_ERR "%s: No con, but init state check "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001760 "requested - %s\n", __FUNCTION__, name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001761 flags &= ~SLAB_DEBUG_INITIAL;
1762 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001763#if FORCED_DEBUG
1764 /*
1765 * Enable redzoning and last user accounting, except for caches with
1766 * large objects, if the increased size would increase the object size
1767 * above the next power of two: caches with object sizes just above a
1768 * power of two have a significant amount of internal fragmentation.
1769 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001770 if ((size < 4096
1771 || fls(size - 1) == fls(size - 1 + 3 * BYTES_PER_WORD)))
1772 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001773 if (!(flags & SLAB_DESTROY_BY_RCU))
1774 flags |= SLAB_POISON;
1775#endif
1776 if (flags & SLAB_DESTROY_BY_RCU)
1777 BUG_ON(flags & SLAB_POISON);
1778#endif
1779 if (flags & SLAB_DESTROY_BY_RCU)
1780 BUG_ON(dtor);
1781
1782 /*
1783 * Always checks flags, a caller might be expecting debug
1784 * support which isn't available.
1785 */
1786 if (flags & ~CREATE_MASK)
1787 BUG();
1788
1789 /* Check that size is in terms of words. This is needed to avoid
1790 * unaligned accesses for some archs when redzoning is used, and makes
1791 * sure any on-slab bufctl's are also correctly aligned.
1792 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001793 if (size & (BYTES_PER_WORD - 1)) {
1794 size += (BYTES_PER_WORD - 1);
1795 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001796 }
1797
1798 /* calculate out the final buffer alignment: */
1799 /* 1) arch recommendation: can be overridden for debug */
1800 if (flags & SLAB_HWCACHE_ALIGN) {
1801 /* Default alignment: as specified by the arch code.
1802 * Except if an object is really small, then squeeze multiple
1803 * objects into one cacheline.
1804 */
1805 ralign = cache_line_size();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001806 while (size <= ralign / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001807 ralign /= 2;
1808 } else {
1809 ralign = BYTES_PER_WORD;
1810 }
1811 /* 2) arch mandated alignment: disables debug if necessary */
1812 if (ralign < ARCH_SLAB_MINALIGN) {
1813 ralign = ARCH_SLAB_MINALIGN;
1814 if (ralign > BYTES_PER_WORD)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001815 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001816 }
1817 /* 3) caller mandated alignment: disables debug if necessary */
1818 if (ralign < align) {
1819 ralign = align;
1820 if (ralign > BYTES_PER_WORD)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001821 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001822 }
1823 /* 4) Store it. Note that the debug code below can reduce
1824 * the alignment to BYTES_PER_WORD.
1825 */
1826 align = ralign;
1827
1828 /* Get cache's description obj. */
Pekka Enberg343e0d72006-02-01 03:05:50 -08001829 cachep = kmem_cache_alloc(&cache_cache, SLAB_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001830 if (!cachep)
Andrew Morton4f12bb42005-11-07 00:58:00 -08001831 goto oops;
Pekka Enberg343e0d72006-02-01 03:05:50 -08001832 memset(cachep, 0, sizeof(struct kmem_cache));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001833
1834#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001835 cachep->obj_size = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001836
1837 if (flags & SLAB_RED_ZONE) {
1838 /* redzoning only works with word aligned caches */
1839 align = BYTES_PER_WORD;
1840
1841 /* add space for red zone words */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001842 cachep->obj_offset += BYTES_PER_WORD;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001843 size += 2 * BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001844 }
1845 if (flags & SLAB_STORE_USER) {
1846 /* user store requires word alignment and
1847 * one word storage behind the end of the real
1848 * object.
1849 */
1850 align = BYTES_PER_WORD;
1851 size += BYTES_PER_WORD;
1852 }
1853#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001854 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001855 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
1856 cachep->obj_offset += PAGE_SIZE - size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001857 size = PAGE_SIZE;
1858 }
1859#endif
1860#endif
1861
1862 /* Determine if the slab management is 'on' or 'off' slab. */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001863 if (size >= (PAGE_SIZE >> 3))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001864 /*
1865 * Size is large, assume best to place the slab management obj
1866 * off-slab (should allow better packing of objs).
1867 */
1868 flags |= CFLGS_OFF_SLAB;
1869
1870 size = ALIGN(size, align);
1871
1872 if ((flags & SLAB_RECLAIM_ACCOUNT) && size <= PAGE_SIZE) {
1873 /*
1874 * A VFS-reclaimable slab tends to have most allocations
1875 * as GFP_NOFS and we really don't want to have to be allocating
1876 * higher-order pages when we are unable to shrink dcache.
1877 */
1878 cachep->gfporder = 0;
1879 cache_estimate(cachep->gfporder, size, align, flags,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001880 &left_over, &cachep->num);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001881 } else
1882 left_over = calculate_slab_order(cachep, size, align, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001883
1884 if (!cachep->num) {
1885 printk("kmem_cache_create: couldn't create cache %s.\n", name);
1886 kmem_cache_free(&cache_cache, cachep);
1887 cachep = NULL;
Andrew Morton4f12bb42005-11-07 00:58:00 -08001888 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001889 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001890 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
1891 + sizeof(struct slab), align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001892
1893 /*
1894 * If the slab has been placed off-slab, and we have enough space then
1895 * move it on-slab. This is at the expense of any extra colouring.
1896 */
1897 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
1898 flags &= ~CFLGS_OFF_SLAB;
1899 left_over -= slab_size;
1900 }
1901
1902 if (flags & CFLGS_OFF_SLAB) {
1903 /* really off slab. No need for manual alignment */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001904 slab_size =
1905 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001906 }
1907
1908 cachep->colour_off = cache_line_size();
1909 /* Offset must be a multiple of the alignment. */
1910 if (cachep->colour_off < align)
1911 cachep->colour_off = align;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001912 cachep->colour = left_over / cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001913 cachep->slab_size = slab_size;
1914 cachep->flags = flags;
1915 cachep->gfpflags = 0;
1916 if (flags & SLAB_CACHE_DMA)
1917 cachep->gfpflags |= GFP_DMA;
1918 spin_lock_init(&cachep->spinlock);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001919 cachep->buffer_size = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001920
1921 if (flags & CFLGS_OFF_SLAB)
Victor Fuscob2d55072005-09-10 00:26:36 -07001922 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001923 cachep->ctor = ctor;
1924 cachep->dtor = dtor;
1925 cachep->name = name;
1926
Linus Torvalds1da177e2005-04-16 15:20:36 -07001927
1928 if (g_cpucache_up == FULL) {
1929 enable_cpucache(cachep);
1930 } else {
1931 if (g_cpucache_up == NONE) {
1932 /* Note: the first kmem_cache_create must create
1933 * the cache that's used by kmalloc(24), otherwise
1934 * the creation of further caches will BUG().
1935 */
Christoph Lametere498be72005-09-09 13:03:32 -07001936 cachep->array[smp_processor_id()] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001937 &initarray_generic.cache;
Christoph Lametere498be72005-09-09 13:03:32 -07001938
1939 /* If the cache that's used by
1940 * kmalloc(sizeof(kmem_list3)) is the first cache,
1941 * then we need to set up all its list3s, otherwise
1942 * the creation of further caches will BUG().
1943 */
1944 set_up_list3s(cachep, SIZE_AC);
1945 if (INDEX_AC == INDEX_L3)
1946 g_cpucache_up = PARTIAL_L3;
1947 else
1948 g_cpucache_up = PARTIAL_AC;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001949 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07001950 cachep->array[smp_processor_id()] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001951 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001952
1953 if (g_cpucache_up == PARTIAL_AC) {
1954 set_up_list3s(cachep, SIZE_L3);
1955 g_cpucache_up = PARTIAL_L3;
1956 } else {
1957 int node;
1958 for_each_online_node(node) {
1959
1960 cachep->nodelists[node] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001961 kmalloc_node(sizeof
1962 (struct kmem_list3),
1963 GFP_KERNEL, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001964 BUG_ON(!cachep->nodelists[node]);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001965 kmem_list3_init(cachep->
1966 nodelists[node]);
Christoph Lametere498be72005-09-09 13:03:32 -07001967 }
1968 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001969 }
Christoph Lametere498be72005-09-09 13:03:32 -07001970 cachep->nodelists[numa_node_id()]->next_reap =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001971 jiffies + REAPTIMEOUT_LIST3 +
1972 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -07001973
Pekka Enberg9a2dba42006-02-01 03:05:49 -08001974 BUG_ON(!cpu_cache_get(cachep));
1975 cpu_cache_get(cachep)->avail = 0;
1976 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1977 cpu_cache_get(cachep)->batchcount = 1;
1978 cpu_cache_get(cachep)->touched = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001979 cachep->batchcount = 1;
1980 cachep->limit = BOOT_CPUCACHE_ENTRIES;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001981 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001982
Linus Torvalds1da177e2005-04-16 15:20:36 -07001983 /* cache setup completed, link it into the list */
1984 list_add(&cachep->next, &cache_chain);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001985 oops:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001986 if (!cachep && (flags & SLAB_PANIC))
1987 panic("kmem_cache_create(): failed to create slab `%s'\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001988 name);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001989 mutex_unlock(&cache_chain_mutex);
Ravikiran G Thirumalaif0188f42006-02-10 01:51:13 -08001990 unlock_cpu_hotplug();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001991 return cachep;
1992}
1993EXPORT_SYMBOL(kmem_cache_create);
1994
1995#if DEBUG
1996static void check_irq_off(void)
1997{
1998 BUG_ON(!irqs_disabled());
1999}
2000
2001static void check_irq_on(void)
2002{
2003 BUG_ON(irqs_disabled());
2004}
2005
Pekka Enberg343e0d72006-02-01 03:05:50 -08002006static void check_spinlock_acquired(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002007{
2008#ifdef CONFIG_SMP
2009 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002010 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002011#endif
2012}
Christoph Lametere498be72005-09-09 13:03:32 -07002013
Pekka Enberg343e0d72006-02-01 03:05:50 -08002014static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07002015{
2016#ifdef CONFIG_SMP
2017 check_irq_off();
2018 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2019#endif
2020}
2021
Linus Torvalds1da177e2005-04-16 15:20:36 -07002022#else
2023#define check_irq_off() do { } while(0)
2024#define check_irq_on() do { } while(0)
2025#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07002026#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002027#endif
2028
2029/*
2030 * Waits for all CPUs to execute func().
