blob: 3d3b5a46854f5dcfef4bf8a920c81c1101d05ac8 [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
71 * The global cache-chain is protected by the semaphore 'cache_chain_sem'.
72 * 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>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700106
107#include <asm/uaccess.h>
108#include <asm/cacheflush.h>
109#include <asm/tlbflush.h>
110#include <asm/page.h>
111
112/*
113 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL,
114 * SLAB_RED_ZONE & SLAB_POISON.
115 * 0 for faster, smaller code (especially in the critical paths).
116 *
117 * STATS - 1 to collect stats for /proc/slabinfo.
118 * 0 for faster, smaller code (especially in the critical paths).
119 *
120 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
121 */
122
123#ifdef CONFIG_DEBUG_SLAB
124#define DEBUG 1
125#define STATS 1
126#define FORCED_DEBUG 1
127#else
128#define DEBUG 0
129#define STATS 0
130#define FORCED_DEBUG 0
131#endif
132
133
134/* Shouldn't this be in a header file somewhere? */
135#define BYTES_PER_WORD sizeof(void *)
136
137#ifndef cache_line_size
138#define cache_line_size() L1_CACHE_BYTES
139#endif
140
141#ifndef ARCH_KMALLOC_MINALIGN
142/*
143 * Enforce a minimum alignment for the kmalloc caches.
144 * Usually, the kmalloc caches are cache_line_size() aligned, except when
145 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
146 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
147 * alignment larger than BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that.
148 * Note that this flag disables some debug features.
149 */
150#define ARCH_KMALLOC_MINALIGN 0
151#endif
152
153#ifndef ARCH_SLAB_MINALIGN
154/*
155 * Enforce a minimum alignment for all caches.
156 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
157 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
158 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
159 * some debug features.
160 */
161#define ARCH_SLAB_MINALIGN 0
162#endif
163
164#ifndef ARCH_KMALLOC_FLAGS
165#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
166#endif
167
168/* Legal flag mask for kmem_cache_create(). */
169#if DEBUG
170# define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \
171 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
172 SLAB_NO_REAP | SLAB_CACHE_DMA | \
173 SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \
174 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
175 SLAB_DESTROY_BY_RCU)
176#else
177# define CREATE_MASK (SLAB_HWCACHE_ALIGN | SLAB_NO_REAP | \
178 SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \
179 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
180 SLAB_DESTROY_BY_RCU)
181#endif
182
183/*
184 * kmem_bufctl_t:
185 *
186 * Bufctl's are used for linking objs within a slab
187 * linked offsets.
188 *
189 * This implementation relies on "struct page" for locating the cache &
190 * slab an object belongs to.
191 * This allows the bufctl structure to be small (one int), but limits
192 * the number of objects a slab (not a cache) can contain when off-slab
193 * bufctls are used. The limit is the size of the largest general cache
194 * that does not use off-slab slabs.
195 * For 32bit archs with 4 kB pages, is this 56.
196 * This is not serious, as it is only for large objects, when it is unwise
197 * to have too many per slab.
198 * Note: This limit can be raised by introducing a general cache whose size
199 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
200 */
201
Kyle Moffettfa5b08d2005-09-03 15:55:03 -0700202typedef unsigned int kmem_bufctl_t;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700203#define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
204#define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
205#define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-2)
206
207/* Max number of objs-per-slab for caches which use off-slab slabs.
208 * Needed to avoid a possible looping condition in cache_grow().
209 */
210static unsigned long offslab_limit;
211
212/*
213 * struct slab
214 *
215 * Manages the objs in a slab. Placed either at the beginning of mem allocated
216 * for a slab, or allocated from an general cache.
217 * Slabs are chained into three list: fully used, partial, fully free slabs.
218 */
219struct slab {
220 struct list_head list;
221 unsigned long colouroff;
222 void *s_mem; /* including colour offset */
223 unsigned int inuse; /* num of objs active in slab */
224 kmem_bufctl_t free;
Christoph Lametere498be72005-09-09 13:03:32 -0700225 unsigned short nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700226};
227
228/*
229 * struct slab_rcu
230 *
231 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
232 * arrange for kmem_freepages to be called via RCU. This is useful if
233 * we need to approach a kernel structure obliquely, from its address
234 * obtained without the usual locking. We can lock the structure to
235 * stabilize it and check it's still at the given address, only if we
236 * can be sure that the memory has not been meanwhile reused for some
237 * other kind of object (which our subsystem's lock might corrupt).
238 *
239 * rcu_read_lock before reading the address, then rcu_read_unlock after
240 * taking the spinlock within the structure expected at that address.
241 *
242 * We assume struct slab_rcu can overlay struct slab when destroying.
243 */
244struct slab_rcu {
245 struct rcu_head head;
246 kmem_cache_t *cachep;
247 void *addr;
248};
249
250/*
251 * struct array_cache
252 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700253 * Purpose:
254 * - LIFO ordering, to hand out cache-warm objects from _alloc
255 * - reduce the number of linked list operations
256 * - reduce spinlock operations
257 *
258 * The limit is stored in the per-cpu structure to reduce the data cache
259 * footprint.
260 *
261 */
262struct array_cache {
263 unsigned int avail;
264 unsigned int limit;
265 unsigned int batchcount;
266 unsigned int touched;
Christoph Lametere498be72005-09-09 13:03:32 -0700267 spinlock_t lock;
268 void *entry[0]; /*
269 * Must have this definition in here for the proper
270 * alignment of array_cache. Also simplifies accessing
271 * the entries.
272 * [0] is for gcc 2.95. It should really be [].
273 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700274};
275
276/* bootstrap: The caches do not work without cpuarrays anymore,
277 * but the cpuarrays are allocated from the generic caches...
278 */
279#define BOOT_CPUCACHE_ENTRIES 1
280struct arraycache_init {
281 struct array_cache cache;
282 void * entries[BOOT_CPUCACHE_ENTRIES];
283};
284
285/*
Christoph Lametere498be72005-09-09 13:03:32 -0700286 * The slab lists for all objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700287 */
288struct kmem_list3 {
289 struct list_head slabs_partial; /* partial list first, better asm code */
290 struct list_head slabs_full;
291 struct list_head slabs_free;
292 unsigned long free_objects;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700293 unsigned long next_reap;
Christoph Lametere498be72005-09-09 13:03:32 -0700294 int free_touched;
295 unsigned int free_limit;
296 spinlock_t list_lock;
297 struct array_cache *shared; /* shared per node */
298 struct array_cache **alien; /* on other nodes */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700299};
300
Christoph Lametere498be72005-09-09 13:03:32 -0700301/*
302 * Need this for bootstrapping a per node allocator.
303 */
304#define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
305struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
306#define CACHE_CACHE 0
307#define SIZE_AC 1
308#define SIZE_L3 (1 + MAX_NUMNODES)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700309
Christoph Lametere498be72005-09-09 13:03:32 -0700310/*
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700311 * This function must be completely optimized away if
Christoph Lametere498be72005-09-09 13:03:32 -0700312 * a constant is passed to it. Mostly the same as
313 * what is in linux/slab.h except it returns an
314 * index.
315 */
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700316static __always_inline int index_of(const size_t size)
Christoph Lametere498be72005-09-09 13:03:32 -0700317{
318 if (__builtin_constant_p(size)) {
319 int i = 0;
320
321#define CACHE(x) \
322 if (size <=x) \
323 return i; \
324 else \
325 i++;
326#include "linux/kmalloc_sizes.h"
327#undef CACHE
328 {
329 extern void __bad_size(void);
330 __bad_size();
331 }
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700332 } else
333 BUG();
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
340static inline void kmem_list3_init(struct kmem_list3 *parent)
341{
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 */
370
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 */
373 struct array_cache *array[NR_CPUS];
374 unsigned int batchcount;
375 unsigned int limit;
Christoph Lametere498be72005-09-09 13:03:32 -0700376 unsigned int shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700377 unsigned int objsize;
Christoph Lametere498be72005-09-09 13:03:32 -0700378/* 2) touched by every alloc & free from the backend */
379 struct kmem_list3 *nodelists[MAX_NUMNODES];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700380 unsigned int flags; /* constant flags */
381 unsigned int num; /* # of objs per slab */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700382 spinlock_t spinlock;
383
384/* 3) cache_grow/shrink */
385 /* order of pgs per slab (2^n) */
386 unsigned int gfporder;
387
388 /* force GFP flags, e.g. GFP_DMA */
Al Viro6daa0e22005-10-21 03:18:50 -0400389 gfp_t gfpflags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700390
391 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 */
397
398 /* constructor func */
399 void (*ctor)(void *, kmem_cache_t *, unsigned long);
400
401 /* de-constructor func */
402 void (*dtor)(void *, kmem_cache_t *, unsigned long);
403
404/* 4) cache creation/removal */
405 const char *name;
406 struct list_head next;
407
408/* 5) statistics */
409#if STATS
410 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;
Christoph Lametere498be72005-09-09 13:03:32 -0700418 unsigned long node_frees;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700419 atomic_t allochit;
420 atomic_t allocmiss;
421 atomic_t freehit;
422 atomic_t freemiss;
423#endif
424#if DEBUG
425 int dbghead;
426 int reallen;
427#endif
428};
429
430#define CFLGS_OFF_SLAB (0x80000000UL)
431#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
432
433#define BATCHREFILL_LIMIT 16
434/* Optimization question: fewer reaps means less
435 * probability for unnessary cpucache drain/refill cycles.
436 *
Adrian Bunkdc6f3f22005-11-08 16:44:08 +0100437 * OTOH the cpuarrays can contain lots of objects,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700438 * which could lock up otherwise freeable slabs.
439 */
440#define REAPTIMEOUT_CPUC (2*HZ)
441#define REAPTIMEOUT_LIST3 (4*HZ)
442
443#if STATS
444#define STATS_INC_ACTIVE(x) ((x)->num_active++)
445#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
446#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
447#define STATS_INC_GROWN(x) ((x)->grown++)
448#define STATS_INC_REAPED(x) ((x)->reaped++)
449#define STATS_SET_HIGH(x) do { if ((x)->num_active > (x)->high_mark) \
450 (x)->high_mark = (x)->num_active; \
451 } while (0)
452#define STATS_INC_ERR(x) ((x)->errors++)
453#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
Christoph Lametere498be72005-09-09 13:03:32 -0700454#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700455#define STATS_SET_FREEABLE(x, i) \
456 do { if ((x)->max_freeable < i) \
457 (x)->max_freeable = i; \
458 } while (0)
459
460#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
461#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
462#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
463#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
464#else
465#define STATS_INC_ACTIVE(x) do { } while (0)
466#define STATS_DEC_ACTIVE(x) do { } while (0)
467#define STATS_INC_ALLOCED(x) do { } while (0)
468#define STATS_INC_GROWN(x) do { } while (0)
469#define STATS_INC_REAPED(x) do { } while (0)
470#define STATS_SET_HIGH(x) do { } while (0)
471#define STATS_INC_ERR(x) do { } while (0)
472#define STATS_INC_NODEALLOCS(x) do { } while (0)
Christoph Lametere498be72005-09-09 13:03:32 -0700473#define STATS_INC_NODEFREES(x) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700474#define STATS_SET_FREEABLE(x, i) \
475 do { } while (0)
476
477#define STATS_INC_ALLOCHIT(x) do { } while (0)
478#define STATS_INC_ALLOCMISS(x) do { } while (0)
479#define STATS_INC_FREEHIT(x) do { } while (0)
480#define STATS_INC_FREEMISS(x) do { } while (0)
481#endif
482
483#if DEBUG
484/* Magic nums for obj red zoning.
485 * Placed in the first word before and the first word after an obj.
486 */
487#define RED_INACTIVE 0x5A2CF071UL /* when obj is inactive */
488#define RED_ACTIVE 0x170FC2A5UL /* when obj is active */
489
490/* ...and for poisoning */
491#define POISON_INUSE 0x5a /* for use-uninitialised poisoning */
492#define POISON_FREE 0x6b /* for use-after-free poisoning */
493#define POISON_END 0xa5 /* end-byte of poisoning */
494
495/* memory layout of objects:
496 * 0 : objp
497 * 0 .. cachep->dbghead - BYTES_PER_WORD - 1: padding. This ensures that
498 * the end of an object is aligned with the end of the real
499 * allocation. Catches writes behind the end of the allocation.
500 * cachep->dbghead - BYTES_PER_WORD .. cachep->dbghead - 1:
501 * redzone word.
502 * cachep->dbghead: The real object.
503 * cachep->objsize - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
504 * cachep->objsize - 1* BYTES_PER_WORD: last caller address [BYTES_PER_WORD long]
505 */
506static int obj_dbghead(kmem_cache_t *cachep)
507{
508 return cachep->dbghead;
509}
510
511static int obj_reallen(kmem_cache_t *cachep)
512{
513 return cachep->reallen;
514}
515
516static unsigned long *dbg_redzone1(kmem_cache_t *cachep, void *objp)
517{
518 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
519 return (unsigned long*) (objp+obj_dbghead(cachep)-BYTES_PER_WORD);
520}
521
522static unsigned long *dbg_redzone2(kmem_cache_t *cachep, void *objp)
523{
524 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
525 if (cachep->flags & SLAB_STORE_USER)
526 return (unsigned long*) (objp+cachep->objsize-2*BYTES_PER_WORD);
527 return (unsigned long*) (objp+cachep->objsize-BYTES_PER_WORD);
528}
529
530static void **dbg_userword(kmem_cache_t *cachep, void *objp)
531{
532 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
533 return (void**)(objp+cachep->objsize-BYTES_PER_WORD);
534}
535
536#else
537
538#define obj_dbghead(x) 0
539#define obj_reallen(cachep) (cachep->objsize)
540#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;})
541#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;})
542#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
543
544#endif
545
546/*
547 * Maximum size of an obj (in 2^order pages)
548 * and absolute limit for the gfp order.
549 */
550#if defined(CONFIG_LARGE_ALLOCS)
551#define MAX_OBJ_ORDER 13 /* up to 32Mb */
552#define MAX_GFP_ORDER 13 /* up to 32Mb */
553#elif defined(CONFIG_MMU)
554#define MAX_OBJ_ORDER 5 /* 32 pages */
555#define MAX_GFP_ORDER 5 /* 32 pages */
556#else
557#define MAX_OBJ_ORDER 8 /* up to 1Mb */
558#define MAX_GFP_ORDER 8 /* up to 1Mb */
559#endif
560
561/*
562 * Do not go above this order unless 0 objects fit into the slab.
563 */
564#define BREAK_GFP_ORDER_HI 1
565#define BREAK_GFP_ORDER_LO 0
566static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
567
Pekka Enberg065d41c2005-11-13 16:06:46 -0800568/* Functions for storing/retrieving the cachep and or slab from the
Linus Torvalds1da177e2005-04-16 15:20:36 -0700569 * global 'mem_map'. These are used to find the slab an obj belongs to.
570 * With kfree(), these are used to find the cache which an obj belongs to.
571 */
Pekka Enberg065d41c2005-11-13 16:06:46 -0800572static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
573{
574 page->lru.next = (struct list_head *)cache;
575}
576
577static inline struct kmem_cache *page_get_cache(struct page *page)
578{
579 return (struct kmem_cache *)page->lru.next;
580}
581
582static inline void page_set_slab(struct page *page, struct slab *slab)
583{
584 page->lru.prev = (struct list_head *)slab;
585}
586
587static inline struct slab *page_get_slab(struct page *page)
588{
589 return (struct slab *)page->lru.prev;
590}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700591
592/* These are the default caches for kmalloc. Custom caches can have other sizes. */
593struct cache_sizes malloc_sizes[] = {
594#define CACHE(x) { .cs_size = (x) },
595#include <linux/kmalloc_sizes.h>
596 CACHE(ULONG_MAX)
597#undef CACHE
598};
599EXPORT_SYMBOL(malloc_sizes);
600
601/* Must match cache_sizes above. Out of line to keep cache footprint low. */
602struct cache_names {
603 char *name;
604 char *name_dma;
605};
606
607static struct cache_names __initdata cache_names[] = {
608#define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
609#include <linux/kmalloc_sizes.h>
610 { NULL, }
611#undef CACHE
612};
613
614static struct arraycache_init initarray_cache __initdata =
615 { { 0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
616static struct arraycache_init initarray_generic =
617 { { 0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
618
619/* internal cache of cache description objs */
620static kmem_cache_t cache_cache = {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700621 .batchcount = 1,
622 .limit = BOOT_CPUCACHE_ENTRIES,
Christoph Lametere498be72005-09-09 13:03:32 -0700623 .shared = 1,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700624 .objsize = sizeof(kmem_cache_t),
625 .flags = SLAB_NO_REAP,
626 .spinlock = SPIN_LOCK_UNLOCKED,
627 .name = "kmem_cache",
628#if DEBUG
629 .reallen = sizeof(kmem_cache_t),
630#endif
631};
632
633/* Guard access to the cache-chain. */
634static struct semaphore cache_chain_sem;
635static struct list_head cache_chain;
636
637/*
638 * vm_enough_memory() looks at this to determine how many
639 * slab-allocated pages are possibly freeable under pressure
640 *
641 * SLAB_RECLAIM_ACCOUNT turns this on per-slab
642 */
643atomic_t slab_reclaim_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700644
645/*
646 * chicken and egg problem: delay the per-cpu array allocation
647 * until the general caches are up.
