blob: ba288b3877d1c27b198488c867625e8ad78c19a8 [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.
58 * Several members in kmem_cache_t and struct slab never change, they
59 * 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;
247 kmem_cache_t *cachep;
248 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;
297 spinlock_t list_lock;
298 struct array_cache *shared; /* shared per node */
299 struct array_cache **alien; /* on other nodes */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700300};
301
Christoph Lametere498be72005-09-09 13:03:32 -0700302/*
303 * Need this for bootstrapping a per node allocator.
304 */
305#define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
306struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
307#define CACHE_CACHE 0
308#define SIZE_AC 1
309#define SIZE_L3 (1 + MAX_NUMNODES)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700310
Christoph Lametere498be72005-09-09 13:03:32 -0700311/*
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700312 * This function must be completely optimized away if
Christoph Lametere498be72005-09-09 13:03:32 -0700313 * a constant is passed to it. Mostly the same as
314 * what is in linux/slab.h except it returns an
315 * index.
316 */
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700317static __always_inline int index_of(const size_t size)
Christoph Lametere498be72005-09-09 13:03:32 -0700318{
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800319 extern void __bad_size(void);
320
Christoph Lametere498be72005-09-09 13:03:32 -0700321 if (__builtin_constant_p(size)) {
322 int i = 0;
323
324#define CACHE(x) \
325 if (size <=x) \
326 return i; \
327 else \
328 i++;
329#include "linux/kmalloc_sizes.h"
330#undef CACHE
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800331 __bad_size();
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700332 } else
Steven Rostedt5ec8a842006-02-01 03:05:44 -0800333 __bad_size();
Christoph Lametere498be72005-09-09 13:03:32 -0700334 return 0;
335}
336
337#define INDEX_AC index_of(sizeof(struct arraycache_init))
338#define INDEX_L3 index_of(sizeof(struct kmem_list3))
339
Pekka Enberg5295a742006-02-01 03:05:48 -0800340static void kmem_list3_init(struct kmem_list3 *parent)
Christoph Lametere498be72005-09-09 13:03:32 -0700341{
342 INIT_LIST_HEAD(&parent->slabs_full);
343 INIT_LIST_HEAD(&parent->slabs_partial);
344 INIT_LIST_HEAD(&parent->slabs_free);
345 parent->shared = NULL;
346 parent->alien = NULL;
347 spin_lock_init(&parent->list_lock);
348 parent->free_objects = 0;
349 parent->free_touched = 0;
350}
351
352#define MAKE_LIST(cachep, listp, slab, nodeid) \
353 do { \
354 INIT_LIST_HEAD(listp); \
355 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
356 } while (0)
357
358#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
359 do { \
360 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
361 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
362 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
363 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700364
365/*
366 * kmem_cache_t
367 *
368 * manages a cache.
369 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800370
Pekka J Enberg2109a2d2005-11-07 00:58:01 -0800371struct kmem_cache {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700372/* 1) per-cpu data, touched during every alloc/free */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800373 struct array_cache *array[NR_CPUS];
374 unsigned int batchcount;
375 unsigned int limit;
376 unsigned int shared;
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800377 unsigned int buffer_size;
Christoph Lametere498be72005-09-09 13:03:32 -0700378/* 2) touched by every alloc & free from the backend */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800379 struct kmem_list3 *nodelists[MAX_NUMNODES];
380 unsigned int flags; /* constant flags */
381 unsigned int num; /* # of objs per slab */
382 spinlock_t spinlock;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700383
384/* 3) cache_grow/shrink */
385 /* order of pgs per slab (2^n) */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800386 unsigned int gfporder;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700387
388 /* force GFP flags, e.g. GFP_DMA */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800389 gfp_t gfpflags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700390
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800391 size_t colour; /* cache colouring range */
392 unsigned int colour_off; /* colour offset */
393 unsigned int colour_next; /* cache colouring */
394 kmem_cache_t *slabp_cache;
395 unsigned int slab_size;
396 unsigned int dflags; /* dynamic flags */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700397
398 /* constructor func */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800399 void (*ctor) (void *, kmem_cache_t *, unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700400
401 /* de-constructor func */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800402 void (*dtor) (void *, kmem_cache_t *, unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700403
404/* 4) cache creation/removal */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800405 const char *name;
406 struct list_head next;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700407
408/* 5) statistics */
409#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800410 unsigned long num_active;
411 unsigned long num_allocations;
412 unsigned long high_mark;
413 unsigned long grown;
414 unsigned long reaped;
415 unsigned long errors;
416 unsigned long max_freeable;
417 unsigned long node_allocs;
418 unsigned long node_frees;
419 atomic_t allochit;
420 atomic_t allocmiss;
421 atomic_t freehit;
422 atomic_t freemiss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700423#endif
424#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800425 /*
426 * If debugging is enabled, then the allocator can add additional
427 * fields and/or padding to every object. buffer_size contains the total
428 * object size including these internal fields, the following two
429 * variables contain the offset to the user object and its size.
430 */
431 int obj_offset;
432 int obj_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700433#endif
434};
435
436#define CFLGS_OFF_SLAB (0x80000000UL)
437#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
438
439#define BATCHREFILL_LIMIT 16
440/* Optimization question: fewer reaps means less
441 * probability for unnessary cpucache drain/refill cycles.
442 *
Adrian Bunkdc6f3f22005-11-08 16:44:08 +0100443 * OTOH the cpuarrays can contain lots of objects,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700444 * which could lock up otherwise freeable slabs.
445 */
446#define REAPTIMEOUT_CPUC (2*HZ)
447#define REAPTIMEOUT_LIST3 (4*HZ)
448
449#if STATS
450#define STATS_INC_ACTIVE(x) ((x)->num_active++)
451#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
452#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
453#define STATS_INC_GROWN(x) ((x)->grown++)
454#define STATS_INC_REAPED(x) ((x)->reaped++)
455#define STATS_SET_HIGH(x) do { if ((x)->num_active > (x)->high_mark) \
456 (x)->high_mark = (x)->num_active; \
457 } while (0)
458#define STATS_INC_ERR(x) ((x)->errors++)
459#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
Christoph Lametere498be72005-09-09 13:03:32 -0700460#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700461#define STATS_SET_FREEABLE(x, i) \
462 do { if ((x)->max_freeable < i) \
463 (x)->max_freeable = i; \
464 } while (0)
465
466#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
467#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
468#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
469#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
470#else
471#define STATS_INC_ACTIVE(x) do { } while (0)
472#define STATS_DEC_ACTIVE(x) do { } while (0)
473#define STATS_INC_ALLOCED(x) do { } while (0)
474#define STATS_INC_GROWN(x) do { } while (0)
475#define STATS_INC_REAPED(x) do { } while (0)
476#define STATS_SET_HIGH(x) do { } while (0)
477#define STATS_INC_ERR(x) do { } while (0)
478#define STATS_INC_NODEALLOCS(x) do { } while (0)
Christoph Lametere498be72005-09-09 13:03:32 -0700479#define STATS_INC_NODEFREES(x) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700480#define STATS_SET_FREEABLE(x, i) \
481 do { } while (0)
482
483#define STATS_INC_ALLOCHIT(x) do { } while (0)
484#define STATS_INC_ALLOCMISS(x) do { } while (0)
485#define STATS_INC_FREEHIT(x) do { } while (0)
486#define STATS_INC_FREEMISS(x) do { } while (0)
487#endif
488
489#if DEBUG
490/* Magic nums for obj red zoning.
491 * Placed in the first word before and the first word after an obj.
492 */
493#define RED_INACTIVE 0x5A2CF071UL /* when obj is inactive */
494#define RED_ACTIVE 0x170FC2A5UL /* when obj is active */
495
496/* ...and for poisoning */
497#define POISON_INUSE 0x5a /* for use-uninitialised poisoning */
498#define POISON_FREE 0x6b /* for use-after-free poisoning */
499#define POISON_END 0xa5 /* end-byte of poisoning */
500
501/* memory layout of objects:
502 * 0 : objp
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800503 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
Linus Torvalds1da177e2005-04-16 15:20:36 -0700504 * the end of an object is aligned with the end of the real
505 * allocation. Catches writes behind the end of the allocation.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800506 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700507 * redzone word.
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800508 * cachep->obj_offset: The real object.
509 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
510 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address [BYTES_PER_WORD long]
Linus Torvalds1da177e2005-04-16 15:20:36 -0700511 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800512static int obj_offset(kmem_cache_t *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700513{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800514 return cachep->obj_offset;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700515}
516
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800517static int obj_size(kmem_cache_t *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700518{
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800519 return cachep->obj_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700520}
521
522static unsigned long *dbg_redzone1(kmem_cache_t *cachep, void *objp)
523{
524 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800525 return (unsigned long*) (objp+obj_offset(cachep)-BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700526}
527
528static unsigned long *dbg_redzone2(kmem_cache_t *cachep, void *objp)
529{
530 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
531 if (cachep->flags & SLAB_STORE_USER)
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800532 return (unsigned long *)(objp + cachep->buffer_size -
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800533 2 * BYTES_PER_WORD);
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800534 return (unsigned long *)(objp + cachep->buffer_size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700535}
536
537static void **dbg_userword(kmem_cache_t *cachep, void *objp)
538{
539 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800540 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700541}
542
543#else
544
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800545#define obj_offset(x) 0
546#define obj_size(cachep) (cachep->buffer_size)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700547#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;})
548#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;})
549#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
550
551#endif
552
553/*
554 * Maximum size of an obj (in 2^order pages)
555 * and absolute limit for the gfp order.
556 */
557#if defined(CONFIG_LARGE_ALLOCS)
558#define MAX_OBJ_ORDER 13 /* up to 32Mb */
559#define MAX_GFP_ORDER 13 /* up to 32Mb */
560#elif defined(CONFIG_MMU)
561#define MAX_OBJ_ORDER 5 /* 32 pages */
562#define MAX_GFP_ORDER 5 /* 32 pages */
563#else
564#define MAX_OBJ_ORDER 8 /* up to 1Mb */
565#define MAX_GFP_ORDER 8 /* up to 1Mb */
566#endif
567
568/*
569 * Do not go above this order unless 0 objects fit into the slab.
570 */
571#define BREAK_GFP_ORDER_HI 1
572#define BREAK_GFP_ORDER_LO 0
573static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
574
Pekka Enberg065d41c2005-11-13 16:06:46 -0800575/* Functions for storing/retrieving the cachep and or slab from the
Linus Torvalds1da177e2005-04-16 15:20:36 -0700576 * global 'mem_map'. These are used to find the slab an obj belongs to.
577 * With kfree(), these are used to find the cache which an obj belongs to.
578 */
Pekka Enberg065d41c2005-11-13 16:06:46 -0800579static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
580{
581 page->lru.next = (struct list_head *)cache;
582}
583
584static inline struct kmem_cache *page_get_cache(struct page *page)
585{
586 return (struct kmem_cache *)page->lru.next;
587}
588
589static inline void page_set_slab(struct page *page, struct slab *slab)
590{
591 page->lru.prev = (struct list_head *)slab;
592}
593
594static inline struct slab *page_get_slab(struct page *page)
595{
596 return (struct slab *)page->lru.prev;
597}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700598
599/* These are the default caches for kmalloc. Custom caches can have other sizes. */
600struct cache_sizes malloc_sizes[] = {
601#define CACHE(x) { .cs_size = (x) },
602#include <linux/kmalloc_sizes.h>
603 CACHE(ULONG_MAX)
604#undef CACHE
605};
606EXPORT_SYMBOL(malloc_sizes);
607
608/* Must match cache_sizes above. Out of line to keep cache footprint low. */
609struct cache_names {
610 char *name;
611 char *name_dma;
612};
613
614static struct cache_names __initdata cache_names[] = {
615#define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
616#include <linux/kmalloc_sizes.h>
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800617 {NULL,}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700618#undef CACHE
619};
620
621static struct arraycache_init initarray_cache __initdata =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800622 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700623static struct arraycache_init initarray_generic =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800624 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700625
626/* internal cache of cache description objs */
627static kmem_cache_t cache_cache = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800628 .batchcount = 1,
629 .limit = BOOT_CPUCACHE_ENTRIES,
630 .shared = 1,
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800631 .buffer_size = sizeof(kmem_cache_t),
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800632 .flags = SLAB_NO_REAP,
633 .spinlock = SPIN_LOCK_UNLOCKED,
634 .name = "kmem_cache",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700635#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -0800636 .obj_size = sizeof(kmem_cache_t),
Linus Torvalds1da177e2005-04-16 15:20:36 -0700637#endif
638};
639
640/* Guard access to the cache-chain. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800641static DEFINE_MUTEX(cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700642static struct list_head cache_chain;
643
644/*
645 * vm_enough_memory() looks at this to determine how many
646 * slab-allocated pages are possibly freeable under pressure
647 *
648 * SLAB_RECLAIM_ACCOUNT turns this on per-slab
649 */
650atomic_t slab_reclaim_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700651
652/*
653 * chicken and egg problem: delay the per-cpu array allocation
654 * until the general caches are up.
655 */
656static enum {
657 NONE,
Christoph Lametere498be72005-09-09 13:03:32 -0700658 PARTIAL_AC,
659 PARTIAL_L3,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700660 FULL
661} g_cpucache_up;
662
663static DEFINE_PER_CPU(struct work_struct, reap_work);
664
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800665static void free_block(kmem_cache_t *cachep, void **objpp, int len, int node);
666static void enable_cpucache(kmem_cache_t *cachep);
667static void cache_reap(void *unused);
Christoph Lametere498be72005-09-09 13:03:32 -0700668static int __node_shrink(kmem_cache_t *cachep, int node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700669
670static inline struct array_cache *ac_data(kmem_cache_t *cachep)
671{
672 return cachep->array[smp_processor_id()];
673}
674
Al Virodd0fc662005-10-07 07:46:04 +0100675static inline kmem_cache_t *__find_general_cachep(size_t size, gfp_t gfpflags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700676{
677 struct cache_sizes *csizep = malloc_sizes;
678
679#if DEBUG
680 /* This happens if someone tries to call
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800681 * kmem_cache_create(), or __kmalloc(), before
682 * the generic caches are initialized.
683 */
Alok Katariac7e43c72005-09-14 12:17:53 -0700684 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700685#endif
686 while (size > csizep->cs_size)
687 csizep++;
688
689 /*
Martin Hicks0abf40c2005-09-03 15:54:54 -0700690 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
Linus Torvalds1da177e2005-04-16 15:20:36 -0700691 * has cs_{dma,}cachep==NULL. Thus no special case
692 * for large kmalloc calls required.
693 */
694 if (unlikely(gfpflags & GFP_DMA))
695 return csizep->cs_dmacachep;
696 return csizep->cs_cachep;
697}
698
Al Virodd0fc662005-10-07 07:46:04 +0100699kmem_cache_t *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700700{
701 return __find_general_cachep(size, gfpflags);
702}
703EXPORT_SYMBOL(kmem_find_general_cachep);
704
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800705static size_t slab_mgmt_size(size_t nr_objs, size_t align)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700706{
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800707 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
708}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700709
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800710/* Calculate the number of objects and left-over bytes for a given
711 buffer size. */
712static void cache_estimate(unsigned long gfporder, size_t buffer_size,
713 size_t align, int flags, size_t *left_over,
714 unsigned int *num)
715{
716 int nr_objs;
717 size_t mgmt_size;
718 size_t slab_size = PAGE_SIZE << gfporder;
719
720 /*
721 * The slab management structure can be either off the slab or
722 * on it. For the latter case, the memory allocated for a
723 * slab is used for:
724 *
725 * - The struct slab
726 * - One kmem_bufctl_t for each object
727 * - Padding to respect alignment of @align
728 * - @buffer_size bytes for each object
729 *
730 * If the slab management structure is off the slab, then the
731 * alignment will already be calculated into the size. Because
732 * the slabs are all pages aligned, the objects will be at the
733 * correct alignment when allocated.