2031 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002032static void smp_call_function_all_cpus(void (*func)(void *arg), void *arg)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002033{
2034 check_irq_on();
2035 preempt_disable();
2036
2037 local_irq_disable();
2038 func(arg);
2039 local_irq_enable();
2040
2041 if (smp_call_function(func, arg, 1, 1))
2042 BUG();
2043
2044 preempt_enable();
2045}
2046
Pekka Enberg343e0d72006-02-01 03:05:50 -08002047static void drain_array_locked(struct kmem_cache *cachep, struct array_cache *ac,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002048 int force, int node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002049
2050static void do_drain(void *arg)
2051{
Pekka Enberg343e0d72006-02-01 03:05:50 -08002052 struct kmem_cache *cachep = (struct kmem_cache *) arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002053 struct array_cache *ac;
Christoph Lameterff694162005-09-22 21:44:02 -07002054 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002055
2056 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002057 ac = cpu_cache_get(cachep);
Christoph Lameterff694162005-09-22 21:44:02 -07002058 spin_lock(&cachep->nodelists[node]->list_lock);
2059 free_block(cachep, ac->entry, ac->avail, node);
2060 spin_unlock(&cachep->nodelists[node]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002061 ac->avail = 0;
2062}
2063
Pekka Enberg343e0d72006-02-01 03:05:50 -08002064static void drain_cpu_caches(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002065{
Christoph Lametere498be72005-09-09 13:03:32 -07002066 struct kmem_list3 *l3;
2067 int node;
2068
Linus Torvalds1da177e2005-04-16 15:20:36 -07002069 smp_call_function_all_cpus(do_drain, cachep);
2070 check_irq_on();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002071 for_each_online_node(node) {
Christoph Lametere498be72005-09-09 13:03:32 -07002072 l3 = cachep->nodelists[node];
2073 if (l3) {
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08002074 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002075 drain_array_locked(cachep, l3->shared, 1, node);
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08002076 spin_unlock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002077 if (l3->alien)
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08002078 drain_alien_cache(cachep, l3->alien);
Christoph Lametere498be72005-09-09 13:03:32 -07002079 }
2080 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002081}
2082
Pekka Enberg343e0d72006-02-01 03:05:50 -08002083static int __node_shrink(struct kmem_cache *cachep, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002084{
2085 struct slab *slabp;
Christoph Lametere498be72005-09-09 13:03:32 -07002086 struct kmem_list3 *l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002087 int ret;
2088
Christoph Lametere498be72005-09-09 13:03:32 -07002089 for (;;) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002090 struct list_head *p;
2091
Christoph Lametere498be72005-09-09 13:03:32 -07002092 p = l3->slabs_free.prev;
2093 if (p == &l3->slabs_free)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002094 break;
2095
Christoph Lametere498be72005-09-09 13:03:32 -07002096 slabp = list_entry(l3->slabs_free.prev, struct slab, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002097#if DEBUG
2098 if (slabp->inuse)
2099 BUG();
2100#endif
2101 list_del(&slabp->list);
2102
Christoph Lametere498be72005-09-09 13:03:32 -07002103 l3->free_objects -= cachep->num;
2104 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002105 slab_destroy(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07002106 spin_lock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002107 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002108 ret = !list_empty(&l3->slabs_full) || !list_empty(&l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002109 return ret;
2110}
2111
Pekka Enberg343e0d72006-02-01 03:05:50 -08002112static int __cache_shrink(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07002113{
2114 int ret = 0, i = 0;
2115 struct kmem_list3 *l3;
2116
2117 drain_cpu_caches(cachep);
2118
2119 check_irq_on();
2120 for_each_online_node(i) {
2121 l3 = cachep->nodelists[i];
2122 if (l3) {
2123 spin_lock_irq(&l3->list_lock);
2124 ret += __node_shrink(cachep, i);
2125 spin_unlock_irq(&l3->list_lock);
2126 }
2127 }
2128 return (ret ? 1 : 0);
2129}
2130
Linus Torvalds1da177e2005-04-16 15:20:36 -07002131/**
2132 * kmem_cache_shrink - Shrink a cache.
2133 * @cachep: The cache to shrink.
2134 *
2135 * Releases as many slabs as possible for a cache.
2136 * To help debugging, a zero exit status indicates all slabs were released.
2137 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002138int kmem_cache_shrink(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002139{
2140 if (!cachep || in_interrupt())
2141 BUG();
2142
2143 return __cache_shrink(cachep);
2144}
2145EXPORT_SYMBOL(kmem_cache_shrink);
2146
2147/**
2148 * kmem_cache_destroy - delete a cache
2149 * @cachep: the cache to destroy
2150 *
Pekka Enberg343e0d72006-02-01 03:05:50 -08002151 * Remove a struct kmem_cache object from the slab cache.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002152 * Returns 0 on success.
2153 *
2154 * It is expected this function will be called by a module when it is
2155 * unloaded. This will remove the cache completely, and avoid a duplicate
2156 * cache being allocated each time a module is loaded and unloaded, if the
2157 * module doesn't have persistent in-kernel storage across loads and unloads.
2158 *
2159 * The cache must be empty before calling this function.
2160 *
2161 * The caller must guarantee that noone will allocate memory from the cache
2162 * during the kmem_cache_destroy().
2163 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002164int kmem_cache_destroy(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002165{
2166 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07002167 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002168
2169 if (!cachep || in_interrupt())
2170 BUG();
2171
2172 /* Don't let CPUs to come and go */
2173 lock_cpu_hotplug();
2174
2175 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002176 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002177 /*
2178 * the chain is never empty, cache_cache is never destroyed
2179 */
2180 list_del(&cachep->next);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002181 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002182
2183 if (__cache_shrink(cachep)) {
2184 slab_error(cachep, "Can't free all objects");
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002185 mutex_lock(&cache_chain_mutex);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002186 list_add(&cachep->next, &cache_chain);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002187 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002188 unlock_cpu_hotplug();
2189 return 1;
2190 }
2191
2192 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
Paul E. McKenneyfbd568a3e2005-05-01 08:59:04 -07002193 synchronize_rcu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002194
Christoph Lametere498be72005-09-09 13:03:32 -07002195 for_each_online_cpu(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002196 kfree(cachep->array[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002197
2198 /* NUMA: free the list3 structures */
Christoph Lametere498be72005-09-09 13:03:32 -07002199 for_each_online_node(i) {
2200 if ((l3 = cachep->nodelists[i])) {
2201 kfree(l3->shared);
2202 free_alien_cache(l3->alien);
2203 kfree(l3);
2204 }
2205 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002206 kmem_cache_free(&cache_cache, cachep);
2207
2208 unlock_cpu_hotplug();
2209
2210 return 0;
2211}
2212EXPORT_SYMBOL(kmem_cache_destroy);
2213
2214/* Get the memory for a slab management obj. */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002215static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002216 int colour_off, gfp_t local_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002217{
2218 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002219
Linus Torvalds1da177e2005-04-16 15:20:36 -07002220 if (OFF_SLAB(cachep)) {
2221 /* Slab management obj is off-slab. */
2222 slabp = kmem_cache_alloc(cachep->slabp_cache, local_flags);
2223 if (!slabp)
2224 return NULL;
2225 } else {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002226 slabp = objp + colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002227 colour_off += cachep->slab_size;
2228 }
2229 slabp->inuse = 0;
2230 slabp->colouroff = colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002231 slabp->s_mem = objp + colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002232
2233 return slabp;
2234}
2235
2236static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2237{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002238 return (kmem_bufctl_t *) (slabp + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002239}
2240
Pekka Enberg343e0d72006-02-01 03:05:50 -08002241static void cache_init_objs(struct kmem_cache *cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002242 struct slab *slabp, unsigned long ctor_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002243{
2244 int i;
2245
2246 for (i = 0; i < cachep->num; i++) {
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002247 void *objp = slabp->s_mem + cachep->buffer_size * i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002248#if DEBUG
2249 /* need to poison the objs? */
2250 if (cachep->flags & SLAB_POISON)
2251 poison_obj(cachep, objp, POISON_FREE);
2252 if (cachep->flags & SLAB_STORE_USER)
2253 *dbg_userword(cachep, objp) = NULL;
2254
2255 if (cachep->flags & SLAB_RED_ZONE) {
2256 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2257 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2258 }
2259 /*
2260 * Constructors are not allowed to allocate memory from
2261 * the same cache which they are a constructor for.