648 */
649static enum {
650 NONE,
Christoph Lametere498be72005-09-09 13:03:32 -0700651 PARTIAL_AC,
652 PARTIAL_L3,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700653 FULL
654} g_cpucache_up;
655
656static DEFINE_PER_CPU(struct work_struct, reap_work);
657
Christoph Lameterff694162005-09-22 21:44:02 -0700658static void free_block(kmem_cache_t* cachep, void** objpp, int len, int node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700659static void enable_cpucache (kmem_cache_t *cachep);
660static void cache_reap (void *unused);
Christoph Lametere498be72005-09-09 13:03:32 -0700661static int __node_shrink(kmem_cache_t *cachep, int node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700662
663static inline struct array_cache *ac_data(kmem_cache_t *cachep)
664{
665 return cachep->array[smp_processor_id()];
666}
667
Al Virodd0fc662005-10-07 07:46:04 +0100668static inline kmem_cache_t *__find_general_cachep(size_t size, gfp_t gfpflags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700669{
670 struct cache_sizes *csizep = malloc_sizes;
671
672#if DEBUG
673 /* This happens if someone tries to call
674 * kmem_cache_create(), or __kmalloc(), before
675 * the generic caches are initialized.
676 */
Alok Katariac7e43c72005-09-14 12:17:53 -0700677 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700678#endif
679 while (size > csizep->cs_size)
680 csizep++;
681
682 /*
Martin Hicks0abf40c2005-09-03 15:54:54 -0700683 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
Linus Torvalds1da177e2005-04-16 15:20:36 -0700684 * has cs_{dma,}cachep==NULL. Thus no special case
685 * for large kmalloc calls required.
686 */
687 if (unlikely(gfpflags & GFP_DMA))
688 return csizep->cs_dmacachep;
689 return csizep->cs_cachep;
690}
691
Al Virodd0fc662005-10-07 07:46:04 +0100692kmem_cache_t *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700693{
694 return __find_general_cachep(size, gfpflags);
695}
696EXPORT_SYMBOL(kmem_find_general_cachep);
697
Linus Torvalds1da177e2005-04-16 15:20:36 -0700698/* Cal the num objs, wastage, and bytes left over for a given slab size. */
699static void cache_estimate(unsigned long gfporder, size_t size, size_t align,
700 int flags, size_t *left_over, unsigned int *num)
701{
702 int i;
703 size_t wastage = PAGE_SIZE<<gfporder;
704 size_t extra = 0;
705 size_t base = 0;
706
707 if (!(flags & CFLGS_OFF_SLAB)) {
708 base = sizeof(struct slab);
709 extra = sizeof(kmem_bufctl_t);
710 }
711 i = 0;
712 while (i*size + ALIGN(base+i*extra, align) <= wastage)
713 i++;
714 if (i > 0)
715 i--;
716
717 if (i > SLAB_LIMIT)
718 i = SLAB_LIMIT;
719
720 *num = i;
721 wastage -= i*size;
722 wastage -= ALIGN(base+i*extra, align);
723 *left_over = wastage;
724}
725
726#define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
727
728static void __slab_error(const char *function, kmem_cache_t *cachep, char *msg)
729{
730 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
731 function, cachep->name, msg);
732 dump_stack();
733}
734
735/*
736 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
737 * via the workqueue/eventd.
738 * Add the CPU number into the expiration time to minimize the possibility of
739 * the CPUs getting into lockstep and contending for the global cache chain
740 * lock.
741 */
742static void __devinit start_cpu_timer(int cpu)
743{
744 struct work_struct *reap_work = &per_cpu(reap_work, cpu);
745
746 /*
747 * When this gets called from do_initcalls via cpucache_init(),
748 * init_workqueues() has already run, so keventd will be setup
749 * at that time.
750 */
751 if (keventd_up() && reap_work->func == NULL) {
752 INIT_WORK(reap_work, cache_reap, NULL);
753 schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu);
754 }
755}
756
Christoph Lametere498be72005-09-09 13:03:32 -0700757static struct array_cache *alloc_arraycache(int node, int entries,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700758 int batchcount)
759{
760 int memsize = sizeof(void*)*entries+sizeof(struct array_cache);
761 struct array_cache *nc = NULL;
762
Christoph Lametere498be72005-09-09 13:03:32 -0700763 nc = kmalloc_node(memsize, GFP_KERNEL, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700764 if (nc) {
765 nc->avail = 0;
766 nc->limit = entries;
767 nc->batchcount = batchcount;
768 nc->touched = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700769 spin_lock_init(&nc->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700770 }
771 return nc;
772}
773
Christoph Lametere498be72005-09-09 13:03:32 -0700774#ifdef CONFIG_NUMA
775static inline struct array_cache **alloc_alien_cache(int node, int limit)
776{
777 struct array_cache **ac_ptr;
778 int memsize = sizeof(void*)*MAX_NUMNODES;
779 int i;
780
781 if (limit > 1)
782 limit = 12;
783 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
784 if (ac_ptr) {
785 for_each_node(i) {
786 if (i == node || !node_online(i)) {
787 ac_ptr[i] = NULL;
788 continue;
789 }
790 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
791 if (!ac_ptr[i]) {
792 for (i--; i <=0; i--)
793 kfree(ac_ptr[i]);
794 kfree(ac_ptr);
795 return NULL;
796 }
797 }
798 }
799 return ac_ptr;
800}
801
802static inline void free_alien_cache(struct array_cache **ac_ptr)
803{
804 int i;
805
806 if (!ac_ptr)
807 return;
808
809 for_each_node(i)
810 kfree(ac_ptr[i]);
811
812 kfree(ac_ptr);
813}
814
815static inline void __drain_alien_cache(kmem_cache_t *cachep, struct array_cache *ac, int node)
816{
817 struct kmem_list3 *rl3 = cachep->nodelists[node];
818
819 if (ac->avail) {
820 spin_lock(&rl3->list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -0700821 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -0700822 ac->avail = 0;
823 spin_unlock(&rl3->list_lock);
824 }
825}
826
827static void drain_alien_cache(kmem_cache_t *cachep, struct kmem_list3 *l3)
828{
829 int i=0;
830 struct array_cache *ac;
831 unsigned long flags;
832
833 for_each_online_node(i) {
834 ac = l3->alien[i];
835 if (ac) {
836 spin_lock_irqsave(&ac->lock, flags);
837 __drain_alien_cache(cachep, ac, i);
838 spin_unlock_irqrestore(&ac->lock, flags);
839 }
840 }
841}
842#else
843#define alloc_alien_cache(node, limit) do { } while (0)
844#define free_alien_cache(ac_ptr) do { } while (0)
845#define drain_alien_cache(cachep, l3) do { } while (0)
846#endif
847
Linus Torvalds1da177e2005-04-16 15:20:36 -0700848static int __devinit cpuup_callback(struct notifier_block *nfb,
849 unsigned long action, void *hcpu)
850{
851 long cpu = (long)hcpu;
852 kmem_cache_t* cachep;
Christoph Lametere498be72005-09-09 13:03:32 -0700853 struct kmem_list3 *l3 = NULL;
854 int node = cpu_to_node(cpu);
855 int memsize = sizeof(struct kmem_list3);
856 struct array_cache *nc = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700857
858 switch (action) {
859 case CPU_UP_PREPARE:
860 down(&cache_chain_sem);
Christoph Lametere498be72005-09-09 13:03:32 -0700861 /* we need to do this right in the beginning since
862 * alloc_arraycache's are going to use this list.
863 * kmalloc_node allows us to add the slab to the right
864 * kmem_list3 and not this cpu's kmem_list3
865 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700866
Christoph Lametere498be72005-09-09 13:03:32 -0700867 list_for_each_entry(cachep, &cache_chain, next) {
868 /* setup the size64 kmemlist for cpu before we can
869 * begin anything. Make sure some other cpu on this
870 * node has not already allocated this
871 */
872 if (!cachep->nodelists[node]) {
873 if (!(l3 = kmalloc_node(memsize,
874 GFP_KERNEL, node)))
875 goto bad;
876 kmem_list3_init(l3);
877 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
878 ((unsigned long)cachep)%REAPTIMEOUT_LIST3;
879
880 cachep->nodelists[node] = l3;
881 }
882
883 spin_lock_irq(&cachep->nodelists[node]->list_lock);
884 cachep->nodelists[node]->free_limit =
885 (1 + nr_cpus_node(node)) *
886 cachep->batchcount + cachep->num;
887 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
888 }
889
890 /* Now we can go ahead with allocating the shared array's
891 & array cache's */
892 list_for_each_entry(cachep, &cache_chain, next) {
893 nc = alloc_arraycache(node, cachep->limit,
894 cachep->batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700895 if (!nc)
896 goto bad;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700897 cachep->array[cpu] = nc;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700898
Christoph Lametere498be72005-09-09 13:03:32 -0700899 l3 = cachep->nodelists[node];
900 BUG_ON(!l3);
901 if (!l3->shared) {
902 if (!(nc = alloc_arraycache(node,
903 cachep->shared*cachep->batchcount,
904 0xbaadf00d)))
905 goto bad;
906
907 /* we are serialised from CPU_DEAD or
908 CPU_UP_CANCELLED by the cpucontrol lock */
909 l3->shared = nc;
910 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700911 }
912 up(&cache_chain_sem);
913 break;
914 case CPU_ONLINE:
915 start_cpu_timer(cpu);
916 break;
917#ifdef CONFIG_HOTPLUG_CPU
918 case CPU_DEAD:
919 /* fall thru */
920 case CPU_UP_CANCELED:
921 down(&cache_chain_sem);
922
923 list_for_each_entry(cachep, &cache_chain, next) {
924 struct array_cache *nc;
Christoph Lametere498be72005-09-09 13:03:32 -0700925 cpumask_t mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700926
Christoph Lametere498be72005-09-09 13:03:32 -0700927 mask = node_to_cpumask(node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700928 spin_lock_irq(&cachep->spinlock);
929 /* cpu is dead; no one can alloc from it. */
930 nc = cachep->array[cpu];
931 cachep->array[cpu] = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -0700932 l3 = cachep->nodelists[node];
933
934 if (!l3)
935 goto unlock_cache;
936
937 spin_lock(&l3->list_lock);
938
939 /* Free limit for this kmem_list3 */
940 l3->free_limit -= cachep->batchcount;
941 if (nc)
Christoph Lameterff694162005-09-22 21:44:02 -0700942 free_block(cachep, nc->entry, nc->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -0700943
944 if (!cpus_empty(mask)) {
945 spin_unlock(&l3->list_lock);
946 goto unlock_cache;
947 }
948
949 if (l3->shared) {
950 free_block(cachep, l3->shared->entry,
Christoph Lameterff694162005-09-22 21:44:02 -0700951 l3->shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -0700952 kfree(l3->shared);
953 l3->shared = NULL;
954 }
955 if (l3->alien) {
956 drain_alien_cache(cachep, l3);
957 free_alien_cache(l3->alien);
958 l3->alien = NULL;
959 }
960
961 /* free slabs belonging to this node */
962 if (__node_shrink(cachep, node)) {
963 cachep->nodelists[node] = NULL;
964 spin_unlock(&l3->list_lock);
965 kfree(l3);
966 } else {
967 spin_unlock(&l3->list_lock);
968 }
969unlock_cache:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700970 spin_unlock_irq(&cachep->spinlock);
971 kfree(nc);
972 }
973 up(&cache_chain_sem);
974 break;
975#endif
976 }
977 return NOTIFY_OK;
978bad:
979 up(&cache_chain_sem);
980 return NOTIFY_BAD;
981}
982
983static struct notifier_block cpucache_notifier = { &cpuup_callback, NULL, 0 };
984
Christoph Lametere498be72005-09-09 13:03:32 -0700985/*
986 * swap the static kmem_list3 with kmalloced memory
987 */
988static void init_list(kmem_cache_t *cachep, struct kmem_list3 *list,
989 int nodeid)
990{
991 struct kmem_list3 *ptr;
992
993 BUG_ON(cachep->nodelists[nodeid] != list);
994 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
995 BUG_ON(!ptr);
996
997 local_irq_disable();
998 memcpy(ptr, list, sizeof(struct kmem_list3));
999 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1000 cachep->nodelists[nodeid] = ptr;
1001 local_irq_enable();
1002}
1003
Linus Torvalds1da177e2005-04-16 15:20:36 -07001004/* Initialisation.
1005 * Called after the gfp() functions have been enabled, and before smp_init().
1006 */
1007void __init kmem_cache_init(void)
1008{
1009 size_t left_over;
1010 struct cache_sizes *sizes;
1011 struct cache_names *names;
Christoph Lametere498be72005-09-09 13:03:32 -07001012 int i;
1013
1014 for (i = 0; i < NUM_INIT_LISTS; i++) {
1015 kmem_list3_init(&initkmem_list3[i]);
1016 if (i < MAX_NUMNODES)
1017 cache_cache.nodelists[i] = NULL;
1018 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001019
1020 /*
1021 * Fragmentation resistance on low memory - only use bigger
1022 * page orders on machines with more than 32MB of memory.
1023 */
1024 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1025 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1026
Linus Torvalds1da177e2005-04-16 15:20:36 -07001027 /* Bootstrap is tricky, because several objects are allocated
1028 * from caches that do not exist yet:
1029 * 1) initialize the cache_cache cache: it contains the kmem_cache_t
1030 * structures of all caches, except cache_cache itself: cache_cache
1031 * is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001032 * Initially an __init data area is used for the head array and the
1033 * kmem_list3 structures, it's replaced with a kmalloc allocated
1034 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001035 * 2) Create the first kmalloc cache.
Christoph Lametere498be72005-09-09 13:03:32 -07001036 * The kmem_cache_t for the new cache is allocated normally.
1037 * An __init data area is used for the head array.
1038 * 3) Create the remaining kmalloc caches, with minimally sized
1039 * head arrays.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001040 * 4) Replace the __init data head arrays for cache_cache and the first
1041 * kmalloc cache with kmalloc allocated arrays.
Christoph Lametere498be72005-09-09 13:03:32 -07001042 * 5) Replace the __init data for kmem_list3 for cache_cache and
1043 * the other cache's with kmalloc allocated memory.
1044 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001045 */
1046
1047 /* 1) create the cache_cache */
1048 init_MUTEX(&cache_chain_sem);
1049 INIT_LIST_HEAD(&cache_chain);
1050 list_add(&cache_cache.next, &cache_chain);
1051 cache_cache.colour_off = cache_line_size();
1052 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
Christoph Lametere498be72005-09-09 13:03:32 -07001053 cache_cache.nodelists[numa_node_id()] = &initkmem_list3[CACHE_CACHE];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001054
1055 cache_cache.objsize = ALIGN(cache_cache.objsize, cache_line_size());
1056
1057 cache_estimate(0, cache_cache.objsize, cache_line_size(), 0,
1058 &left_over, &cache_cache.num);
1059 if (!cache_cache.num)
1060 BUG();
1061
1062 cache_cache.colour = left_over/cache_cache.colour_off;
1063 cache_cache.colour_next = 0;
1064 cache_cache.slab_size = ALIGN(cache_cache.num*sizeof(kmem_bufctl_t) +
1065 sizeof(struct slab), cache_line_size());
1066
1067 /* 2+3) create the kmalloc caches */
1068 sizes = malloc_sizes;
1069 names = cache_names;
1070
Christoph Lametere498be72005-09-09 13:03:32 -07001071 /* Initialize the caches that provide memory for the array cache
1072 * and the kmem_list3 structures first.