734 */
735 if (flags & CFLGS_OFF_SLAB) {
736 mgmt_size = 0;
737 nr_objs = slab_size / buffer_size;
738
739 if (nr_objs > SLAB_LIMIT)
740 nr_objs = SLAB_LIMIT;
741 } else {
742 /*
743 * Ignore padding for the initial guess. The padding
744 * is at most @align-1 bytes, and @buffer_size is at
745 * least @align. In the worst case, this result will
746 * be one greater than the number of objects that fit
747 * into the memory allocation when taking the padding
748 * into account.
749 */
750 nr_objs = (slab_size - sizeof(struct slab)) /
751 (buffer_size + sizeof(kmem_bufctl_t));
752
753 /*
754 * This calculated number will be either the right
755 * amount, or one greater than what we want.
756 */
757 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
758 > slab_size)
759 nr_objs--;
760
761 if (nr_objs > SLAB_LIMIT)
762 nr_objs = SLAB_LIMIT;
763
764 mgmt_size = slab_mgmt_size(nr_objs, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700765 }
Steven Rostedtfbaccac2006-02-01 03:05:45 -0800766 *num = nr_objs;
767 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700768}
769
770#define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
771
772static void __slab_error(const char *function, kmem_cache_t *cachep, char *msg)
773{
774 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800775 function, cachep->name, msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700776 dump_stack();
777}
778
779/*
780 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
781 * via the workqueue/eventd.
782 * Add the CPU number into the expiration time to minimize the possibility of
783 * the CPUs getting into lockstep and contending for the global cache chain
784 * lock.
785 */
786static void __devinit start_cpu_timer(int cpu)
787{
788 struct work_struct *reap_work = &per_cpu(reap_work, cpu);
789
790 /*
791 * When this gets called from do_initcalls via cpucache_init(),
792 * init_workqueues() has already run, so keventd will be setup
793 * at that time.
794 */
795 if (keventd_up() && reap_work->func == NULL) {
796 INIT_WORK(reap_work, cache_reap, NULL);
797 schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu);
798 }
799}
800
Christoph Lametere498be72005-09-09 13:03:32 -0700801static struct array_cache *alloc_arraycache(int node, int entries,
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800802 int batchcount)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700803{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800804 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700805 struct array_cache *nc = NULL;
806
Christoph Lametere498be72005-09-09 13:03:32 -0700807 nc = kmalloc_node(memsize, GFP_KERNEL, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700808 if (nc) {
809 nc->avail = 0;
810 nc->limit = entries;
811 nc->batchcount = batchcount;
812 nc->touched = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700813 spin_lock_init(&nc->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700814 }
815 return nc;
816}
817
Christoph Lametere498be72005-09-09 13:03:32 -0700818#ifdef CONFIG_NUMA
Christoph Lameterdc85da12006-01-18 17:42:36 -0800819static void *__cache_alloc_node(kmem_cache_t *, gfp_t, int);
820
Pekka Enberg5295a742006-02-01 03:05:48 -0800821static struct array_cache **alloc_alien_cache(int node, int limit)
Christoph Lametere498be72005-09-09 13:03:32 -0700822{
823 struct array_cache **ac_ptr;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800824 int memsize = sizeof(void *) * MAX_NUMNODES;
Christoph Lametere498be72005-09-09 13:03:32 -0700825 int i;
826
827 if (limit > 1)
828 limit = 12;
829 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
830 if (ac_ptr) {
831 for_each_node(i) {
832 if (i == node || !node_online(i)) {
833 ac_ptr[i] = NULL;
834 continue;
835 }
836 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
837 if (!ac_ptr[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800838 for (i--; i <= 0; i--)
Christoph Lametere498be72005-09-09 13:03:32 -0700839 kfree(ac_ptr[i]);
840 kfree(ac_ptr);
841 return NULL;
842 }
843 }
844 }
845 return ac_ptr;
846}
847
Pekka Enberg5295a742006-02-01 03:05:48 -0800848static void free_alien_cache(struct array_cache **ac_ptr)
Christoph Lametere498be72005-09-09 13:03:32 -0700849{
850 int i;
851
852 if (!ac_ptr)
853 return;
854
855 for_each_node(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800856 kfree(ac_ptr[i]);
Christoph Lametere498be72005-09-09 13:03:32 -0700857
858 kfree(ac_ptr);
859}
860
Pekka Enberg5295a742006-02-01 03:05:48 -0800861static void __drain_alien_cache(kmem_cache_t *cachep,
862 struct array_cache *ac, int node)
Christoph Lametere498be72005-09-09 13:03:32 -0700863{
864 struct kmem_list3 *rl3 = cachep->nodelists[node];
865
866 if (ac->avail) {
867 spin_lock(&rl3->list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -0700868 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -0700869 ac->avail = 0;
870 spin_unlock(&rl3->list_lock);
871 }
872}
873
874static void drain_alien_cache(kmem_cache_t *cachep, struct kmem_list3 *l3)
875{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800876 int i = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700877 struct array_cache *ac;
878 unsigned long flags;
879
880 for_each_online_node(i) {
881 ac = l3->alien[i];
882 if (ac) {
883 spin_lock_irqsave(&ac->lock, flags);
884 __drain_alien_cache(cachep, ac, i);
885 spin_unlock_irqrestore(&ac->lock, flags);
886 }
887 }
888}
889#else
890#define alloc_alien_cache(node, limit) do { } while (0)
891#define free_alien_cache(ac_ptr) do { } while (0)
892#define drain_alien_cache(cachep, l3) do { } while (0)
893#endif
894
Linus Torvalds1da177e2005-04-16 15:20:36 -0700895static int __devinit cpuup_callback(struct notifier_block *nfb,
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800896 unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700897{
898 long cpu = (long)hcpu;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800899 kmem_cache_t *cachep;
Christoph Lametere498be72005-09-09 13:03:32 -0700900 struct kmem_list3 *l3 = NULL;
901 int node = cpu_to_node(cpu);
902 int memsize = sizeof(struct kmem_list3);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700903
904 switch (action) {
905 case CPU_UP_PREPARE:
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800906 mutex_lock(&cache_chain_mutex);
Christoph Lametere498be72005-09-09 13:03:32 -0700907 /* we need to do this right in the beginning since
908 * alloc_arraycache's are going to use this list.
909 * kmalloc_node allows us to add the slab to the right
910 * kmem_list3 and not this cpu's kmem_list3
911 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700912
Christoph Lametere498be72005-09-09 13:03:32 -0700913 list_for_each_entry(cachep, &cache_chain, next) {
914 /* setup the size64 kmemlist for cpu before we can
915 * begin anything. Make sure some other cpu on this
916 * node has not already allocated this
917 */
918 if (!cachep->nodelists[node]) {
919 if (!(l3 = kmalloc_node(memsize,
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800920 GFP_KERNEL, node)))
Christoph Lametere498be72005-09-09 13:03:32 -0700921 goto bad;
922 kmem_list3_init(l3);
923 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800924 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -0700925
926 cachep->nodelists[node] = l3;
927 }
928
929 spin_lock_irq(&cachep->nodelists[node]->list_lock);
930 cachep->nodelists[node]->free_limit =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800931 (1 + nr_cpus_node(node)) *
932 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -0700933 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
934 }
935
936 /* Now we can go ahead with allocating the shared array's
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800937 & array cache's */
Christoph Lametere498be72005-09-09 13:03:32 -0700938 list_for_each_entry(cachep, &cache_chain, next) {
Tobias Klausercd105df2006-01-08 01:00:59 -0800939 struct array_cache *nc;
940
Christoph Lametere498be72005-09-09 13:03:32 -0700941 nc = alloc_arraycache(node, cachep->limit,
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800942 cachep->batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700943 if (!nc)
944 goto bad;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700945 cachep->array[cpu] = nc;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700946
Christoph Lametere498be72005-09-09 13:03:32 -0700947 l3 = cachep->nodelists[node];
948 BUG_ON(!l3);
949 if (!l3->shared) {
950 if (!(nc = alloc_arraycache(node,
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800951 cachep->shared *
952 cachep->batchcount,
953 0xbaadf00d)))
954 goto bad;
Christoph Lametere498be72005-09-09 13:03:32 -0700955
956 /* we are serialised from CPU_DEAD or
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800957 CPU_UP_CANCELLED by the cpucontrol lock */
Christoph Lametere498be72005-09-09 13:03:32 -0700958 l3->shared = nc;
959 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700960 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800961 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700962 break;
963 case CPU_ONLINE:
964 start_cpu_timer(cpu);
965 break;
966#ifdef CONFIG_HOTPLUG_CPU
967 case CPU_DEAD:
968 /* fall thru */
969 case CPU_UP_CANCELED:
Ingo Molnarfc0abb12006-01-18 17:42:33 -0800970 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700971
972 list_for_each_entry(cachep, &cache_chain, next) {
973 struct array_cache *nc;
Christoph Lametere498be72005-09-09 13:03:32 -0700974 cpumask_t mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700975
Christoph Lametere498be72005-09-09 13:03:32 -0700976 mask = node_to_cpumask(node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700977 spin_lock_irq(&cachep->spinlock);
978 /* cpu is dead; no one can alloc from it. */
979 nc = cachep->array[cpu];
980 cachep->array[cpu] = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -0700981 l3 = cachep->nodelists[node];
982
983 if (!l3)
984 goto unlock_cache;
985
986 spin_lock(&l3->list_lock);
987
988 /* Free limit for this kmem_list3 */
989 l3->free_limit -= cachep->batchcount;
990 if (nc)
Christoph Lameterff694162005-09-22 21:44:02 -0700991 free_block(cachep, nc->entry, nc->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -0700992
993 if (!cpus_empty(mask)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800994 spin_unlock(&l3->list_lock);
995 goto unlock_cache;
996 }
Christoph Lametere498be72005-09-09 13:03:32 -0700997
998 if (l3->shared) {
999 free_block(cachep, l3->shared->entry,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001000 l3->shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001001 kfree(l3->shared);
1002 l3->shared = NULL;
1003 }
1004 if (l3->alien) {
1005 drain_alien_cache(cachep, l3);
1006 free_alien_cache(l3->alien);
1007 l3->alien = NULL;
1008 }
1009
1010 /* free slabs belonging to this node */
1011 if (__node_shrink(cachep, node)) {
1012 cachep->nodelists[node] = NULL;
1013 spin_unlock(&l3->list_lock);
1014 kfree(l3);
1015 } else {
1016 spin_unlock(&l3->list_lock);
1017 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001018 unlock_cache:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001019 spin_unlock_irq(&cachep->spinlock);
1020 kfree(nc);
1021 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001022 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001023 break;
1024#endif
1025 }
1026 return NOTIFY_OK;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001027 bad:
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001028 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001029 return NOTIFY_BAD;
1030}
1031
1032static struct notifier_block cpucache_notifier = { &cpuup_callback, NULL, 0 };
1033
Christoph Lametere498be72005-09-09 13:03:32 -07001034/*
1035 * swap the static kmem_list3 with kmalloced memory
1036 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001037static void init_list(kmem_cache_t *cachep, struct kmem_list3 *list, int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07001038{
1039 struct kmem_list3 *ptr;
1040
1041 BUG_ON(cachep->nodelists[nodeid] != list);
1042 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1043 BUG_ON(!ptr);
1044
1045 local_irq_disable();
1046 memcpy(ptr, list, sizeof(struct kmem_list3));
1047 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1048 cachep->nodelists[nodeid] = ptr;
1049 local_irq_enable();
1050}
1051
Linus Torvalds1da177e2005-04-16 15:20:36 -07001052/* Initialisation.
1053 * Called after the gfp() functions have been enabled, and before smp_init().
1054 */
1055void __init kmem_cache_init(void)
1056{
1057 size_t left_over;
1058 struct cache_sizes *sizes;
1059 struct cache_names *names;
Christoph Lametere498be72005-09-09 13:03:32 -07001060 int i;
1061
1062 for (i = 0; i < NUM_INIT_LISTS; i++) {
1063 kmem_list3_init(&initkmem_list3[i]);
1064 if (i < MAX_NUMNODES)
1065 cache_cache.nodelists[i] = NULL;
1066 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001067
1068 /*
1069 * Fragmentation resistance on low memory - only use bigger
1070 * page orders on machines with more than 32MB of memory.
1071 */
1072 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1073 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1074
Linus Torvalds1da177e2005-04-16 15:20:36 -07001075 /* Bootstrap is tricky, because several objects are allocated
1076 * from caches that do not exist yet:
1077 * 1) initialize the cache_cache cache: it contains the kmem_cache_t
1078 * structures of all caches, except cache_cache itself: cache_cache
1079 * is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001080 * Initially an __init data area is used for the head array and the
1081 * kmem_list3 structures, it's replaced with a kmalloc allocated
1082 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001083 * 2) Create the first kmalloc cache.
Christoph Lametere498be72005-09-09 13:03:32 -07001084 * The kmem_cache_t for the new cache is allocated normally.
1085 * An __init data area is used for the head array.
1086 * 3) Create the remaining kmalloc caches, with minimally sized
1087 * head arrays.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001088 * 4) Replace the __init data head arrays for cache_cache and the first
1089 * kmalloc cache with kmalloc allocated arrays.
Christoph Lametere498be72005-09-09 13:03:32 -07001090 * 5) Replace the __init data for kmem_list3 for cache_cache and
1091 * the other cache's with kmalloc allocated memory.
1092 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001093 */
1094
1095 /* 1) create the cache_cache */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001096 INIT_LIST_HEAD(&cache_chain);
1097 list_add(&cache_cache.next, &cache_chain);
1098 cache_cache.colour_off = cache_line_size();
1099 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
Christoph Lametere498be72005-09-09 13:03:32 -07001100 cache_cache.nodelists[numa_node_id()] = &initkmem_list3[CACHE_CACHE];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001101
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001102 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size, cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001103
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001104 cache_estimate(0, cache_cache.buffer_size, cache_line_size(), 0,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001105 &left_over, &cache_cache.num);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001106 if (!cache_cache.num)
1107 BUG();
1108
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001109 cache_cache.colour = left_over / cache_cache.colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001110 cache_cache.colour_next = 0;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001111 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1112 sizeof(struct slab), cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001113
1114 /* 2+3) create the kmalloc caches */
1115 sizes = malloc_sizes;
1116 names = cache_names;
1117
Christoph Lametere498be72005-09-09 13:03:32 -07001118 /* Initialize the caches that provide memory for the array cache
1119 * and the kmem_list3 structures first.
1120 * Without this, further allocations will bug
1121 */
1122
1123 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001124 sizes[INDEX_AC].cs_size,
1125 ARCH_KMALLOC_MINALIGN,
1126 (ARCH_KMALLOC_FLAGS |
1127 SLAB_PANIC), NULL, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07001128
1129 if (INDEX_AC != INDEX_L3)
1130 sizes[INDEX_L3].cs_cachep =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001131 kmem_cache_create(names[INDEX_L3].name,
1132 sizes[INDEX_L3].cs_size,
1133 ARCH_KMALLOC_MINALIGN,
1134 (ARCH_KMALLOC_FLAGS | SLAB_PANIC), NULL,
1135 NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07001136
Linus Torvalds1da177e2005-04-16 15:20:36 -07001137 while (sizes->cs_size != ULONG_MAX) {
Christoph Lametere498be72005-09-09 13:03:32 -07001138 /*
1139 * For performance, all the general caches are L1 aligned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001140 * This should be particularly beneficial on SMP boxes, as it
1141 * eliminates "false sharing".