2262 * Otherwise, deadlock. They must also be threaded.
2263 */
2264 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002265 cachep->ctor(objp + obj_offset(cachep), cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002266 ctor_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002267
2268 if (cachep->flags & SLAB_RED_ZONE) {
2269 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2270 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002271 " end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002272 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2273 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002274 " start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002275 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002276 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002277 && cachep->flags & SLAB_POISON)
2278 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002279 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002280#else
2281 if (cachep->ctor)
2282 cachep->ctor(objp, cachep, ctor_flags);
2283#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002284 slab_bufctl(slabp)[i] = i + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002285 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002286 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002287 slabp->free = 0;
2288}
2289
Pekka Enberg343e0d72006-02-01 03:05:50 -08002290static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002291{
2292 if (flags & SLAB_DMA) {
2293 if (!(cachep->gfpflags & GFP_DMA))
2294 BUG();
2295 } else {
2296 if (cachep->gfpflags & GFP_DMA)
2297 BUG();
2298 }
2299}
2300
Pekka Enberg343e0d72006-02-01 03:05:50 -08002301static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp, int nodeid)
Matthew Dobson78d382d2006-02-01 03:05:47 -08002302{
2303 void *objp = slabp->s_mem + (slabp->free * cachep->buffer_size);
2304 kmem_bufctl_t next;
2305
2306 slabp->inuse++;
2307 next = slab_bufctl(slabp)[slabp->free];
2308#if DEBUG
2309 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2310 WARN_ON(slabp->nodeid != nodeid);
2311#endif
2312 slabp->free = next;
2313
2314 return objp;
2315}
2316
Pekka Enberg343e0d72006-02-01 03:05:50 -08002317static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp, void *objp,
Matthew Dobson78d382d2006-02-01 03:05:47 -08002318 int nodeid)
2319{
2320 unsigned int objnr = (unsigned)(objp-slabp->s_mem) / cachep->buffer_size;
2321
2322#if DEBUG
2323 /* Verify that the slab belongs to the intended node */
2324 WARN_ON(slabp->nodeid != nodeid);
2325
2326 if (slab_bufctl(slabp)[objnr] != BUFCTL_FREE) {
2327 printk(KERN_ERR "slab: double free detected in cache "
2328 "'%s', objp %p\n", cachep->name, objp);
2329 BUG();
2330 }
2331#endif
2332 slab_bufctl(slabp)[objnr] = slabp->free;
2333 slabp->free = objnr;
2334 slabp->inuse--;
2335}
2336
Pekka Enberg343e0d72006-02-01 03:05:50 -08002337static void set_slab_attr(struct kmem_cache *cachep, struct slab *slabp, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002338{
2339 int i;
2340 struct page *page;
2341
2342 /* Nasty!!!!!! I hope this is OK. */
2343 i = 1 << cachep->gfporder;
2344 page = virt_to_page(objp);
2345 do {
Pekka Enberg065d41c2005-11-13 16:06:46 -08002346 page_set_cache(page, cachep);
2347 page_set_slab(page, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002348 page++;
2349 } while (--i);
2350}
2351
2352/*
2353 * Grow (by 1) the number of slabs within a cache. This is called by
2354 * kmem_cache_alloc() when there are no active objs left in a cache.
2355 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002356static int cache_grow(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002357{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002358 struct slab *slabp;
2359 void *objp;
2360 size_t offset;
2361 gfp_t local_flags;
2362 unsigned long ctor_flags;
Christoph Lametere498be72005-09-09 13:03:32 -07002363 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002364
2365 /* Be lazy and only check for valid flags here,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002366 * keeping it out of the critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002367 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002368 if (flags & ~(SLAB_DMA | SLAB_LEVEL_MASK | SLAB_NO_GROW))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002369 BUG();
2370 if (flags & SLAB_NO_GROW)
2371 return 0;
2372
2373 ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2374 local_flags = (flags & SLAB_LEVEL_MASK);
2375 if (!(local_flags & __GFP_WAIT))
2376 /*
2377 * Not allowed to sleep. Need to tell a constructor about
2378 * this - it might need to know...
2379 */
2380 ctor_flags |= SLAB_CTOR_ATOMIC;
2381
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002382 /* Take the l3 list lock to change the colour_next on this node */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002383 check_irq_off();
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002384 l3 = cachep->nodelists[nodeid];
2385 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002386
2387 /* Get colour for the slab, and cal the next value. */
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002388 offset = l3->colour_next;
2389 l3->colour_next++;
2390 if (l3->colour_next >= cachep->colour)
2391 l3->colour_next = 0;
2392 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002393
Ravikiran G Thirumalai2e1217c2006-02-04 23:27:56 -08002394 offset *= cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002395
2396 if (local_flags & __GFP_WAIT)
2397 local_irq_enable();
2398
2399 /*
2400 * The test for missing atomic flag is performed here, rather than
2401 * the more obvious place, simply to reduce the critical path length
2402 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2403 * will eventually be caught here (where it matters).
2404 */
2405 kmem_flagcheck(cachep, flags);
2406
Christoph Lametere498be72005-09-09 13:03:32 -07002407 /* Get mem for the objs.
2408 * Attempt to allocate a physical page from 'nodeid',
2409 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002410 if (!(objp = kmem_getpages(cachep, flags, nodeid)))
2411 goto failed;
2412
2413 /* Get slab management. */
2414 if (!(slabp = alloc_slabmgmt(cachep, objp, offset, local_flags)))
2415 goto opps1;
2416
Christoph Lametere498be72005-09-09 13:03:32 -07002417 slabp->nodeid = nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002418 set_slab_attr(cachep, slabp, objp);
2419
2420 cache_init_objs(cachep, slabp, ctor_flags);
2421
2422 if (local_flags & __GFP_WAIT)
2423 local_irq_disable();
2424 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002425 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002426
2427 /* Make slab active. */
Christoph Lametere498be72005-09-09 13:03:32 -07002428 list_add_tail(&slabp->list, &(l3->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002429 STATS_INC_GROWN(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002430 l3->free_objects += cachep->num;
2431 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002432 return 1;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002433 opps1:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002434 kmem_freepages(cachep, objp);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002435 failed:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002436 if (local_flags & __GFP_WAIT)
2437 local_irq_disable();
2438 return 0;
2439}
2440
2441#if DEBUG
2442
2443/*
2444 * Perform extra freeing checks:
2445 * - detect bad pointers.
2446 * - POISON/RED_ZONE checking
2447 * - destructor calls, for caches with POISON+dtor
2448 */
2449static void kfree_debugcheck(const void *objp)
2450{
2451 struct page *page;
2452
2453 if (!virt_addr_valid(objp)) {
2454 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002455 (unsigned long)objp);
2456 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002457 }
2458 page = virt_to_page(objp);
2459 if (!PageSlab(page)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002460 printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n",
2461 (unsigned long)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002462 BUG();
2463 }
2464}
2465
Pekka Enberg343e0d72006-02-01 03:05:50 -08002466static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002467 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002468{
2469 struct page *page;
2470 unsigned int objnr;
2471 struct slab *slabp;
2472
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002473 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002474 kfree_debugcheck(objp);
2475 page = virt_to_page(objp);
2476
Pekka Enberg065d41c2005-11-13 16:06:46 -08002477 if (page_get_cache(page) != cachep) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002478 printk(KERN_ERR
2479 "mismatch in kmem_cache_free: expected cache %p, got %p\n",
2480 page_get_cache(page), cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002481 printk(KERN_ERR "%p is %s.\n", cachep, cachep->name);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002482 printk(KERN_ERR "%p is %s.\n", page_get_cache(page),
2483 page_get_cache(page)->name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002484 WARN_ON(1);
2485 }
Pekka Enberg065d41c2005-11-13 16:06:46 -08002486 slabp = page_get_slab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002487
2488 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002489 if (*dbg_redzone1(cachep, objp) != RED_ACTIVE
2490 || *dbg_redzone2(cachep, objp) != RED_ACTIVE) {
2491 slab_error(cachep,
2492 "double free, or memory outside"
2493 " object was overwritten");
2494 printk(KERN_ERR
2495 "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n",
2496 objp, *dbg_redzone1(cachep, objp),
2497 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002498 }
2499 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2500 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2501 }
2502 if (cachep->flags & SLAB_STORE_USER)
2503 *dbg_userword(cachep, objp) = caller;
2504
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002505 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002506
2507 BUG_ON(objnr >= cachep->num);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002508 BUG_ON(objp != slabp->s_mem + objnr * cachep->buffer_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002509
2510 if (cachep->flags & SLAB_DEBUG_INITIAL) {
2511 /* Need to call the slab's constructor so the
2512 * caller can perform a verify of its state (debugging).
2513 * Called without the cache-lock held.