1073 * Without this, further allocations will bug
1074 */
1075
1076 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
1077 sizes[INDEX_AC].cs_size, ARCH_KMALLOC_MINALIGN,
1078 (ARCH_KMALLOC_FLAGS | SLAB_PANIC), NULL, NULL);
1079
1080 if (INDEX_AC != INDEX_L3)
1081 sizes[INDEX_L3].cs_cachep =
1082 kmem_cache_create(names[INDEX_L3].name,
1083 sizes[INDEX_L3].cs_size, ARCH_KMALLOC_MINALIGN,
1084 (ARCH_KMALLOC_FLAGS | SLAB_PANIC), NULL, NULL);
1085
Linus Torvalds1da177e2005-04-16 15:20:36 -07001086 while (sizes->cs_size != ULONG_MAX) {
Christoph Lametere498be72005-09-09 13:03:32 -07001087 /*
1088 * For performance, all the general caches are L1 aligned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001089 * This should be particularly beneficial on SMP boxes, as it
1090 * eliminates "false sharing".
1091 * Note for systems short on memory removing the alignment will
Christoph Lametere498be72005-09-09 13:03:32 -07001092 * allow tighter packing of the smaller caches.
1093 */
1094 if(!sizes->cs_cachep)
1095 sizes->cs_cachep = kmem_cache_create(names->name,
1096 sizes->cs_size, ARCH_KMALLOC_MINALIGN,
1097 (ARCH_KMALLOC_FLAGS | SLAB_PANIC), NULL, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001098
1099 /* Inc off-slab bufctl limit until the ceiling is hit. */
1100 if (!(OFF_SLAB(sizes->cs_cachep))) {
1101 offslab_limit = sizes->cs_size-sizeof(struct slab);
1102 offslab_limit /= sizeof(kmem_bufctl_t);
1103 }
1104
1105 sizes->cs_dmacachep = kmem_cache_create(names->name_dma,
1106 sizes->cs_size, ARCH_KMALLOC_MINALIGN,
1107 (ARCH_KMALLOC_FLAGS | SLAB_CACHE_DMA | SLAB_PANIC),
1108 NULL, NULL);
1109
1110 sizes++;
1111 names++;
1112 }
1113 /* 4) Replace the bootstrap head arrays */
1114 {
1115 void * ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001116
Linus Torvalds1da177e2005-04-16 15:20:36 -07001117 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001118
Linus Torvalds1da177e2005-04-16 15:20:36 -07001119 local_irq_disable();
1120 BUG_ON(ac_data(&cache_cache) != &initarray_cache.cache);
Christoph Lametere498be72005-09-09 13:03:32 -07001121 memcpy(ptr, ac_data(&cache_cache),
1122 sizeof(struct arraycache_init));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001123 cache_cache.array[smp_processor_id()] = ptr;
1124 local_irq_enable();
Christoph Lametere498be72005-09-09 13:03:32 -07001125
Linus Torvalds1da177e2005-04-16 15:20:36 -07001126 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001127
Linus Torvalds1da177e2005-04-16 15:20:36 -07001128 local_irq_disable();
Christoph Lametere498be72005-09-09 13:03:32 -07001129 BUG_ON(ac_data(malloc_sizes[INDEX_AC].cs_cachep)
1130 != &initarray_generic.cache);
1131 memcpy(ptr, ac_data(malloc_sizes[INDEX_AC].cs_cachep),
Linus Torvalds1da177e2005-04-16 15:20:36 -07001132 sizeof(struct arraycache_init));
Christoph Lametere498be72005-09-09 13:03:32 -07001133 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1134 ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001135 local_irq_enable();
1136 }
Christoph Lametere498be72005-09-09 13:03:32 -07001137 /* 5) Replace the bootstrap kmem_list3's */
1138 {
1139 int node;
1140 /* Replace the static kmem_list3 structures for the boot cpu */
1141 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE],
1142 numa_node_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001143
Christoph Lametere498be72005-09-09 13:03:32 -07001144 for_each_online_node(node) {
1145 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1146 &initkmem_list3[SIZE_AC+node], node);
1147
1148 if (INDEX_AC != INDEX_L3) {
1149 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1150 &initkmem_list3[SIZE_L3+node],
1151 node);
1152 }
1153 }
1154 }
1155
1156 /* 6) resize the head arrays to their final sizes */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001157 {
1158 kmem_cache_t *cachep;
1159 down(&cache_chain_sem);
1160 list_for_each_entry(cachep, &cache_chain, next)
1161 enable_cpucache(cachep);
1162 up(&cache_chain_sem);
1163 }
1164
1165 /* Done! */
1166 g_cpucache_up = FULL;
1167
1168 /* Register a cpu startup notifier callback
1169 * that initializes ac_data for all new cpus
1170 */
1171 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001172
1173 /* The reap timers are started later, with a module init call:
1174 * That part of the kernel is not yet operational.
1175 */
1176}
1177
1178static int __init cpucache_init(void)
1179{
1180 int cpu;
1181
1182 /*
1183 * Register the timers that return unneeded
1184 * pages to gfp.
1185 */
Christoph Lametere498be72005-09-09 13:03:32 -07001186 for_each_online_cpu(cpu)
1187 start_cpu_timer(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001188
1189 return 0;
1190}
1191
1192__initcall(cpucache_init);
1193
1194/*
1195 * Interface to system's page allocator. No need to hold the cache-lock.
1196 *
1197 * If we requested dmaable memory, we will get it. Even if we
1198 * did not request dmaable memory, we might get it, but that
1199 * would be relatively rare and ignorable.
1200 */
Al Virodd0fc662005-10-07 07:46:04 +01001201static void *kmem_getpages(kmem_cache_t *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001202{
1203 struct page *page;
1204 void *addr;
1205 int i;
1206
1207 flags |= cachep->gfpflags;
Christoph Lameter50c85a12005-11-13 16:06:47 -08001208 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001209 if (!page)
1210 return NULL;
1211 addr = page_address(page);
1212
1213 i = (1 << cachep->gfporder);
1214 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1215 atomic_add(i, &slab_reclaim_pages);
1216 add_page_state(nr_slab, i);
1217 while (i--) {
1218 SetPageSlab(page);
1219 page++;
1220 }
1221 return addr;
1222}
1223
1224/*
1225 * Interface to system's page release.
1226 */
1227static void kmem_freepages(kmem_cache_t *cachep, void *addr)
1228{
1229 unsigned long i = (1<<cachep->gfporder);
1230 struct page *page = virt_to_page(addr);
1231 const unsigned long nr_freed = i;
1232
1233 while (i--) {
1234 if (!TestClearPageSlab(page))
1235 BUG();
1236 page++;
1237 }
1238 sub_page_state(nr_slab, nr_freed);
1239 if (current->reclaim_state)
1240 current->reclaim_state->reclaimed_slab += nr_freed;
1241 free_pages((unsigned long)addr, cachep->gfporder);
1242 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1243 atomic_sub(1<<cachep->gfporder, &slab_reclaim_pages);
1244}
1245
1246static void kmem_rcu_free(struct rcu_head *head)
1247{
1248 struct slab_rcu *slab_rcu = (struct slab_rcu *) head;
1249 kmem_cache_t *cachep = slab_rcu->cachep;
1250
1251 kmem_freepages(cachep, slab_rcu->addr);
1252 if (OFF_SLAB(cachep))
1253 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1254}
1255
1256#if DEBUG
1257
1258#ifdef CONFIG_DEBUG_PAGEALLOC
1259static void store_stackinfo(kmem_cache_t *cachep, unsigned long *addr,
1260 unsigned long caller)
1261{
1262 int size = obj_reallen(cachep);
1263
1264 addr = (unsigned long *)&((char*)addr)[obj_dbghead(cachep)];
1265
1266 if (size < 5*sizeof(unsigned long))
1267 return;
1268
1269 *addr++=0x12345678;
1270 *addr++=caller;
1271 *addr++=smp_processor_id();
1272 size -= 3*sizeof(unsigned long);
1273 {
1274 unsigned long *sptr = &caller;
1275 unsigned long svalue;
1276
1277 while (!kstack_end(sptr)) {
1278 svalue = *sptr++;
1279 if (kernel_text_address(svalue)) {
1280 *addr++=svalue;
1281 size -= sizeof(unsigned long);
1282 if (size <= sizeof(unsigned long))
1283 break;
1284 }
1285 }
1286
1287 }
1288 *addr++=0x87654321;
1289}
1290#endif
1291
1292static void poison_obj(kmem_cache_t *cachep, void *addr, unsigned char val)
1293{
1294 int size = obj_reallen(cachep);
1295 addr = &((char*)addr)[obj_dbghead(cachep)];
1296
1297 memset(addr, val, size);
1298 *(unsigned char *)(addr+size-1) = POISON_END;
1299}
1300
1301static void dump_line(char *data, int offset, int limit)
1302{
1303 int i;
1304 printk(KERN_ERR "%03x:", offset);
1305 for (i=0;i<limit;i++) {
1306 printk(" %02x", (unsigned char)data[offset+i]);
1307 }
1308 printk("\n");
1309}
1310#endif
1311
1312#if DEBUG
1313
1314static void print_objinfo(kmem_cache_t *cachep, void *objp, int lines)
1315{
1316 int i, size;
1317 char *realobj;
1318
1319 if (cachep->flags & SLAB_RED_ZONE) {
1320 printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n",
1321 *dbg_redzone1(cachep, objp),
1322 *dbg_redzone2(cachep, objp));
1323 }
1324
1325 if (cachep->flags & SLAB_STORE_USER) {
1326 printk(KERN_ERR "Last user: [<%p>]",
1327 *dbg_userword(cachep, objp));
1328 print_symbol("(%s)",
1329 (unsigned long)*dbg_userword(cachep, objp));
1330 printk("\n");
1331 }
1332 realobj = (char*)objp+obj_dbghead(cachep);
1333 size = obj_reallen(cachep);
1334 for (i=0; i<size && lines;i+=16, lines--) {
1335 int limit;
1336 limit = 16;
1337 if (i+limit > size)
1338 limit = size-i;
1339 dump_line(realobj, i, limit);
1340 }
1341}
1342
1343static void check_poison_obj(kmem_cache_t *cachep, void *objp)
1344{
1345 char *realobj;
1346 int size, i;
1347 int lines = 0;
1348
1349 realobj = (char*)objp+obj_dbghead(cachep);
1350 size = obj_reallen(cachep);
1351
1352 for (i=0;i<size;i++) {
1353 char exp = POISON_FREE;
1354 if (i == size-1)
1355 exp = POISON_END;
1356 if (realobj[i] != exp) {
1357 int limit;
1358 /* Mismatch ! */
1359 /* Print header */
1360 if (lines == 0) {
1361 printk(KERN_ERR "Slab corruption: start=%p, len=%d\n",
1362 realobj, size);
1363 print_objinfo(cachep, objp, 0);
1364 }
1365 /* Hexdump the affected line */
1366 i = (i/16)*16;
1367 limit = 16;
1368 if (i+limit > size)
1369 limit = size-i;
1370 dump_line(realobj, i, limit);
1371 i += 16;
1372 lines++;
1373 /* Limit to 5 lines */
1374 if (lines > 5)
1375 break;
1376 }
1377 }
1378 if (lines != 0) {
1379 /* Print some data about the neighboring objects, if they
1380 * exist:
1381 */
Pekka Enberg065d41c2005-11-13 16:06:46 -08001382 struct slab *slabp = page_get_slab(virt_to_page(objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001383 int objnr;
1384
1385 objnr = (objp-slabp->s_mem)/cachep->objsize;
1386 if (objnr) {
1387 objp = slabp->s_mem+(objnr-1)*cachep->objsize;
1388 realobj = (char*)objp+obj_dbghead(cachep);
1389 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1390 realobj, size);
1391 print_objinfo(cachep, objp, 2);
1392 }
1393 if (objnr+1 < cachep->num) {
1394 objp = slabp->s_mem+(objnr+1)*cachep->objsize;
1395 realobj = (char*)objp+obj_dbghead(cachep);
1396 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1397 realobj, size);
1398 print_objinfo(cachep, objp, 2);
1399 }
1400 }
1401}
1402#endif
1403
1404/* Destroy all the objs in a slab, and release the mem back to the system.
1405 * Before calling the slab must have been unlinked from the cache.
1406 * The cache-lock is not held/needed.
1407 */
1408static void slab_destroy (kmem_cache_t *cachep, struct slab *slabp)
1409{
1410 void *addr = slabp->s_mem - slabp->colouroff;
1411
1412#if DEBUG
1413 int i;
1414 for (i = 0; i < cachep->num; i++) {
1415 void *objp = slabp->s_mem + cachep->objsize * i;
1416
1417 if (cachep->flags & SLAB_POISON) {
1418#ifdef CONFIG_DEBUG_PAGEALLOC
1419 if ((cachep->objsize%PAGE_SIZE)==0 && OFF_SLAB(cachep))
1420 kernel_map_pages(virt_to_page(objp), cachep->objsize/PAGE_SIZE,1);
1421 else
1422 check_poison_obj(cachep, objp);
1423#else
1424 check_poison_obj(cachep, objp);
1425#endif
1426 }
1427 if (cachep->flags & SLAB_RED_ZONE) {
1428 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1429 slab_error(cachep, "start of a freed object "
1430 "was overwritten");
1431 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1432 slab_error(cachep, "end of a freed object "
1433 "was overwritten");
1434 }
1435 if (cachep->dtor && !(cachep->flags & SLAB_POISON))
1436 (cachep->dtor)(objp+obj_dbghead(cachep), cachep, 0);
1437 }
1438#else
1439 if (cachep->dtor) {
1440 int i;
1441 for (i = 0; i < cachep->num; i++) {
1442 void* objp = slabp->s_mem+cachep->objsize*i;
1443 (cachep->dtor)(objp, cachep, 0);
1444 }
1445 }
1446#endif
1447
1448 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1449 struct slab_rcu *slab_rcu;
1450
1451 slab_rcu = (struct slab_rcu *) slabp;
1452 slab_rcu->cachep = cachep;
1453 slab_rcu->addr = addr;
1454 call_rcu(&slab_rcu->head, kmem_rcu_free);
1455 } else {
1456 kmem_freepages(cachep, addr);
1457 if (OFF_SLAB(cachep))
1458 kmem_cache_free(cachep->slabp_cache, slabp);
1459 }
1460}
1461
Christoph Lametere498be72005-09-09 13:03:32 -07001462/* For setting up all the kmem_list3s for cache whose objsize is same
1463 as size of kmem_list3. */
1464static inline void set_up_list3s(kmem_cache_t *cachep, int index)
1465{
1466 int node;
1467
1468 for_each_online_node(node) {
1469 cachep->nodelists[node] = &initkmem_list3[index+node];
1470 cachep->nodelists[node]->next_reap = jiffies +
1471 REAPTIMEOUT_LIST3 +
1472 ((unsigned long)cachep)%REAPTIMEOUT_LIST3;
1473 }
1474}
1475
Linus Torvalds1da177e2005-04-16 15:20:36 -07001476/**
1477 * kmem_cache_create - Create a cache.
1478 * @name: A string which is used in /proc/slabinfo to identify this cache.
1479 * @size: The size of objects to be created in this cache.
1480 * @align: The required alignment for the objects.
1481 * @flags: SLAB flags
1482 * @ctor: A constructor for the objects.
1483 * @dtor: A destructor for the objects.
1484 *
1485 * Returns a ptr to the cache on success, NULL on failure.
1486 * Cannot be called within a int, but can be interrupted.
1487 * The @ctor is run when new pages are allocated by the cache
1488 * and the @dtor is run before the pages are handed back.
1489 *
1490 * @name must be valid until the cache is destroyed. This implies that
1491 * the module calling this has to destroy the cache before getting
1492 * unloaded.
1493 *
1494 * The flags are
1495 *
1496 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1497 * to catch references to uninitialised memory.
1498 *
1499 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1500 * for buffer overruns.
1501 *
1502 * %SLAB_NO_REAP - Don't automatically reap this cache when we're under
1503 * memory pressure.
1504 *
1505 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1506 * cacheline. This can be beneficial if you're counting cycles as closely
1507 * as davem.