1142 * Note for systems short on memory removing the alignment will
Christoph Lametere498be72005-09-09 13:03:32 -07001143 * allow tighter packing of the smaller caches.
1144 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001145 if (!sizes->cs_cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07001146 sizes->cs_cachep = kmem_cache_create(names->name,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001147 sizes->cs_size,
1148 ARCH_KMALLOC_MINALIGN,
1149 (ARCH_KMALLOC_FLAGS
1150 | SLAB_PANIC),
1151 NULL, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001152
1153 /* Inc off-slab bufctl limit until the ceiling is hit. */
1154 if (!(OFF_SLAB(sizes->cs_cachep))) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001155 offslab_limit = sizes->cs_size - sizeof(struct slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001156 offslab_limit /= sizeof(kmem_bufctl_t);
1157 }
1158
1159 sizes->cs_dmacachep = kmem_cache_create(names->name_dma,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001160 sizes->cs_size,
1161 ARCH_KMALLOC_MINALIGN,
1162 (ARCH_KMALLOC_FLAGS |
1163 SLAB_CACHE_DMA |
1164 SLAB_PANIC), NULL,
1165 NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001166
1167 sizes++;
1168 names++;
1169 }
1170 /* 4) Replace the bootstrap head arrays */
1171 {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001172 void *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001173
Linus Torvalds1da177e2005-04-16 15:20:36 -07001174 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001175
Linus Torvalds1da177e2005-04-16 15:20:36 -07001176 local_irq_disable();
1177 BUG_ON(ac_data(&cache_cache) != &initarray_cache.cache);
Christoph Lametere498be72005-09-09 13:03:32 -07001178 memcpy(ptr, ac_data(&cache_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001179 sizeof(struct arraycache_init));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001180 cache_cache.array[smp_processor_id()] = ptr;
1181 local_irq_enable();
Christoph Lametere498be72005-09-09 13:03:32 -07001182
Linus Torvalds1da177e2005-04-16 15:20:36 -07001183 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001184
Linus Torvalds1da177e2005-04-16 15:20:36 -07001185 local_irq_disable();
Christoph Lametere498be72005-09-09 13:03:32 -07001186 BUG_ON(ac_data(malloc_sizes[INDEX_AC].cs_cachep)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001187 != &initarray_generic.cache);
Christoph Lametere498be72005-09-09 13:03:32 -07001188 memcpy(ptr, ac_data(malloc_sizes[INDEX_AC].cs_cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001189 sizeof(struct arraycache_init));
Christoph Lametere498be72005-09-09 13:03:32 -07001190 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001191 ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001192 local_irq_enable();
1193 }
Christoph Lametere498be72005-09-09 13:03:32 -07001194 /* 5) Replace the bootstrap kmem_list3's */
1195 {
1196 int node;
1197 /* Replace the static kmem_list3 structures for the boot cpu */
1198 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE],
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001199 numa_node_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001200
Christoph Lametere498be72005-09-09 13:03:32 -07001201 for_each_online_node(node) {
1202 init_list(malloc_sizes[INDEX_AC].cs_cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001203 &initkmem_list3[SIZE_AC + node], node);
Christoph Lametere498be72005-09-09 13:03:32 -07001204
1205 if (INDEX_AC != INDEX_L3) {
1206 init_list(malloc_sizes[INDEX_L3].cs_cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001207 &initkmem_list3[SIZE_L3 + node],
1208 node);
Christoph Lametere498be72005-09-09 13:03:32 -07001209 }
1210 }
1211 }
1212
1213 /* 6) resize the head arrays to their final sizes */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001214 {
1215 kmem_cache_t *cachep;
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001216 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001217 list_for_each_entry(cachep, &cache_chain, next)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001218 enable_cpucache(cachep);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001219 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001220 }
1221
1222 /* Done! */
1223 g_cpucache_up = FULL;
1224
1225 /* Register a cpu startup notifier callback
1226 * that initializes ac_data for all new cpus
1227 */
1228 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001229
1230 /* The reap timers are started later, with a module init call:
1231 * That part of the kernel is not yet operational.
1232 */
1233}
1234
1235static int __init cpucache_init(void)
1236{
1237 int cpu;
1238
1239 /*
1240 * Register the timers that return unneeded
1241 * pages to gfp.
1242 */
Christoph Lametere498be72005-09-09 13:03:32 -07001243 for_each_online_cpu(cpu)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001244 start_cpu_timer(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001245
1246 return 0;
1247}
1248
1249__initcall(cpucache_init);
1250
1251/*
1252 * Interface to system's page allocator. No need to hold the cache-lock.
1253 *
1254 * If we requested dmaable memory, we will get it. Even if we
1255 * did not request dmaable memory, we might get it, but that
1256 * would be relatively rare and ignorable.
1257 */
Al Virodd0fc662005-10-07 07:46:04 +01001258static void *kmem_getpages(kmem_cache_t *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001259{
1260 struct page *page;
1261 void *addr;
1262 int i;
1263
1264 flags |= cachep->gfpflags;
Christoph Lameter50c85a12005-11-13 16:06:47 -08001265 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001266 if (!page)
1267 return NULL;
1268 addr = page_address(page);
1269
1270 i = (1 << cachep->gfporder);
1271 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1272 atomic_add(i, &slab_reclaim_pages);
1273 add_page_state(nr_slab, i);
1274 while (i--) {
1275 SetPageSlab(page);
1276 page++;
1277 }
1278 return addr;
1279}
1280
1281/*
1282 * Interface to system's page release.
1283 */
1284static void kmem_freepages(kmem_cache_t *cachep, void *addr)
1285{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001286 unsigned long i = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001287 struct page *page = virt_to_page(addr);
1288 const unsigned long nr_freed = i;
1289
1290 while (i--) {
1291 if (!TestClearPageSlab(page))
1292 BUG();
1293 page++;
1294 }
1295 sub_page_state(nr_slab, nr_freed);
1296 if (current->reclaim_state)
1297 current->reclaim_state->reclaimed_slab += nr_freed;
1298 free_pages((unsigned long)addr, cachep->gfporder);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001299 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1300 atomic_sub(1 << cachep->gfporder, &slab_reclaim_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001301}
1302
1303static void kmem_rcu_free(struct rcu_head *head)
1304{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001305 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001306 kmem_cache_t *cachep = slab_rcu->cachep;
1307
1308 kmem_freepages(cachep, slab_rcu->addr);
1309 if (OFF_SLAB(cachep))
1310 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1311}
1312
1313#if DEBUG
1314
1315#ifdef CONFIG_DEBUG_PAGEALLOC
1316static void store_stackinfo(kmem_cache_t *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001317 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001318{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001319 int size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001320
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001321 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001322
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001323 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001324 return;
1325
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001326 *addr++ = 0x12345678;
1327 *addr++ = caller;
1328 *addr++ = smp_processor_id();
1329 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001330 {
1331 unsigned long *sptr = &caller;
1332 unsigned long svalue;
1333
1334 while (!kstack_end(sptr)) {
1335 svalue = *sptr++;
1336 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001337 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001338 size -= sizeof(unsigned long);
1339 if (size <= sizeof(unsigned long))
1340 break;
1341 }
1342 }
1343
1344 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001345 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001346}
1347#endif
1348
1349static void poison_obj(kmem_cache_t *cachep, void *addr, unsigned char val)
1350{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001351 int size = obj_size(cachep);
1352 addr = &((char *)addr)[obj_offset(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001353
1354 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001355 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001356}
1357
1358static void dump_line(char *data, int offset, int limit)
1359{
1360 int i;
1361 printk(KERN_ERR "%03x:", offset);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001362 for (i = 0; i < limit; i++) {
1363 printk(" %02x", (unsigned char)data[offset + i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001364 }
1365 printk("\n");
1366}
1367#endif
1368
1369#if DEBUG
1370
1371static void print_objinfo(kmem_cache_t *cachep, void *objp, int lines)
1372{
1373 int i, size;
1374 char *realobj;
1375
1376 if (cachep->flags & SLAB_RED_ZONE) {
1377 printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001378 *dbg_redzone1(cachep, objp),
1379 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001380 }
1381
1382 if (cachep->flags & SLAB_STORE_USER) {
1383 printk(KERN_ERR "Last user: [<%p>]",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001384 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001385 print_symbol("(%s)",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001386 (unsigned long)*dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001387 printk("\n");
1388 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001389 realobj = (char *)objp + obj_offset(cachep);
1390 size = obj_size(cachep);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001391 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001392 int limit;
1393 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001394 if (i + limit > size)
1395 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001396 dump_line(realobj, i, limit);
1397 }
1398}
1399
1400static void check_poison_obj(kmem_cache_t *cachep, void *objp)
1401{
1402 char *realobj;
1403 int size, i;
1404 int lines = 0;
1405
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001406 realobj = (char *)objp + obj_offset(cachep);
1407 size = obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001408
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001409 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001410 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001411 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001412 exp = POISON_END;
1413 if (realobj[i] != exp) {
1414 int limit;
1415 /* Mismatch ! */
1416 /* Print header */
1417 if (lines == 0) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001418 printk(KERN_ERR
1419 "Slab corruption: start=%p, len=%d\n",
1420 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001421 print_objinfo(cachep, objp, 0);
1422 }
1423 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001424 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001425 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001426 if (i + limit > size)
1427 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001428 dump_line(realobj, i, limit);
1429 i += 16;
1430 lines++;
1431 /* Limit to 5 lines */
1432 if (lines > 5)
1433 break;
1434 }
1435 }
1436 if (lines != 0) {
1437 /* Print some data about the neighboring objects, if they
1438 * exist:
1439 */
Pekka Enberg065d41c2005-11-13 16:06:46 -08001440 struct slab *slabp = page_get_slab(virt_to_page(objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001441 int objnr;
1442
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001443 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001444 if (objnr) {
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001445 objp = slabp->s_mem + (objnr - 1) * cachep->buffer_size;
1446 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001447 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001448 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001449 print_objinfo(cachep, objp, 2);
1450 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001451 if (objnr + 1 < cachep->num) {
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001452 objp = slabp->s_mem + (objnr + 1) * cachep->buffer_size;
1453 realobj = (char *)objp + obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001454 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001455 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001456 print_objinfo(cachep, objp, 2);
1457 }
1458 }
1459}
1460#endif
1461
Linus Torvalds1da177e2005-04-16 15:20:36 -07001462#if DEBUG
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001463/**
1464 * slab_destroy_objs - call the registered destructor for each object in
1465 * a slab that is to be destroyed.
1466 */
1467static void slab_destroy_objs(kmem_cache_t *cachep, struct slab *slabp)
1468{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001469 int i;
1470 for (i = 0; i < cachep->num; i++) {
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001471 void *objp = slabp->s_mem + cachep->buffer_size * i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001472
1473 if (cachep->flags & SLAB_POISON) {
1474#ifdef CONFIG_DEBUG_PAGEALLOC
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001475 if ((cachep->buffer_size % PAGE_SIZE) == 0
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001476 && OFF_SLAB(cachep))
1477 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001478 cachep->buffer_size / PAGE_SIZE,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001479 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001480 else
1481 check_poison_obj(cachep, objp);
1482#else
1483 check_poison_obj(cachep, objp);
1484#endif
1485 }
1486 if (cachep->flags & SLAB_RED_ZONE) {
1487 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1488 slab_error(cachep, "start of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001489 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001490 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1491 slab_error(cachep, "end of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001492 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001493 }
1494 if (cachep->dtor && !(cachep->flags & SLAB_POISON))
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001495 (cachep->dtor) (objp + obj_offset(cachep), cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001496 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001497}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001498#else
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001499static void slab_destroy_objs(kmem_cache_t *cachep, struct slab *slabp)
1500{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001501 if (cachep->dtor) {
1502 int i;
1503 for (i = 0; i < cachep->num; i++) {
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001504 void *objp = slabp->s_mem + cachep->buffer_size * i;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001505 (cachep->dtor) (objp, cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001506 }
1507 }
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001508}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001509#endif
1510
Matthew Dobson12dd36f2006-02-01 03:05:46 -08001511/**
1512 * Destroy all the objs in a slab, and release the mem back to the system.
1513 * Before calling the slab must have been unlinked from the cache.
1514 * The cache-lock is not held/needed.
1515 */
1516static void slab_destroy(kmem_cache_t *cachep, struct slab *slabp)
1517{
1518 void *addr = slabp->s_mem - slabp->colouroff;
1519
1520 slab_destroy_objs(cachep, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001521 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1522 struct slab_rcu *slab_rcu;
1523
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001524 slab_rcu = (struct slab_rcu *)slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001525 slab_rcu->cachep = cachep;
1526 slab_rcu->addr = addr;
1527 call_rcu(&slab_rcu->head, kmem_rcu_free);
1528 } else {
1529 kmem_freepages(cachep, addr);
1530 if (OFF_SLAB(cachep))
1531 kmem_cache_free(cachep->slabp_cache, slabp);
1532 }
1533}
1534
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001535/* For setting up all the kmem_list3s for cache whose buffer_size is same
Christoph Lametere498be72005-09-09 13:03:32 -07001536 as size of kmem_list3. */
Pekka Enberg5295a742006-02-01 03:05:48 -08001537static void set_up_list3s(kmem_cache_t *cachep, int index)
Christoph Lametere498be72005-09-09 13:03:32 -07001538{
1539 int node;
1540
1541 for_each_online_node(node) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001542 cachep->nodelists[node] = &initkmem_list3[index + node];
Christoph Lametere498be72005-09-09 13:03:32 -07001543 cachep->nodelists[node]->next_reap = jiffies +
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001544 REAPTIMEOUT_LIST3 +
1545 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -07001546 }
1547}
1548
Linus Torvalds1da177e2005-04-16 15:20:36 -07001549/**
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001550 * calculate_slab_order - calculate size (page order) of slabs and the number
1551 * of objects per slab.
1552 *
1553 * This could be made much more intelligent. For now, try to avoid using
1554 * high order pages for slabs. When the gfp() functions are more friendly
1555 * towards high-order requests, this should be changed.
1556 */
1557static inline size_t calculate_slab_order(kmem_cache_t *cachep, size_t size,
1558 size_t align, gfp_t flags)
1559{
1560 size_t left_over = 0;
1561
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001562 for (;; cachep->gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001563 unsigned int num;
1564 size_t remainder;
1565
1566 if (cachep->gfporder > MAX_GFP_ORDER) {
1567 cachep->num = 0;
1568 break;
1569 }
1570
1571 cache_estimate(cachep->gfporder, size, align, flags,
1572 &remainder, &num);
1573 if (!num)
1574 continue;
1575 /* More than offslab_limit objects will cause problems */
1576 if (flags & CFLGS_OFF_SLAB && cachep->num > offslab_limit)
1577 break;
1578
1579 cachep->num = num;
1580 left_over = remainder;
1581
1582 /*
1583 * Large number of objects is good, but very large slabs are
1584 * currently bad for the gfp()s.
1585 */
1586 if (cachep->gfporder >= slab_break_gfp_order)
1587 break;
1588
1589 if ((left_over * 8) <= (PAGE_SIZE << cachep->gfporder))
1590 /* Acceptable internal fragmentation */
1591 break;
1592 }
1593 return left_over;
1594}
1595
1596/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07001597 * kmem_cache_create - Create a cache.