2514 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002515 cachep->ctor(objp + obj_offset(cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002516 cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002517 }
2518 if (cachep->flags & SLAB_POISON && cachep->dtor) {
2519 /* we want to cache poison the object,
2520 * call the destruction callback
2521 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002522 cachep->dtor(objp + obj_offset(cachep), cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002523 }
2524 if (cachep->flags & SLAB_POISON) {
2525#ifdef CONFIG_DEBUG_PAGEALLOC
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002526 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002527 store_stackinfo(cachep, objp, (unsigned long)caller);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002528 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002529 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002530 } else {
2531 poison_obj(cachep, objp, POISON_FREE);
2532 }
2533#else
2534 poison_obj(cachep, objp, POISON_FREE);
2535#endif
2536 }
2537 return objp;
2538}
2539
Pekka Enberg343e0d72006-02-01 03:05:50 -08002540static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002541{
2542 kmem_bufctl_t i;
2543 int entries = 0;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002544
Linus Torvalds1da177e2005-04-16 15:20:36 -07002545 /* Check slab's freelist to see if this obj is there. */
2546 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2547 entries++;
2548 if (entries > cachep->num || i >= cachep->num)
2549 goto bad;
2550 }
2551 if (entries != cachep->num - slabp->inuse) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002552 bad:
2553 printk(KERN_ERR
2554 "slab: Internal list corruption detected in cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2555 cachep->name, cachep->num, slabp, slabp->inuse);
2556 for (i = 0;
2557 i < sizeof(slabp) + cachep->num * sizeof(kmem_bufctl_t);
2558 i++) {
2559 if ((i % 16) == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002560 printk("\n%03x:", i);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002561 printk(" %02x", ((unsigned char *)slabp)[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002562 }
2563 printk("\n");
2564 BUG();
2565 }
2566}
2567#else
2568#define kfree_debugcheck(x) do { } while(0)
2569#define cache_free_debugcheck(x,objp,z) (objp)
2570#define check_slabp(x,y) do { } while(0)
2571#endif
2572
Pekka Enberg343e0d72006-02-01 03:05:50 -08002573static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002574{
2575 int batchcount;
2576 struct kmem_list3 *l3;
2577 struct array_cache *ac;
2578
2579 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002580 ac = cpu_cache_get(cachep);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002581 retry:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002582 batchcount = ac->batchcount;
2583 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2584 /* if there was little recent activity on this
2585 * cache, then perform only a partial refill.
2586 * Otherwise we could generate refill bouncing.
2587 */
2588 batchcount = BATCHREFILL_LIMIT;
2589 }
Christoph Lametere498be72005-09-09 13:03:32 -07002590 l3 = cachep->nodelists[numa_node_id()];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002591
Christoph Lametere498be72005-09-09 13:03:32 -07002592 BUG_ON(ac->avail > 0 || !l3);
2593 spin_lock(&l3->list_lock);
2594
Linus Torvalds1da177e2005-04-16 15:20:36 -07002595 if (l3->shared) {
2596 struct array_cache *shared_array = l3->shared;
2597 if (shared_array->avail) {
2598 if (batchcount > shared_array->avail)
2599 batchcount = shared_array->avail;
2600 shared_array->avail -= batchcount;
2601 ac->avail = batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07002602 memcpy(ac->entry,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002603 &(shared_array->entry[shared_array->avail]),
2604 sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002605 shared_array->touched = 1;
2606 goto alloc_done;
2607 }
2608 }
2609 while (batchcount > 0) {
2610 struct list_head *entry;
2611 struct slab *slabp;
2612 /* Get slab alloc is to come from. */
2613 entry = l3->slabs_partial.next;
2614 if (entry == &l3->slabs_partial) {
2615 l3->free_touched = 1;
2616 entry = l3->slabs_free.next;
2617 if (entry == &l3->slabs_free)
2618 goto must_grow;
2619 }
2620
2621 slabp = list_entry(entry, struct slab, list);
2622 check_slabp(cachep, slabp);
2623 check_spinlock_acquired(cachep);
2624 while (slabp->inuse < cachep->num && batchcount--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002625 STATS_INC_ALLOCED(cachep);
2626 STATS_INC_ACTIVE(cachep);
2627 STATS_SET_HIGH(cachep);
2628
Matthew Dobson78d382d2006-02-01 03:05:47 -08002629 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
2630 numa_node_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002631 }
2632 check_slabp(cachep, slabp);
2633
2634 /* move slabp to correct slabp list: */
2635 list_del(&slabp->list);
2636 if (slabp->free == BUFCTL_END)
2637 list_add(&slabp->list, &l3->slabs_full);
2638 else
2639 list_add(&slabp->list, &l3->slabs_partial);
2640 }
2641
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002642 must_grow:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002643 l3->free_objects -= ac->avail;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002644 alloc_done:
Christoph Lametere498be72005-09-09 13:03:32 -07002645 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002646
2647 if (unlikely(!ac->avail)) {
2648 int x;
Christoph Lametere498be72005-09-09 13:03:32 -07002649 x = cache_grow(cachep, flags, numa_node_id());
2650
Linus Torvalds1da177e2005-04-16 15:20:36 -07002651 // cache_grow can reenable interrupts, then ac could change.
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002652 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002653 if (!x && ac->avail == 0) // no objects in sight? abort
2654 return NULL;
2655
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002656 if (!ac->avail) // objects refilled by interrupt?
Linus Torvalds1da177e2005-04-16 15:20:36 -07002657 goto retry;
2658 }
2659 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07002660 return ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002661}
2662
2663static inline void
Pekka Enberg343e0d72006-02-01 03:05:50 -08002664cache_alloc_debugcheck_before(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002665{
2666 might_sleep_if(flags & __GFP_WAIT);
2667#if DEBUG
2668 kmem_flagcheck(cachep, flags);
2669#endif
2670}
2671
2672#if DEBUG
Pekka Enberg343e0d72006-02-01 03:05:50 -08002673static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep, gfp_t flags,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002674 void *objp, void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002675{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002676 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002677 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002678 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002679#ifdef CONFIG_DEBUG_PAGEALLOC
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002680 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002681 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002682 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002683 else
2684 check_poison_obj(cachep, objp);
2685#else
2686 check_poison_obj(cachep, objp);
2687#endif
2688 poison_obj(cachep, objp, POISON_INUSE);
2689 }
2690 if (cachep->flags & SLAB_STORE_USER)
2691 *dbg_userword(cachep, objp) = caller;
2692
2693 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002694 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE
2695 || *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2696 slab_error(cachep,
2697 "double free, or memory outside"
2698 " object was overwritten");
2699 printk(KERN_ERR
2700 "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n",
2701 objp, *dbg_redzone1(cachep, objp),
2702 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002703 }
2704 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2705 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2706 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002707 objp += obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002708 if (cachep->ctor && cachep->flags & SLAB_POISON) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002709 unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002710
2711 if (!(flags & __GFP_WAIT))
2712 ctor_flags |= SLAB_CTOR_ATOMIC;
2713
2714 cachep->ctor(objp, cachep, ctor_flags);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002715 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002716 return objp;
2717}
2718#else
2719#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2720#endif
2721
Pekka Enberg343e0d72006-02-01 03:05:50 -08002722static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002723{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002724 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002725 struct array_cache *ac;
2726
Christoph Lameterdc85da12006-01-18 17:42:36 -08002727#ifdef CONFIG_NUMA
Christoph Lameter86c562a2006-01-18 17:42:37 -08002728 if (unlikely(current->mempolicy && !in_interrupt())) {
Christoph Lameterdc85da12006-01-18 17:42:36 -08002729 int nid = slab_node(current->mempolicy);
2730
2731 if (nid != numa_node_id())
2732 return __cache_alloc_node(cachep, flags, nid);
2733 }
2734#endif
2735
Alok N Kataria5c382302005-09-27 21:45:46 -07002736 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002737 ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002738 if (likely(ac->avail)) {
2739 STATS_INC_ALLOCHIT(cachep);
2740 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07002741 objp = ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002742 } else {
2743 STATS_INC_ALLOCMISS(cachep);
2744 objp = cache_alloc_refill(cachep, flags);
2745 }
Alok N Kataria5c382302005-09-27 21:45:46 -07002746 return objp;
2747}
2748
Pekka Enberg7fd6b142006-02-01 03:05:52 -08002749static __always_inline void *
2750__cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
Alok N Kataria5c382302005-09-27 21:45:46 -07002751{
2752 unsigned long save_flags;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002753 void *objp;
Alok N Kataria5c382302005-09-27 21:45:46 -07002754
2755 cache_alloc_debugcheck_before(cachep, flags);
2756
2757 local_irq_save(save_flags);
2758 objp = ____cache_alloc(cachep, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002759 local_irq_restore(save_flags);
Eric Dumazet34342e82005-09-03 15:55:06 -07002760 objp = cache_alloc_debugcheck_after(cachep, flags, objp,
Pekka Enberg7fd6b142006-02-01 03:05:52 -08002761 caller);
Eric Dumazet34342e82005-09-03 15:55:06 -07002762 prefetchw(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002763 return objp;
2764}
2765
Christoph Lametere498be72005-09-09 13:03:32 -07002766#ifdef CONFIG_NUMA
2767/*
2768 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002769 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002770static void *__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07002771{
2772 struct list_head *entry;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002773 struct slab *slabp;
2774 struct kmem_list3 *l3;
2775 void *obj;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002776 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002777
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002778 l3 = cachep->nodelists[nodeid];
2779 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07002780
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002781 retry:
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08002782 check_irq_off();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002783 spin_lock(&l3->list_lock);
2784 entry = l3->slabs_partial.next;
2785 if (entry == &l3->slabs_partial) {
2786 l3->free_touched = 1;
2787 entry = l3->slabs_free.next;
2788 if (entry == &l3->slabs_free)
2789 goto must_grow;
2790 }
Christoph Lametere498be72005-09-09 13:03:32 -07002791
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002792 slabp = list_entry(entry, struct slab, list);
2793 check_spinlock_acquired_node(cachep, nodeid);
2794 check_slabp(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07002795
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002796 STATS_INC_NODEALLOCS(cachep);
2797 STATS_INC_ACTIVE(cachep);
2798 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002799
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002800 BUG_ON(slabp->inuse == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07002801
Matthew Dobson78d382d2006-02-01 03:05:47 -08002802 obj = slab_get_obj(cachep, slabp, nodeid);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002803 check_slabp(cachep, slabp);
2804 l3->free_objects--;
2805 /* move slabp to correct slabp list: */
2806 list_del(&slabp->list);
Christoph Lametere498be72005-09-09 13:03:32 -07002807
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002808 if (slabp->free == BUFCTL_END) {
2809 list_add(&slabp->list, &l3->slabs_full);
2810 } else {
2811 list_add(&slabp->list, &l3->slabs_partial);
2812 }
Christoph Lametere498be72005-09-09 13:03:32 -07002813
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002814 spin_unlock(&l3->list_lock);
2815 goto done;
Christoph Lametere498be72005-09-09 13:03:32 -07002816
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002817 must_grow:
2818 spin_unlock(&l3->list_lock);
2819 x = cache_grow(cachep, flags, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07002820
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002821 if (!x)
2822 return NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07002823
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002824 goto retry;
2825 done:
2826 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07002827}
2828#endif
2829
2830/*
2831 * Caller needs to acquire correct kmem_list's list_lock
2832 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002833static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002834 int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002835{
2836 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07002837 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002838
2839 for (i = 0; i < nr_objects; i++) {
2840 void *objp = objpp[i];
2841 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002842
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08002843 slabp = virt_to_slab(objp);
Christoph Lameterff694162005-09-22 21:44:02 -07002844 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002845 list_del(&slabp->list);
Christoph Lameterff694162005-09-22 21:44:02 -07002846 check_spinlock_acquired_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002847 check_slabp(cachep, slabp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002848 slab_put_obj(cachep, slabp, objp, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002849 STATS_DEC_ACTIVE(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002850 l3->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002851 check_slabp(cachep, slabp);
2852
2853 /* fixup slab chains */
2854 if (slabp->inuse == 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07002855 if (l3->free_objects > l3->free_limit) {
2856 l3->free_objects -= cachep->num;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002857 slab_destroy(cachep, slabp);
2858 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07002859 list_add(&slabp->list, &l3->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002860 }
2861 } else {
2862 /* Unconditionally move a slab to the end of the
2863 * partial list on free - maximum time for the
2864 * other objects to be freed, too.
