blob: 9374293a301297edef94491494db8e58f591ba82 [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
Linus Torvalds1da177e2005-04-16 15:20:36 -0700133/* Shouldn't this be in a header file somewhere? */
134#define BYTES_PER_WORD sizeof(void *)
135
136#ifndef cache_line_size
137#define cache_line_size() L1_CACHE_BYTES
138#endif
139
140#ifndef ARCH_KMALLOC_MINALIGN
141/*
142 * Enforce a minimum alignment for the kmalloc caches.
143 * Usually, the kmalloc caches are cache_line_size() aligned, except when
144 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
145 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
146 * alignment larger than BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that.
147 * Note that this flag disables some debug features.
148 */
149#define ARCH_KMALLOC_MINALIGN 0
150#endif
151
152#ifndef ARCH_SLAB_MINALIGN
153/*
154 * Enforce a minimum alignment for all caches.
155 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
156 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
157 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
158 * some debug features.
159 */
160#define ARCH_SLAB_MINALIGN 0
161#endif
162
163#ifndef ARCH_KMALLOC_FLAGS
164#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
165#endif
166
167/* Legal flag mask for kmem_cache_create(). */
168#if DEBUG
169# define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \
170 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
171 SLAB_NO_REAP | SLAB_CACHE_DMA | \
172 SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \
173 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
174 SLAB_DESTROY_BY_RCU)
175#else
176# define CREATE_MASK (SLAB_HWCACHE_ALIGN | SLAB_NO_REAP | \
177 SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \
178 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
179 SLAB_DESTROY_BY_RCU)
180#endif
181
182/*
183 * kmem_bufctl_t:
184 *
185 * Bufctl's are used for linking objs within a slab
186 * linked offsets.
187 *
188 * This implementation relies on "struct page" for locating the cache &
189 * slab an object belongs to.
190 * This allows the bufctl structure to be small (one int), but limits
191 * the number of objects a slab (not a cache) can contain when off-slab
192 * bufctls are used. The limit is the size of the largest general cache
193 * that does not use off-slab slabs.
194 * For 32bit archs with 4 kB pages, is this 56.
195 * This is not serious, as it is only for large objects, when it is unwise
196 * to have too many per slab.
197 * Note: This limit can be raised by introducing a general cache whose size
198 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
199 */
200
Kyle Moffettfa5b08d2005-09-03 15:55:03 -0700201typedef unsigned int kmem_bufctl_t;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700202#define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
203#define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
204#define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-2)
205
206/* Max number of objs-per-slab for caches which use off-slab slabs.
207 * Needed to avoid a possible looping condition in cache_grow().
208 */
209static unsigned long offslab_limit;
210
211/*
212 * struct slab
213 *
214 * Manages the objs in a slab. Placed either at the beginning of mem allocated
215 * for a slab, or allocated from an general cache.
216 * Slabs are chained into three list: fully used, partial, fully free slabs.
217 */
218struct slab {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800219 struct list_head list;
220 unsigned long colouroff;
221 void *s_mem; /* including colour offset */
222 unsigned int inuse; /* num of objs active in slab */
223 kmem_bufctl_t free;
224 unsigned short nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700225};
226
227/*
228 * struct slab_rcu
229 *
230 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
231 * arrange for kmem_freepages to be called via RCU. This is useful if
232 * we need to approach a kernel structure obliquely, from its address
233 * obtained without the usual locking. We can lock the structure to
234 * stabilize it and check it's still at the given address, only if we
235 * can be sure that the memory has not been meanwhile reused for some
236 * other kind of object (which our subsystem's lock might corrupt).
237 *
238 * rcu_read_lock before reading the address, then rcu_read_unlock after
239 * taking the spinlock within the structure expected at that address.
240 *
241 * We assume struct slab_rcu can overlay struct slab when destroying.
242 */
243struct slab_rcu {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800244 struct rcu_head head;
245 kmem_cache_t *cachep;
246 void *addr;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700247};
248
249/*
250 * struct array_cache
251 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700252 * Purpose:
253 * - LIFO ordering, to hand out cache-warm objects from _alloc
254 * - reduce the number of linked list operations
255 * - reduce spinlock operations
256 *
257 * The limit is stored in the per-cpu structure to reduce the data cache
258 * footprint.
259 *
260 */
261struct array_cache {
262 unsigned int avail;
263 unsigned int limit;
264 unsigned int batchcount;
265 unsigned int touched;
Christoph Lametere498be72005-09-09 13:03:32 -0700266 spinlock_t lock;
267 void *entry[0]; /*
268 * Must have this definition in here for the proper
269 * alignment of array_cache. Also simplifies accessing
270 * the entries.