1508 */
1509kmem_cache_t *
1510kmem_cache_create (const char *name, size_t size, size_t align,
1511 unsigned long flags, void (*ctor)(void*, kmem_cache_t *, unsigned long),
1512 void (*dtor)(void*, kmem_cache_t *, unsigned long))
1513{
1514 size_t left_over, slab_size, ralign;
1515 kmem_cache_t *cachep = NULL;
Andrew Morton4f12bb42005-11-07 00:58:00 -08001516 struct list_head *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001517
1518 /*
1519 * Sanity checks... these are all serious usage bugs.
1520 */
1521 if ((!name) ||
1522 in_interrupt() ||
1523 (size < BYTES_PER_WORD) ||
1524 (size > (1<<MAX_OBJ_ORDER)*PAGE_SIZE) ||
1525 (dtor && !ctor)) {
1526 printk(KERN_ERR "%s: Early error in slab %s\n",
1527 __FUNCTION__, name);
1528 BUG();
1529 }
1530
Andrew Morton4f12bb42005-11-07 00:58:00 -08001531 down(&cache_chain_sem);
1532
1533 list_for_each(p, &cache_chain) {
1534 kmem_cache_t *pc = list_entry(p, kmem_cache_t, next);
1535 mm_segment_t old_fs = get_fs();
1536 char tmp;
1537 int res;
1538
1539 /*
1540 * This happens when the module gets unloaded and doesn't
1541 * destroy its slab cache and no-one else reuses the vmalloc
1542 * area of the module. Print a warning.
1543 */
1544 set_fs(KERNEL_DS);
1545 res = __get_user(tmp, pc->name);
1546 set_fs(old_fs);
1547 if (res) {
1548 printk("SLAB: cache with size %d has lost its name\n",
1549 pc->objsize);
1550 continue;
1551 }
1552
1553 if (!strcmp(pc->name,name)) {
1554 printk("kmem_cache_create: duplicate cache %s\n", name);
1555 dump_stack();
1556 goto oops;
1557 }
1558 }
1559
Linus Torvalds1da177e2005-04-16 15:20:36 -07001560#if DEBUG
1561 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
1562 if ((flags & SLAB_DEBUG_INITIAL) && !ctor) {
1563 /* No constructor, but inital state check requested */
1564 printk(KERN_ERR "%s: No con, but init state check "
1565 "requested - %s\n", __FUNCTION__, name);
1566 flags &= ~SLAB_DEBUG_INITIAL;
1567 }
1568
1569#if FORCED_DEBUG
1570 /*
1571 * Enable redzoning and last user accounting, except for caches with
1572 * large objects, if the increased size would increase the object size
1573 * above the next power of two: caches with object sizes just above a
1574 * power of two have a significant amount of internal fragmentation.
1575 */
1576 if ((size < 4096 || fls(size-1) == fls(size-1+3*BYTES_PER_WORD)))
1577 flags |= SLAB_RED_ZONE|SLAB_STORE_USER;
1578 if (!(flags & SLAB_DESTROY_BY_RCU))
1579 flags |= SLAB_POISON;
1580#endif
1581 if (flags & SLAB_DESTROY_BY_RCU)
1582 BUG_ON(flags & SLAB_POISON);
1583#endif
1584 if (flags & SLAB_DESTROY_BY_RCU)
1585 BUG_ON(dtor);
1586
1587 /*
1588 * Always checks flags, a caller might be expecting debug
1589 * support which isn't available.
1590 */
1591 if (flags & ~CREATE_MASK)
1592 BUG();
1593
1594 /* Check that size is in terms of words. This is needed to avoid
1595 * unaligned accesses for some archs when redzoning is used, and makes
1596 * sure any on-slab bufctl's are also correctly aligned.
1597 */
1598 if (size & (BYTES_PER_WORD-1)) {
1599 size += (BYTES_PER_WORD-1);
1600 size &= ~(BYTES_PER_WORD-1);
1601 }
1602
1603 /* calculate out the final buffer alignment: */
1604 /* 1) arch recommendation: can be overridden for debug */
1605 if (flags & SLAB_HWCACHE_ALIGN) {
1606 /* Default alignment: as specified by the arch code.
1607 * Except if an object is really small, then squeeze multiple
1608 * objects into one cacheline.
1609 */
1610 ralign = cache_line_size();
1611 while (size <= ralign/2)
1612 ralign /= 2;
1613 } else {
1614 ralign = BYTES_PER_WORD;
1615 }
1616 /* 2) arch mandated alignment: disables debug if necessary */
1617 if (ralign < ARCH_SLAB_MINALIGN) {
1618 ralign = ARCH_SLAB_MINALIGN;
1619 if (ralign > BYTES_PER_WORD)
1620 flags &= ~(SLAB_RED_ZONE|SLAB_STORE_USER);
1621 }
1622 /* 3) caller mandated alignment: disables debug if necessary */
1623 if (ralign < align) {
1624 ralign = align;
1625 if (ralign > BYTES_PER_WORD)
1626 flags &= ~(SLAB_RED_ZONE|SLAB_STORE_USER);
1627 }
1628 /* 4) Store it. Note that the debug code below can reduce
1629 * the alignment to BYTES_PER_WORD.
1630 */
1631 align = ralign;
1632
1633 /* Get cache's description obj. */
1634 cachep = (kmem_cache_t *) kmem_cache_alloc(&cache_cache, SLAB_KERNEL);
1635 if (!cachep)
Andrew Morton4f12bb42005-11-07 00:58:00 -08001636 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001637 memset(cachep, 0, sizeof(kmem_cache_t));
1638
1639#if DEBUG
1640 cachep->reallen = size;
1641
1642 if (flags & SLAB_RED_ZONE) {
1643 /* redzoning only works with word aligned caches */
1644 align = BYTES_PER_WORD;
1645
1646 /* add space for red zone words */
1647 cachep->dbghead += BYTES_PER_WORD;
1648 size += 2*BYTES_PER_WORD;
1649 }
1650 if (flags & SLAB_STORE_USER) {
1651 /* user store requires word alignment and
1652 * one word storage behind the end of the real
1653 * object.
1654 */
1655 align = BYTES_PER_WORD;
1656 size += BYTES_PER_WORD;
1657 }
1658#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Christoph Lametere498be72005-09-09 13:03:32 -07001659 if (size >= malloc_sizes[INDEX_L3+1].cs_size && cachep->reallen > cache_line_size() && size < PAGE_SIZE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001660 cachep->dbghead += PAGE_SIZE - size;
1661 size = PAGE_SIZE;
1662 }
1663#endif
1664#endif
1665
1666 /* Determine if the slab management is 'on' or 'off' slab. */
1667 if (size >= (PAGE_SIZE>>3))
1668 /*
1669 * Size is large, assume best to place the slab management obj
1670 * off-slab (should allow better packing of objs).
1671 */
1672 flags |= CFLGS_OFF_SLAB;
1673
1674 size = ALIGN(size, align);
1675
1676 if ((flags & SLAB_RECLAIM_ACCOUNT) && size <= PAGE_SIZE) {
1677 /*
1678 * A VFS-reclaimable slab tends to have most allocations
1679 * as GFP_NOFS and we really don't want to have to be allocating
1680 * higher-order pages when we are unable to shrink dcache.
1681 */
1682 cachep->gfporder = 0;
1683 cache_estimate(cachep->gfporder, size, align, flags,
1684 &left_over, &cachep->num);
1685 } else {
1686 /*
1687 * Calculate size (in pages) of slabs, and the num of objs per
1688 * slab. This could be made much more intelligent. For now,
1689 * try to avoid using high page-orders for slabs. When the
1690 * gfp() funcs are more friendly towards high-order requests,
1691 * this should be changed.
1692 */
1693 do {
1694 unsigned int break_flag = 0;
1695cal_wastage:
1696 cache_estimate(cachep->gfporder, size, align, flags,
1697 &left_over, &cachep->num);
1698 if (break_flag)
1699 break;
1700 if (cachep->gfporder >= MAX_GFP_ORDER)
1701 break;
1702 if (!cachep->num)
1703 goto next;
1704 if (flags & CFLGS_OFF_SLAB &&
1705 cachep->num > offslab_limit) {
1706 /* This num of objs will cause problems. */
1707 cachep->gfporder--;
1708 break_flag++;
1709 goto cal_wastage;
1710 }
1711
1712 /*
1713 * Large num of objs is good, but v. large slabs are
1714 * currently bad for the gfp()s.
1715 */
1716 if (cachep->gfporder >= slab_break_gfp_order)
1717 break;
1718
1719 if ((left_over*8) <= (PAGE_SIZE<<cachep->gfporder))
1720 break; /* Acceptable internal fragmentation. */
1721next:
1722 cachep->gfporder++;
1723 } while (1);
1724 }
1725
1726 if (!cachep->num) {
1727 printk("kmem_cache_create: couldn't create cache %s.\n", name);
1728 kmem_cache_free(&cache_cache, cachep);
1729 cachep = NULL;
Andrew Morton4f12bb42005-11-07 00:58:00 -08001730 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001731 }
1732 slab_size = ALIGN(cachep->num*sizeof(kmem_bufctl_t)
1733 + sizeof(struct slab), align);
1734
1735 /*
1736 * If the slab has been placed off-slab, and we have enough space then
1737 * move it on-slab. This is at the expense of any extra colouring.
1738 */
1739 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
1740 flags &= ~CFLGS_OFF_SLAB;
1741 left_over -= slab_size;
1742 }
1743
1744 if (flags & CFLGS_OFF_SLAB) {
1745 /* really off slab. No need for manual alignment */
1746 slab_size = cachep->num*sizeof(kmem_bufctl_t)+sizeof(struct slab);
1747 }
1748
1749 cachep->colour_off = cache_line_size();
1750 /* Offset must be a multiple of the alignment. */
1751 if (cachep->colour_off < align)
1752 cachep->colour_off = align;
1753 cachep->colour = left_over/cachep->colour_off;
1754 cachep->slab_size = slab_size;
1755 cachep->flags = flags;
1756 cachep->gfpflags = 0;
1757 if (flags & SLAB_CACHE_DMA)
1758 cachep->gfpflags |= GFP_DMA;
1759 spin_lock_init(&cachep->spinlock);
1760 cachep->objsize = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001761
1762 if (flags & CFLGS_OFF_SLAB)
Victor Fuscob2d55072005-09-10 00:26:36 -07001763 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001764 cachep->ctor = ctor;
1765 cachep->dtor = dtor;
1766 cachep->name = name;
1767
1768 /* Don't let CPUs to come and go */
1769 lock_cpu_hotplug();
1770
1771 if (g_cpucache_up == FULL) {
1772 enable_cpucache(cachep);
1773 } else {
1774 if (g_cpucache_up == NONE) {
1775 /* Note: the first kmem_cache_create must create
1776 * the cache that's used by kmalloc(24), otherwise
1777 * the creation of further caches will BUG().
1778 */
Christoph Lametere498be72005-09-09 13:03:32 -07001779 cachep->array[smp_processor_id()] =
1780 &initarray_generic.cache;
1781
1782 /* If the cache that's used by
1783 * kmalloc(sizeof(kmem_list3)) is the first cache,
1784 * then we need to set up all its list3s, otherwise
1785 * the creation of further caches will BUG().
1786 */
1787 set_up_list3s(cachep, SIZE_AC);
1788 if (INDEX_AC == INDEX_L3)
1789 g_cpucache_up = PARTIAL_L3;
1790 else
1791 g_cpucache_up = PARTIAL_AC;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001792 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07001793 cachep->array[smp_processor_id()] =
1794 kmalloc(sizeof(struct arraycache_init),
1795 GFP_KERNEL);
1796
1797 if (g_cpucache_up == PARTIAL_AC) {
1798 set_up_list3s(cachep, SIZE_L3);
1799 g_cpucache_up = PARTIAL_L3;
1800 } else {
1801 int node;
1802 for_each_online_node(node) {
1803
1804 cachep->nodelists[node] =
1805 kmalloc_node(sizeof(struct kmem_list3),
1806 GFP_KERNEL, node);
1807 BUG_ON(!cachep->nodelists[node]);
1808 kmem_list3_init(cachep->nodelists[node]);
1809 }
1810 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001811 }
Christoph Lametere498be72005-09-09 13:03:32 -07001812 cachep->nodelists[numa_node_id()]->next_reap =
1813 jiffies + REAPTIMEOUT_LIST3 +
1814 ((unsigned long)cachep)%REAPTIMEOUT_LIST3;
1815
Linus Torvalds1da177e2005-04-16 15:20:36 -07001816 BUG_ON(!ac_data(cachep));
1817 ac_data(cachep)->avail = 0;
1818 ac_data(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1819 ac_data(cachep)->batchcount = 1;
1820 ac_data(cachep)->touched = 0;
1821 cachep->batchcount = 1;
1822 cachep->limit = BOOT_CPUCACHE_ENTRIES;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001823 }
1824
Linus Torvalds1da177e2005-04-16 15:20:36 -07001825 /* cache setup completed, link it into the list */
1826 list_add(&cachep->next, &cache_chain);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001827 unlock_cpu_hotplug();
Andrew Morton4f12bb42005-11-07 00:58:00 -08001828oops:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001829 if (!cachep && (flags & SLAB_PANIC))
1830 panic("kmem_cache_create(): failed to create slab `%s'\n",
1831 name);
Andrew Morton4f12bb42005-11-07 00:58:00 -08001832 up(&cache_chain_sem);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001833 return cachep;
1834}
1835EXPORT_SYMBOL(kmem_cache_create);
1836
1837#if DEBUG
1838static void check_irq_off(void)
1839{
1840 BUG_ON(!irqs_disabled());
1841}
1842
1843static void check_irq_on(void)
1844{
1845 BUG_ON(irqs_disabled());
1846}
1847
1848static void check_spinlock_acquired(kmem_cache_t *cachep)
1849{
1850#ifdef CONFIG_SMP
1851 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07001852 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001853#endif
1854}
Christoph Lametere498be72005-09-09 13:03:32 -07001855
1856static inline void check_spinlock_acquired_node(kmem_cache_t *cachep, int node)
1857{
1858#ifdef CONFIG_SMP
1859 check_irq_off();
1860 assert_spin_locked(&cachep->nodelists[node]->list_lock);
1861#endif
1862}
1863
Linus Torvalds1da177e2005-04-16 15:20:36 -07001864#else
1865#define check_irq_off() do { } while(0)
1866#define check_irq_on() do { } while(0)
1867#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07001868#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001869#endif
1870
1871/*
1872 * Waits for all CPUs to execute func().