2865 */
Christoph Lametere498be72005-09-09 13:03:32 -07002866 list_add_tail(&slabp->list, &l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002867 }
2868 }
2869}
2870
Pekka Enberg343e0d72006-02-01 03:05:50 -08002871static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002872{
2873 int batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07002874 struct kmem_list3 *l3;
Christoph Lameterff694162005-09-22 21:44:02 -07002875 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002876
2877 batchcount = ac->batchcount;
2878#if DEBUG
2879 BUG_ON(!batchcount || batchcount > ac->avail);
2880#endif
2881 check_irq_off();
Christoph Lameterff694162005-09-22 21:44:02 -07002882 l3 = cachep->nodelists[node];
Christoph Lametere498be72005-09-09 13:03:32 -07002883 spin_lock(&l3->list_lock);
2884 if (l3->shared) {
2885 struct array_cache *shared_array = l3->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002886 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002887 if (max) {
2888 if (batchcount > max)
2889 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07002890 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002891 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002892 shared_array->avail += batchcount;
2893 goto free_done;
2894 }
2895 }
2896
Christoph Lameterff694162005-09-22 21:44:02 -07002897 free_block(cachep, ac->entry, batchcount, node);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002898 free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002899#if STATS
2900 {
2901 int i = 0;
2902 struct list_head *p;
2903
Christoph Lametere498be72005-09-09 13:03:32 -07002904 p = l3->slabs_free.next;
2905 while (p != &(l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002906 struct slab *slabp;
2907
2908 slabp = list_entry(p, struct slab, list);
2909 BUG_ON(slabp->inuse);
2910
2911 i++;
2912 p = p->next;
2913 }
2914 STATS_SET_FREEABLE(cachep, i);
2915 }
2916#endif
Christoph Lametere498be72005-09-09 13:03:32 -07002917 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002918 ac->avail -= batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07002919 memmove(ac->entry, &(ac->entry[batchcount]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002920 sizeof(void *) * ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002921}
2922
2923/*
2924 * __cache_free
2925 * Release an obj back to its cache. If the obj has a constructed
2926 * state, it must be in this state _before_ it is released.
2927 *
2928 * Called with disabled ints.
2929 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002930static inline void __cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002931{
Pekka Enberg9a2dba42006-02-01 03:05:49 -08002932 struct array_cache *ac = cpu_cache_get(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002933
2934 check_irq_off();
2935 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
2936
Christoph Lametere498be72005-09-09 13:03:32 -07002937 /* Make sure we are not freeing a object from another
2938 * node to the array cache on this cpu.
2939 */
2940#ifdef CONFIG_NUMA
2941 {
2942 struct slab *slabp;
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08002943 slabp = virt_to_slab(objp);
Christoph Lametere498be72005-09-09 13:03:32 -07002944 if (unlikely(slabp->nodeid != numa_node_id())) {
2945 struct array_cache *alien = NULL;
2946 int nodeid = slabp->nodeid;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002947 struct kmem_list3 *l3 =
2948 cachep->nodelists[numa_node_id()];
Christoph Lametere498be72005-09-09 13:03:32 -07002949
2950 STATS_INC_NODEFREES(cachep);
2951 if (l3->alien && l3->alien[nodeid]) {
2952 alien = l3->alien[nodeid];
2953 spin_lock(&alien->lock);
2954 if (unlikely(alien->avail == alien->limit))
2955 __drain_alien_cache(cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002956 alien, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07002957 alien->entry[alien->avail++] = objp;
2958 spin_unlock(&alien->lock);
2959 } else {
2960 spin_lock(&(cachep->nodelists[nodeid])->
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002961 list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -07002962 free_block(cachep, &objp, 1, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07002963 spin_unlock(&(cachep->nodelists[nodeid])->
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002964 list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002965 }
2966 return;
2967 }
2968 }
2969#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002970 if (likely(ac->avail < ac->limit)) {
2971 STATS_INC_FREEHIT(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002972 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002973 return;
2974 } else {
2975 STATS_INC_FREEMISS(cachep);
2976 cache_flusharray(cachep, ac);
Christoph Lametere498be72005-09-09 13:03:32 -07002977 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002978 }
2979}
2980
2981/**
2982 * kmem_cache_alloc - Allocate an object
2983 * @cachep: The cache to allocate from.
2984 * @flags: See kmalloc().
2985 *
2986 * Allocate an object from this cache. The flags are only relevant
2987 * if the cache has no available objects.
2988 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08002989void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002990{
Pekka Enberg7fd6b142006-02-01 03:05:52 -08002991 return __cache_alloc(cachep, flags, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002992}
2993EXPORT_SYMBOL(kmem_cache_alloc);
2994
2995/**
2996 * kmem_ptr_validate - check if an untrusted pointer might
2997 * be a slab entry.
2998 * @cachep: the cache we're checking against
2999 * @ptr: pointer to validate
3000 *
3001 * This verifies that the untrusted pointer looks sane:
3002 * it is _not_ a guarantee that the pointer is actually
3003 * part of the slab cache in question, but it at least
3004 * validates that the pointer can be dereferenced and
3005 * looks half-way sane.
3006 *
3007 * Currently only used for dentry validation.
3008 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003009int fastcall kmem_ptr_validate(struct kmem_cache *cachep, void *ptr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003010{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003011 unsigned long addr = (unsigned long)ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003012 unsigned long min_addr = PAGE_OFFSET;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003013 unsigned long align_mask = BYTES_PER_WORD - 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003014 unsigned long size = cachep->buffer_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003015 struct page *page;
3016
3017 if (unlikely(addr < min_addr))
3018 goto out;
3019 if (unlikely(addr > (unsigned long)high_memory - size))
3020 goto out;
3021 if (unlikely(addr & align_mask))
3022 goto out;
3023 if (unlikely(!kern_addr_valid(addr)))
3024 goto out;
3025 if (unlikely(!kern_addr_valid(addr + size - 1)))
3026 goto out;
3027 page = virt_to_page(ptr);
3028 if (unlikely(!PageSlab(page)))
3029 goto out;
Pekka Enberg065d41c2005-11-13 16:06:46 -08003030 if (unlikely(page_get_cache(page) != cachep))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003031 goto out;
3032 return 1;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003033 out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003034 return 0;
3035}
3036
3037#ifdef CONFIG_NUMA
3038/**
3039 * kmem_cache_alloc_node - Allocate an object on the specified node
3040 * @cachep: The cache to allocate from.
3041 * @flags: See kmalloc().
3042 * @nodeid: node number of the target node.
3043 *
3044 * Identical to kmem_cache_alloc, except that this function is slow
3045 * and can sleep. And it will allocate memory on the given node, which
3046 * can improve the performance for cpu bound structures.