271 * [0] is for gcc 2.95. It should really be [].
272 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700273};
274
275/* bootstrap: The caches do not work without cpuarrays anymore,
276 * but the cpuarrays are allocated from the generic caches...
277 */
278#define BOOT_CPUCACHE_ENTRIES 1
279struct arraycache_init {
280 struct array_cache cache;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800281 void *entries[BOOT_CPUCACHE_ENTRIES];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700282};
283
284/*
Christoph Lametere498be72005-09-09 13:03:32 -0700285 * The slab lists for all objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700286 */
287struct kmem_list3 {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800288 struct list_head slabs_partial; /* partial list first, better asm code */
289 struct list_head slabs_full;
290 struct list_head slabs_free;
291 unsigned long free_objects;
292 unsigned long next_reap;
293 int free_touched;
294 unsigned int free_limit;
295 spinlock_t list_lock;
296 struct array_cache *shared; /* shared per node */
297 struct array_cache **alien; /* on other nodes */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700298};
299
Christoph Lametere498be72005-09-09 13:03:32 -0700300/*
301 * Need this for bootstrapping a per node allocator.
302 */
303#define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
304struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
305#define CACHE_CACHE 0
306#define SIZE_AC 1
307#define SIZE_L3 (1 + MAX_NUMNODES)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700308
Christoph Lametere498be72005-09-09 13:03:32 -0700309/*
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700310 * This function must be completely optimized away if
Christoph Lametere498be72005-09-09 13:03:32 -0700311 * a constant is passed to it. Mostly the same as
312 * what is in linux/slab.h except it returns an
313 * index.
314 */
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700315static __always_inline int index_of(const size_t size)
Christoph Lametere498be72005-09-09 13:03:32 -0700316{
317 if (__builtin_constant_p(size)) {
318 int i = 0;
319
320#define CACHE(x) \
321 if (size <=x) \
322 return i; \
323 else \
324 i++;
325#include "linux/kmalloc_sizes.h"
326#undef CACHE
327 {
328 extern void __bad_size(void);
329 __bad_size();
330 }
Ivan Kokshaysky7243cc02005-09-22 21:43:58 -0700331 } else
332 BUG();
Christoph Lametere498be72005-09-09 13:03:32 -0700333 return 0;
334}
335
336#define INDEX_AC index_of(sizeof(struct arraycache_init))
337#define INDEX_L3 index_of(sizeof(struct kmem_list3))
338
339static inline void kmem_list3_init(struct kmem_list3 *parent)
340{
341 INIT_LIST_HEAD(&parent->slabs_full);
342 INIT_LIST_HEAD(&parent->slabs_partial);
343 INIT_LIST_HEAD(&parent->slabs_free);
344 parent->shared = NULL;
345 parent->alien = NULL;
346 spin_lock_init(&parent->list_lock);
347 parent->free_objects = 0;
348 parent->free_touched = 0;
349}
350
351#define MAKE_LIST(cachep, listp, slab, nodeid) \
352 do { \
353 INIT_LIST_HEAD(listp); \
354 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
355 } while (0)
356
357#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
358 do { \
359 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
360 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
361 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
362 } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700363
364/*
365 * kmem_cache_t
366 *
367 * manages a cache.
368 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800369
Pekka J Enberg2109a2d2005-11-07 00:58:01 -0800370struct kmem_cache {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700371/* 1) per-cpu data, touched during every alloc/free */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800372 struct array_cache *array[NR_CPUS];
373 unsigned int batchcount;
374 unsigned int limit;
375 unsigned int shared;
376 unsigned int objsize;
Christoph Lametere498be72005-09-09 13:03:32 -0700377/* 2) touched by every alloc & free from the backend */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800378 struct kmem_list3 *nodelists[MAX_NUMNODES];
379 unsigned int flags; /* constant flags */
380 unsigned int num; /* # of objs per slab */
381 spinlock_t spinlock;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700382
383/* 3) cache_grow/shrink */
384 /* order of pgs per slab (2^n) */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800385 unsigned int gfporder;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700386
387 /* force GFP flags, e.g. GFP_DMA */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800388 gfp_t gfpflags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700389
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800390 size_t colour; /* cache colouring range */
391 unsigned int colour_off; /* colour offset */
392 unsigned int colour_next; /* cache colouring */
393 kmem_cache_t *slabp_cache;
394 unsigned int slab_size;
395 unsigned int dflags; /* dynamic flags */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700396
397 /* constructor func */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800398 void (*ctor) (void *, kmem_cache_t *, unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700399
400 /* de-constructor func */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800401 void (*dtor) (void *, kmem_cache_t *, unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700402
403/* 4) cache creation/removal */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800404 const char *name;
405 struct list_head next;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700406
407/* 5) statistics */
408#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800409 unsigned long num_active;
410 unsigned long num_allocations;
411 unsigned long high_mark;
412 unsigned long grown;
413 unsigned long reaped;
414 unsigned long errors;
415 unsigned long max_freeable;
416 unsigned long node_allocs;
417 unsigned long node_frees;
418 atomic_t allochit;
419 atomic_t allocmiss;
420 atomic_t freehit;
421 atomic_t freemiss;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700422#endif
423#if DEBUG
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800424 int dbghead;
425 int reallen;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700426#endif
427};
428
429#define CFLGS_OFF_SLAB (0x80000000UL)
430#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
431
432#define BATCHREFILL_LIMIT 16
433/* Optimization question: fewer reaps means less
434 * probability for unnessary cpucache drain/refill cycles.