1873 */
1874static void smp_call_function_all_cpus(void (*func) (void *arg), void *arg)
1875{
1876 check_irq_on();
1877 preempt_disable();
1878
1879 local_irq_disable();
1880 func(arg);
1881 local_irq_enable();
1882
1883 if (smp_call_function(func, arg, 1, 1))
1884 BUG();
1885
1886 preempt_enable();
1887}
1888
1889static void drain_array_locked(kmem_cache_t* cachep,
Christoph Lametere498be72005-09-09 13:03:32 -07001890 struct array_cache *ac, int force, int node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001891
1892static void do_drain(void *arg)
1893{
1894 kmem_cache_t *cachep = (kmem_cache_t*)arg;
1895 struct array_cache *ac;
Christoph Lameterff694162005-09-22 21:44:02 -07001896 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001897
1898 check_irq_off();
1899 ac = ac_data(cachep);
Christoph Lameterff694162005-09-22 21:44:02 -07001900 spin_lock(&cachep->nodelists[node]->list_lock);
1901 free_block(cachep, ac->entry, ac->avail, node);
1902 spin_unlock(&cachep->nodelists[node]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001903 ac->avail = 0;
1904}
1905
1906static void drain_cpu_caches(kmem_cache_t *cachep)
1907{
Christoph Lametere498be72005-09-09 13:03:32 -07001908 struct kmem_list3 *l3;
1909 int node;
1910
Linus Torvalds1da177e2005-04-16 15:20:36 -07001911 smp_call_function_all_cpus(do_drain, cachep);
1912 check_irq_on();
1913 spin_lock_irq(&cachep->spinlock);
Christoph Lametere498be72005-09-09 13:03:32 -07001914 for_each_online_node(node) {
1915 l3 = cachep->nodelists[node];
1916 if (l3) {
1917 spin_lock(&l3->list_lock);
1918 drain_array_locked(cachep, l3->shared, 1, node);
1919 spin_unlock(&l3->list_lock);
1920 if (l3->alien)
1921 drain_alien_cache(cachep, l3);
1922 }
1923 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001924 spin_unlock_irq(&cachep->spinlock);
1925}
1926
Christoph Lametere498be72005-09-09 13:03:32 -07001927static int __node_shrink(kmem_cache_t *cachep, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001928{
1929 struct slab *slabp;
Christoph Lametere498be72005-09-09 13:03:32 -07001930 struct kmem_list3 *l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001931 int ret;
1932
Christoph Lametere498be72005-09-09 13:03:32 -07001933 for (;;) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001934 struct list_head *p;
1935
Christoph Lametere498be72005-09-09 13:03:32 -07001936 p = l3->slabs_free.prev;
1937 if (p == &l3->slabs_free)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001938 break;
1939
Christoph Lametere498be72005-09-09 13:03:32 -07001940 slabp = list_entry(l3->slabs_free.prev, struct slab, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001941#if DEBUG
1942 if (slabp->inuse)
1943 BUG();
1944#endif
1945 list_del(&slabp->list);
1946
Christoph Lametere498be72005-09-09 13:03:32 -07001947 l3->free_objects -= cachep->num;
1948 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001949 slab_destroy(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07001950 spin_lock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001951 }
Christoph Lametere498be72005-09-09 13:03:32 -07001952 ret = !list_empty(&l3->slabs_full) ||
1953 !list_empty(&l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001954 return ret;
1955}
1956
Christoph Lametere498be72005-09-09 13:03:32 -07001957static int __cache_shrink(kmem_cache_t *cachep)
1958{
1959 int ret = 0, i = 0;
1960 struct kmem_list3 *l3;
1961
1962 drain_cpu_caches(cachep);
1963
1964 check_irq_on();
1965 for_each_online_node(i) {
1966 l3 = cachep->nodelists[i];
1967 if (l3) {
1968 spin_lock_irq(&l3->list_lock);
1969 ret += __node_shrink(cachep, i);
1970 spin_unlock_irq(&l3->list_lock);
1971 }
1972 }
1973 return (ret ? 1 : 0);
1974}
1975
Linus Torvalds1da177e2005-04-16 15:20:36 -07001976/**
1977 * kmem_cache_shrink - Shrink a cache.
1978 * @cachep: The cache to shrink.
1979 *
1980 * Releases as many slabs as possible for a cache.
1981 * To help debugging, a zero exit status indicates all slabs were released.
1982 */
1983int kmem_cache_shrink(kmem_cache_t *cachep)
1984{
1985 if (!cachep || in_interrupt())
1986 BUG();
1987
1988 return __cache_shrink(cachep);
1989}
1990EXPORT_SYMBOL(kmem_cache_shrink);
1991
1992/**
1993 * kmem_cache_destroy - delete a cache
1994 * @cachep: the cache to destroy
1995 *
1996 * Remove a kmem_cache_t object from the slab cache.
1997 * Returns 0 on success.
1998 *
1999 * It is expected this function will be called by a module when it is
2000 * unloaded. This will remove the cache completely, and avoid a duplicate
2001 * cache being allocated each time a module is loaded and unloaded, if the
2002 * module doesn't have persistent in-kernel storage across loads and unloads.
2003 *
2004 * The cache must be empty before calling this function.
2005 *
2006 * The caller must guarantee that noone will allocate memory from the cache
2007 * during the kmem_cache_destroy().
2008 */
2009int kmem_cache_destroy(kmem_cache_t * cachep)
2010{
2011 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07002012 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002013
2014 if (!cachep || in_interrupt())
2015 BUG();
2016
2017 /* Don't let CPUs to come and go */
2018 lock_cpu_hotplug();
2019
2020 /* Find the cache in the chain of caches. */
2021 down(&cache_chain_sem);
2022 /*
2023 * the chain is never empty, cache_cache is never destroyed
2024 */
2025 list_del(&cachep->next);
2026 up(&cache_chain_sem);
2027
2028 if (__cache_shrink(cachep)) {
2029 slab_error(cachep, "Can't free all objects");
2030 down(&cache_chain_sem);
2031 list_add(&cachep->next,&cache_chain);
2032 up(&cache_chain_sem);
2033 unlock_cpu_hotplug();
2034 return 1;
2035 }
2036
2037 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
Paul E. McKenneyfbd568a3e2005-05-01 08:59:04 -07002038 synchronize_rcu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002039
Christoph Lametere498be72005-09-09 13:03:32 -07002040 for_each_online_cpu(i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002041 kfree(cachep->array[i]);
2042
2043 /* NUMA: free the list3 structures */
Christoph Lametere498be72005-09-09 13:03:32 -07002044 for_each_online_node(i) {
2045 if ((l3 = cachep->nodelists[i])) {
2046 kfree(l3->shared);
2047 free_alien_cache(l3->alien);
2048 kfree(l3);
2049 }
2050 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002051 kmem_cache_free(&cache_cache, cachep);
2052
2053 unlock_cpu_hotplug();
2054
2055 return 0;
2056}
2057EXPORT_SYMBOL(kmem_cache_destroy);
2058
2059/* Get the memory for a slab management obj. */
Christoph Lametere498be72005-09-09 13:03:32 -07002060static struct slab* alloc_slabmgmt(kmem_cache_t *cachep, void *objp,
Al Virodd0fc662005-10-07 07:46:04 +01002061 int colour_off, gfp_t local_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002062{
2063 struct slab *slabp;
2064
2065 if (OFF_SLAB(cachep)) {
2066 /* Slab management obj is off-slab. */
2067 slabp = kmem_cache_alloc(cachep->slabp_cache, local_flags);
2068 if (!slabp)
2069 return NULL;
2070 } else {
2071 slabp = objp+colour_off;
2072 colour_off += cachep->slab_size;
2073 }
2074 slabp->inuse = 0;
2075 slabp->colouroff = colour_off;
2076 slabp->s_mem = objp+colour_off;
2077
2078 return slabp;
2079}
2080
2081static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2082{
2083 return (kmem_bufctl_t *)(slabp+1);
2084}
2085
2086static void cache_init_objs(kmem_cache_t *cachep,
2087 struct slab *slabp, unsigned long ctor_flags)
2088{
2089 int i;
2090
2091 for (i = 0; i < cachep->num; i++) {
Christoph Lametere498be72005-09-09 13:03:32 -07002092 void *objp = slabp->s_mem+cachep->objsize*i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002093#if DEBUG
2094 /* need to poison the objs? */
2095 if (cachep->flags & SLAB_POISON)
2096 poison_obj(cachep, objp, POISON_FREE);
2097 if (cachep->flags & SLAB_STORE_USER)
2098 *dbg_userword(cachep, objp) = NULL;
2099
2100 if (cachep->flags & SLAB_RED_ZONE) {
2101 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2102 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2103 }
2104 /*
2105 * Constructors are not allowed to allocate memory from
2106 * the same cache which they are a constructor for.
2107 * Otherwise, deadlock. They must also be threaded.
2108 */
2109 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2110 cachep->ctor(objp+obj_dbghead(cachep), cachep, ctor_flags);
2111
2112 if (cachep->flags & SLAB_RED_ZONE) {
2113 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2114 slab_error(cachep, "constructor overwrote the"
2115 " end of an object");
2116 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2117 slab_error(cachep, "constructor overwrote the"
2118 " start of an object");
2119 }
2120 if ((cachep->objsize % PAGE_SIZE) == 0 && OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2121 kernel_map_pages(virt_to_page(objp), cachep->objsize/PAGE_SIZE, 0);
2122#else
2123 if (cachep->ctor)
2124 cachep->ctor(objp, cachep, ctor_flags);
2125#endif
2126 slab_bufctl(slabp)[i] = i+1;
2127 }
2128 slab_bufctl(slabp)[i-1] = BUFCTL_END;
2129 slabp->free = 0;
2130}
2131
Al Viro6daa0e22005-10-21 03:18:50 -04002132static void kmem_flagcheck(kmem_cache_t *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002133{
2134 if (flags & SLAB_DMA) {
2135 if (!(cachep->gfpflags & GFP_DMA))
2136 BUG();
2137 } else {
2138 if (cachep->gfpflags & GFP_DMA)
2139 BUG();
2140 }
2141}
2142
2143static void set_slab_attr(kmem_cache_t *cachep, struct slab *slabp, void *objp)
2144{
2145 int i;
2146 struct page *page;
2147
2148 /* Nasty!!!!!! I hope this is OK. */
2149 i = 1 << cachep->gfporder;
2150 page = virt_to_page(objp);
2151 do {
Pekka Enberg065d41c2005-11-13 16:06:46 -08002152 page_set_cache(page, cachep);
2153 page_set_slab(page, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002154 page++;
2155 } while (--i);
2156}
2157
2158/*
2159 * Grow (by 1) the number of slabs within a cache. This is called by
2160 * kmem_cache_alloc() when there are no active objs left in a cache.
2161 */
Al Virodd0fc662005-10-07 07:46:04 +01002162static int cache_grow(kmem_cache_t *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002163{
2164 struct slab *slabp;
2165 void *objp;
2166 size_t offset;
Al Viro6daa0e22005-10-21 03:18:50 -04002167 gfp_t local_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002168 unsigned long ctor_flags;
Christoph Lametere498be72005-09-09 13:03:32 -07002169 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002170
2171 /* Be lazy and only check for valid flags here,
2172 * keeping it out of the critical path in kmem_cache_alloc().
2173 */
2174 if (flags & ~(SLAB_DMA|SLAB_LEVEL_MASK|SLAB_NO_GROW))
2175 BUG();
2176 if (flags & SLAB_NO_GROW)
2177 return 0;
2178
2179 ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2180 local_flags = (flags & SLAB_LEVEL_MASK);
2181 if (!(local_flags & __GFP_WAIT))
2182 /*
2183 * Not allowed to sleep. Need to tell a constructor about
2184 * this - it might need to know...
2185 */
2186 ctor_flags |= SLAB_CTOR_ATOMIC;
2187
2188 /* About to mess with non-constant members - lock. */
2189 check_irq_off();
2190 spin_lock(&cachep->spinlock);
2191
2192 /* Get colour for the slab, and cal the next value. */
2193 offset = cachep->colour_next;
2194 cachep->colour_next++;
2195 if (cachep->colour_next >= cachep->colour)
2196 cachep->colour_next = 0;
2197 offset *= cachep->colour_off;
2198
2199 spin_unlock(&cachep->spinlock);
2200
Christoph Lametere498be72005-09-09 13:03:32 -07002201 check_irq_off();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002202 if (local_flags & __GFP_WAIT)
2203 local_irq_enable();
2204
2205 /*
2206 * The test for missing atomic flag is performed here, rather than
2207 * the more obvious place, simply to reduce the critical path length
2208 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2209 * will eventually be caught here (where it matters).
2210 */
2211 kmem_flagcheck(cachep, flags);
2212
Christoph Lametere498be72005-09-09 13:03:32 -07002213 /* Get mem for the objs.
2214 * Attempt to allocate a physical page from 'nodeid',
2215 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002216 if (!(objp = kmem_getpages(cachep, flags, nodeid)))
2217 goto failed;
2218
2219 /* Get slab management. */
2220 if (!(slabp = alloc_slabmgmt(cachep, objp, offset, local_flags)))
2221 goto opps1;
2222
Christoph Lametere498be72005-09-09 13:03:32 -07002223 slabp->nodeid = nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002224 set_slab_attr(cachep, slabp, objp);
2225
2226 cache_init_objs(cachep, slabp, ctor_flags);
2227
2228 if (local_flags & __GFP_WAIT)
2229 local_irq_disable();
2230 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002231 l3 = cachep->nodelists[nodeid];
2232 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002233
2234 /* Make slab active. */
Christoph Lametere498be72005-09-09 13:03:32 -07002235 list_add_tail(&slabp->list, &(l3->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002236 STATS_INC_GROWN(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002237 l3->free_objects += cachep->num;
2238 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002239 return 1;
2240opps1:
2241 kmem_freepages(cachep, objp);
2242failed:
2243 if (local_flags & __GFP_WAIT)
2244 local_irq_disable();
2245 return 0;
2246}
2247
2248#if DEBUG
2249
2250/*
2251 * Perform extra freeing checks:
2252 * - detect bad pointers.
2253 * - POISON/RED_ZONE checking
2254 * - destructor calls, for caches with POISON+dtor
2255 */
2256static void kfree_debugcheck(const void *objp)
2257{
2258 struct page *page;
2259
2260 if (!virt_addr_valid(objp)) {
2261 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2262 (unsigned long)objp);
2263 BUG();
2264 }
2265 page = virt_to_page(objp);
2266 if (!PageSlab(page)) {
2267 printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n", (unsigned long)objp);
2268 BUG();
2269 }
2270}
2271
2272static void *cache_free_debugcheck(kmem_cache_t *cachep, void *objp,
2273 void *caller)
2274{
2275 struct page *page;
2276 unsigned int objnr;
2277 struct slab *slabp;
2278
2279 objp -= obj_dbghead(cachep);
2280 kfree_debugcheck(objp);
2281 page = virt_to_page(objp);
2282
Pekka Enberg065d41c2005-11-13 16:06:46 -08002283 if (page_get_cache(page) != cachep) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002284 printk(KERN_ERR "mismatch in kmem_cache_free: expected cache %p, got %p\n",
Pekka Enberg065d41c2005-11-13 16:06:46 -08002285 page_get_cache(page),cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002286 printk(KERN_ERR "%p is %s.\n", cachep, cachep->name);
Pekka Enberg065d41c2005-11-13 16:06:46 -08002287 printk(KERN_ERR "%p is %s.\n", page_get_cache(page), page_get_cache(page)->name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002288 WARN_ON(1);
2289 }
Pekka Enberg065d41c2005-11-13 16:06:46 -08002290 slabp = page_get_slab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002291
2292 if (cachep->flags & SLAB_RED_ZONE) {
2293 if (*dbg_redzone1(cachep, objp) != RED_ACTIVE || *dbg_redzone2(cachep, objp) != RED_ACTIVE) {
2294 slab_error(cachep, "double free, or memory outside"
2295 " object was overwritten");
2296 printk(KERN_ERR "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n",
2297 objp, *dbg_redzone1(cachep, objp), *dbg_redzone2(cachep, objp));
2298 }
2299 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2300 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2301 }
2302 if (cachep->flags & SLAB_STORE_USER)
2303 *dbg_userword(cachep, objp) = caller;
2304
2305 objnr = (objp-slabp->s_mem)/cachep->objsize;
2306
2307 BUG_ON(objnr >= cachep->num);
2308 BUG_ON(objp != slabp->s_mem + objnr*cachep->objsize);
2309
2310 if (cachep->flags & SLAB_DEBUG_INITIAL) {
2311 /* Need to call the slab's constructor so the
2312 * caller can perform a verify of its state (debugging).
2313 * Called without the cache-lock held.
2314 */
2315 cachep->ctor(objp+obj_dbghead(cachep),
2316 cachep, SLAB_CTOR_CONSTRUCTOR|SLAB_CTOR_VERIFY);
2317 }
2318 if (cachep->flags & SLAB_POISON && cachep->dtor) {
2319 /* we want to cache poison the object,
2320 * call the destruction callback
2321 */
2322 cachep->dtor(objp+obj_dbghead(cachep), cachep, 0);
2323 }
2324 if (cachep->flags & SLAB_POISON) {
2325#ifdef CONFIG_DEBUG_PAGEALLOC
2326 if ((cachep->objsize % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) {
2327 store_stackinfo(cachep, objp, (unsigned long)caller);
2328 kernel_map_pages(virt_to_page(objp), cachep->objsize/PAGE_SIZE, 0);
2329 } else {
2330 poison_obj(cachep, objp, POISON_FREE);
2331 }
2332#else
2333 poison_obj(cachep, objp, POISON_FREE);
2334#endif
2335 }
2336 return objp;
2337}
2338
2339static void check_slabp(kmem_cache_t *cachep, struct slab *slabp)
2340{
2341 kmem_bufctl_t i;
2342 int entries = 0;
2343
Linus Torvalds1da177e2005-04-16 15:20:36 -07002344 /* Check slab's freelist to see if this obj is there. */
2345 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2346 entries++;
2347 if (entries > cachep->num || i >= cachep->num)
2348 goto bad;
2349 }
2350 if (entries != cachep->num - slabp->inuse) {
2351bad:
2352 printk(KERN_ERR "slab: Internal list corruption detected in cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2353 cachep->name, cachep->num, slabp, slabp->inuse);
2354 for (i=0;i<sizeof(slabp)+cachep->num*sizeof(kmem_bufctl_t);i++) {
2355 if ((i%16)==0)
2356 printk("\n%03x:", i);
2357 printk(" %02x", ((unsigned char*)slabp)[i]);
2358 }
2359 printk("\n");
2360 BUG();
2361 }
2362}
2363#else
2364#define kfree_debugcheck(x) do { } while(0)
2365#define cache_free_debugcheck(x,objp,z) (objp)
2366#define check_slabp(x,y) do { } while(0)
2367#endif
2368
Al Virodd0fc662005-10-07 07:46:04 +01002369static void *cache_alloc_refill(kmem_cache_t *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002370{
2371 int batchcount;
2372 struct kmem_list3 *l3;
2373 struct array_cache *ac;
2374
2375 check_irq_off();
2376 ac = ac_data(cachep);
2377retry:
2378 batchcount = ac->batchcount;
2379 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2380 /* if there was little recent activity on this
2381 * cache, then perform only a partial refill.