Christoph Lametere498be72005-09-09 13:03:32 -07003047 * New and improved: it will now make sure that the object gets
3048 * put on the correct node list so that there is no false sharing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003049 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003050void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003051{
Christoph Lametere498be72005-09-09 13:03:32 -07003052 unsigned long save_flags;
3053 void *ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003054
Christoph Lametere498be72005-09-09 13:03:32 -07003055 cache_alloc_debugcheck_before(cachep, flags);
3056 local_irq_save(save_flags);
Christoph Lameter18f820f2006-02-01 03:05:43 -08003057
3058 if (nodeid == -1 || nodeid == numa_node_id() ||
3059 !cachep->nodelists[nodeid])
Alok N Kataria5c382302005-09-27 21:45:46 -07003060 ptr = ____cache_alloc(cachep, flags);
3061 else
3062 ptr = __cache_alloc_node(cachep, flags, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07003063 local_irq_restore(save_flags);
Christoph Lameter18f820f2006-02-01 03:05:43 -08003064
3065 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr,
3066 __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003067
Christoph Lametere498be72005-09-09 13:03:32 -07003068 return ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003069}
3070EXPORT_SYMBOL(kmem_cache_alloc_node);
3071
Al Virodd0fc662005-10-07 07:46:04 +01003072void *kmalloc_node(size_t size, gfp_t flags, int node)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003073{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003074 struct kmem_cache *cachep;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003075
3076 cachep = kmem_find_general_cachep(size, flags);
3077 if (unlikely(cachep == NULL))
3078 return NULL;
3079 return kmem_cache_alloc_node(cachep, flags, node);
3080}
3081EXPORT_SYMBOL(kmalloc_node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003082#endif
3083
3084/**
3085 * kmalloc - allocate memory
3086 * @size: how many bytes of memory are required.
3087 * @flags: the type of memory to allocate.
3088 *
3089 * kmalloc is the normal method of allocating memory
3090 * in the kernel.
3091 *
3092 * The @flags argument may be one of:
3093 *
3094 * %GFP_USER - Allocate memory on behalf of user. May sleep.
3095 *
3096 * %GFP_KERNEL - Allocate normal kernel ram. May sleep.
3097 *
3098 * %GFP_ATOMIC - Allocation will not sleep. Use inside interrupt handlers.
3099 *
3100 * Additionally, the %GFP_DMA flag may be set to indicate the memory
3101 * must be suitable for DMA. This can mean different things on different
3102 * platforms. For example, on i386, it means that the memory must come
3103 * from the first 16MB.
3104 */
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003105static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3106 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003107{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003108 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003109
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003110 /* If you want to save a few bytes .text space: replace
3111 * __ with kmem_.
3112 * Then kmalloc uses the uninlined functions instead of the inline
3113 * functions.
3114 */
3115 cachep = __find_general_cachep(size, flags);
Andrew Mortondbdb9042005-09-23 13:24:10 -07003116 if (unlikely(cachep == NULL))
3117 return NULL;
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003118 return __cache_alloc(cachep, flags, caller);
3119}
3120
3121#ifndef CONFIG_DEBUG_SLAB
3122
3123void *__kmalloc(size_t size, gfp_t flags)
3124{
3125 return __do_kmalloc(size, flags, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003126}
3127EXPORT_SYMBOL(__kmalloc);
3128
Pekka Enberg7fd6b142006-02-01 03:05:52 -08003129#else
3130
3131void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller)
3132{
3133 return __do_kmalloc(size, flags, caller);
3134}
3135EXPORT_SYMBOL(__kmalloc_track_caller);
3136
3137#endif
3138
Linus Torvalds1da177e2005-04-16 15:20:36 -07003139#ifdef CONFIG_SMP
3140/**
3141 * __alloc_percpu - allocate one copy of the object for every present
3142 * cpu in the system, zeroing them.
3143 * Objects should be dereferenced using the per_cpu_ptr macro only.
3144 *
3145 * @size: how many bytes of memory are required.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003146 */
Pekka Enbergf9f75002006-01-08 01:00:33 -08003147void *__alloc_percpu(size_t size)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003148{
3149 int i;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003150 struct percpu_data *pdata = kmalloc(sizeof(*pdata), GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003151
3152 if (!pdata)
3153 return NULL;
3154
Christoph Lametere498be72005-09-09 13:03:32 -07003155 /*
3156 * Cannot use for_each_online_cpu since a cpu may come online
3157 * and we have no way of figuring out how to fix the array
3158 * that we have allocated then....
3159 */
3160 for_each_cpu(i) {
3161 int node = cpu_to_node(i);
3162
3163 if (node_online(node))
3164 pdata->ptrs[i] = kmalloc_node(size, GFP_KERNEL, node);
3165 else
3166 pdata->ptrs[i] = kmalloc(size, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003167
3168 if (!pdata->ptrs[i])
3169 goto unwind_oom;
3170 memset(pdata->ptrs[i], 0, size);
3171 }
3172
3173 /* Catch derefs w/o wrappers */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003174 return (void *)(~(unsigned long)pdata);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003175
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003176 unwind_oom:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003177 while (--i >= 0) {
3178 if (!cpu_possible(i))
3179 continue;
3180 kfree(pdata->ptrs[i]);
3181 }
3182 kfree(pdata);
3183 return NULL;
3184}
3185EXPORT_SYMBOL(__alloc_percpu);
3186#endif
3187
3188/**
3189 * kmem_cache_free - Deallocate an object
3190 * @cachep: The cache the allocation was from.
3191 * @objp: The previously allocated object.
3192 *
3193 * Free an object which was previously allocated from this
3194 * cache.
3195 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003196void kmem_cache_free(struct kmem_cache *cachep, void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003197{
3198 unsigned long flags;
3199
3200 local_irq_save(flags);
3201 __cache_free(cachep, objp);
3202 local_irq_restore(flags);
3203}
3204EXPORT_SYMBOL(kmem_cache_free);
3205
3206/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003207 * kfree - free previously allocated memory
3208 * @objp: pointer returned by kmalloc.
3209 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003210 * If @objp is NULL, no operation is performed.
3211 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003212 * Don't free memory not originally allocated by kmalloc()
3213 * or you will run into trouble.
3214 */
3215void kfree(const void *objp)
3216{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003217 struct kmem_cache *c;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003218 unsigned long flags;
3219
3220 if (unlikely(!objp))
3221 return;
3222 local_irq_save(flags);
3223 kfree_debugcheck(objp);
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08003224 c = virt_to_cache(objp);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003225 mutex_debug_check_no_locks_freed(objp, obj_size(c));
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003226 __cache_free(c, (void *)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003227 local_irq_restore(flags);
3228}
3229EXPORT_SYMBOL(kfree);
3230
3231#ifdef CONFIG_SMP
3232/**
3233 * free_percpu - free previously allocated percpu memory
3234 * @objp: pointer returned by alloc_percpu.
3235 *
3236 * Don't free memory not originally allocated by alloc_percpu()
3237 * The complemented objp is to check for that.
3238 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003239void free_percpu(const void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003240{
3241 int i;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003242 struct percpu_data *p = (struct percpu_data *)(~(unsigned long)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003243
Christoph Lametere498be72005-09-09 13:03:32 -07003244 /*
3245 * We allocate for all cpus so we cannot use for online cpu here.
3246 */
3247 for_each_cpu(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003248 kfree(p->ptrs[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003249 kfree(p);
3250}
3251EXPORT_SYMBOL(free_percpu);
3252#endif
3253
Pekka Enberg343e0d72006-02-01 03:05:50 -08003254unsigned int kmem_cache_size(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003255{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003256 return obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003257}
3258EXPORT_SYMBOL(kmem_cache_size);
3259
Pekka Enberg343e0d72006-02-01 03:05:50 -08003260const char *kmem_cache_name(struct kmem_cache *cachep)
Arnaldo Carvalho de Melo19449722005-06-18 22:46:19 -07003261{
3262 return cachep->name;
3263}
3264EXPORT_SYMBOL_GPL(kmem_cache_name);
3265
Christoph Lametere498be72005-09-09 13:03:32 -07003266/*
3267 * This initializes kmem_list3 for all nodes.