1598 * @name: A string which is used in /proc/slabinfo to identify this cache.
1599 * @size: The size of objects to be created in this cache.
1600 * @align: The required alignment for the objects.
1601 * @flags: SLAB flags
1602 * @ctor: A constructor for the objects.
1603 * @dtor: A destructor for the objects.
1604 *
1605 * Returns a ptr to the cache on success, NULL on failure.
1606 * Cannot be called within a int, but can be interrupted.
1607 * The @ctor is run when new pages are allocated by the cache
1608 * and the @dtor is run before the pages are handed back.
1609 *
1610 * @name must be valid until the cache is destroyed. This implies that
1611 * the module calling this has to destroy the cache before getting
1612 * unloaded.
1613 *
1614 * The flags are
1615 *
1616 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1617 * to catch references to uninitialised memory.
1618 *
1619 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1620 * for buffer overruns.
1621 *
1622 * %SLAB_NO_REAP - Don't automatically reap this cache when we're under
1623 * memory pressure.
1624 *
1625 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1626 * cacheline. This can be beneficial if you're counting cycles as closely
1627 * as davem.
1628 */
1629kmem_cache_t *
1630kmem_cache_create (const char *name, size_t size, size_t align,
1631 unsigned long flags, void (*ctor)(void*, kmem_cache_t *, unsigned long),
1632 void (*dtor)(void*, kmem_cache_t *, unsigned long))
1633{
1634 size_t left_over, slab_size, ralign;
1635 kmem_cache_t *cachep = NULL;
Andrew Morton4f12bb42005-11-07 00:58:00 -08001636 struct list_head *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001637
1638 /*
1639 * Sanity checks... these are all serious usage bugs.
1640 */
1641 if ((!name) ||
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001642 in_interrupt() ||
1643 (size < BYTES_PER_WORD) ||
1644 (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) {
1645 printk(KERN_ERR "%s: Early error in slab %s\n",
1646 __FUNCTION__, name);
1647 BUG();
1648 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001649
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001650 mutex_lock(&cache_chain_mutex);
Andrew Morton4f12bb42005-11-07 00:58:00 -08001651
1652 list_for_each(p, &cache_chain) {
1653 kmem_cache_t *pc = list_entry(p, kmem_cache_t, next);
1654 mm_segment_t old_fs = get_fs();
1655 char tmp;
1656 int res;
1657
1658 /*
1659 * This happens when the module gets unloaded and doesn't
1660 * destroy its slab cache and no-one else reuses the vmalloc
1661 * area of the module. Print a warning.
1662 */
1663 set_fs(KERNEL_DS);
1664 res = __get_user(tmp, pc->name);
1665 set_fs(old_fs);
1666 if (res) {
1667 printk("SLAB: cache with size %d has lost its name\n",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001668 pc->buffer_size);
Andrew Morton4f12bb42005-11-07 00:58:00 -08001669 continue;
1670 }
1671
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001672 if (!strcmp(pc->name, name)) {
Andrew Morton4f12bb42005-11-07 00:58:00 -08001673 printk("kmem_cache_create: duplicate cache %s\n", name);
1674 dump_stack();
1675 goto oops;
1676 }
1677 }
1678
Linus Torvalds1da177e2005-04-16 15:20:36 -07001679#if DEBUG
1680 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
1681 if ((flags & SLAB_DEBUG_INITIAL) && !ctor) {
1682 /* No constructor, but inital state check requested */
1683 printk(KERN_ERR "%s: No con, but init state check "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001684 "requested - %s\n", __FUNCTION__, name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001685 flags &= ~SLAB_DEBUG_INITIAL;
1686 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001687#if FORCED_DEBUG
1688 /*
1689 * Enable redzoning and last user accounting, except for caches with
1690 * large objects, if the increased size would increase the object size
1691 * above the next power of two: caches with object sizes just above a
1692 * power of two have a significant amount of internal fragmentation.
1693 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001694 if ((size < 4096
1695 || fls(size - 1) == fls(size - 1 + 3 * BYTES_PER_WORD)))
1696 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001697 if (!(flags & SLAB_DESTROY_BY_RCU))
1698 flags |= SLAB_POISON;
1699#endif
1700 if (flags & SLAB_DESTROY_BY_RCU)
1701 BUG_ON(flags & SLAB_POISON);
1702#endif
1703 if (flags & SLAB_DESTROY_BY_RCU)
1704 BUG_ON(dtor);
1705
1706 /*
1707 * Always checks flags, a caller might be expecting debug
1708 * support which isn't available.
1709 */
1710 if (flags & ~CREATE_MASK)
1711 BUG();
1712
1713 /* Check that size is in terms of words. This is needed to avoid
1714 * unaligned accesses for some archs when redzoning is used, and makes
1715 * sure any on-slab bufctl's are also correctly aligned.
1716 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001717 if (size & (BYTES_PER_WORD - 1)) {
1718 size += (BYTES_PER_WORD - 1);
1719 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001720 }
1721
1722 /* calculate out the final buffer alignment: */
1723 /* 1) arch recommendation: can be overridden for debug */
1724 if (flags & SLAB_HWCACHE_ALIGN) {
1725 /* Default alignment: as specified by the arch code.
1726 * Except if an object is really small, then squeeze multiple
1727 * objects into one cacheline.
1728 */
1729 ralign = cache_line_size();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001730 while (size <= ralign / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001731 ralign /= 2;
1732 } else {
1733 ralign = BYTES_PER_WORD;
1734 }
1735 /* 2) arch mandated alignment: disables debug if necessary */
1736 if (ralign < ARCH_SLAB_MINALIGN) {
1737 ralign = ARCH_SLAB_MINALIGN;
1738 if (ralign > BYTES_PER_WORD)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001739 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001740 }
1741 /* 3) caller mandated alignment: disables debug if necessary */
1742 if (ralign < align) {
1743 ralign = align;
1744 if (ralign > BYTES_PER_WORD)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001745 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001746 }
1747 /* 4) Store it. Note that the debug code below can reduce
1748 * the alignment to BYTES_PER_WORD.
1749 */
1750 align = ralign;
1751
1752 /* Get cache's description obj. */
1753 cachep = (kmem_cache_t *) kmem_cache_alloc(&cache_cache, SLAB_KERNEL);
1754 if (!cachep)
Andrew Morton4f12bb42005-11-07 00:58:00 -08001755 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001756 memset(cachep, 0, sizeof(kmem_cache_t));
1757
1758#if DEBUG
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001759 cachep->obj_size = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001760
1761 if (flags & SLAB_RED_ZONE) {
1762 /* redzoning only works with word aligned caches */
1763 align = BYTES_PER_WORD;
1764
1765 /* add space for red zone words */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001766 cachep->obj_offset += BYTES_PER_WORD;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001767 size += 2 * BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001768 }
1769 if (flags & SLAB_STORE_USER) {
1770 /* user store requires word alignment and
1771 * one word storage behind the end of the real
1772 * object.
1773 */
1774 align = BYTES_PER_WORD;
1775 size += BYTES_PER_WORD;
1776 }
1777#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001778 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001779 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
1780 cachep->obj_offset += PAGE_SIZE - size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001781 size = PAGE_SIZE;
1782 }
1783#endif
1784#endif
1785
1786 /* Determine if the slab management is 'on' or 'off' slab. */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001787 if (size >= (PAGE_SIZE >> 3))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001788 /*
1789 * Size is large, assume best to place the slab management obj
1790 * off-slab (should allow better packing of objs).
1791 */
1792 flags |= CFLGS_OFF_SLAB;
1793
1794 size = ALIGN(size, align);
1795
1796 if ((flags & SLAB_RECLAIM_ACCOUNT) && size <= PAGE_SIZE) {
1797 /*
1798 * A VFS-reclaimable slab tends to have most allocations
1799 * as GFP_NOFS and we really don't want to have to be allocating
1800 * higher-order pages when we are unable to shrink dcache.
1801 */
1802 cachep->gfporder = 0;
1803 cache_estimate(cachep->gfporder, size, align, flags,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001804 &left_over, &cachep->num);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001805 } else
1806 left_over = calculate_slab_order(cachep, size, align, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001807
1808 if (!cachep->num) {
1809 printk("kmem_cache_create: couldn't create cache %s.\n", name);
1810 kmem_cache_free(&cache_cache, cachep);
1811 cachep = NULL;
Andrew Morton4f12bb42005-11-07 00:58:00 -08001812 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001813 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001814 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
1815 + sizeof(struct slab), align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001816
1817 /*
1818 * If the slab has been placed off-slab, and we have enough space then
1819 * move it on-slab. This is at the expense of any extra colouring.
1820 */
1821 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
1822 flags &= ~CFLGS_OFF_SLAB;
1823 left_over -= slab_size;
1824 }
1825
1826 if (flags & CFLGS_OFF_SLAB) {
1827 /* really off slab. No need for manual alignment */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001828 slab_size =
1829 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001830 }
1831
1832 cachep->colour_off = cache_line_size();
1833 /* Offset must be a multiple of the alignment. */
1834 if (cachep->colour_off < align)
1835 cachep->colour_off = align;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001836 cachep->colour = left_over / cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001837 cachep->slab_size = slab_size;
1838 cachep->flags = flags;
1839 cachep->gfpflags = 0;
1840 if (flags & SLAB_CACHE_DMA)
1841 cachep->gfpflags |= GFP_DMA;
1842 spin_lock_init(&cachep->spinlock);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08001843 cachep->buffer_size = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001844
1845 if (flags & CFLGS_OFF_SLAB)
Victor Fuscob2d55072005-09-10 00:26:36 -07001846 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001847 cachep->ctor = ctor;
1848 cachep->dtor = dtor;
1849 cachep->name = name;
1850
1851 /* Don't let CPUs to come and go */
1852 lock_cpu_hotplug();
1853
1854 if (g_cpucache_up == FULL) {
1855 enable_cpucache(cachep);
1856 } else {
1857 if (g_cpucache_up == NONE) {
1858 /* Note: the first kmem_cache_create must create
1859 * the cache that's used by kmalloc(24), otherwise
1860 * the creation of further caches will BUG().
1861 */
Christoph Lametere498be72005-09-09 13:03:32 -07001862 cachep->array[smp_processor_id()] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001863 &initarray_generic.cache;
Christoph Lametere498be72005-09-09 13:03:32 -07001864
1865 /* If the cache that's used by
1866 * kmalloc(sizeof(kmem_list3)) is the first cache,
1867 * then we need to set up all its list3s, otherwise
1868 * the creation of further caches will BUG().
1869 */
1870 set_up_list3s(cachep, SIZE_AC);
1871 if (INDEX_AC == INDEX_L3)
1872 g_cpucache_up = PARTIAL_L3;
1873 else
1874 g_cpucache_up = PARTIAL_AC;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001875 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07001876 cachep->array[smp_processor_id()] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001877 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001878
1879 if (g_cpucache_up == PARTIAL_AC) {
1880 set_up_list3s(cachep, SIZE_L3);
1881 g_cpucache_up = PARTIAL_L3;
1882 } else {
1883 int node;
1884 for_each_online_node(node) {
1885
1886 cachep->nodelists[node] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001887 kmalloc_node(sizeof
1888 (struct kmem_list3),
1889 GFP_KERNEL, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001890 BUG_ON(!cachep->nodelists[node]);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001891 kmem_list3_init(cachep->
1892 nodelists[node]);
Christoph Lametere498be72005-09-09 13:03:32 -07001893 }
1894 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001895 }
Christoph Lametere498be72005-09-09 13:03:32 -07001896 cachep->nodelists[numa_node_id()]->next_reap =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001897 jiffies + REAPTIMEOUT_LIST3 +
1898 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -07001899
Linus Torvalds1da177e2005-04-16 15:20:36 -07001900 BUG_ON(!ac_data(cachep));
1901 ac_data(cachep)->avail = 0;
1902 ac_data(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1903 ac_data(cachep)->batchcount = 1;
1904 ac_data(cachep)->touched = 0;
1905 cachep->batchcount = 1;
1906 cachep->limit = BOOT_CPUCACHE_ENTRIES;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001907 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001908
Linus Torvalds1da177e2005-04-16 15:20:36 -07001909 /* cache setup completed, link it into the list */
1910 list_add(&cachep->next, &cache_chain);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001911 unlock_cpu_hotplug();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001912 oops:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001913 if (!cachep && (flags & SLAB_PANIC))
1914 panic("kmem_cache_create(): failed to create slab `%s'\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001915 name);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08001916 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001917 return cachep;
1918}
1919EXPORT_SYMBOL(kmem_cache_create);
1920
1921#if DEBUG
1922static void check_irq_off(void)
1923{
1924 BUG_ON(!irqs_disabled());
1925}
1926
1927static void check_irq_on(void)
1928{
1929 BUG_ON(irqs_disabled());
1930}
1931
1932static void check_spinlock_acquired(kmem_cache_t *cachep)
1933{
1934#ifdef CONFIG_SMP
1935 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07001936 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001937#endif
1938}
Christoph Lametere498be72005-09-09 13:03:32 -07001939
Pekka Enberg5295a742006-02-01 03:05:48 -08001940static void check_spinlock_acquired_node(kmem_cache_t *cachep, int node)
Christoph Lametere498be72005-09-09 13:03:32 -07001941{
1942#ifdef CONFIG_SMP
1943 check_irq_off();
1944 assert_spin_locked(&cachep->nodelists[node]->list_lock);
1945#endif
1946}
1947
Linus Torvalds1da177e2005-04-16 15:20:36 -07001948#else
1949#define check_irq_off() do { } while(0)
1950#define check_irq_on() do { } while(0)
1951#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07001952#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001953#endif
1954
1955/*
1956 * Waits for all CPUs to execute func().