2382 * Otherwise we could generate refill bouncing.
2383 */
2384 batchcount = BATCHREFILL_LIMIT;
2385 }
Christoph Lametere498be72005-09-09 13:03:32 -07002386 l3 = cachep->nodelists[numa_node_id()];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002387
Christoph Lametere498be72005-09-09 13:03:32 -07002388 BUG_ON(ac->avail > 0 || !l3);
2389 spin_lock(&l3->list_lock);
2390
Linus Torvalds1da177e2005-04-16 15:20:36 -07002391 if (l3->shared) {
2392 struct array_cache *shared_array = l3->shared;
2393 if (shared_array->avail) {
2394 if (batchcount > shared_array->avail)
2395 batchcount = shared_array->avail;
2396 shared_array->avail -= batchcount;
2397 ac->avail = batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07002398 memcpy(ac->entry,
2399 &(shared_array->entry[shared_array->avail]),
2400 sizeof(void*)*batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002401 shared_array->touched = 1;
2402 goto alloc_done;
2403 }
2404 }
2405 while (batchcount > 0) {
2406 struct list_head *entry;
2407 struct slab *slabp;
2408 /* Get slab alloc is to come from. */
2409 entry = l3->slabs_partial.next;
2410 if (entry == &l3->slabs_partial) {
2411 l3->free_touched = 1;
2412 entry = l3->slabs_free.next;
2413 if (entry == &l3->slabs_free)
2414 goto must_grow;
2415 }
2416
2417 slabp = list_entry(entry, struct slab, list);
2418 check_slabp(cachep, slabp);
2419 check_spinlock_acquired(cachep);
2420 while (slabp->inuse < cachep->num && batchcount--) {
2421 kmem_bufctl_t next;
2422 STATS_INC_ALLOCED(cachep);
2423 STATS_INC_ACTIVE(cachep);
2424 STATS_SET_HIGH(cachep);
2425
2426 /* get obj pointer */
Christoph Lametere498be72005-09-09 13:03:32 -07002427 ac->entry[ac->avail++] = slabp->s_mem +
2428 slabp->free*cachep->objsize;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002429
2430 slabp->inuse++;
2431 next = slab_bufctl(slabp)[slabp->free];
2432#if DEBUG
2433 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
Christoph Lameter09ad4bb2005-10-29 18:15:52 -07002434 WARN_ON(numa_node_id() != slabp->nodeid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002435#endif
2436 slabp->free = next;
2437 }
2438 check_slabp(cachep, slabp);
2439
2440 /* move slabp to correct slabp list: */
2441 list_del(&slabp->list);
2442 if (slabp->free == BUFCTL_END)
2443 list_add(&slabp->list, &l3->slabs_full);
2444 else
2445 list_add(&slabp->list, &l3->slabs_partial);
2446 }
2447
2448must_grow:
2449 l3->free_objects -= ac->avail;
2450alloc_done:
Christoph Lametere498be72005-09-09 13:03:32 -07002451 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002452
2453 if (unlikely(!ac->avail)) {
2454 int x;
Christoph Lametere498be72005-09-09 13:03:32 -07002455 x = cache_grow(cachep, flags, numa_node_id());
2456
Linus Torvalds1da177e2005-04-16 15:20:36 -07002457 // cache_grow can reenable interrupts, then ac could change.
2458 ac = ac_data(cachep);
2459 if (!x && ac->avail == 0) // no objects in sight? abort
2460 return NULL;
2461
2462 if (!ac->avail) // objects refilled by interrupt?
2463 goto retry;
2464 }
2465 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07002466 return ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002467}
2468
2469static inline void
Al Virodd0fc662005-10-07 07:46:04 +01002470cache_alloc_debugcheck_before(kmem_cache_t *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002471{
2472 might_sleep_if(flags & __GFP_WAIT);
2473#if DEBUG
2474 kmem_flagcheck(cachep, flags);
2475#endif
2476}
2477
2478#if DEBUG
2479static void *
2480cache_alloc_debugcheck_after(kmem_cache_t *cachep,
Al Virodd0fc662005-10-07 07:46:04 +01002481 gfp_t flags, void *objp, void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002482{
2483 if (!objp)
2484 return objp;
2485 if (cachep->flags & SLAB_POISON) {
2486#ifdef CONFIG_DEBUG_PAGEALLOC
2487 if ((cachep->objsize % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
2488 kernel_map_pages(virt_to_page(objp), cachep->objsize/PAGE_SIZE, 1);
2489 else
2490 check_poison_obj(cachep, objp);
2491#else
2492 check_poison_obj(cachep, objp);
2493#endif
2494 poison_obj(cachep, objp, POISON_INUSE);
2495 }
2496 if (cachep->flags & SLAB_STORE_USER)
2497 *dbg_userword(cachep, objp) = caller;
2498
2499 if (cachep->flags & SLAB_RED_ZONE) {
2500 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE || *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2501 slab_error(cachep, "double free, or memory outside"
2502 " object was overwritten");
2503 printk(KERN_ERR "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n",
2504 objp, *dbg_redzone1(cachep, objp), *dbg_redzone2(cachep, objp));
2505 }
2506 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2507 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2508 }
2509 objp += obj_dbghead(cachep);
2510 if (cachep->ctor && cachep->flags & SLAB_POISON) {
2511 unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2512
2513 if (!(flags & __GFP_WAIT))
2514 ctor_flags |= SLAB_CTOR_ATOMIC;
2515
2516 cachep->ctor(objp, cachep, ctor_flags);
2517 }
2518 return objp;
2519}
2520#else
2521#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2522#endif
2523
Al Virodd0fc662005-10-07 07:46:04 +01002524static inline void *____cache_alloc(kmem_cache_t *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002525{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002526 void* objp;
2527 struct array_cache *ac;
2528
Alok N Kataria5c382302005-09-27 21:45:46 -07002529 check_irq_off();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002530 ac = ac_data(cachep);
2531 if (likely(ac->avail)) {
2532 STATS_INC_ALLOCHIT(cachep);
2533 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07002534 objp = ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002535 } else {
2536 STATS_INC_ALLOCMISS(cachep);
2537 objp = cache_alloc_refill(cachep, flags);
2538 }
Alok N Kataria5c382302005-09-27 21:45:46 -07002539 return objp;
2540}
2541
Al Virodd0fc662005-10-07 07:46:04 +01002542static inline void *__cache_alloc(kmem_cache_t *cachep, gfp_t flags)
Alok N Kataria5c382302005-09-27 21:45:46 -07002543{
2544 unsigned long save_flags;
2545 void* objp;
2546
2547 cache_alloc_debugcheck_before(cachep, flags);
2548
2549 local_irq_save(save_flags);
2550 objp = ____cache_alloc(cachep, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002551 local_irq_restore(save_flags);
Eric Dumazet34342e82005-09-03 15:55:06 -07002552 objp = cache_alloc_debugcheck_after(cachep, flags, objp,
2553 __builtin_return_address(0));
2554 prefetchw(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002555 return objp;
2556}
2557
Christoph Lametere498be72005-09-09 13:03:32 -07002558#ifdef CONFIG_NUMA
2559/*
2560 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002561 */
Al Viro6daa0e22005-10-21 03:18:50 -04002562static void *__cache_alloc_node(kmem_cache_t *cachep, gfp_t flags, int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07002563{
2564 struct list_head *entry;
2565 struct slab *slabp;
2566 struct kmem_list3 *l3;
2567 void *obj;
2568 kmem_bufctl_t next;
2569 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002570
Christoph Lametere498be72005-09-09 13:03:32 -07002571 l3 = cachep->nodelists[nodeid];
2572 BUG_ON(!l3);
2573
2574retry:
2575 spin_lock(&l3->list_lock);
2576 entry = l3->slabs_partial.next;
2577 if (entry == &l3->slabs_partial) {
2578 l3->free_touched = 1;
2579 entry = l3->slabs_free.next;
2580 if (entry == &l3->slabs_free)
2581 goto must_grow;
2582 }
2583
2584 slabp = list_entry(entry, struct slab, list);
2585 check_spinlock_acquired_node(cachep, nodeid);
2586 check_slabp(cachep, slabp);
2587
2588 STATS_INC_NODEALLOCS(cachep);
2589 STATS_INC_ACTIVE(cachep);
2590 STATS_SET_HIGH(cachep);
2591
2592 BUG_ON(slabp->inuse == cachep->num);
2593
2594 /* get obj pointer */
2595 obj = slabp->s_mem + slabp->free*cachep->objsize;
2596 slabp->inuse++;
2597 next = slab_bufctl(slabp)[slabp->free];
2598#if DEBUG
2599 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2600#endif
2601 slabp->free = next;
2602 check_slabp(cachep, slabp);
2603 l3->free_objects--;
2604 /* move slabp to correct slabp list: */
2605 list_del(&slabp->list);
2606
2607 if (slabp->free == BUFCTL_END) {
2608 list_add(&slabp->list, &l3->slabs_full);
2609 } else {
2610 list_add(&slabp->list, &l3->slabs_partial);
2611 }
2612
2613 spin_unlock(&l3->list_lock);
2614 goto done;
2615
2616must_grow:
2617 spin_unlock(&l3->list_lock);
2618 x = cache_grow(cachep, flags, nodeid);
2619
2620 if (!x)
2621 return NULL;
2622
2623 goto retry;
2624done:
2625 return obj;
2626}
2627#endif
2628
2629/*
2630 * Caller needs to acquire correct kmem_list's list_lock
2631 */
Christoph Lameterff694162005-09-22 21:44:02 -07002632static void free_block(kmem_cache_t *cachep, void **objpp, int nr_objects, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002633{
2634 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07002635 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002636
2637 for (i = 0; i < nr_objects; i++) {
2638 void *objp = objpp[i];
2639 struct slab *slabp;
2640 unsigned int objnr;
2641
Pekka Enberg065d41c2005-11-13 16:06:46 -08002642 slabp = page_get_slab(virt_to_page(objp));
Christoph Lameterff694162005-09-22 21:44:02 -07002643 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002644 list_del(&slabp->list);
2645 objnr = (objp - slabp->s_mem) / cachep->objsize;
Christoph Lameterff694162005-09-22 21:44:02 -07002646 check_spinlock_acquired_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002647 check_slabp(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07002648
Linus Torvalds1da177e2005-04-16 15:20:36 -07002649#if DEBUG
Christoph Lameter09ad4bb2005-10-29 18:15:52 -07002650 /* Verify that the slab belongs to the intended node */
2651 WARN_ON(slabp->nodeid != node);
2652
Linus Torvalds1da177e2005-04-16 15:20:36 -07002653 if (slab_bufctl(slabp)[objnr] != BUFCTL_FREE) {
Christoph Lametere498be72005-09-09 13:03:32 -07002654 printk(KERN_ERR "slab: double free detected in cache "
2655 "'%s', objp %p\n", cachep->name, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002656 BUG();
2657 }
2658#endif
2659 slab_bufctl(slabp)[objnr] = slabp->free;
2660 slabp->free = objnr;
2661 STATS_DEC_ACTIVE(cachep);
2662 slabp->inuse--;
Christoph Lametere498be72005-09-09 13:03:32 -07002663 l3->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002664 check_slabp(cachep, slabp);
2665
2666 /* fixup slab chains */
2667 if (slabp->inuse == 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07002668 if (l3->free_objects > l3->free_limit) {
2669 l3->free_objects -= cachep->num;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002670 slab_destroy(cachep, slabp);
2671 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07002672 list_add(&slabp->list, &l3->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002673 }
2674 } else {
2675 /* Unconditionally move a slab to the end of the
2676 * partial list on free - maximum time for the
2677 * other objects to be freed, too.
2678 */
Christoph Lametere498be72005-09-09 13:03:32 -07002679 list_add_tail(&slabp->list, &l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002680 }
2681 }
2682}
2683
2684static void cache_flusharray(kmem_cache_t *cachep, struct array_cache *ac)
2685{
2686 int batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07002687 struct kmem_list3 *l3;
Christoph Lameterff694162005-09-22 21:44:02 -07002688 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002689
2690 batchcount = ac->batchcount;
2691#if DEBUG
2692 BUG_ON(!batchcount || batchcount > ac->avail);
2693#endif
2694 check_irq_off();
Christoph Lameterff694162005-09-22 21:44:02 -07002695 l3 = cachep->nodelists[node];
Christoph Lametere498be72005-09-09 13:03:32 -07002696 spin_lock(&l3->list_lock);
2697 if (l3->shared) {
2698 struct array_cache *shared_array = l3->shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002699 int max = shared_array->limit-shared_array->avail;
2700 if (max) {
2701 if (batchcount > max)
2702 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07002703 memcpy(&(shared_array->entry[shared_array->avail]),
2704 ac->entry,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002705 sizeof(void*)*batchcount);
2706 shared_array->avail += batchcount;
2707 goto free_done;
2708 }
2709 }
2710
Christoph Lameterff694162005-09-22 21:44:02 -07002711 free_block(cachep, ac->entry, batchcount, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002712free_done:
2713#if STATS
2714 {
2715 int i = 0;
2716 struct list_head *p;
2717
Christoph Lametere498be72005-09-09 13:03:32 -07002718 p = l3->slabs_free.next;
2719 while (p != &(l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002720 struct slab *slabp;
2721
2722 slabp = list_entry(p, struct slab, list);
2723 BUG_ON(slabp->inuse);
2724
2725 i++;
2726 p = p->next;
2727 }
2728 STATS_SET_FREEABLE(cachep, i);
2729 }
2730#endif
Christoph Lametere498be72005-09-09 13:03:32 -07002731 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002732 ac->avail -= batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07002733 memmove(ac->entry, &(ac->entry[batchcount]),
Linus Torvalds1da177e2005-04-16 15:20:36 -07002734 sizeof(void*)*ac->avail);
2735}
2736
Christoph Lametere498be72005-09-09 13:03:32 -07002737
Linus Torvalds1da177e2005-04-16 15:20:36 -07002738/*
2739 * __cache_free
2740 * Release an obj back to its cache. If the obj has a constructed
2741 * state, it must be in this state _before_ it is released.
2742 *
2743 * Called with disabled ints.
2744 */
2745static inline void __cache_free(kmem_cache_t *cachep, void *objp)
2746{
2747 struct array_cache *ac = ac_data(cachep);
2748
2749 check_irq_off();
2750 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
2751
Christoph Lametere498be72005-09-09 13:03:32 -07002752 /* Make sure we are not freeing a object from another
2753 * node to the array cache on this cpu.