435 *
Adrian Bunkdc6f3f22005-11-08 16:44:08 +0100436 * OTOH the cpuarrays can contain lots of objects,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700437 * which could lock up otherwise freeable slabs.
438 */
439#define REAPTIMEOUT_CPUC (2*HZ)
440#define REAPTIMEOUT_LIST3 (4*HZ)
441
442#if STATS
443#define STATS_INC_ACTIVE(x) ((x)->num_active++)
444#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
445#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
446#define STATS_INC_GROWN(x) ((x)->grown++)
447#define STATS_INC_REAPED(x) ((x)->reaped++)
448#define STATS_SET_HIGH(x) do { if ((x)->num_active > (x)->high_mark) \
449 (x)->high_mark = (x)->num_active; \
450 } while (0)
451#define STATS_INC_ERR(x) ((x)->errors++)
452#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
Christoph Lametere498be72005-09-09 13:03:32 -0700453#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700454#define STATS_SET_FREEABLE(x, i) \
455 do { if ((x)->max_freeable < i) \
456 (x)->max_freeable = i; \
457 } while (0)
458
459#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
460#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
461#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
462#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
463#else
464#define STATS_INC_ACTIVE(x) do { } while (0)
465#define STATS_DEC_ACTIVE(x) do { } while (0)
466#define STATS_INC_ALLOCED(x) do { } while (0)
467#define STATS_INC_GROWN(x) do { } while (0)
468#define STATS_INC_REAPED(x) do { } while (0)
469#define STATS_SET_HIGH(x) do { } while (0)
470#define STATS_INC_ERR(x) do { } while (0)
471#define STATS_INC_NODEALLOCS(x) do { } while (0)
Christoph Lametere498be72005-09-09 13:03:32 -0700472#define STATS_INC_NODEFREES(x) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700473#define STATS_SET_FREEABLE(x, i) \
474 do { } while (0)
475
476#define STATS_INC_ALLOCHIT(x) do { } while (0)
477#define STATS_INC_ALLOCMISS(x) do { } while (0)
478#define STATS_INC_FREEHIT(x) do { } while (0)
479#define STATS_INC_FREEMISS(x) do { } while (0)
480#endif
481
482#if DEBUG
483/* Magic nums for obj red zoning.
484 * Placed in the first word before and the first word after an obj.
485 */
486#define RED_INACTIVE 0x5A2CF071UL /* when obj is inactive */
487#define RED_ACTIVE 0x170FC2A5UL /* when obj is active */
488
489/* ...and for poisoning */
490#define POISON_INUSE 0x5a /* for use-uninitialised poisoning */
491#define POISON_FREE 0x6b /* for use-after-free poisoning */
492#define POISON_END 0xa5 /* end-byte of poisoning */
493
494/* memory layout of objects:
495 * 0 : objp
496 * 0 .. cachep->dbghead - BYTES_PER_WORD - 1: padding. This ensures that
497 * the end of an object is aligned with the end of the real
498 * allocation. Catches writes behind the end of the allocation.
499 * cachep->dbghead - BYTES_PER_WORD .. cachep->dbghead - 1:
500 * redzone word.
501 * cachep->dbghead: The real object.