3268 */
Pekka Enberg343e0d72006-02-01 03:05:50 -08003269static int alloc_kmemlist(struct kmem_cache *cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07003270{
3271 int node;
3272 struct kmem_list3 *l3;
3273 int err = 0;
3274
3275 for_each_online_node(node) {
3276 struct array_cache *nc = NULL, *new;
3277 struct array_cache **new_alien = NULL;
3278#ifdef CONFIG_NUMA
3279 if (!(new_alien = alloc_alien_cache(node, cachep->limit)))
3280 goto fail;
3281#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003282 if (!(new = alloc_arraycache(node, (cachep->shared *
3283 cachep->batchcount),
3284 0xbaadf00d)))
Christoph Lametere498be72005-09-09 13:03:32 -07003285 goto fail;
3286 if ((l3 = cachep->nodelists[node])) {
3287
3288 spin_lock_irq(&l3->list_lock);
3289
3290 if ((nc = cachep->nodelists[node]->shared))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003291 free_block(cachep, nc->entry, nc->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07003292
3293 l3->shared = new;
3294 if (!cachep->nodelists[node]->alien) {
3295 l3->alien = new_alien;
3296 new_alien = NULL;
3297 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003298 l3->free_limit = (1 + nr_cpus_node(node)) *
3299 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003300 spin_unlock_irq(&l3->list_lock);
3301 kfree(nc);
3302 free_alien_cache(new_alien);
3303 continue;
3304 }
3305 if (!(l3 = kmalloc_node(sizeof(struct kmem_list3),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003306 GFP_KERNEL, node)))
Christoph Lametere498be72005-09-09 13:03:32 -07003307 goto fail;
3308
3309 kmem_list3_init(l3);
3310 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003311 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -07003312 l3->shared = new;
3313 l3->alien = new_alien;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003314 l3->free_limit = (1 + nr_cpus_node(node)) *
3315 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003316 cachep->nodelists[node] = l3;
3317 }
3318 return err;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003319 fail:
Christoph Lametere498be72005-09-09 13:03:32 -07003320 err = -ENOMEM;
3321 return err;
3322}
3323
Linus Torvalds1da177e2005-04-16 15:20:36 -07003324struct ccupdate_struct {
Pekka Enberg343e0d72006-02-01 03:05:50 -08003325 struct kmem_cache *cachep;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003326 struct array_cache *new[NR_CPUS];
3327};
3328
3329static void do_ccupdate_local(void *info)
3330{
3331 struct ccupdate_struct *new = (struct ccupdate_struct *)info;
3332 struct array_cache *old;
3333
3334 check_irq_off();
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003335 old = cpu_cache_get(new->cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003336
Linus Torvalds1da177e2005-04-16 15:20:36 -07003337 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3338 new->new[smp_processor_id()] = old;
3339}
3340
Pekka Enberg343e0d72006-02-01 03:05:50 -08003341static int do_tune_cpucache(struct kmem_cache *cachep, int limit, int batchcount,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003342 int shared)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003343{
3344 struct ccupdate_struct new;
Christoph Lametere498be72005-09-09 13:03:32 -07003345 int i, err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003346
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003347 memset(&new.new, 0, sizeof(new.new));
Christoph Lametere498be72005-09-09 13:03:32 -07003348 for_each_online_cpu(i) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003349 new.new[i] =
3350 alloc_arraycache(cpu_to_node(i), limit, batchcount);
Christoph Lametere498be72005-09-09 13:03:32 -07003351 if (!new.new[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003352 for (i--; i >= 0; i--)
3353 kfree(new.new[i]);
Christoph Lametere498be72005-09-09 13:03:32 -07003354 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003355 }
3356 }
3357 new.cachep = cachep;
3358
3359 smp_call_function_all_cpus(do_ccupdate_local, (void *)&new);
Christoph Lametere498be72005-09-09 13:03:32 -07003360
Linus Torvalds1da177e2005-04-16 15:20:36 -07003361 check_irq_on();
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003362 spin_lock(&cachep->spinlock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003363 cachep->batchcount = batchcount;
3364 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07003365 cachep->shared = shared;
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003366 spin_unlock(&cachep->spinlock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003367
Christoph Lametere498be72005-09-09 13:03:32 -07003368 for_each_online_cpu(i) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003369 struct array_cache *ccold = new.new[i];
3370 if (!ccold)
3371 continue;
Christoph Lametere498be72005-09-09 13:03:32 -07003372 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -07003373 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
Christoph Lametere498be72005-09-09 13:03:32 -07003374 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003375 kfree(ccold);
3376 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003377
Christoph Lametere498be72005-09-09 13:03:32 -07003378 err = alloc_kmemlist(cachep);
3379 if (err) {
3380 printk(KERN_ERR "alloc_kmemlist failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003381 cachep->name, -err);
Christoph Lametere498be72005-09-09 13:03:32 -07003382 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003383 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003384 return 0;
3385}
3386
Pekka Enberg343e0d72006-02-01 03:05:50 -08003387static void enable_cpucache(struct kmem_cache *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003388{
3389 int err;
3390 int limit, shared;
3391
3392 /* The head array serves three purposes:
3393 * - create a LIFO ordering, i.e. return objects that are cache-warm
3394 * - reduce the number of spinlock operations.
3395 * - reduce the number of linked list operations on the slab and
3396 * bufctl chains: array operations are cheaper.
3397 * The numbers are guessed, we should auto-tune as described by
3398 * Bonwick.
3399 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003400 if (cachep->buffer_size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003401 limit = 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003402 else if (cachep->buffer_size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003403 limit = 8;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003404 else if (cachep->buffer_size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003405 limit = 24;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003406 else if (cachep->buffer_size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003407 limit = 54;
3408 else
3409 limit = 120;
3410
3411 /* Cpu bound tasks (e.g. network routing) can exhibit cpu bound
3412 * allocation behaviour: Most allocs on one cpu, most free operations
3413 * on another cpu. For these cases, an efficient object passing between
3414 * cpus is necessary. This is provided by a shared array. The array
3415 * replaces Bonwick's magazine layer.
3416 * On uniprocessor, it's functionally equivalent (but less efficient)
3417 * to a larger limit. Thus disabled by default.
3418 */
3419 shared = 0;
3420#ifdef CONFIG_SMP
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003421 if (cachep->buffer_size <= PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003422 shared = 8;
3423#endif
3424
3425#if DEBUG
3426 /* With debugging enabled, large batchcount lead to excessively
3427 * long periods with disabled local interrupts. Limit the
3428 * batchcount
3429 */
3430 if (limit > 32)
3431 limit = 32;
3432#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003433 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003434 if (err)
3435 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003436 cachep->name, -err);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003437}
3438
Pekka Enberg343e0d72006-02-01 03:05:50 -08003439static void drain_array_locked(struct kmem_cache *cachep, struct array_cache *ac,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003440 int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003441{
3442 int tofree;
3443
Christoph Lametere498be72005-09-09 13:03:32 -07003444 check_spinlock_acquired_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003445 if (ac->touched && !force) {
3446 ac->touched = 0;
3447 } else if (ac->avail) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003448 tofree = force ? ac->avail : (ac->limit + 4) / 5;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003449 if (tofree > ac->avail) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003450 tofree = (ac->avail + 1) / 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003451 }
Christoph Lameterff694162005-09-22 21:44:02 -07003452 free_block(cachep, ac->entry, tofree, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003453 ac->avail -= tofree;
Christoph Lametere498be72005-09-09 13:03:32 -07003454 memmove(ac->entry, &(ac->entry[tofree]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003455 sizeof(void *) * ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003456 }
3457}
3458
3459/**
3460 * cache_reap - Reclaim memory from caches.
Randy Dunlap1e5d5332005-11-07 01:01:06 -08003461 * @unused: unused parameter
Linus Torvalds1da177e2005-04-16 15:20:36 -07003462 *
3463 * Called from workqueue/eventd every few seconds.
3464 * Purpose:
3465 * - clear the per-cpu caches for this CPU.
3466 * - return freeable pages to the main free memory pool.
3467 *
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003468 * If we cannot acquire the cache chain mutex then just give up - we'll
Linus Torvalds1da177e2005-04-16 15:20:36 -07003469 * try again on the next iteration.
3470 */
3471static void cache_reap(void *unused)
3472{
3473 struct list_head *walk;
Christoph Lametere498be72005-09-09 13:03:32 -07003474 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003475
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003476 if (!mutex_trylock(&cache_chain_mutex)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003477 /* Give up. Setup the next iteration. */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003478 schedule_delayed_work(&__get_cpu_var(reap_work),
3479 REAPTIMEOUT_CPUC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003480 return;
3481 }
3482
3483 list_for_each(walk, &cache_chain) {
Pekka Enberg343e0d72006-02-01 03:05:50 -08003484 struct kmem_cache *searchp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003485 struct list_head *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003486 int tofree;
3487 struct slab *slabp;
3488
Pekka Enberg343e0d72006-02-01 03:05:50 -08003489 searchp = list_entry(walk, struct kmem_cache, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003490
3491 if (searchp->flags & SLAB_NO_REAP)
3492 goto next;
3493
3494 check_irq_on();
3495
Christoph Lametere498be72005-09-09 13:03:32 -07003496 l3 = searchp->nodelists[numa_node_id()];
3497 if (l3->alien)
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08003498 drain_alien_cache(searchp, l3->alien);
Christoph Lametere498be72005-09-09 13:03:32 -07003499 spin_lock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003500
Pekka Enberg9a2dba42006-02-01 03:05:49 -08003501 drain_array_locked(searchp, cpu_cache_get(searchp), 0,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003502 numa_node_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -07003503
Christoph Lametere498be72005-09-09 13:03:32 -07003504 if (time_after(l3->next_reap, jiffies))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003505 goto next_unlock;
3506
Christoph Lametere498be72005-09-09 13:03:32 -07003507 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003508
Christoph Lametere498be72005-09-09 13:03:32 -07003509 if (l3->shared)
3510 drain_array_locked(searchp, l3->shared, 0,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003511 numa_node_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -07003512
Christoph Lametere498be72005-09-09 13:03:32 -07003513 if (l3->free_touched) {
3514 l3->free_touched = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003515 goto next_unlock;
3516 }
3517
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003518 tofree =
3519 (l3->free_limit + 5 * searchp->num -
3520 1) / (5 * searchp->num);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003521 do {
Christoph Lametere498be72005-09-09 13:03:32 -07003522 p = l3->slabs_free.next;
3523 if (p == &(l3->slabs_free))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003524 break;
3525
3526 slabp = list_entry(p, struct slab, list);
3527 BUG_ON(slabp->inuse);
3528 list_del(&slabp->list);
3529 STATS_INC_REAPED(searchp);
3530
3531 /* Safe to drop the lock. The slab is no longer
3532 * linked to the cache.