1957 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001958static void smp_call_function_all_cpus(void (*func)(void *arg), void *arg)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001959{
1960 check_irq_on();
1961 preempt_disable();
1962
1963 local_irq_disable();
1964 func(arg);
1965 local_irq_enable();
1966
1967 if (smp_call_function(func, arg, 1, 1))
1968 BUG();
1969
1970 preempt_enable();
1971}
1972
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001973static void drain_array_locked(kmem_cache_t *cachep, struct array_cache *ac,
1974 int force, int node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001975
1976static void do_drain(void *arg)
1977{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001978 kmem_cache_t *cachep = (kmem_cache_t *) arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001979 struct array_cache *ac;
Christoph Lameterff694162005-09-22 21:44:02 -07001980 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001981
1982 check_irq_off();
1983 ac = ac_data(cachep);
Christoph Lameterff694162005-09-22 21:44:02 -07001984 spin_lock(&cachep->nodelists[node]->list_lock);
1985 free_block(cachep, ac->entry, ac->avail, node);
1986 spin_unlock(&cachep->nodelists[node]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001987 ac->avail = 0;
1988}
1989
1990static void drain_cpu_caches(kmem_cache_t *cachep)
1991{
Christoph Lametere498be72005-09-09 13:03:32 -07001992 struct kmem_list3 *l3;
1993 int node;
1994
Linus Torvalds1da177e2005-04-16 15:20:36 -07001995 smp_call_function_all_cpus(do_drain, cachep);
1996 check_irq_on();
1997 spin_lock_irq(&cachep->spinlock);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001998 for_each_online_node(node) {
Christoph Lametere498be72005-09-09 13:03:32 -07001999 l3 = cachep->nodelists[node];
2000 if (l3) {
2001 spin_lock(&l3->list_lock);
2002 drain_array_locked(cachep, l3->shared, 1, node);
2003 spin_unlock(&l3->list_lock);
2004 if (l3->alien)
2005 drain_alien_cache(cachep, l3);
2006 }
2007 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002008 spin_unlock_irq(&cachep->spinlock);
2009}
2010
Christoph Lametere498be72005-09-09 13:03:32 -07002011static int __node_shrink(kmem_cache_t *cachep, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002012{
2013 struct slab *slabp;
Christoph Lametere498be72005-09-09 13:03:32 -07002014 struct kmem_list3 *l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002015 int ret;
2016
Christoph Lametere498be72005-09-09 13:03:32 -07002017 for (;;) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002018 struct list_head *p;
2019
Christoph Lametere498be72005-09-09 13:03:32 -07002020 p = l3->slabs_free.prev;
2021 if (p == &l3->slabs_free)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002022 break;
2023
Christoph Lametere498be72005-09-09 13:03:32 -07002024 slabp = list_entry(l3->slabs_free.prev, struct slab, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002025#if DEBUG
2026 if (slabp->inuse)
2027 BUG();
2028#endif
2029 list_del(&slabp->list);
2030
Christoph Lametere498be72005-09-09 13:03:32 -07002031 l3->free_objects -= cachep->num;
2032 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002033 slab_destroy(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07002034 spin_lock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002035 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002036 ret = !list_empty(&l3->slabs_full) || !list_empty(&l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002037 return ret;
2038}
2039
Christoph Lametere498be72005-09-09 13:03:32 -07002040static int __cache_shrink(kmem_cache_t *cachep)
2041{
2042 int ret = 0, i = 0;
2043 struct kmem_list3 *l3;
2044
2045 drain_cpu_caches(cachep);
2046
2047 check_irq_on();
2048 for_each_online_node(i) {
2049 l3 = cachep->nodelists[i];
2050 if (l3) {
2051 spin_lock_irq(&l3->list_lock);
2052 ret += __node_shrink(cachep, i);
2053 spin_unlock_irq(&l3->list_lock);
2054 }
2055 }
2056 return (ret ? 1 : 0);
2057}
2058
Linus Torvalds1da177e2005-04-16 15:20:36 -07002059/**
2060 * kmem_cache_shrink - Shrink a cache.
2061 * @cachep: The cache to shrink.
2062 *
2063 * Releases as many slabs as possible for a cache.
2064 * To help debugging, a zero exit status indicates all slabs were released.
2065 */
2066int kmem_cache_shrink(kmem_cache_t *cachep)
2067{
2068 if (!cachep || in_interrupt())
2069 BUG();
2070
2071 return __cache_shrink(cachep);
2072}
2073EXPORT_SYMBOL(kmem_cache_shrink);
2074
2075/**
2076 * kmem_cache_destroy - delete a cache
2077 * @cachep: the cache to destroy
2078 *
2079 * Remove a kmem_cache_t object from the slab cache.
2080 * Returns 0 on success.
2081 *
2082 * It is expected this function will be called by a module when it is
2083 * unloaded. This will remove the cache completely, and avoid a duplicate
2084 * cache being allocated each time a module is loaded and unloaded, if the
2085 * module doesn't have persistent in-kernel storage across loads and unloads.
2086 *
2087 * The cache must be empty before calling this function.
2088 *
2089 * The caller must guarantee that noone will allocate memory from the cache
2090 * during the kmem_cache_destroy().
2091 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002092int kmem_cache_destroy(kmem_cache_t *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002093{
2094 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07002095 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002096
2097 if (!cachep || in_interrupt())
2098 BUG();
2099
2100 /* Don't let CPUs to come and go */
2101 lock_cpu_hotplug();
2102
2103 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002104 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002105 /*
2106 * the chain is never empty, cache_cache is never destroyed
2107 */
2108 list_del(&cachep->next);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002109 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002110
2111 if (__cache_shrink(cachep)) {
2112 slab_error(cachep, "Can't free all objects");
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002113 mutex_lock(&cache_chain_mutex);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002114 list_add(&cachep->next, &cache_chain);
Ingo Molnarfc0abb12006-01-18 17:42:33 -08002115 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002116 unlock_cpu_hotplug();
2117 return 1;
2118 }
2119
2120 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
Paul E. McKenneyfbd568a3e2005-05-01 08:59:04 -07002121 synchronize_rcu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002122
Christoph Lametere498be72005-09-09 13:03:32 -07002123 for_each_online_cpu(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002124 kfree(cachep->array[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002125
2126 /* NUMA: free the list3 structures */
Christoph Lametere498be72005-09-09 13:03:32 -07002127 for_each_online_node(i) {
2128 if ((l3 = cachep->nodelists[i])) {
2129 kfree(l3->shared);
2130 free_alien_cache(l3->alien);
2131 kfree(l3);
2132 }
2133 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002134 kmem_cache_free(&cache_cache, cachep);
2135
2136 unlock_cpu_hotplug();
2137
2138 return 0;
2139}
2140EXPORT_SYMBOL(kmem_cache_destroy);
2141
2142/* Get the memory for a slab management obj. */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002143static struct slab *alloc_slabmgmt(kmem_cache_t *cachep, void *objp,
2144 int colour_off, gfp_t local_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002145{
2146 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002147
Linus Torvalds1da177e2005-04-16 15:20:36 -07002148 if (OFF_SLAB(cachep)) {
2149 /* Slab management obj is off-slab. */
2150 slabp = kmem_cache_alloc(cachep->slabp_cache, local_flags);
2151 if (!slabp)
2152 return NULL;
2153 } else {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002154 slabp = objp + colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002155 colour_off += cachep->slab_size;
2156 }
2157 slabp->inuse = 0;
2158 slabp->colouroff = colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002159 slabp->s_mem = objp + colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002160
2161 return slabp;
2162}
2163
2164static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2165{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002166 return (kmem_bufctl_t *) (slabp + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002167}
2168
2169static void cache_init_objs(kmem_cache_t *cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002170 struct slab *slabp, unsigned long ctor_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002171{
2172 int i;
2173
2174 for (i = 0; i < cachep->num; i++) {
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002175 void *objp = slabp->s_mem + cachep->buffer_size * i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002176#if DEBUG
2177 /* need to poison the objs? */
2178 if (cachep->flags & SLAB_POISON)
2179 poison_obj(cachep, objp, POISON_FREE);
2180 if (cachep->flags & SLAB_STORE_USER)
2181 *dbg_userword(cachep, objp) = NULL;
2182
2183 if (cachep->flags & SLAB_RED_ZONE) {
2184 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2185 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2186 }
2187 /*
2188 * Constructors are not allowed to allocate memory from
2189 * the same cache which they are a constructor for.
2190 * Otherwise, deadlock. They must also be threaded.
2191 */
2192 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002193 cachep->ctor(objp + obj_offset(cachep), cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002194 ctor_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002195
2196 if (cachep->flags & SLAB_RED_ZONE) {
2197 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2198 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002199 " end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002200 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2201 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002202 " start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002203 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002204 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002205 && cachep->flags & SLAB_POISON)
2206 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002207 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002208#else
2209 if (cachep->ctor)
2210 cachep->ctor(objp, cachep, ctor_flags);
2211#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002212 slab_bufctl(slabp)[i] = i + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002213 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002214 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002215 slabp->free = 0;
2216}
2217
Al Viro6daa0e22005-10-21 03:18:50 -04002218static void kmem_flagcheck(kmem_cache_t *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002219{
2220 if (flags & SLAB_DMA) {
2221 if (!(cachep->gfpflags & GFP_DMA))
2222 BUG();
2223 } else {
2224 if (cachep->gfpflags & GFP_DMA)
2225 BUG();
2226 }
2227}
2228
Matthew Dobson78d382d2006-02-01 03:05:47 -08002229static void *slab_get_obj(kmem_cache_t *cachep, struct slab *slabp, int nodeid)
2230{
2231 void *objp = slabp->s_mem + (slabp->free * cachep->buffer_size);
2232 kmem_bufctl_t next;
2233
2234 slabp->inuse++;
2235 next = slab_bufctl(slabp)[slabp->free];
2236#if DEBUG
2237 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2238 WARN_ON(slabp->nodeid != nodeid);
2239#endif
2240 slabp->free = next;
2241
2242 return objp;
2243}
2244
2245static void slab_put_obj(kmem_cache_t *cachep, struct slab *slabp, void *objp,
2246 int nodeid)
2247{
2248 unsigned int objnr = (unsigned)(objp-slabp->s_mem) / cachep->buffer_size;
2249
2250#if DEBUG
2251 /* Verify that the slab belongs to the intended node */
2252 WARN_ON(slabp->nodeid != nodeid);
2253
2254 if (slab_bufctl(slabp)[objnr] != BUFCTL_FREE) {
2255 printk(KERN_ERR "slab: double free detected in cache "
2256 "'%s', objp %p\n", cachep->name, objp);
2257 BUG();
2258 }
2259#endif
2260 slab_bufctl(slabp)[objnr] = slabp->free;
2261 slabp->free = objnr;
2262 slabp->inuse--;
2263}
2264
Linus Torvalds1da177e2005-04-16 15:20:36 -07002265static void set_slab_attr(kmem_cache_t *cachep, struct slab *slabp, void *objp)
2266{
2267 int i;
2268 struct page *page;
2269
2270 /* Nasty!!!!!! I hope this is OK. */
2271 i = 1 << cachep->gfporder;
2272 page = virt_to_page(objp);
2273 do {
Pekka Enberg065d41c2005-11-13 16:06:46 -08002274 page_set_cache(page, cachep);
2275 page_set_slab(page, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002276 page++;
2277 } while (--i);
2278}
2279
2280/*
2281 * Grow (by 1) the number of slabs within a cache. This is called by
2282 * kmem_cache_alloc() when there are no active objs left in a cache.
2283 */
Al Virodd0fc662005-10-07 07:46:04 +01002284static int cache_grow(kmem_cache_t *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002285{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002286 struct slab *slabp;
2287 void *objp;
2288 size_t offset;
2289 gfp_t local_flags;
2290 unsigned long ctor_flags;
Christoph Lametere498be72005-09-09 13:03:32 -07002291 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002292
2293 /* Be lazy and only check for valid flags here,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002294 * keeping it out of the critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002295 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002296 if (flags & ~(SLAB_DMA | SLAB_LEVEL_MASK | SLAB_NO_GROW))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002297 BUG();
2298 if (flags & SLAB_NO_GROW)
2299 return 0;
2300
2301 ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2302 local_flags = (flags & SLAB_LEVEL_MASK);
2303 if (!(local_flags & __GFP_WAIT))
2304 /*
2305 * Not allowed to sleep. Need to tell a constructor about
2306 * this - it might need to know...
2307 */
2308 ctor_flags |= SLAB_CTOR_ATOMIC;
2309
2310 /* About to mess with non-constant members - lock. */
2311 check_irq_off();
2312 spin_lock(&cachep->spinlock);
2313
2314 /* Get colour for the slab, and cal the next value. */
2315 offset = cachep->colour_next;
2316 cachep->colour_next++;
2317 if (cachep->colour_next >= cachep->colour)
2318 cachep->colour_next = 0;
2319 offset *= cachep->colour_off;
2320
2321 spin_unlock(&cachep->spinlock);
2322
Christoph Lametere498be72005-09-09 13:03:32 -07002323 check_irq_off();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002324 if (local_flags & __GFP_WAIT)
2325 local_irq_enable();
2326
2327 /*
2328 * The test for missing atomic flag is performed here, rather than
2329 * the more obvious place, simply to reduce the critical path length
2330 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2331 * will eventually be caught here (where it matters).
2332 */
2333 kmem_flagcheck(cachep, flags);
2334
Christoph Lametere498be72005-09-09 13:03:32 -07002335 /* Get mem for the objs.
2336 * Attempt to allocate a physical page from 'nodeid',
2337 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002338 if (!(objp = kmem_getpages(cachep, flags, nodeid)))
2339 goto failed;
2340
2341 /* Get slab management. */
2342 if (!(slabp = alloc_slabmgmt(cachep, objp, offset, local_flags)))
2343 goto opps1;
2344
Christoph Lametere498be72005-09-09 13:03:32 -07002345 slabp->nodeid = nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002346 set_slab_attr(cachep, slabp, objp);
2347
2348 cache_init_objs(cachep, slabp, ctor_flags);
2349
2350 if (local_flags & __GFP_WAIT)
2351 local_irq_disable();
2352 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002353 l3 = cachep->nodelists[nodeid];
2354 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002355
2356 /* Make slab active. */
Christoph Lametere498be72005-09-09 13:03:32 -07002357 list_add_tail(&slabp->list, &(l3->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002358 STATS_INC_GROWN(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002359 l3->free_objects += cachep->num;
2360 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002361 return 1;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002362 opps1:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002363 kmem_freepages(cachep, objp);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002364 failed:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002365 if (local_flags & __GFP_WAIT)
2366 local_irq_disable();
2367 return 0;
2368}
2369
2370#if DEBUG
2371
2372/*
2373 * Perform extra freeing checks:
2374 * - detect bad pointers.
2375 * - POISON/RED_ZONE checking
2376 * - destructor calls, for caches with POISON+dtor
2377 */
2378static void kfree_debugcheck(const void *objp)
2379{
2380 struct page *page;
2381
2382 if (!virt_addr_valid(objp)) {
2383 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002384 (unsigned long)objp);
2385 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002386 }
2387 page = virt_to_page(objp);
2388 if (!PageSlab(page)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002389 printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n",
2390 (unsigned long)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002391 BUG();
2392 }
2393}
2394
2395static void *cache_free_debugcheck(kmem_cache_t *cachep, void *objp,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002396 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002397{
2398 struct page *page;
2399 unsigned int objnr;
2400 struct slab *slabp;
2401
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002402 objp -= obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002403 kfree_debugcheck(objp);
2404 page = virt_to_page(objp);
2405
Pekka Enberg065d41c2005-11-13 16:06:46 -08002406 if (page_get_cache(page) != cachep) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002407 printk(KERN_ERR
2408 "mismatch in kmem_cache_free: expected cache %p, got %p\n",
2409 page_get_cache(page), cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002410 printk(KERN_ERR "%p is %s.\n", cachep, cachep->name);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002411 printk(KERN_ERR "%p is %s.\n", page_get_cache(page),
2412 page_get_cache(page)->name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002413 WARN_ON(1);
2414 }
Pekka Enberg065d41c2005-11-13 16:06:46 -08002415 slabp = page_get_slab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002416
2417 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002418 if (*dbg_redzone1(cachep, objp) != RED_ACTIVE
2419 || *dbg_redzone2(cachep, objp) != RED_ACTIVE) {
2420 slab_error(cachep,
2421 "double free, or memory outside"
2422 " object was overwritten");
2423 printk(KERN_ERR
2424 "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n",
2425 objp, *dbg_redzone1(cachep, objp),
2426 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002427 }
2428 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2429 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2430 }
2431 if (cachep->flags & SLAB_STORE_USER)
2432 *dbg_userword(cachep, objp) = caller;
2433
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002434 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002435
2436 BUG_ON(objnr >= cachep->num);
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002437 BUG_ON(objp != slabp->s_mem + objnr * cachep->buffer_size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002438
2439 if (cachep->flags & SLAB_DEBUG_INITIAL) {
2440 /* Need to call the slab's constructor so the
2441 * caller can perform a verify of its state (debugging).
2442 * Called without the cache-lock held.