502 * cachep->objsize - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
503 * cachep->objsize - 1* BYTES_PER_WORD: last caller address [BYTES_PER_WORD long]
504 */
505static int obj_dbghead(kmem_cache_t *cachep)
506{
507 return cachep->dbghead;
508}
509
510static int obj_reallen(kmem_cache_t *cachep)
511{
512 return cachep->reallen;
513}
514
515static unsigned long *dbg_redzone1(kmem_cache_t *cachep, void *objp)
516{
517 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
518 return (unsigned long*) (objp+obj_dbghead(cachep)-BYTES_PER_WORD);
519}
520
521static unsigned long *dbg_redzone2(kmem_cache_t *cachep, void *objp)
522{
523 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
524 if (cachep->flags & SLAB_STORE_USER)
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800525 return (unsigned long *)(objp + cachep->objsize -
526 2 * BYTES_PER_WORD);
527 return (unsigned long *)(objp + cachep->objsize - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700528}
529
530static void **dbg_userword(kmem_cache_t *cachep, void *objp)
531{
532 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800533 return (void **)(objp + cachep->objsize - BYTES_PER_WORD);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700534}
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>
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800610 {NULL,}
Linus Torvalds1da177e2005-04-16 15:20:36 -0700611#undef CACHE
612};
613
614static struct arraycache_init initarray_cache __initdata =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800615 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700616static struct arraycache_init initarray_generic =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800617 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
Linus Torvalds1da177e2005-04-16 15:20:36 -0700618
619/* internal cache of cache description objs */
620static kmem_cache_t cache_cache = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800621 .batchcount = 1,
622 .limit = BOOT_CPUCACHE_ENTRIES,
623 .shared = 1,
624 .objsize = sizeof(kmem_cache_t),
625 .flags = SLAB_NO_REAP,
626 .spinlock = SPIN_LOCK_UNLOCKED,
627 .name = "kmem_cache",
Linus Torvalds1da177e2005-04-16 15:20:36 -0700628#if DEBUG
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800629 .reallen = sizeof(kmem_cache_t),
Linus Torvalds1da177e2005-04-16 15:20:36 -0700630#endif
631};
632
633/* Guard access to the cache-chain. */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800634static struct semaphore cache_chain_sem;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700635static 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
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800658static void free_block(kmem_cache_t *cachep, void **objpp, int len, int node);
659static 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
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800674 * 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,
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800700 int flags, size_t *left_over, unsigned int *num)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700701{
702 int i;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800703 size_t wastage = PAGE_SIZE << gfporder;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700704 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;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800712 while (i * size + ALIGN(base + i * extra, align) <= wastage)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700713 i++;
714 if (i > 0)
715 i--;
716
717 if (i > SLAB_LIMIT)
718 i = SLAB_LIMIT;
719
720 *num = i;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800721 wastage -= i * size;
722 wastage -= ALIGN(base + i * extra, align);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700723 *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",
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800731 function, cachep->name, msg);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700732 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,
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800758 int batchcount)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700759{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800760 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700761 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;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800778 int memsize = sizeof(void *) * MAX_NUMNODES;
Christoph Lametere498be72005-09-09 13:03:32 -0700779 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]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800792 for (i--; i <= 0; i--)
Christoph Lametere498be72005-09-09 13:03:32 -0700793 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)
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800810 kfree(ac_ptr[i]);
Christoph Lametere498be72005-09-09 13:03:32 -0700811
812 kfree(ac_ptr);
813}
814
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800815static inline void __drain_alien_cache(kmem_cache_t *cachep,
816 struct array_cache *ac, int node)
Christoph Lametere498be72005-09-09 13:03:32 -0700817{
818 struct kmem_list3 *rl3 = cachep->nodelists[node];
819
820 if (ac->avail) {
821 spin_lock(&rl3->list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -0700822 free_block(cachep, ac->entry, ac->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -0700823 ac->avail = 0;
824 spin_unlock(&rl3->list_lock);
825 }
826}
827
828static void drain_alien_cache(kmem_cache_t *cachep, struct kmem_list3 *l3)
829{
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800830 int i = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700831 struct array_cache *ac;
832 unsigned long flags;
833
834 for_each_online_node(i) {
835 ac = l3->alien[i];
836 if (ac) {
837 spin_lock_irqsave(&ac->lock, flags);
838 __drain_alien_cache(cachep, ac, i);
839 spin_unlock_irqrestore(&ac->lock, flags);