3533 * searchp cannot disappear, we hold
3534 * cache_chain_lock
3535 */
Christoph Lametere498be72005-09-09 13:03:32 -07003536 l3->free_objects -= searchp->num;
3537 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003538 slab_destroy(searchp, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07003539 spin_lock_irq(&l3->list_lock);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003540 } while (--tofree > 0);
3541 next_unlock:
Christoph Lametere498be72005-09-09 13:03:32 -07003542 spin_unlock_irq(&l3->list_lock);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003543 next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003544 cond_resched();
3545 }
3546 check_irq_on();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003547 mutex_unlock(&cache_chain_mutex);
Christoph Lameter4ae7c032005-06-21 17:14:57 -07003548 drain_remote_pages();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003549 /* Setup the next iteration */
Manfred Spraulcd61ef62005-11-07 00:58:02 -08003550 schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003551}
3552
3553#ifdef CONFIG_PROC_FS
3554
Pekka Enberg85289f92006-01-08 01:00:36 -08003555static void print_slabinfo_header(struct seq_file *m)
3556{
3557 /*
3558 * Output format version, so at least we can change it
3559 * without _too_ many complaints.
3560 */
3561#if STATS
3562 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
3563#else
3564 seq_puts(m, "slabinfo - version: 2.1\n");
3565#endif
3566 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
3567 "<objperslab> <pagesperslab>");
3568 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
3569 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
3570#if STATS
3571 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
3572 "<error> <maxfreeable> <nodeallocs> <remotefrees>");
3573 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
3574#endif
3575 seq_putc(m, '\n');
3576}
3577
Linus Torvalds1da177e2005-04-16 15:20:36 -07003578static void *s_start(struct seq_file *m, loff_t *pos)
3579{
3580 loff_t n = *pos;
3581 struct list_head *p;
3582
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003583 mutex_lock(&cache_chain_mutex);
Pekka Enberg85289f92006-01-08 01:00:36 -08003584 if (!n)
3585 print_slabinfo_header(m);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003586 p = cache_chain.next;
3587 while (n--) {
3588 p = p->next;
3589 if (p == &cache_chain)
3590 return NULL;
3591 }
Pekka Enberg343e0d72006-02-01 03:05:50 -08003592 return list_entry(p, struct kmem_cache, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003593}
3594
3595static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3596{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003597 struct kmem_cache *cachep = p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003598 ++*pos;
3599 return cachep->next.next == &cache_chain ? NULL
Pekka Enberg343e0d72006-02-01 03:05:50 -08003600 : list_entry(cachep->next.next, struct kmem_cache, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003601}
3602
3603static void s_stop(struct seq_file *m, void *p)
3604{
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003605 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003606}
3607
3608static int s_show(struct seq_file *m, void *p)
3609{
Pekka Enberg343e0d72006-02-01 03:05:50 -08003610 struct kmem_cache *cachep = p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003611 struct list_head *q;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003612 struct slab *slabp;
3613 unsigned long active_objs;
3614 unsigned long num_objs;
3615 unsigned long active_slabs = 0;
3616 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07003617 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003618 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003619 int node;
3620 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003621
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003622 spin_lock(&cachep->spinlock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003623 active_objs = 0;
3624 num_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07003625 for_each_online_node(node) {
3626 l3 = cachep->nodelists[node];
3627 if (!l3)
3628 continue;
3629
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003630 check_irq_on();
3631 spin_lock_irq(&l3->list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07003632
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003633 list_for_each(q, &l3->slabs_full) {
Christoph Lametere498be72005-09-09 13:03:32 -07003634 slabp = list_entry(q, struct slab, list);
3635 if (slabp->inuse != cachep->num && !error)
3636 error = "slabs_full accounting error";
3637 active_objs += cachep->num;
3638 active_slabs++;
3639 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003640 list_for_each(q, &l3->slabs_partial) {
Christoph Lametere498be72005-09-09 13:03:32 -07003641 slabp = list_entry(q, struct slab, list);
3642 if (slabp->inuse == cachep->num && !error)
3643 error = "slabs_partial inuse accounting error";
3644 if (!slabp->inuse && !error)
3645 error = "slabs_partial/inuse accounting error";
3646 active_objs += slabp->inuse;
3647 active_slabs++;
3648 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003649 list_for_each(q, &l3->slabs_free) {
Christoph Lametere498be72005-09-09 13:03:32 -07003650 slabp = list_entry(q, struct slab, list);
3651 if (slabp->inuse && !error)
3652 error = "slabs_free/inuse accounting error";
3653 num_slabs++;
3654 }
3655 free_objects += l3->free_objects;
Ravikiran G Thirumalai4484ebf2006-02-04 23:27:59 -08003656 if (l3->shared)
3657 shared_avail += l3->shared->avail;
Christoph Lametere498be72005-09-09 13:03:32 -07003658
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003659 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003660 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003661 num_slabs += active_slabs;
3662 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003663 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003664 error = "free_objects accounting error";
3665
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003666 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003667 if (error)
3668 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
3669
3670 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003671 name, active_objs, num_objs, cachep->buffer_size,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003672 cachep->num, (1 << cachep->gfporder));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003673 seq_printf(m, " : tunables %4u %4u %4u",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003674 cachep->limit, cachep->batchcount, cachep->shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003675 seq_printf(m, " : slabdata %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003676 active_slabs, num_slabs, shared_avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003677#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003678 { /* list3 stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003679 unsigned long high = cachep->high_mark;
3680 unsigned long allocs = cachep->num_allocations;
3681 unsigned long grown = cachep->grown;
3682 unsigned long reaped = cachep->reaped;
3683 unsigned long errors = cachep->errors;
3684 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003685 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07003686 unsigned long node_frees = cachep->node_frees;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003687
Christoph Lametere498be72005-09-09 13:03:32 -07003688 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003689 %4lu %4lu %4lu %4lu", allocs, high, grown, reaped, errors, max_freeable, node_allocs, node_frees);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003690 }
3691 /* cpu stats */
3692 {
3693 unsigned long allochit = atomic_read(&cachep->allochit);
3694 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
3695 unsigned long freehit = atomic_read(&cachep->freehit);
3696 unsigned long freemiss = atomic_read(&cachep->freemiss);
3697
3698 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003699 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003700 }
3701#endif
3702 seq_putc(m, '\n');
Ravikiran G Thirumalaica3b9b92006-02-04 23:27:58 -08003703 spin_unlock(&cachep->spinlock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003704 return 0;
3705}
3706
3707/*
3708 * slabinfo_op - iterator that generates /proc/slabinfo
3709 *
3710 * Output layout:
3711 * cache-name
3712 * num-active-objs
3713 * total-objs
3714 * object size
3715 * num-active-slabs
3716 * total-slabs
3717 * num-pages-per-slab
3718 * + further values on SMP and with statistics enabled
3719 */
3720
3721struct seq_operations slabinfo_op = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003722 .start = s_start,
3723 .next = s_next,
3724 .stop = s_stop,
3725 .show = s_show,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003726};
3727
3728#define MAX_SLABINFO_WRITE 128
3729/**
3730 * slabinfo_write - Tuning for the slab allocator
3731 * @file: unused
3732 * @buffer: user buffer
3733 * @count: data length
3734 * @ppos: unused
3735 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003736ssize_t slabinfo_write(struct file *file, const char __user * buffer,
3737 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003738{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003739 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003740 int limit, batchcount, shared, res;
3741 struct list_head *p;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003742
Linus Torvalds1da177e2005-04-16 15:20:36 -07003743 if (count > MAX_SLABINFO_WRITE)
3744 return -EINVAL;
3745 if (copy_from_user(&kbuf, buffer, count))
3746 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003747 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07003748
3749 tmp = strchr(kbuf, ' ');
3750 if (!tmp)
3751 return -EINVAL;
3752 *tmp = '\0';
3753 tmp++;
3754 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
3755 return -EINVAL;
3756
3757 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003758 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003759 res = -EINVAL;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003760 list_for_each(p, &cache_chain) {
Pekka Enberg343e0d72006-02-01 03:05:50 -08003761 struct kmem_cache *cachep = list_entry(p, struct kmem_cache,
3762 next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003763
3764 if (!strcmp(cachep->name, kbuf)) {
3765 if (limit < 1 ||
3766 batchcount < 1 ||
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003767 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07003768 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003769 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07003770 res = do_tune_cpucache(cachep, limit,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003771 batchcount, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003772 }
3773 break;
3774 }
3775 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003776 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003777 if (res >= 0)
3778 res = count;
3779 return res;
3780}
3781#endif
3782
Manfred Spraul00e145b2005-09-03 15:55:07 -07003783/**
3784 * ksize - get the actual amount of memory allocated for a given object
3785 * @objp: Pointer to the object
3786 *
3787 * kmalloc may internally round up allocations and return more memory
3788 * than requested. ksize() can be used to determine the actual amount of
3789 * memory allocated. The caller may use this additional memory, even though
3790 * a smaller amount of memory was initially specified with the kmalloc call.
3791 * The caller must guarantee that objp points to a valid object previously
3792 * allocated with either kmalloc() or kmem_cache_alloc(). The object
3793 * must not be freed during the duration of the call.
3794 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003795unsigned int ksize(const void *objp)
3796{
Manfred Spraul00e145b2005-09-03 15:55:07 -07003797 if (unlikely(objp == NULL))
3798 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003799
Pekka Enberg6ed5eb22006-02-01 03:05:49 -08003800 return obj_size(virt_to_cache(objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003801}