2443 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002444 cachep->ctor(objp + obj_offset(cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002445 cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002446 }
2447 if (cachep->flags & SLAB_POISON && cachep->dtor) {
2448 /* we want to cache poison the object,
2449 * call the destruction callback
2450 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002451 cachep->dtor(objp + obj_offset(cachep), cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002452 }
2453 if (cachep->flags & SLAB_POISON) {
2454#ifdef CONFIG_DEBUG_PAGEALLOC
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002455 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002456 store_stackinfo(cachep, objp, (unsigned long)caller);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002457 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002458 cachep->buffer_size / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002459 } else {
2460 poison_obj(cachep, objp, POISON_FREE);
2461 }
2462#else
2463 poison_obj(cachep, objp, POISON_FREE);
2464#endif
2465 }
2466 return objp;
2467}
2468
2469static void check_slabp(kmem_cache_t *cachep, struct slab *slabp)
2470{
2471 kmem_bufctl_t i;
2472 int entries = 0;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002473
Linus Torvalds1da177e2005-04-16 15:20:36 -07002474 /* Check slab's freelist to see if this obj is there. */
2475 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2476 entries++;
2477 if (entries > cachep->num || i >= cachep->num)
2478 goto bad;
2479 }
2480 if (entries != cachep->num - slabp->inuse) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002481 bad:
2482 printk(KERN_ERR
2483 "slab: Internal list corruption detected in cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2484 cachep->name, cachep->num, slabp, slabp->inuse);
2485 for (i = 0;
2486 i < sizeof(slabp) + cachep->num * sizeof(kmem_bufctl_t);
2487 i++) {
2488 if ((i % 16) == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002489 printk("\n%03x:", i);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002490 printk(" %02x", ((unsigned char *)slabp)[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002491 }
2492 printk("\n");
2493 BUG();
2494 }
2495}
2496#else
2497#define kfree_debugcheck(x) do { } while(0)
2498#define cache_free_debugcheck(x,objp,z) (objp)
2499#define check_slabp(x,y) do { } while(0)
2500#endif
2501
Al Virodd0fc662005-10-07 07:46:04 +01002502static void *cache_alloc_refill(kmem_cache_t *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002503{
2504 int batchcount;
2505 struct kmem_list3 *l3;
2506 struct array_cache *ac;
2507
2508 check_irq_off();
2509 ac = ac_data(cachep);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002510 retry:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002511 batchcount = ac->batchcount;
2512 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2513 /* if there was little recent activity on this
2514 * cache, then perform only a partial refill.
2515 * Otherwise we could generate refill bouncing.
2516 */
2517 batchcount = BATCHREFILL_LIMIT;
2518 }
Christoph Lametere498be72005-09-09 13:03:32 -07002519 l3 = cachep->nodelists[numa_node_id()];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002520
Christoph Lametere498be72005-09-09 13:03:32 -07002521 BUG_ON(ac->avail > 0 || !l3);
2522 spin_lock(&l3->list_lock);
2523
Linus Torvalds1da177e2005-04-16 15:20:36 -07002524 if (l3->shared) {
2525 struct array_cache *shared_array = l3->shared;
2526 if (shared_array->avail) {
2527 if (batchcount > shared_array->avail)
2528 batchcount = shared_array->avail;
2529 shared_array->avail -= batchcount;
2530 ac->avail = batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07002531 memcpy(ac->entry,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002532 &(shared_array->entry[shared_array->avail]),
2533 sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002534 shared_array->touched = 1;
2535 goto alloc_done;
2536 }
2537 }
2538 while (batchcount > 0) {
2539 struct list_head *entry;
2540 struct slab *slabp;
2541 /* Get slab alloc is to come from. */
2542 entry = l3->slabs_partial.next;
2543 if (entry == &l3->slabs_partial) {
2544 l3->free_touched = 1;
2545 entry = l3->slabs_free.next;
2546 if (entry == &l3->slabs_free)
2547 goto must_grow;
2548 }
2549
2550 slabp = list_entry(entry, struct slab, list);
2551 check_slabp(cachep, slabp);
2552 check_spinlock_acquired(cachep);
2553 while (slabp->inuse < cachep->num && batchcount--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002554 STATS_INC_ALLOCED(cachep);
2555 STATS_INC_ACTIVE(cachep);
2556 STATS_SET_HIGH(cachep);
2557
Matthew Dobson78d382d2006-02-01 03:05:47 -08002558 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
2559 numa_node_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002560 }
2561 check_slabp(cachep, slabp);
2562
2563 /* move slabp to correct slabp list: */
2564 list_del(&slabp->list);
2565 if (slabp->free == BUFCTL_END)
2566 list_add(&slabp->list, &l3->slabs_full);
2567 else
2568 list_add(&slabp->list, &l3->slabs_partial);
2569 }
2570
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002571 must_grow:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002572 l3->free_objects -= ac->avail;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002573 alloc_done:
Christoph Lametere498be72005-09-09 13:03:32 -07002574 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002575
2576 if (unlikely(!ac->avail)) {
2577 int x;
Christoph Lametere498be72005-09-09 13:03:32 -07002578 x = cache_grow(cachep, flags, numa_node_id());
2579
Linus Torvalds1da177e2005-04-16 15:20:36 -07002580 // cache_grow can reenable interrupts, then ac could change.
2581 ac = ac_data(cachep);
2582 if (!x && ac->avail == 0) // no objects in sight? abort
2583 return NULL;
2584
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002585 if (!ac->avail) // objects refilled by interrupt?
Linus Torvalds1da177e2005-04-16 15:20:36 -07002586 goto retry;
2587 }
2588 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07002589 return ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002590}
2591
2592static inline void
Al Virodd0fc662005-10-07 07:46:04 +01002593cache_alloc_debugcheck_before(kmem_cache_t *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002594{
2595 might_sleep_if(flags & __GFP_WAIT);
2596#if DEBUG
2597 kmem_flagcheck(cachep, flags);
2598#endif
2599}
2600
2601#if DEBUG
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002602static void *cache_alloc_debugcheck_after(kmem_cache_t *cachep, gfp_t flags,
2603 void *objp, void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002604{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002605 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002606 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002607 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002608#ifdef CONFIG_DEBUG_PAGEALLOC
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002609 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002610 kernel_map_pages(virt_to_page(objp),
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002611 cachep->buffer_size / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002612 else
2613 check_poison_obj(cachep, objp);
2614#else
2615 check_poison_obj(cachep, objp);
2616#endif
2617 poison_obj(cachep, objp, POISON_INUSE);
2618 }
2619 if (cachep->flags & SLAB_STORE_USER)
2620 *dbg_userword(cachep, objp) = caller;
2621
2622 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002623 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE
2624 || *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2625 slab_error(cachep,
2626 "double free, or memory outside"
2627 " object was overwritten");
2628 printk(KERN_ERR
2629 "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n",
2630 objp, *dbg_redzone1(cachep, objp),
2631 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002632 }
2633 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2634 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2635 }
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002636 objp += obj_offset(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002637 if (cachep->ctor && cachep->flags & SLAB_POISON) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002638 unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002639
2640 if (!(flags & __GFP_WAIT))
2641 ctor_flags |= SLAB_CTOR_ATOMIC;
2642
2643 cachep->ctor(objp, cachep, ctor_flags);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002644 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002645 return objp;
2646}
2647#else
2648#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2649#endif
2650
Al Virodd0fc662005-10-07 07:46:04 +01002651static inline void *____cache_alloc(kmem_cache_t *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002652{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002653 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002654 struct array_cache *ac;
2655
Christoph Lameterdc85da12006-01-18 17:42:36 -08002656#ifdef CONFIG_NUMA
Christoph Lameter86c562a2006-01-18 17:42:37 -08002657 if (unlikely(current->mempolicy && !in_interrupt())) {
Christoph Lameterdc85da12006-01-18 17:42:36 -08002658 int nid = slab_node(current->mempolicy);
2659
2660 if (nid != numa_node_id())
2661 return __cache_alloc_node(cachep, flags, nid);
2662 }
2663#endif
2664
Alok N Kataria5c382302005-09-27 21:45:46 -07002665 check_irq_off();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002666 ac = ac_data(cachep);
2667 if (likely(ac->avail)) {
2668 STATS_INC_ALLOCHIT(cachep);
2669 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07002670 objp = ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002671 } else {
2672 STATS_INC_ALLOCMISS(cachep);
2673 objp = cache_alloc_refill(cachep, flags);
2674 }
Alok N Kataria5c382302005-09-27 21:45:46 -07002675 return objp;
2676}
2677
Al Virodd0fc662005-10-07 07:46:04 +01002678static inline void *__cache_alloc(kmem_cache_t *cachep, gfp_t flags)
Alok N Kataria5c382302005-09-27 21:45:46 -07002679{
2680 unsigned long save_flags;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002681 void *objp;
Alok N Kataria5c382302005-09-27 21:45:46 -07002682
2683 cache_alloc_debugcheck_before(cachep, flags);
2684
2685 local_irq_save(save_flags);
2686 objp = ____cache_alloc(cachep, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002687 local_irq_restore(save_flags);
Eric Dumazet34342e82005-09-03 15:55:06 -07002688 objp = cache_alloc_debugcheck_after(cachep, flags, objp,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002689 __builtin_return_address(0));
Eric Dumazet34342e82005-09-03 15:55:06 -07002690 prefetchw(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002691 return objp;
2692}
2693
Christoph Lametere498be72005-09-09 13:03:32 -07002694#ifdef CONFIG_NUMA
2695/*
2696 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002697 */
Al Viro6daa0e22005-10-21 03:18:50 -04002698static void *__cache_alloc_node(kmem_cache_t *cachep, gfp_t flags, int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07002699{
2700 struct list_head *entry;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002701 struct slab *slabp;
2702 struct kmem_list3 *l3;
2703 void *obj;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002704 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002705
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002706 l3 = cachep->nodelists[nodeid];
2707 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07002708
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002709 retry:
2710 spin_lock(&l3->list_lock);
2711 entry = l3->slabs_partial.next;
2712 if (entry == &l3->slabs_partial) {
2713 l3->free_touched = 1;
2714 entry = l3->slabs_free.next;
2715 if (entry == &l3->slabs_free)
2716 goto must_grow;
2717 }
Christoph Lametere498be72005-09-09 13:03:32 -07002718
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002719 slabp = list_entry(entry, struct slab, list);
2720 check_spinlock_acquired_node(cachep, nodeid);
2721 check_slabp(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07002722
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002723 STATS_INC_NODEALLOCS(cachep);
2724 STATS_INC_ACTIVE(cachep);
2725 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002726
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002727 BUG_ON(slabp->inuse == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07002728
Matthew Dobson78d382d2006-02-01 03:05:47 -08002729 obj = slab_get_obj(cachep, slabp, nodeid);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002730 check_slabp(cachep, slabp);
2731 l3->free_objects--;
2732 /* move slabp to correct slabp list: */
2733 list_del(&slabp->list);
Christoph Lametere498be72005-09-09 13:03:32 -07002734
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002735 if (slabp->free == BUFCTL_END) {
2736 list_add(&slabp->list, &l3->slabs_full);
2737 } else {
2738 list_add(&slabp->list, &l3->slabs_partial);
2739 }
Christoph Lametere498be72005-09-09 13:03:32 -07002740
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002741 spin_unlock(&l3->list_lock);
2742 goto done;
Christoph Lametere498be72005-09-09 13:03:32 -07002743
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002744 must_grow:
2745 spin_unlock(&l3->list_lock);
2746 x = cache_grow(cachep, flags, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07002747
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002748 if (!x)
2749 return NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07002750
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002751 goto retry;
2752 done:
2753 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07002754}
2755#endif
2756
2757/*
2758 * Caller needs to acquire correct kmem_list's list_lock
2759 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002760static void free_block(kmem_cache_t *cachep, void **objpp, int nr_objects,
2761 int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002762{
2763 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07002764 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002765
2766 for (i = 0; i < nr_objects; i++) {
2767 void *objp = objpp[i];
2768 struct slab *slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002769
Pekka Enberg065d41c2005-11-13 16:06:46 -08002770 slabp = page_get_slab(virt_to_page(objp));
Christoph Lameterff694162005-09-22 21:44:02 -07002771 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002772 list_del(&slabp->list);
Christoph Lameterff694162005-09-22 21:44:02 -07002773 check_spinlock_acquired_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002774 check_slabp(cachep, slabp);
Matthew Dobson78d382d2006-02-01 03:05:47 -08002775 slab_put_obj(cachep, slabp, objp, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002776 STATS_DEC_ACTIVE(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002777 l3->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002778 check_slabp(cachep, slabp);
2779
2780 /* fixup slab chains */
2781 if (slabp->inuse == 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07002782 if (l3->free_objects > l3->free_limit) {
2783 l3->free_objects -= cachep->num;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002784 slab_destroy(cachep, slabp);
2785 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07002786 list_add(&slabp->list, &l3->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002787 }
2788 } else {
2789 /* Unconditionally move a slab to the end of the
2790 * partial list on free - maximum time for the
2791 * other objects to be freed, too.
2792 */
Christoph Lametere498be72005-09-09 13:03:32 -07002793 list_add_tail(&slabp->list, &l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002794 }
2795 }
2796}
2797
2798static void cache_flusharray(kmem_cache_t *cachep, struct array_cache *ac)
2799{
2800 int batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07002801 struct kmem_list3 *l3;
Christoph Lameterff694162005-09-22 21:44:02 -07002802 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002803
2804 batchcount = ac->batchcount;
2805#if DEBUG
2806 BUG_ON(!batchcount || batchcount > ac->avail);
2807#endif
2808 check_irq_off();
Christoph Lameterff694162005-09-22 21:44:02 -07002809 l3 = cachep->nodelists[node];
Christoph Lametere498be72005-09-09 13:03:32 -07002810 spin_lock(&l3->list_lock);
2811 if (l3->shared) {
2812 struct array_cache *shared_array = l3->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002813 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002814 if (max) {
2815 if (batchcount > max)
2816 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07002817 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002818 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002819 shared_array->avail += batchcount;
2820 goto free_done;
2821 }
2822 }
2823
Christoph Lameterff694162005-09-22 21:44:02 -07002824 free_block(cachep, ac->entry, batchcount, node);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002825 free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002826#if STATS
2827 {
2828 int i = 0;
2829 struct list_head *p;
2830
Christoph Lametere498be72005-09-09 13:03:32 -07002831 p = l3->slabs_free.next;
2832 while (p != &(l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002833 struct slab *slabp;
2834
2835 slabp = list_entry(p, struct slab, list);
2836 BUG_ON(slabp->inuse);
2837
2838 i++;
2839 p = p->next;
2840 }
2841 STATS_SET_FREEABLE(cachep, i);
2842 }
2843#endif
Christoph Lametere498be72005-09-09 13:03:32 -07002844 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002845 ac->avail -= batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07002846 memmove(ac->entry, &(ac->entry[batchcount]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002847 sizeof(void *) * ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002848}
2849
2850/*
2851 * __cache_free
2852 * Release an obj back to its cache. If the obj has a constructed
2853 * state, it must be in this state _before_ it is released.
2854 *
2855 * Called with disabled ints.
2856 */
2857static inline void __cache_free(kmem_cache_t *cachep, void *objp)
2858{
2859 struct array_cache *ac = ac_data(cachep);
2860
2861 check_irq_off();
2862 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
2863
Christoph Lametere498be72005-09-09 13:03:32 -07002864 /* Make sure we are not freeing a object from another
2865 * node to the array cache on this cpu.