840 }
841 }
842}
843#else
844#define alloc_alien_cache(node, limit) do { } while (0)
845#define free_alien_cache(ac_ptr) do { } while (0)
846#define drain_alien_cache(cachep, l3) do { } while (0)
847#endif
848
Linus Torvalds1da177e2005-04-16 15:20:36 -0700849static int __devinit cpuup_callback(struct notifier_block *nfb,
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800850 unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700851{
852 long cpu = (long)hcpu;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800853 kmem_cache_t *cachep;
Christoph Lametere498be72005-09-09 13:03:32 -0700854 struct kmem_list3 *l3 = NULL;
855 int node = cpu_to_node(cpu);
856 int memsize = sizeof(struct kmem_list3);
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,
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800874 GFP_KERNEL, node)))
Christoph Lametere498be72005-09-09 13:03:32 -0700875 goto bad;
876 kmem_list3_init(l3);
877 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800878 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -0700879
880 cachep->nodelists[node] = l3;
881 }
882
883 spin_lock_irq(&cachep->nodelists[node]->list_lock);
884 cachep->nodelists[node]->free_limit =
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800885 (1 + nr_cpus_node(node)) *
886 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -0700887 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
888 }
889
890 /* Now we can go ahead with allocating the shared array's
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800891 & array cache's */
Christoph Lametere498be72005-09-09 13:03:32 -0700892 list_for_each_entry(cachep, &cache_chain, next) {
Tobias Klausercd105df2006-01-08 01:00:59 -0800893 struct array_cache *nc;
894
Christoph Lametere498be72005-09-09 13:03:32 -0700895 nc = alloc_arraycache(node, cachep->limit,
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800896 cachep->batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700897 if (!nc)
898 goto bad;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700899 cachep->array[cpu] = nc;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700900
Christoph Lametere498be72005-09-09 13:03:32 -0700901 l3 = cachep->nodelists[node];
902 BUG_ON(!l3);
903 if (!l3->shared) {
904 if (!(nc = alloc_arraycache(node,
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800905 cachep->shared *
906 cachep->batchcount,
907 0xbaadf00d)))
908 goto bad;
Christoph Lametere498be72005-09-09 13:03:32 -0700909
910 /* we are serialised from CPU_DEAD or
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800911 CPU_UP_CANCELLED by the cpucontrol lock */
Christoph Lametere498be72005-09-09 13:03:32 -0700912 l3->shared = nc;
913 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700914 }
915 up(&cache_chain_sem);
916 break;
917 case CPU_ONLINE:
918 start_cpu_timer(cpu);
919 break;
920#ifdef CONFIG_HOTPLUG_CPU
921 case CPU_DEAD:
922 /* fall thru */
923 case CPU_UP_CANCELED:
924 down(&cache_chain_sem);
925
926 list_for_each_entry(cachep, &cache_chain, next) {
927 struct array_cache *nc;
Christoph Lametere498be72005-09-09 13:03:32 -0700928 cpumask_t mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700929
Christoph Lametere498be72005-09-09 13:03:32 -0700930 mask = node_to_cpumask(node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700931 spin_lock_irq(&cachep->spinlock);
932 /* cpu is dead; no one can alloc from it. */
933 nc = cachep->array[cpu];
934 cachep->array[cpu] = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -0700935 l3 = cachep->nodelists[node];
936
937 if (!l3)
938 goto unlock_cache;
939
940 spin_lock(&l3->list_lock);
941
942 /* Free limit for this kmem_list3 */
943 l3->free_limit -= cachep->batchcount;
944 if (nc)
Christoph Lameterff694162005-09-22 21:44:02 -0700945 free_block(cachep, nc->entry, nc->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -0700946
947 if (!cpus_empty(mask)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800948 spin_unlock(&l3->list_lock);
949 goto unlock_cache;
950 }
Christoph Lametere498be72005-09-09 13:03:32 -0700951
952 if (l3->shared) {
953 free_block(cachep, l3->shared->entry,
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800954 l3->shared->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -0700955 kfree(l3->shared);
956 l3->shared = NULL;
957 }
958 if (l3->alien) {
959 drain_alien_cache(cachep, l3);
960 free_alien_cache(l3->alien);
961 l3->alien = NULL;
962 }
963
964 /* free slabs belonging to this node */
965 if (__node_shrink(cachep, node)) {
966 cachep->nodelists[node] = NULL;
967 spin_unlock(&l3->list_lock);
968 kfree(l3);
969 } else {
970 spin_unlock(&l3->list_lock);
971 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800972 unlock_cache:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700973 spin_unlock_irq(&cachep->spinlock);
974 kfree(nc);
975 }
976 up(&cache_chain_sem);
977 break;
978#endif
979 }
980 return NOTIFY_OK;
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800981 bad:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700982 up(&cache_chain_sem);
983 return NOTIFY_BAD;
984}
985
986static struct notifier_block cpucache_notifier = { &cpuup_callback, NULL, 0 };
987
Christoph Lametere498be72005-09-09 13:03:32 -0700988/*
989 * swap the static kmem_list3 with kmalloced memory
990 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -0800991static void init_list(kmem_cache_t *cachep, struct kmem_list3 *list, int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -0700992{
993 struct kmem_list3 *ptr;
994
995 BUG_ON(cachep->nodelists[nodeid] != list);
996 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
997 BUG_ON(!ptr);
998
999 local_irq_disable();
1000 memcpy(ptr, list, sizeof(struct kmem_list3));
1001 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1002 cachep->nodelists[nodeid] = ptr;
1003 local_irq_enable();
1004}
1005
Linus Torvalds1da177e2005-04-16 15:20:36 -07001006/* Initialisation.