2754 */
2755#ifdef CONFIG_NUMA
2756 {
2757 struct slab *slabp;
Pekka Enberg065d41c2005-11-13 16:06:46 -08002758 slabp = page_get_slab(virt_to_page(objp));
Christoph Lametere498be72005-09-09 13:03:32 -07002759 if (unlikely(slabp->nodeid != numa_node_id())) {
2760 struct array_cache *alien = NULL;
2761 int nodeid = slabp->nodeid;
2762 struct kmem_list3 *l3 = cachep->nodelists[numa_node_id()];
2763
2764 STATS_INC_NODEFREES(cachep);
2765 if (l3->alien && l3->alien[nodeid]) {
2766 alien = l3->alien[nodeid];
2767 spin_lock(&alien->lock);
2768 if (unlikely(alien->avail == alien->limit))
2769 __drain_alien_cache(cachep,
2770 alien, nodeid);
2771 alien->entry[alien->avail++] = objp;
2772 spin_unlock(&alien->lock);
2773 } else {
2774 spin_lock(&(cachep->nodelists[nodeid])->
2775 list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -07002776 free_block(cachep, &objp, 1, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07002777 spin_unlock(&(cachep->nodelists[nodeid])->
2778 list_lock);
2779 }
2780 return;
2781 }
2782 }
2783#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002784 if (likely(ac->avail < ac->limit)) {
2785 STATS_INC_FREEHIT(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002786 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002787 return;
2788 } else {
2789 STATS_INC_FREEMISS(cachep);
2790 cache_flusharray(cachep, ac);
Christoph Lametere498be72005-09-09 13:03:32 -07002791 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002792 }
2793}
2794
2795/**
2796 * kmem_cache_alloc - Allocate an object
2797 * @cachep: The cache to allocate from.
2798 * @flags: See kmalloc().
2799 *
2800 * Allocate an object from this cache. The flags are only relevant
2801 * if the cache has no available objects.
2802 */
Al Virodd0fc662005-10-07 07:46:04 +01002803void *kmem_cache_alloc(kmem_cache_t *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002804{
2805 return __cache_alloc(cachep, flags);
2806}
2807EXPORT_SYMBOL(kmem_cache_alloc);
2808
2809/**
2810 * kmem_ptr_validate - check if an untrusted pointer might
2811 * be a slab entry.
2812 * @cachep: the cache we're checking against
2813 * @ptr: pointer to validate
2814 *
2815 * This verifies that the untrusted pointer looks sane:
2816 * it is _not_ a guarantee that the pointer is actually
2817 * part of the slab cache in question, but it at least
2818 * validates that the pointer can be dereferenced and
2819 * looks half-way sane.
2820 *
2821 * Currently only used for dentry validation.
2822 */
2823int fastcall kmem_ptr_validate(kmem_cache_t *cachep, void *ptr)
2824{
2825 unsigned long addr = (unsigned long) ptr;
2826 unsigned long min_addr = PAGE_OFFSET;
2827 unsigned long align_mask = BYTES_PER_WORD-1;
2828 unsigned long size = cachep->objsize;
2829 struct page *page;
2830
2831 if (unlikely(addr < min_addr))
2832 goto out;
2833 if (unlikely(addr > (unsigned long)high_memory - size))
2834 goto out;
2835 if (unlikely(addr & align_mask))
2836 goto out;
2837 if (unlikely(!kern_addr_valid(addr)))
2838 goto out;
2839 if (unlikely(!kern_addr_valid(addr + size - 1)))
2840 goto out;
2841 page = virt_to_page(ptr);
2842 if (unlikely(!PageSlab(page)))
2843 goto out;
Pekka Enberg065d41c2005-11-13 16:06:46 -08002844 if (unlikely(page_get_cache(page) != cachep))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002845 goto out;
2846 return 1;
2847out:
2848 return 0;
2849}
2850
2851#ifdef CONFIG_NUMA
2852/**
2853 * kmem_cache_alloc_node - Allocate an object on the specified node
2854 * @cachep: The cache to allocate from.
2855 * @flags: See kmalloc().
2856 * @nodeid: node number of the target node.
2857 *
2858 * Identical to kmem_cache_alloc, except that this function is slow
2859 * and can sleep. And it will allocate memory on the given node, which
2860 * can improve the performance for cpu bound structures.
Christoph Lametere498be72005-09-09 13:03:32 -07002861 * New and improved: it will now make sure that the object gets
2862 * put on the correct node list so that there is no false sharing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002863 */
Al Virodd0fc662005-10-07 07:46:04 +01002864void *kmem_cache_alloc_node(kmem_cache_t *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002865{
Christoph Lametere498be72005-09-09 13:03:32 -07002866 unsigned long save_flags;
2867 void *ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002868
Christoph Lameterff694162005-09-22 21:44:02 -07002869 if (nodeid == -1)
Christoph Lametere498be72005-09-09 13:03:32 -07002870 return __cache_alloc(cachep, flags);
Christoph Lameter83b78bd2005-07-06 10:47:07 -07002871
Christoph Lametere498be72005-09-09 13:03:32 -07002872 if (unlikely(!cachep->nodelists[nodeid])) {
2873 /* Fall back to __cache_alloc if we run into trouble */
2874 printk(KERN_WARNING "slab: not allocating in inactive node %d for cache %s\n", nodeid, cachep->name);
2875 return __cache_alloc(cachep,flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002876 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002877
Christoph Lametere498be72005-09-09 13:03:32 -07002878 cache_alloc_debugcheck_before(cachep, flags);
2879 local_irq_save(save_flags);
Alok N Kataria5c382302005-09-27 21:45:46 -07002880 if (nodeid == numa_node_id())
2881 ptr = ____cache_alloc(cachep, flags);
2882 else
2883 ptr = __cache_alloc_node(cachep, flags, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07002884 local_irq_restore(save_flags);
2885 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002886
Christoph Lametere498be72005-09-09 13:03:32 -07002887 return ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002888}
2889EXPORT_SYMBOL(kmem_cache_alloc_node);
2890
Al Virodd0fc662005-10-07 07:46:04 +01002891void *kmalloc_node(size_t size, gfp_t flags, int node)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07002892{
2893 kmem_cache_t *cachep;
2894
2895 cachep = kmem_find_general_cachep(size, flags);
2896 if (unlikely(cachep == NULL))
2897 return NULL;
2898 return kmem_cache_alloc_node(cachep, flags, node);
2899}
2900EXPORT_SYMBOL(kmalloc_node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002901#endif
2902
2903/**
2904 * kmalloc - allocate memory
2905 * @size: how many bytes of memory are required.
2906 * @flags: the type of memory to allocate.
2907 *
2908 * kmalloc is the normal method of allocating memory
2909 * in the kernel.
2910 *
2911 * The @flags argument may be one of:
2912 *
2913 * %GFP_USER - Allocate memory on behalf of user. May sleep.
2914 *
2915 * %GFP_KERNEL - Allocate normal kernel ram. May sleep.
2916 *
2917 * %GFP_ATOMIC - Allocation will not sleep. Use inside interrupt handlers.
2918 *
2919 * Additionally, the %GFP_DMA flag may be set to indicate the memory
2920 * must be suitable for DMA. This can mean different things on different
2921 * platforms. For example, on i386, it means that the memory must come
2922 * from the first 16MB.
2923 */
Al Virodd0fc662005-10-07 07:46:04 +01002924void *__kmalloc(size_t size, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002925{
2926 kmem_cache_t *cachep;
2927
Manfred Spraul97e2bde2005-05-01 08:58:38 -07002928 /* If you want to save a few bytes .text space: replace
2929 * __ with kmem_.
2930 * Then kmalloc uses the uninlined functions instead of the inline
2931 * functions.
2932 */
2933 cachep = __find_general_cachep(size, flags);
Andrew Mortondbdb9042005-09-23 13:24:10 -07002934 if (unlikely(cachep == NULL))
2935 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002936 return __cache_alloc(cachep, flags);
2937}
2938EXPORT_SYMBOL(__kmalloc);
2939
2940#ifdef CONFIG_SMP
2941/**
2942 * __alloc_percpu - allocate one copy of the object for every present
2943 * cpu in the system, zeroing them.
2944 * Objects should be dereferenced using the per_cpu_ptr macro only.
2945 *
2946 * @size: how many bytes of memory are required.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002947 */
Pekka Enbergf9f75002006-01-08 01:00:33 -08002948void *__alloc_percpu(size_t size)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002949{
2950 int i;
2951 struct percpu_data *pdata = kmalloc(sizeof (*pdata), GFP_KERNEL);
2952
2953 if (!pdata)
2954 return NULL;
2955
Christoph Lametere498be72005-09-09 13:03:32 -07002956 /*
2957 * Cannot use for_each_online_cpu since a cpu may come online
2958 * and we have no way of figuring out how to fix the array
2959 * that we have allocated then....
2960 */
2961 for_each_cpu(i) {
2962 int node = cpu_to_node(i);
2963
2964 if (node_online(node))
2965 pdata->ptrs[i] = kmalloc_node(size, GFP_KERNEL, node);
2966 else
2967 pdata->ptrs[i] = kmalloc(size, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002968
2969 if (!pdata->ptrs[i])
2970 goto unwind_oom;
2971 memset(pdata->ptrs[i], 0, size);
2972 }
2973
2974 /* Catch derefs w/o wrappers */
2975 return (void *) (~(unsigned long) pdata);
2976
2977unwind_oom:
2978 while (--i >= 0) {
2979 if (!cpu_possible(i))
2980 continue;
2981 kfree(pdata->ptrs[i]);
2982 }
2983 kfree(pdata);
2984 return NULL;
2985}
2986EXPORT_SYMBOL(__alloc_percpu);
2987#endif
2988
2989/**
2990 * kmem_cache_free - Deallocate an object
2991 * @cachep: The cache the allocation was from.
2992 * @objp: The previously allocated object.
2993 *
2994 * Free an object which was previously allocated from this
2995 * cache.
2996 */
2997void kmem_cache_free(kmem_cache_t *cachep, void *objp)
2998{
2999 unsigned long flags;
3000
3001 local_irq_save(flags);
3002 __cache_free(cachep, objp);
3003 local_irq_restore(flags);
3004}
3005EXPORT_SYMBOL(kmem_cache_free);
3006
3007/**
Pekka J Enbergdd392712005-09-06 15:18:31 -07003008 * kzalloc - allocate memory. The memory is set to zero.
3009 * @size: how many bytes of memory are required.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003010 * @flags: the type of memory to allocate.
3011 */
Al Virodd0fc662005-10-07 07:46:04 +01003012void *kzalloc(size_t size, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003013{
Pekka J Enbergdd392712005-09-06 15:18:31 -07003014 void *ret = kmalloc(size, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003015 if (ret)
Pekka J Enbergdd392712005-09-06 15:18:31 -07003016 memset(ret, 0, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003017 return ret;
3018}
Pekka J Enbergdd392712005-09-06 15:18:31 -07003019EXPORT_SYMBOL(kzalloc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003020
3021/**
3022 * kfree - free previously allocated memory
3023 * @objp: pointer returned by kmalloc.
3024 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003025 * If @objp is NULL, no operation is performed.
3026 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003027 * Don't free memory not originally allocated by kmalloc()
3028 * or you will run into trouble.
3029 */
3030void kfree(const void *objp)
3031{
3032 kmem_cache_t *c;
3033 unsigned long flags;
3034
3035 if (unlikely(!objp))
3036 return;
3037 local_irq_save(flags);
3038 kfree_debugcheck(objp);
Pekka Enberg065d41c2005-11-13 16:06:46 -08003039 c = page_get_cache(virt_to_page(objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003040 __cache_free(c, (void*)objp);
3041 local_irq_restore(flags);
3042}
3043EXPORT_SYMBOL(kfree);
3044
3045#ifdef CONFIG_SMP
3046/**
3047 * free_percpu - free previously allocated percpu memory
3048 * @objp: pointer returned by alloc_percpu.
3049 *
3050 * Don't free memory not originally allocated by alloc_percpu()
3051 * The complemented objp is to check for that.
3052 */
3053void
3054free_percpu(const void *objp)
3055{
3056 int i;
3057 struct percpu_data *p = (struct percpu_data *) (~(unsigned long) objp);
3058
Christoph Lametere498be72005-09-09 13:03:32 -07003059 /*
3060 * We allocate for all cpus so we cannot use for online cpu here.
3061 */
3062 for_each_cpu(i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003063 kfree(p->ptrs[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003064 kfree(p);
3065}
3066EXPORT_SYMBOL(free_percpu);
3067#endif
3068
3069unsigned int kmem_cache_size(kmem_cache_t *cachep)
3070{
3071 return obj_reallen(cachep);
3072}
3073EXPORT_SYMBOL(kmem_cache_size);
3074
Arnaldo Carvalho de Melo19449722005-06-18 22:46:19 -07003075const char *kmem_cache_name(kmem_cache_t *cachep)
3076{
3077 return cachep->name;
3078}
3079EXPORT_SYMBOL_GPL(kmem_cache_name);
3080
Christoph Lametere498be72005-09-09 13:03:32 -07003081/*
3082 * This initializes kmem_list3 for all nodes.
3083 */
3084static int alloc_kmemlist(kmem_cache_t *cachep)
3085{
3086 int node;
3087 struct kmem_list3 *l3;
3088 int err = 0;
3089
3090 for_each_online_node(node) {
3091 struct array_cache *nc = NULL, *new;
3092 struct array_cache **new_alien = NULL;
3093#ifdef CONFIG_NUMA
3094 if (!(new_alien = alloc_alien_cache(node, cachep->limit)))
3095 goto fail;
3096#endif
3097 if (!(new = alloc_arraycache(node, (cachep->shared*
3098 cachep->batchcount), 0xbaadf00d)))
3099 goto fail;
3100 if ((l3 = cachep->nodelists[node])) {
3101
3102 spin_lock_irq(&l3->list_lock);
3103
3104 if ((nc = cachep->nodelists[node]->shared))
3105 free_block(cachep, nc->entry,
Christoph Lameterff694162005-09-22 21:44:02 -07003106 nc->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07003107
3108 l3->shared = new;
3109 if (!cachep->nodelists[node]->alien) {
3110 l3->alien = new_alien;
3111 new_alien = NULL;
3112 }
3113 l3->free_limit = (1 + nr_cpus_node(node))*
3114 cachep->batchcount + cachep->num;
3115 spin_unlock_irq(&l3->list_lock);
3116 kfree(nc);
3117 free_alien_cache(new_alien);
3118 continue;
3119 }
3120 if (!(l3 = kmalloc_node(sizeof(struct kmem_list3),
3121 GFP_KERNEL, node)))
3122 goto fail;
3123
3124 kmem_list3_init(l3);
3125 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3126 ((unsigned long)cachep)%REAPTIMEOUT_LIST3;
3127 l3->shared = new;
3128 l3->alien = new_alien;
3129 l3->free_limit = (1 + nr_cpus_node(node))*
3130 cachep->batchcount + cachep->num;
3131 cachep->nodelists[node] = l3;
3132 }
3133 return err;
3134fail:
3135 err = -ENOMEM;
3136 return err;
3137}
3138
Linus Torvalds1da177e2005-04-16 15:20:36 -07003139struct ccupdate_struct {
3140 kmem_cache_t *cachep;
3141 struct array_cache *new[NR_CPUS];
3142};
3143
3144static void do_ccupdate_local(void *info)
3145{
3146 struct ccupdate_struct *new = (struct ccupdate_struct *)info;
3147 struct array_cache *old;
3148
3149 check_irq_off();
3150 old = ac_data(new->cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003151
Linus Torvalds1da177e2005-04-16 15:20:36 -07003152 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3153 new->new[smp_processor_id()] = old;
3154}
3155
3156
3157static int do_tune_cpucache(kmem_cache_t *cachep, int limit, int batchcount,
3158 int shared)
3159{
3160 struct ccupdate_struct new;
Christoph Lametere498be72005-09-09 13:03:32 -07003161 int i, err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003162
3163 memset(&new.new,0,sizeof(new.new));
Christoph Lametere498be72005-09-09 13:03:32 -07003164 for_each_online_cpu(i) {
3165 new.new[i] = alloc_arraycache(cpu_to_node(i), limit, batchcount);
3166 if (!new.new[i]) {
3167 for (i--; i >= 0; i--) kfree(new.new[i]);
3168 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003169 }
3170 }
3171 new.cachep = cachep;
3172
3173 smp_call_function_all_cpus(do_ccupdate_local, (void *)&new);
Christoph Lametere498be72005-09-09 13:03:32 -07003174
Linus Torvalds1da177e2005-04-16 15:20:36 -07003175 check_irq_on();
3176 spin_lock_irq(&cachep->spinlock);
3177 cachep->batchcount = batchcount;
3178 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07003179 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003180 spin_unlock_irq(&cachep->spinlock);
3181
Christoph Lametere498be72005-09-09 13:03:32 -07003182 for_each_online_cpu(i) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003183 struct array_cache *ccold = new.new[i];
3184 if (!ccold)
3185 continue;
Christoph Lametere498be72005-09-09 13:03:32 -07003186 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -07003187 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
Christoph Lametere498be72005-09-09 13:03:32 -07003188 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003189 kfree(ccold);
3190 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003191
Christoph Lametere498be72005-09-09 13:03:32 -07003192 err = alloc_kmemlist(cachep);
3193 if (err) {
3194 printk(KERN_ERR "alloc_kmemlist failed for %s, error %d.\n",
3195 cachep->name, -err);
3196 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003197 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003198 return 0;
3199}
3200
3201
3202static void enable_cpucache(kmem_cache_t *cachep)
3203{
3204 int err;
3205 int limit, shared;
3206
3207 /* The head array serves three purposes:
3208 * - create a LIFO ordering, i.e. return objects that are cache-warm
3209 * - reduce the number of spinlock operations.