2866 */
2867#ifdef CONFIG_NUMA
2868 {
2869 struct slab *slabp;
Pekka Enberg065d41c2005-11-13 16:06:46 -08002870 slabp = page_get_slab(virt_to_page(objp));
Christoph Lametere498be72005-09-09 13:03:32 -07002871 if (unlikely(slabp->nodeid != numa_node_id())) {
2872 struct array_cache *alien = NULL;
2873 int nodeid = slabp->nodeid;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002874 struct kmem_list3 *l3 =
2875 cachep->nodelists[numa_node_id()];
Christoph Lametere498be72005-09-09 13:03:32 -07002876
2877 STATS_INC_NODEFREES(cachep);
2878 if (l3->alien && l3->alien[nodeid]) {
2879 alien = l3->alien[nodeid];
2880 spin_lock(&alien->lock);
2881 if (unlikely(alien->avail == alien->limit))
2882 __drain_alien_cache(cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002883 alien, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07002884 alien->entry[alien->avail++] = objp;
2885 spin_unlock(&alien->lock);
2886 } else {
2887 spin_lock(&(cachep->nodelists[nodeid])->
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002888 list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -07002889 free_block(cachep, &objp, 1, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07002890 spin_unlock(&(cachep->nodelists[nodeid])->
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002891 list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002892 }
2893 return;
2894 }
2895 }
2896#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002897 if (likely(ac->avail < ac->limit)) {
2898 STATS_INC_FREEHIT(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002899 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002900 return;
2901 } else {
2902 STATS_INC_FREEMISS(cachep);
2903 cache_flusharray(cachep, ac);
Christoph Lametere498be72005-09-09 13:03:32 -07002904 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002905 }
2906}
2907
2908/**
2909 * kmem_cache_alloc - Allocate an object
2910 * @cachep: The cache to allocate from.
2911 * @flags: See kmalloc().
2912 *
2913 * Allocate an object from this cache. The flags are only relevant
2914 * if the cache has no available objects.
2915 */
Al Virodd0fc662005-10-07 07:46:04 +01002916void *kmem_cache_alloc(kmem_cache_t *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002917{
2918 return __cache_alloc(cachep, flags);
2919}
2920EXPORT_SYMBOL(kmem_cache_alloc);
2921
2922/**
2923 * kmem_ptr_validate - check if an untrusted pointer might
2924 * be a slab entry.
2925 * @cachep: the cache we're checking against
2926 * @ptr: pointer to validate
2927 *
2928 * This verifies that the untrusted pointer looks sane:
2929 * it is _not_ a guarantee that the pointer is actually
2930 * part of the slab cache in question, but it at least
2931 * validates that the pointer can be dereferenced and
2932 * looks half-way sane.
2933 *
2934 * Currently only used for dentry validation.
2935 */
2936int fastcall kmem_ptr_validate(kmem_cache_t *cachep, void *ptr)
2937{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002938 unsigned long addr = (unsigned long)ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002939 unsigned long min_addr = PAGE_OFFSET;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002940 unsigned long align_mask = BYTES_PER_WORD - 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08002941 unsigned long size = cachep->buffer_size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002942 struct page *page;
2943
2944 if (unlikely(addr < min_addr))
2945 goto out;
2946 if (unlikely(addr > (unsigned long)high_memory - size))
2947 goto out;
2948 if (unlikely(addr & align_mask))
2949 goto out;
2950 if (unlikely(!kern_addr_valid(addr)))
2951 goto out;
2952 if (unlikely(!kern_addr_valid(addr + size - 1)))
2953 goto out;
2954 page = virt_to_page(ptr);
2955 if (unlikely(!PageSlab(page)))
2956 goto out;
Pekka Enberg065d41c2005-11-13 16:06:46 -08002957 if (unlikely(page_get_cache(page) != cachep))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002958 goto out;
2959 return 1;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002960 out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002961 return 0;
2962}
2963
2964#ifdef CONFIG_NUMA
2965/**
2966 * kmem_cache_alloc_node - Allocate an object on the specified node
2967 * @cachep: The cache to allocate from.
2968 * @flags: See kmalloc().
2969 * @nodeid: node number of the target node.
2970 *
2971 * Identical to kmem_cache_alloc, except that this function is slow
2972 * and can sleep. And it will allocate memory on the given node, which
2973 * can improve the performance for cpu bound structures.
Christoph Lametere498be72005-09-09 13:03:32 -07002974 * New and improved: it will now make sure that the object gets
2975 * put on the correct node list so that there is no false sharing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002976 */
Al Virodd0fc662005-10-07 07:46:04 +01002977void *kmem_cache_alloc_node(kmem_cache_t *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002978{
Christoph Lametere498be72005-09-09 13:03:32 -07002979 unsigned long save_flags;
2980 void *ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002981
Christoph Lametere498be72005-09-09 13:03:32 -07002982 cache_alloc_debugcheck_before(cachep, flags);
2983 local_irq_save(save_flags);
Christoph Lameter18f820f2006-02-01 03:05:43 -08002984
2985 if (nodeid == -1 || nodeid == numa_node_id() ||
2986 !cachep->nodelists[nodeid])
Alok N Kataria5c382302005-09-27 21:45:46 -07002987 ptr = ____cache_alloc(cachep, flags);
2988 else
2989 ptr = __cache_alloc_node(cachep, flags, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07002990 local_irq_restore(save_flags);
Christoph Lameter18f820f2006-02-01 03:05:43 -08002991
2992 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr,
2993 __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002994
Christoph Lametere498be72005-09-09 13:03:32 -07002995 return ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002996}
2997EXPORT_SYMBOL(kmem_cache_alloc_node);
2998
Al Virodd0fc662005-10-07 07:46:04 +01002999void *kmalloc_node(size_t size, gfp_t flags, int node)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003000{
3001 kmem_cache_t *cachep;
3002
3003 cachep = kmem_find_general_cachep(size, flags);
3004 if (unlikely(cachep == NULL))
3005 return NULL;
3006 return kmem_cache_alloc_node(cachep, flags, node);
3007}
3008EXPORT_SYMBOL(kmalloc_node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003009#endif
3010
3011/**
3012 * kmalloc - allocate memory
3013 * @size: how many bytes of memory are required.
3014 * @flags: the type of memory to allocate.
3015 *
3016 * kmalloc is the normal method of allocating memory
3017 * in the kernel.
3018 *
3019 * The @flags argument may be one of:
3020 *
3021 * %GFP_USER - Allocate memory on behalf of user. May sleep.
3022 *
3023 * %GFP_KERNEL - Allocate normal kernel ram. May sleep.
3024 *
3025 * %GFP_ATOMIC - Allocation will not sleep. Use inside interrupt handlers.
3026 *
3027 * Additionally, the %GFP_DMA flag may be set to indicate the memory
3028 * must be suitable for DMA. This can mean different things on different
3029 * platforms. For example, on i386, it means that the memory must come
3030 * from the first 16MB.
3031 */
Al Virodd0fc662005-10-07 07:46:04 +01003032void *__kmalloc(size_t size, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003033{
3034 kmem_cache_t *cachep;
3035
Manfred Spraul97e2bde2005-05-01 08:58:38 -07003036 /* If you want to save a few bytes .text space: replace
3037 * __ with kmem_.
3038 * Then kmalloc uses the uninlined functions instead of the inline
3039 * functions.
3040 */
3041 cachep = __find_general_cachep(size, flags);
Andrew Mortondbdb9042005-09-23 13:24:10 -07003042 if (unlikely(cachep == NULL))
3043 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003044 return __cache_alloc(cachep, flags);
3045}
3046EXPORT_SYMBOL(__kmalloc);
3047
3048#ifdef CONFIG_SMP
3049/**
3050 * __alloc_percpu - allocate one copy of the object for every present
3051 * cpu in the system, zeroing them.
3052 * Objects should be dereferenced using the per_cpu_ptr macro only.
3053 *
3054 * @size: how many bytes of memory are required.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003055 */
Pekka Enbergf9f75002006-01-08 01:00:33 -08003056void *__alloc_percpu(size_t size)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003057{
3058 int i;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003059 struct percpu_data *pdata = kmalloc(sizeof(*pdata), GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003060
3061 if (!pdata)
3062 return NULL;
3063
Christoph Lametere498be72005-09-09 13:03:32 -07003064 /*
3065 * Cannot use for_each_online_cpu since a cpu may come online
3066 * and we have no way of figuring out how to fix the array
3067 * that we have allocated then....
3068 */
3069 for_each_cpu(i) {
3070 int node = cpu_to_node(i);
3071
3072 if (node_online(node))
3073 pdata->ptrs[i] = kmalloc_node(size, GFP_KERNEL, node);
3074 else
3075 pdata->ptrs[i] = kmalloc(size, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003076
3077 if (!pdata->ptrs[i])
3078 goto unwind_oom;
3079 memset(pdata->ptrs[i], 0, size);
3080 }
3081
3082 /* Catch derefs w/o wrappers */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003083 return (void *)(~(unsigned long)pdata);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003084
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003085 unwind_oom:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003086 while (--i >= 0) {
3087 if (!cpu_possible(i))
3088 continue;
3089 kfree(pdata->ptrs[i]);
3090 }
3091 kfree(pdata);
3092 return NULL;
3093}
3094EXPORT_SYMBOL(__alloc_percpu);
3095#endif
3096
3097/**
3098 * kmem_cache_free - Deallocate an object
3099 * @cachep: The cache the allocation was from.
3100 * @objp: The previously allocated object.
3101 *
3102 * Free an object which was previously allocated from this
3103 * cache.
3104 */
3105void kmem_cache_free(kmem_cache_t *cachep, void *objp)
3106{
3107 unsigned long flags;
3108
3109 local_irq_save(flags);
3110 __cache_free(cachep, objp);
3111 local_irq_restore(flags);
3112}
3113EXPORT_SYMBOL(kmem_cache_free);
3114
3115/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003116 * kfree - free previously allocated memory
3117 * @objp: pointer returned by kmalloc.
3118 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003119 * If @objp is NULL, no operation is performed.
3120 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003121 * Don't free memory not originally allocated by kmalloc()
3122 * or you will run into trouble.
3123 */
3124void kfree(const void *objp)
3125{
3126 kmem_cache_t *c;
3127 unsigned long flags;
3128
3129 if (unlikely(!objp))
3130 return;
3131 local_irq_save(flags);
3132 kfree_debugcheck(objp);
Pekka Enberg065d41c2005-11-13 16:06:46 -08003133 c = page_get_cache(virt_to_page(objp));
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003134 mutex_debug_check_no_locks_freed(objp, obj_size(c));
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003135 __cache_free(c, (void *)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003136 local_irq_restore(flags);
3137}
3138EXPORT_SYMBOL(kfree);
3139
3140#ifdef CONFIG_SMP
3141/**
3142 * free_percpu - free previously allocated percpu memory
3143 * @objp: pointer returned by alloc_percpu.
3144 *
3145 * Don't free memory not originally allocated by alloc_percpu()
3146 * The complemented objp is to check for that.
3147 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003148void free_percpu(const void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003149{
3150 int i;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003151 struct percpu_data *p = (struct percpu_data *)(~(unsigned long)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003152
Christoph Lametere498be72005-09-09 13:03:32 -07003153 /*
3154 * We allocate for all cpus so we cannot use for online cpu here.
3155 */
3156 for_each_cpu(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003157 kfree(p->ptrs[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003158 kfree(p);
3159}
3160EXPORT_SYMBOL(free_percpu);
3161#endif
3162
3163unsigned int kmem_cache_size(kmem_cache_t *cachep)
3164{
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003165 return obj_size(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003166}
3167EXPORT_SYMBOL(kmem_cache_size);
3168
Arnaldo Carvalho de Melo19449722005-06-18 22:46:19 -07003169const char *kmem_cache_name(kmem_cache_t *cachep)
3170{
3171 return cachep->name;
3172}
3173EXPORT_SYMBOL_GPL(kmem_cache_name);
3174
Christoph Lametere498be72005-09-09 13:03:32 -07003175/*
3176 * This initializes kmem_list3 for all nodes.
3177 */
3178static int alloc_kmemlist(kmem_cache_t *cachep)
3179{
3180 int node;
3181 struct kmem_list3 *l3;
3182 int err = 0;
3183
3184 for_each_online_node(node) {
3185 struct array_cache *nc = NULL, *new;
3186 struct array_cache **new_alien = NULL;
3187#ifdef CONFIG_NUMA
3188 if (!(new_alien = alloc_alien_cache(node, cachep->limit)))
3189 goto fail;
3190#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003191 if (!(new = alloc_arraycache(node, (cachep->shared *
3192 cachep->batchcount),
3193 0xbaadf00d)))
Christoph Lametere498be72005-09-09 13:03:32 -07003194 goto fail;
3195 if ((l3 = cachep->nodelists[node])) {
3196
3197 spin_lock_irq(&l3->list_lock);
3198
3199 if ((nc = cachep->nodelists[node]->shared))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003200 free_block(cachep, nc->entry, nc->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07003201
3202 l3->shared = new;
3203 if (!cachep->nodelists[node]->alien) {
3204 l3->alien = new_alien;
3205 new_alien = NULL;
3206 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003207 l3->free_limit = (1 + nr_cpus_node(node)) *
3208 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003209 spin_unlock_irq(&l3->list_lock);
3210 kfree(nc);
3211 free_alien_cache(new_alien);
3212 continue;
3213 }
3214 if (!(l3 = kmalloc_node(sizeof(struct kmem_list3),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003215 GFP_KERNEL, node)))
Christoph Lametere498be72005-09-09 13:03:32 -07003216 goto fail;
3217
3218 kmem_list3_init(l3);
3219 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003220 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -07003221 l3->shared = new;
3222 l3->alien = new_alien;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003223 l3->free_limit = (1 + nr_cpus_node(node)) *
3224 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003225 cachep->nodelists[node] = l3;
3226 }
3227 return err;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003228 fail:
Christoph Lametere498be72005-09-09 13:03:32 -07003229 err = -ENOMEM;
3230 return err;
3231}
3232
Linus Torvalds1da177e2005-04-16 15:20:36 -07003233struct ccupdate_struct {
3234 kmem_cache_t *cachep;
3235 struct array_cache *new[NR_CPUS];
3236};
3237
3238static void do_ccupdate_local(void *info)
3239{
3240 struct ccupdate_struct *new = (struct ccupdate_struct *)info;
3241 struct array_cache *old;
3242
3243 check_irq_off();
3244 old = ac_data(new->cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003245
Linus Torvalds1da177e2005-04-16 15:20:36 -07003246 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3247 new->new[smp_processor_id()] = old;
3248}
3249
Linus Torvalds1da177e2005-04-16 15:20:36 -07003250static int do_tune_cpucache(kmem_cache_t *cachep, int limit, int batchcount,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003251 int shared)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003252{
3253 struct ccupdate_struct new;
Christoph Lametere498be72005-09-09 13:03:32 -07003254 int i, err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003255
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003256 memset(&new.new, 0, sizeof(new.new));
Christoph Lametere498be72005-09-09 13:03:32 -07003257 for_each_online_cpu(i) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003258 new.new[i] =
3259 alloc_arraycache(cpu_to_node(i), limit, batchcount);
Christoph Lametere498be72005-09-09 13:03:32 -07003260 if (!new.new[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003261 for (i--; i >= 0; i--)
3262 kfree(new.new[i]);
Christoph Lametere498be72005-09-09 13:03:32 -07003263 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003264 }
3265 }
3266 new.cachep = cachep;
3267
3268 smp_call_function_all_cpus(do_ccupdate_local, (void *)&new);
Christoph Lametere498be72005-09-09 13:03:32 -07003269
Linus Torvalds1da177e2005-04-16 15:20:36 -07003270 check_irq_on();
3271 spin_lock_irq(&cachep->spinlock);
3272 cachep->batchcount = batchcount;
3273 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07003274 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003275 spin_unlock_irq(&cachep->spinlock);
3276
Christoph Lametere498be72005-09-09 13:03:32 -07003277 for_each_online_cpu(i) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003278 struct array_cache *ccold = new.new[i];
3279 if (!ccold)
3280 continue;
Christoph Lametere498be72005-09-09 13:03:32 -07003281 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -07003282 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
Christoph Lametere498be72005-09-09 13:03:32 -07003283 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003284 kfree(ccold);
3285 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003286
Christoph Lametere498be72005-09-09 13:03:32 -07003287 err = alloc_kmemlist(cachep);
3288 if (err) {
3289 printk(KERN_ERR "alloc_kmemlist failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003290 cachep->name, -err);
Christoph Lametere498be72005-09-09 13:03:32 -07003291 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003292 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003293 return 0;
3294}
3295
Linus Torvalds1da177e2005-04-16 15:20:36 -07003296static void enable_cpucache(kmem_cache_t *cachep)
3297{
3298 int err;
3299 int limit, shared;
3300
3301 /* The head array serves three purposes:
3302 * - create a LIFO ordering, i.e. return objects that are cache-warm
3303 * - reduce the number of spinlock operations.