1007 * Called after the gfp() functions have been enabled, and before smp_init().
1008 */
1009void __init kmem_cache_init(void)
1010{
1011 size_t left_over;
1012 struct cache_sizes *sizes;
1013 struct cache_names *names;
Christoph Lametere498be72005-09-09 13:03:32 -07001014 int i;
1015
1016 for (i = 0; i < NUM_INIT_LISTS; i++) {
1017 kmem_list3_init(&initkmem_list3[i]);
1018 if (i < MAX_NUMNODES)
1019 cache_cache.nodelists[i] = NULL;
1020 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001021
1022 /*
1023 * Fragmentation resistance on low memory - only use bigger
1024 * page orders on machines with more than 32MB of memory.
1025 */
1026 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1027 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1028
Linus Torvalds1da177e2005-04-16 15:20:36 -07001029 /* Bootstrap is tricky, because several objects are allocated
1030 * from caches that do not exist yet:
1031 * 1) initialize the cache_cache cache: it contains the kmem_cache_t
1032 * structures of all caches, except cache_cache itself: cache_cache
1033 * is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001034 * Initially an __init data area is used for the head array and the
1035 * kmem_list3 structures, it's replaced with a kmalloc allocated
1036 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001037 * 2) Create the first kmalloc cache.
Christoph Lametere498be72005-09-09 13:03:32 -07001038 * The kmem_cache_t for the new cache is allocated normally.
1039 * An __init data area is used for the head array.
1040 * 3) Create the remaining kmalloc caches, with minimally sized
1041 * head arrays.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001042 * 4) Replace the __init data head arrays for cache_cache and the first
1043 * kmalloc cache with kmalloc allocated arrays.
Christoph Lametere498be72005-09-09 13:03:32 -07001044 * 5) Replace the __init data for kmem_list3 for cache_cache and
1045 * the other cache's with kmalloc allocated memory.
1046 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001047 */
1048
1049 /* 1) create the cache_cache */
1050 init_MUTEX(&cache_chain_sem);
1051 INIT_LIST_HEAD(&cache_chain);
1052 list_add(&cache_cache.next, &cache_chain);
1053 cache_cache.colour_off = cache_line_size();
1054 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
Christoph Lametere498be72005-09-09 13:03:32 -07001055 cache_cache.nodelists[numa_node_id()] = &initkmem_list3[CACHE_CACHE];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001056
1057 cache_cache.objsize = ALIGN(cache_cache.objsize, cache_line_size());
1058
1059 cache_estimate(0, cache_cache.objsize, cache_line_size(), 0,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001060 &left_over, &cache_cache.num);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001061 if (!cache_cache.num)
1062 BUG();
1063
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001064 cache_cache.colour = left_over / cache_cache.colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001065 cache_cache.colour_next = 0;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001066 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1067 sizeof(struct slab), cache_line_size());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001068
1069 /* 2+3) create the kmalloc caches */
1070 sizes = malloc_sizes;
1071 names = cache_names;
1072
Christoph Lametere498be72005-09-09 13:03:32 -07001073 /* Initialize the caches that provide memory for the array cache
1074 * and the kmem_list3 structures first.
1075 * Without this, further allocations will bug
1076 */
1077
1078 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001079 sizes[INDEX_AC].cs_size,
1080 ARCH_KMALLOC_MINALIGN,
1081 (ARCH_KMALLOC_FLAGS |
1082 SLAB_PANIC), NULL, NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07001083
1084 if (INDEX_AC != INDEX_L3)
1085 sizes[INDEX_L3].cs_cachep =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001086 kmem_cache_create(names[INDEX_L3].name,
1087 sizes[INDEX_L3].cs_size,
1088 ARCH_KMALLOC_MINALIGN,
1089 (ARCH_KMALLOC_FLAGS | SLAB_PANIC), NULL,
1090 NULL);
Christoph Lametere498be72005-09-09 13:03:32 -07001091
Linus Torvalds1da177e2005-04-16 15:20:36 -07001092 while (sizes->cs_size != ULONG_MAX) {
Christoph Lametere498be72005-09-09 13:03:32 -07001093 /*
1094 * For performance, all the general caches are L1 aligned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001095 * This should be particularly beneficial on SMP boxes, as it
1096 * eliminates "false sharing".