3210 * - reduce the number of linked list operations on the slab and
3211 * bufctl chains: array operations are cheaper.
3212 * The numbers are guessed, we should auto-tune as described by
3213 * Bonwick.
3214 */
3215 if (cachep->objsize > 131072)
3216 limit = 1;
3217 else if (cachep->objsize > PAGE_SIZE)
3218 limit = 8;
3219 else if (cachep->objsize > 1024)
3220 limit = 24;
3221 else if (cachep->objsize > 256)
3222 limit = 54;
3223 else
3224 limit = 120;
3225
3226 /* Cpu bound tasks (e.g. network routing) can exhibit cpu bound
3227 * allocation behaviour: Most allocs on one cpu, most free operations
3228 * on another cpu. For these cases, an efficient object passing between
3229 * cpus is necessary. This is provided by a shared array. The array
3230 * replaces Bonwick's magazine layer.
3231 * On uniprocessor, it's functionally equivalent (but less efficient)
3232 * to a larger limit. Thus disabled by default.
3233 */
3234 shared = 0;
3235#ifdef CONFIG_SMP
3236 if (cachep->objsize <= PAGE_SIZE)
3237 shared = 8;
3238#endif
3239
3240#if DEBUG
3241 /* With debugging enabled, large batchcount lead to excessively
3242 * long periods with disabled local interrupts. Limit the
3243 * batchcount
3244 */
3245 if (limit > 32)
3246 limit = 32;
3247#endif
3248 err = do_tune_cpucache(cachep, limit, (limit+1)/2, shared);
3249 if (err)
3250 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
3251 cachep->name, -err);
3252}
3253
3254static void drain_array_locked(kmem_cache_t *cachep,
Christoph Lametere498be72005-09-09 13:03:32 -07003255 struct array_cache *ac, int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003256{
3257 int tofree;
3258
Christoph Lametere498be72005-09-09 13:03:32 -07003259 check_spinlock_acquired_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003260 if (ac->touched && !force) {
3261 ac->touched = 0;
3262 } else if (ac->avail) {
3263 tofree = force ? ac->avail : (ac->limit+4)/5;
3264 if (tofree > ac->avail) {
3265 tofree = (ac->avail+1)/2;
3266 }
Christoph Lameterff694162005-09-22 21:44:02 -07003267 free_block(cachep, ac->entry, tofree, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003268 ac->avail -= tofree;
Christoph Lametere498be72005-09-09 13:03:32 -07003269 memmove(ac->entry, &(ac->entry[tofree]),
Linus Torvalds1da177e2005-04-16 15:20:36 -07003270 sizeof(void*)*ac->avail);
3271 }
3272}
3273
3274/**
3275 * cache_reap - Reclaim memory from caches.
Randy Dunlap1e5d5332005-11-07 01:01:06 -08003276 * @unused: unused parameter
Linus Torvalds1da177e2005-04-16 15:20:36 -07003277 *
3278 * Called from workqueue/eventd every few seconds.
3279 * Purpose:
3280 * - clear the per-cpu caches for this CPU.
3281 * - return freeable pages to the main free memory pool.
3282 *
3283 * If we cannot acquire the cache chain semaphore then just give up - we'll
3284 * try again on the next iteration.
3285 */
3286static void cache_reap(void *unused)
3287{
3288 struct list_head *walk;
Christoph Lametere498be72005-09-09 13:03:32 -07003289 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003290
3291 if (down_trylock(&cache_chain_sem)) {
3292 /* Give up. Setup the next iteration. */
Manfred Spraulcd61ef62005-11-07 00:58:02 -08003293 schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003294 return;
3295 }
3296
3297 list_for_each(walk, &cache_chain) {
3298 kmem_cache_t *searchp;
3299 struct list_head* p;
3300 int tofree;
3301 struct slab *slabp;
3302
3303 searchp = list_entry(walk, kmem_cache_t, next);
3304
3305 if (searchp->flags & SLAB_NO_REAP)
3306 goto next;
3307
3308 check_irq_on();
3309
Christoph Lametere498be72005-09-09 13:03:32 -07003310 l3 = searchp->nodelists[numa_node_id()];
3311 if (l3->alien)
3312 drain_alien_cache(searchp, l3);
3313 spin_lock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003314
Christoph Lametere498be72005-09-09 13:03:32 -07003315 drain_array_locked(searchp, ac_data(searchp), 0,
3316 numa_node_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -07003317
Christoph Lametere498be72005-09-09 13:03:32 -07003318 if (time_after(l3->next_reap, jiffies))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003319 goto next_unlock;
3320
Christoph Lametere498be72005-09-09 13:03:32 -07003321 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003322
Christoph Lametere498be72005-09-09 13:03:32 -07003323 if (l3->shared)
3324 drain_array_locked(searchp, l3->shared, 0,
3325 numa_node_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -07003326
Christoph Lametere498be72005-09-09 13:03:32 -07003327 if (l3->free_touched) {
3328 l3->free_touched = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003329 goto next_unlock;
3330 }
3331
Christoph Lametere498be72005-09-09 13:03:32 -07003332 tofree = (l3->free_limit+5*searchp->num-1)/(5*searchp->num);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003333 do {
Christoph Lametere498be72005-09-09 13:03:32 -07003334 p = l3->slabs_free.next;
3335 if (p == &(l3->slabs_free))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003336 break;
3337
3338 slabp = list_entry(p, struct slab, list);
3339 BUG_ON(slabp->inuse);
3340 list_del(&slabp->list);
3341 STATS_INC_REAPED(searchp);
3342
3343 /* Safe to drop the lock. The slab is no longer
3344 * linked to the cache.
3345 * searchp cannot disappear, we hold
3346 * cache_chain_lock
3347 */
Christoph Lametere498be72005-09-09 13:03:32 -07003348 l3->free_objects -= searchp->num;
3349 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003350 slab_destroy(searchp, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07003351 spin_lock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003352 } while(--tofree > 0);
3353next_unlock:
Christoph Lametere498be72005-09-09 13:03:32 -07003354 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003355next:
3356 cond_resched();
3357 }
3358 check_irq_on();
3359 up(&cache_chain_sem);
Christoph Lameter4ae7c032005-06-21 17:14:57 -07003360 drain_remote_pages();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003361 /* Setup the next iteration */
Manfred Spraulcd61ef62005-11-07 00:58:02 -08003362 schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003363}
3364
3365#ifdef CONFIG_PROC_FS
3366
Pekka Enberg85289f92006-01-08 01:00:36 -08003367static void print_slabinfo_header(struct seq_file *m)
3368{
3369 /*
3370 * Output format version, so at least we can change it
3371 * without _too_ many complaints.
3372 */
3373#if STATS
3374 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
3375#else
3376 seq_puts(m, "slabinfo - version: 2.1\n");
3377#endif
3378 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
3379 "<objperslab> <pagesperslab>");
3380 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
3381 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
3382#if STATS
3383 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
3384 "<error> <maxfreeable> <nodeallocs> <remotefrees>");
3385 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
3386#endif
3387 seq_putc(m, '\n');
3388}
3389
Linus Torvalds1da177e2005-04-16 15:20:36 -07003390static void *s_start(struct seq_file *m, loff_t *pos)
3391{
3392 loff_t n = *pos;
3393 struct list_head *p;
3394
3395 down(&cache_chain_sem);
Pekka Enberg85289f92006-01-08 01:00:36 -08003396 if (!n)
3397 print_slabinfo_header(m);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003398 p = cache_chain.next;
3399 while (n--) {
3400 p = p->next;
3401 if (p == &cache_chain)
3402 return NULL;
3403 }
3404 return list_entry(p, kmem_cache_t, next);
3405}
3406
3407static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3408{
3409 kmem_cache_t *cachep = p;
3410 ++*pos;
3411 return cachep->next.next == &cache_chain ? NULL
3412 : list_entry(cachep->next.next, kmem_cache_t, next);
3413}
3414
3415static void s_stop(struct seq_file *m, void *p)
3416{
3417 up(&cache_chain_sem);
3418}
3419
3420static int s_show(struct seq_file *m, void *p)
3421{
3422 kmem_cache_t *cachep = p;
3423 struct list_head *q;
3424 struct slab *slabp;
3425 unsigned long active_objs;
3426 unsigned long num_objs;
3427 unsigned long active_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07003428 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
3429 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003430 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003431 int node;
3432 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003433
3434 check_irq_on();
3435 spin_lock_irq(&cachep->spinlock);
3436 active_objs = 0;
3437 num_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07003438 for_each_online_node(node) {
3439 l3 = cachep->nodelists[node];
3440 if (!l3)
3441 continue;
3442
3443 spin_lock(&l3->list_lock);
3444
3445 list_for_each(q,&l3->slabs_full) {
3446 slabp = list_entry(q, struct slab, list);
3447 if (slabp->inuse != cachep->num && !error)
3448 error = "slabs_full accounting error";
3449 active_objs += cachep->num;
3450 active_slabs++;
3451 }
3452 list_for_each(q,&l3->slabs_partial) {
3453 slabp = list_entry(q, struct slab, list);
3454 if (slabp->inuse == cachep->num && !error)
3455 error = "slabs_partial inuse accounting error";
3456 if (!slabp->inuse && !error)
3457 error = "slabs_partial/inuse accounting error";
3458 active_objs += slabp->inuse;
3459 active_slabs++;
3460 }
3461 list_for_each(q,&l3->slabs_free) {
3462 slabp = list_entry(q, struct slab, list);
3463 if (slabp->inuse && !error)
3464 error = "slabs_free/inuse accounting error";
3465 num_slabs++;
3466 }
3467 free_objects += l3->free_objects;
3468 shared_avail += l3->shared->avail;
3469
3470 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003471 }
3472 num_slabs+=active_slabs;
3473 num_objs = num_slabs*cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003474 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003475 error = "free_objects accounting error";
3476
3477 name = cachep->name;
3478 if (error)
3479 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
3480
3481 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
3482 name, active_objs, num_objs, cachep->objsize,
3483 cachep->num, (1<<cachep->gfporder));
3484 seq_printf(m, " : tunables %4u %4u %4u",
3485 cachep->limit, cachep->batchcount,
Christoph Lametere498be72005-09-09 13:03:32 -07003486 cachep->shared);
3487 seq_printf(m, " : slabdata %6lu %6lu %6lu",
3488 active_slabs, num_slabs, shared_avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003489#if STATS
3490 { /* list3 stats */
3491 unsigned long high = cachep->high_mark;
3492 unsigned long allocs = cachep->num_allocations;
3493 unsigned long grown = cachep->grown;
3494 unsigned long reaped = cachep->reaped;
3495 unsigned long errors = cachep->errors;
3496 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003497 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07003498 unsigned long node_frees = cachep->node_frees;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003499
Christoph Lametere498be72005-09-09 13:03:32 -07003500 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
3501 %4lu %4lu %4lu %4lu",
3502 allocs, high, grown, reaped, errors,
3503 max_freeable, node_allocs, node_frees);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003504 }
3505 /* cpu stats */
3506 {
3507 unsigned long allochit = atomic_read(&cachep->allochit);
3508 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
3509 unsigned long freehit = atomic_read(&cachep->freehit);
3510 unsigned long freemiss = atomic_read(&cachep->freemiss);
3511
3512 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
3513 allochit, allocmiss, freehit, freemiss);
3514 }
3515#endif
3516 seq_putc(m, '\n');
3517 spin_unlock_irq(&cachep->spinlock);
3518 return 0;
3519}
3520
3521/*
3522 * slabinfo_op - iterator that generates /proc/slabinfo
3523 *
3524 * Output layout:
3525 * cache-name
3526 * num-active-objs
3527 * total-objs
3528 * object size
3529 * num-active-slabs
3530 * total-slabs
3531 * num-pages-per-slab
3532 * + further values on SMP and with statistics enabled
3533 */
3534
3535struct seq_operations slabinfo_op = {
3536 .start = s_start,
3537 .next = s_next,
3538 .stop = s_stop,
3539 .show = s_show,
3540};
3541
3542#define MAX_SLABINFO_WRITE 128
3543/**
3544 * slabinfo_write - Tuning for the slab allocator
3545 * @file: unused
3546 * @buffer: user buffer
3547 * @count: data length
3548 * @ppos: unused
3549 */
3550ssize_t slabinfo_write(struct file *file, const char __user *buffer,
3551 size_t count, loff_t *ppos)
3552{
3553 char kbuf[MAX_SLABINFO_WRITE+1], *tmp;
3554 int limit, batchcount, shared, res;
3555 struct list_head *p;
3556
3557 if (count > MAX_SLABINFO_WRITE)
3558 return -EINVAL;
3559 if (copy_from_user(&kbuf, buffer, count))
3560 return -EFAULT;
3561 kbuf[MAX_SLABINFO_WRITE] = '\0';
3562
3563 tmp = strchr(kbuf, ' ');
3564 if (!tmp)
3565 return -EINVAL;
3566 *tmp = '\0';
3567 tmp++;
3568 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
3569 return -EINVAL;
3570
3571 /* Find the cache in the chain of caches. */
3572 down(&cache_chain_sem);
3573 res = -EINVAL;
3574 list_for_each(p,&cache_chain) {
3575 kmem_cache_t *cachep = list_entry(p, kmem_cache_t, next);
3576
3577 if (!strcmp(cachep->name, kbuf)) {
3578 if (limit < 1 ||
3579 batchcount < 1 ||
3580 batchcount > limit ||
3581 shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07003582 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003583 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07003584 res = do_tune_cpucache(cachep, limit,
3585 batchcount, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003586 }
3587 break;
3588 }
3589 }
3590 up(&cache_chain_sem);
3591 if (res >= 0)
3592 res = count;
3593 return res;
3594}
3595#endif
3596
Manfred Spraul00e145b2005-09-03 15:55:07 -07003597/**
3598 * ksize - get the actual amount of memory allocated for a given object
3599 * @objp: Pointer to the object
3600 *
3601 * kmalloc may internally round up allocations and return more memory
3602 * than requested. ksize() can be used to determine the actual amount of
3603 * memory allocated. The caller may use this additional memory, even though
3604 * a smaller amount of memory was initially specified with the kmalloc call.
3605 * The caller must guarantee that objp points to a valid object previously
3606 * allocated with either kmalloc() or kmem_cache_alloc(). The object
3607 * must not be freed during the duration of the call.
3608 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003609unsigned int ksize(const void *objp)
3610{
Manfred Spraul00e145b2005-09-03 15:55:07 -07003611 if (unlikely(objp == NULL))
3612 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003613
Pekka Enberg065d41c2005-11-13 16:06:46 -08003614 return obj_reallen(page_get_cache(virt_to_page(objp)));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003615}
Paulo Marques543537b2005-06-23 00:09:02 -07003616
3617
3618/*
3619 * kstrdup - allocate space for and copy an existing string
3620 *
3621 * @s: the string to duplicate
3622 * @gfp: the GFP mask used in the kmalloc() call when allocating memory
3623 */
Al Virodd0fc662005-10-07 07:46:04 +01003624char *kstrdup(const char *s, gfp_t gfp)
Paulo Marques543537b2005-06-23 00:09:02 -07003625{
3626 size_t len;
3627 char *buf;
3628
3629 if (!s)
3630 return NULL;
3631
3632 len = strlen(s) + 1;
3633 buf = kmalloc(len, gfp);
3634 if (buf)
3635 memcpy(buf, s, len);
3636 return buf;
3637}
3638EXPORT_SYMBOL(kstrdup);