3304 * - reduce the number of linked list operations on the slab and
3305 * bufctl chains: array operations are cheaper.
3306 * The numbers are guessed, we should auto-tune as described by
3307 * Bonwick.
3308 */
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003309 if (cachep->buffer_size > 131072)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003310 limit = 1;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003311 else if (cachep->buffer_size > PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003312 limit = 8;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003313 else if (cachep->buffer_size > 1024)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003314 limit = 24;
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003315 else if (cachep->buffer_size > 256)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003316 limit = 54;
3317 else
3318 limit = 120;
3319
3320 /* Cpu bound tasks (e.g. network routing) can exhibit cpu bound
3321 * allocation behaviour: Most allocs on one cpu, most free operations
3322 * on another cpu. For these cases, an efficient object passing between
3323 * cpus is necessary. This is provided by a shared array. The array
3324 * replaces Bonwick's magazine layer.
3325 * On uniprocessor, it's functionally equivalent (but less efficient)
3326 * to a larger limit. Thus disabled by default.
3327 */
3328 shared = 0;
3329#ifdef CONFIG_SMP
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003330 if (cachep->buffer_size <= PAGE_SIZE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003331 shared = 8;
3332#endif
3333
3334#if DEBUG
3335 /* With debugging enabled, large batchcount lead to excessively
3336 * long periods with disabled local interrupts. Limit the
3337 * batchcount
3338 */
3339 if (limit > 32)
3340 limit = 32;
3341#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003342 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003343 if (err)
3344 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003345 cachep->name, -err);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003346}
3347
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003348static void drain_array_locked(kmem_cache_t *cachep, struct array_cache *ac,
3349 int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003350{
3351 int tofree;
3352
Christoph Lametere498be72005-09-09 13:03:32 -07003353 check_spinlock_acquired_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003354 if (ac->touched && !force) {
3355 ac->touched = 0;
3356 } else if (ac->avail) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003357 tofree = force ? ac->avail : (ac->limit + 4) / 5;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003358 if (tofree > ac->avail) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003359 tofree = (ac->avail + 1) / 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003360 }
Christoph Lameterff694162005-09-22 21:44:02 -07003361 free_block(cachep, ac->entry, tofree, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003362 ac->avail -= tofree;
Christoph Lametere498be72005-09-09 13:03:32 -07003363 memmove(ac->entry, &(ac->entry[tofree]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003364 sizeof(void *) * ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003365 }
3366}
3367
3368/**
3369 * cache_reap - Reclaim memory from caches.
Randy Dunlap1e5d5332005-11-07 01:01:06 -08003370 * @unused: unused parameter
Linus Torvalds1da177e2005-04-16 15:20:36 -07003371 *
3372 * Called from workqueue/eventd every few seconds.
3373 * Purpose:
3374 * - clear the per-cpu caches for this CPU.
3375 * - return freeable pages to the main free memory pool.
3376 *
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003377 * If we cannot acquire the cache chain mutex then just give up - we'll
Linus Torvalds1da177e2005-04-16 15:20:36 -07003378 * try again on the next iteration.
3379 */
3380static void cache_reap(void *unused)
3381{
3382 struct list_head *walk;
Christoph Lametere498be72005-09-09 13:03:32 -07003383 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003384
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003385 if (!mutex_trylock(&cache_chain_mutex)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003386 /* Give up. Setup the next iteration. */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003387 schedule_delayed_work(&__get_cpu_var(reap_work),
3388 REAPTIMEOUT_CPUC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003389 return;
3390 }
3391
3392 list_for_each(walk, &cache_chain) {
3393 kmem_cache_t *searchp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003394 struct list_head *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003395 int tofree;
3396 struct slab *slabp;
3397
3398 searchp = list_entry(walk, kmem_cache_t, next);
3399
3400 if (searchp->flags & SLAB_NO_REAP)
3401 goto next;
3402
3403 check_irq_on();
3404
Christoph Lametere498be72005-09-09 13:03:32 -07003405 l3 = searchp->nodelists[numa_node_id()];
3406 if (l3->alien)
3407 drain_alien_cache(searchp, l3);
3408 spin_lock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003409
Christoph Lametere498be72005-09-09 13:03:32 -07003410 drain_array_locked(searchp, ac_data(searchp), 0,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003411 numa_node_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -07003412
Christoph Lametere498be72005-09-09 13:03:32 -07003413 if (time_after(l3->next_reap, jiffies))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003414 goto next_unlock;
3415
Christoph Lametere498be72005-09-09 13:03:32 -07003416 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003417
Christoph Lametere498be72005-09-09 13:03:32 -07003418 if (l3->shared)
3419 drain_array_locked(searchp, l3->shared, 0,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003420 numa_node_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -07003421
Christoph Lametere498be72005-09-09 13:03:32 -07003422 if (l3->free_touched) {
3423 l3->free_touched = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003424 goto next_unlock;
3425 }
3426
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003427 tofree =
3428 (l3->free_limit + 5 * searchp->num -
3429 1) / (5 * searchp->num);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003430 do {
Christoph Lametere498be72005-09-09 13:03:32 -07003431 p = l3->slabs_free.next;
3432 if (p == &(l3->slabs_free))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003433 break;
3434
3435 slabp = list_entry(p, struct slab, list);
3436 BUG_ON(slabp->inuse);
3437 list_del(&slabp->list);
3438 STATS_INC_REAPED(searchp);
3439
3440 /* Safe to drop the lock. The slab is no longer
3441 * linked to the cache.
3442 * searchp cannot disappear, we hold
3443 * cache_chain_lock
3444 */
Christoph Lametere498be72005-09-09 13:03:32 -07003445 l3->free_objects -= searchp->num;
3446 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003447 slab_destroy(searchp, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07003448 spin_lock_irq(&l3->list_lock);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003449 } while (--tofree > 0);
3450 next_unlock:
Christoph Lametere498be72005-09-09 13:03:32 -07003451 spin_unlock_irq(&l3->list_lock);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003452 next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003453 cond_resched();
3454 }
3455 check_irq_on();
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003456 mutex_unlock(&cache_chain_mutex);
Christoph Lameter4ae7c032005-06-21 17:14:57 -07003457 drain_remote_pages();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003458 /* Setup the next iteration */
Manfred Spraulcd61ef62005-11-07 00:58:02 -08003459 schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003460}
3461
3462#ifdef CONFIG_PROC_FS
3463
Pekka Enberg85289f92006-01-08 01:00:36 -08003464static void print_slabinfo_header(struct seq_file *m)
3465{
3466 /*
3467 * Output format version, so at least we can change it
3468 * without _too_ many complaints.
3469 */
3470#if STATS
3471 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
3472#else
3473 seq_puts(m, "slabinfo - version: 2.1\n");
3474#endif
3475 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
3476 "<objperslab> <pagesperslab>");
3477 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
3478 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
3479#if STATS
3480 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
3481 "<error> <maxfreeable> <nodeallocs> <remotefrees>");
3482 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
3483#endif
3484 seq_putc(m, '\n');
3485}
3486
Linus Torvalds1da177e2005-04-16 15:20:36 -07003487static void *s_start(struct seq_file *m, loff_t *pos)
3488{
3489 loff_t n = *pos;
3490 struct list_head *p;
3491
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003492 mutex_lock(&cache_chain_mutex);
Pekka Enberg85289f92006-01-08 01:00:36 -08003493 if (!n)
3494 print_slabinfo_header(m);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003495 p = cache_chain.next;
3496 while (n--) {
3497 p = p->next;
3498 if (p == &cache_chain)
3499 return NULL;
3500 }
3501 return list_entry(p, kmem_cache_t, next);
3502}
3503
3504static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3505{
3506 kmem_cache_t *cachep = p;
3507 ++*pos;
3508 return cachep->next.next == &cache_chain ? NULL
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003509 : list_entry(cachep->next.next, kmem_cache_t, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003510}
3511
3512static void s_stop(struct seq_file *m, void *p)
3513{
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003514 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003515}
3516
3517static int s_show(struct seq_file *m, void *p)
3518{
3519 kmem_cache_t *cachep = p;
3520 struct list_head *q;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003521 struct slab *slabp;
3522 unsigned long active_objs;
3523 unsigned long num_objs;
3524 unsigned long active_slabs = 0;
3525 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07003526 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003527 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003528 int node;
3529 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003530
3531 check_irq_on();
3532 spin_lock_irq(&cachep->spinlock);
3533 active_objs = 0;
3534 num_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07003535 for_each_online_node(node) {
3536 l3 = cachep->nodelists[node];
3537 if (!l3)
3538 continue;
3539
3540 spin_lock(&l3->list_lock);
3541
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003542 list_for_each(q, &l3->slabs_full) {
Christoph Lametere498be72005-09-09 13:03:32 -07003543 slabp = list_entry(q, struct slab, list);
3544 if (slabp->inuse != cachep->num && !error)
3545 error = "slabs_full accounting error";
3546 active_objs += cachep->num;
3547 active_slabs++;
3548 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003549 list_for_each(q, &l3->slabs_partial) {
Christoph Lametere498be72005-09-09 13:03:32 -07003550 slabp = list_entry(q, struct slab, list);
3551 if (slabp->inuse == cachep->num && !error)
3552 error = "slabs_partial inuse accounting error";
3553 if (!slabp->inuse && !error)
3554 error = "slabs_partial/inuse accounting error";
3555 active_objs += slabp->inuse;
3556 active_slabs++;
3557 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003558 list_for_each(q, &l3->slabs_free) {
Christoph Lametere498be72005-09-09 13:03:32 -07003559 slabp = list_entry(q, struct slab, list);
3560 if (slabp->inuse && !error)
3561 error = "slabs_free/inuse accounting error";
3562 num_slabs++;
3563 }
3564 free_objects += l3->free_objects;
3565 shared_avail += l3->shared->avail;
3566
3567 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003568 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003569 num_slabs += active_slabs;
3570 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003571 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003572 error = "free_objects accounting error";
3573
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003574 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003575 if (error)
3576 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
3577
3578 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003579 name, active_objs, num_objs, cachep->buffer_size,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003580 cachep->num, (1 << cachep->gfporder));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003581 seq_printf(m, " : tunables %4u %4u %4u",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003582 cachep->limit, cachep->batchcount, cachep->shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003583 seq_printf(m, " : slabdata %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003584 active_slabs, num_slabs, shared_avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003585#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003586 { /* list3 stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003587 unsigned long high = cachep->high_mark;
3588 unsigned long allocs = cachep->num_allocations;
3589 unsigned long grown = cachep->grown;
3590 unsigned long reaped = cachep->reaped;
3591 unsigned long errors = cachep->errors;
3592 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003593 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07003594 unsigned long node_frees = cachep->node_frees;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003595
Christoph Lametere498be72005-09-09 13:03:32 -07003596 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003597 %4lu %4lu %4lu %4lu", allocs, high, grown, reaped, errors, max_freeable, node_allocs, node_frees);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003598 }
3599 /* cpu stats */
3600 {
3601 unsigned long allochit = atomic_read(&cachep->allochit);
3602 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
3603 unsigned long freehit = atomic_read(&cachep->freehit);
3604 unsigned long freemiss = atomic_read(&cachep->freemiss);
3605
3606 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003607 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003608 }
3609#endif
3610 seq_putc(m, '\n');
3611 spin_unlock_irq(&cachep->spinlock);
3612 return 0;
3613}
3614
3615/*
3616 * slabinfo_op - iterator that generates /proc/slabinfo
3617 *
3618 * Output layout:
3619 * cache-name
3620 * num-active-objs
3621 * total-objs
3622 * object size
3623 * num-active-slabs
3624 * total-slabs
3625 * num-pages-per-slab
3626 * + further values on SMP and with statistics enabled
3627 */
3628
3629struct seq_operations slabinfo_op = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003630 .start = s_start,
3631 .next = s_next,
3632 .stop = s_stop,
3633 .show = s_show,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003634};
3635
3636#define MAX_SLABINFO_WRITE 128
3637/**
3638 * slabinfo_write - Tuning for the slab allocator
3639 * @file: unused
3640 * @buffer: user buffer
3641 * @count: data length
3642 * @ppos: unused
3643 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003644ssize_t slabinfo_write(struct file *file, const char __user * buffer,
3645 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003646{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003647 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003648 int limit, batchcount, shared, res;
3649 struct list_head *p;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003650
Linus Torvalds1da177e2005-04-16 15:20:36 -07003651 if (count > MAX_SLABINFO_WRITE)
3652 return -EINVAL;
3653 if (copy_from_user(&kbuf, buffer, count))
3654 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003655 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07003656
3657 tmp = strchr(kbuf, ' ');
3658 if (!tmp)
3659 return -EINVAL;
3660 *tmp = '\0';
3661 tmp++;
3662 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
3663 return -EINVAL;
3664
3665 /* Find the cache in the chain of caches. */
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003666 mutex_lock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003667 res = -EINVAL;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003668 list_for_each(p, &cache_chain) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003669 kmem_cache_t *cachep = list_entry(p, kmem_cache_t, next);
3670
3671 if (!strcmp(cachep->name, kbuf)) {
3672 if (limit < 1 ||
3673 batchcount < 1 ||
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003674 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07003675 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003676 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07003677 res = do_tune_cpucache(cachep, limit,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003678 batchcount, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003679 }
3680 break;
3681 }
3682 }
Ingo Molnarfc0abb12006-01-18 17:42:33 -08003683 mutex_unlock(&cache_chain_mutex);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003684 if (res >= 0)
3685 res = count;
3686 return res;
3687}
3688#endif
3689
Manfred Spraul00e145b2005-09-03 15:55:07 -07003690/**
3691 * ksize - get the actual amount of memory allocated for a given object
3692 * @objp: Pointer to the object
3693 *
3694 * kmalloc may internally round up allocations and return more memory
3695 * than requested. ksize() can be used to determine the actual amount of
3696 * memory allocated. The caller may use this additional memory, even though
3697 * a smaller amount of memory was initially specified with the kmalloc call.
3698 * The caller must guarantee that objp points to a valid object previously
3699 * allocated with either kmalloc() or kmem_cache_alloc(). The object
3700 * must not be freed during the duration of the call.
3701 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003702unsigned int ksize(const void *objp)
3703{
Manfred Spraul00e145b2005-09-03 15:55:07 -07003704 if (unlikely(objp == NULL))
3705 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003706
Manfred Spraul3dafccf2006-02-01 03:05:42 -08003707 return obj_size(page_get_cache(virt_to_page(objp)));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003708}