1097 * Note for systems short on memory removing the alignment will
Christoph Lametere498be72005-09-09 13:03:32 -07001098 * allow tighter packing of the smaller caches.
1099 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001100 if (!sizes->cs_cachep)
Christoph Lametere498be72005-09-09 13:03:32 -07001101 sizes->cs_cachep = kmem_cache_create(names->name,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001102 sizes->cs_size,
1103 ARCH_KMALLOC_MINALIGN,
1104 (ARCH_KMALLOC_FLAGS
1105 | SLAB_PANIC),
1106 NULL, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001107
1108 /* Inc off-slab bufctl limit until the ceiling is hit. */
1109 if (!(OFF_SLAB(sizes->cs_cachep))) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001110 offslab_limit = sizes->cs_size - sizeof(struct slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001111 offslab_limit /= sizeof(kmem_bufctl_t);
1112 }
1113
1114 sizes->cs_dmacachep = kmem_cache_create(names->name_dma,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001115 sizes->cs_size,
1116 ARCH_KMALLOC_MINALIGN,
1117 (ARCH_KMALLOC_FLAGS |
1118 SLAB_CACHE_DMA |
1119 SLAB_PANIC), NULL,
1120 NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001121
1122 sizes++;
1123 names++;
1124 }
1125 /* 4) Replace the bootstrap head arrays */
1126 {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001127 void *ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001128
Linus Torvalds1da177e2005-04-16 15:20:36 -07001129 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001130
Linus Torvalds1da177e2005-04-16 15:20:36 -07001131 local_irq_disable();
1132 BUG_ON(ac_data(&cache_cache) != &initarray_cache.cache);
Christoph Lametere498be72005-09-09 13:03:32 -07001133 memcpy(ptr, ac_data(&cache_cache),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001134 sizeof(struct arraycache_init));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001135 cache_cache.array[smp_processor_id()] = ptr;
1136 local_irq_enable();
Christoph Lametere498be72005-09-09 13:03:32 -07001137
Linus Torvalds1da177e2005-04-16 15:20:36 -07001138 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001139
Linus Torvalds1da177e2005-04-16 15:20:36 -07001140 local_irq_disable();
Christoph Lametere498be72005-09-09 13:03:32 -07001141 BUG_ON(ac_data(malloc_sizes[INDEX_AC].cs_cachep)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001142 != &initarray_generic.cache);
Christoph Lametere498be72005-09-09 13:03:32 -07001143 memcpy(ptr, ac_data(malloc_sizes[INDEX_AC].cs_cachep),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001144 sizeof(struct arraycache_init));
Christoph Lametere498be72005-09-09 13:03:32 -07001145 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001146 ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001147 local_irq_enable();
1148 }
Christoph Lametere498be72005-09-09 13:03:32 -07001149 /* 5) Replace the bootstrap kmem_list3's */
1150 {
1151 int node;
1152 /* Replace the static kmem_list3 structures for the boot cpu */
1153 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE],
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001154 numa_node_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001155
Christoph Lametere498be72005-09-09 13:03:32 -07001156 for_each_online_node(node) {
1157 init_list(malloc_sizes[INDEX_AC].cs_cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001158 &initkmem_list3[SIZE_AC + node], node);
Christoph Lametere498be72005-09-09 13:03:32 -07001159
1160 if (INDEX_AC != INDEX_L3) {
1161 init_list(malloc_sizes[INDEX_L3].cs_cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001162 &initkmem_list3[SIZE_L3 + node],
1163 node);
Christoph Lametere498be72005-09-09 13:03:32 -07001164 }
1165 }
1166 }
1167
1168 /* 6) resize the head arrays to their final sizes */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001169 {
1170 kmem_cache_t *cachep;
1171 down(&cache_chain_sem);
1172 list_for_each_entry(cachep, &cache_chain, next)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001173 enable_cpucache(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001174 up(&cache_chain_sem);
1175 }
1176
1177 /* Done! */
1178 g_cpucache_up = FULL;
1179
1180 /* Register a cpu startup notifier callback
1181 * that initializes ac_data for all new cpus
1182 */
1183 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001184
1185 /* The reap timers are started later, with a module init call:
1186 * That part of the kernel is not yet operational.
1187 */
1188}
1189
1190static int __init cpucache_init(void)
1191{
1192 int cpu;
1193
1194 /*
1195 * Register the timers that return unneeded
1196 * pages to gfp.
1197 */
Christoph Lametere498be72005-09-09 13:03:32 -07001198 for_each_online_cpu(cpu)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001199 start_cpu_timer(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001200
1201 return 0;
1202}
1203
1204__initcall(cpucache_init);
1205
1206/*
1207 * Interface to system's page allocator. No need to hold the cache-lock.
1208 *
1209 * If we requested dmaable memory, we will get it. Even if we
1210 * did not request dmaable memory, we might get it, but that
1211 * would be relatively rare and ignorable.
1212 */
Al Virodd0fc662005-10-07 07:46:04 +01001213static void *kmem_getpages(kmem_cache_t *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001214{
1215 struct page *page;
1216 void *addr;
1217 int i;
1218
1219 flags |= cachep->gfpflags;
Christoph Lameter50c85a12005-11-13 16:06:47 -08001220 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001221 if (!page)
1222 return NULL;
1223 addr = page_address(page);
1224
1225 i = (1 << cachep->gfporder);
1226 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1227 atomic_add(i, &slab_reclaim_pages);
1228 add_page_state(nr_slab, i);
1229 while (i--) {
1230 SetPageSlab(page);
1231 page++;
1232 }
1233 return addr;
1234}
1235
1236/*
1237 * Interface to system's page release.
1238 */
1239static void kmem_freepages(kmem_cache_t *cachep, void *addr)
1240{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001241 unsigned long i = (1 << cachep->gfporder);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001242 struct page *page = virt_to_page(addr);
1243 const unsigned long nr_freed = i;
1244
1245 while (i--) {
1246 if (!TestClearPageSlab(page))
1247 BUG();
1248 page++;
1249 }
1250 sub_page_state(nr_slab, nr_freed);
1251 if (current->reclaim_state)
1252 current->reclaim_state->reclaimed_slab += nr_freed;
1253 free_pages((unsigned long)addr, cachep->gfporder);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001254 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1255 atomic_sub(1 << cachep->gfporder, &slab_reclaim_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001256}
1257
1258static void kmem_rcu_free(struct rcu_head *head)
1259{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001260 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001261 kmem_cache_t *cachep = slab_rcu->cachep;
1262
1263 kmem_freepages(cachep, slab_rcu->addr);
1264 if (OFF_SLAB(cachep))
1265 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1266}
1267
1268#if DEBUG
1269
1270#ifdef CONFIG_DEBUG_PAGEALLOC
1271static void store_stackinfo(kmem_cache_t *cachep, unsigned long *addr,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001272 unsigned long caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001273{
1274 int size = obj_reallen(cachep);
1275
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001276 addr = (unsigned long *)&((char *)addr)[obj_dbghead(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001277
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001278 if (size < 5 * sizeof(unsigned long))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001279 return;
1280
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001281 *addr++ = 0x12345678;
1282 *addr++ = caller;
1283 *addr++ = smp_processor_id();
1284 size -= 3 * sizeof(unsigned long);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001285 {
1286 unsigned long *sptr = &caller;
1287 unsigned long svalue;
1288
1289 while (!kstack_end(sptr)) {
1290 svalue = *sptr++;
1291 if (kernel_text_address(svalue)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001292 *addr++ = svalue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001293 size -= sizeof(unsigned long);
1294 if (size <= sizeof(unsigned long))
1295 break;
1296 }
1297 }
1298
1299 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001300 *addr++ = 0x87654321;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001301}
1302#endif
1303
1304static void poison_obj(kmem_cache_t *cachep, void *addr, unsigned char val)
1305{
1306 int size = obj_reallen(cachep);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001307 addr = &((char *)addr)[obj_dbghead(cachep)];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001308
1309 memset(addr, val, size);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001310 *(unsigned char *)(addr + size - 1) = POISON_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001311}
1312
1313static void dump_line(char *data, int offset, int limit)
1314{
1315 int i;
1316 printk(KERN_ERR "%03x:", offset);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001317 for (i = 0; i < limit; i++) {
1318 printk(" %02x", (unsigned char)data[offset + i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001319 }
1320 printk("\n");
1321}
1322#endif
1323
1324#if DEBUG
1325
1326static void print_objinfo(kmem_cache_t *cachep, void *objp, int lines)
1327{
1328 int i, size;
1329 char *realobj;
1330
1331 if (cachep->flags & SLAB_RED_ZONE) {
1332 printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001333 *dbg_redzone1(cachep, objp),
1334 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001335 }
1336
1337 if (cachep->flags & SLAB_STORE_USER) {
1338 printk(KERN_ERR "Last user: [<%p>]",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001339 *dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001340 print_symbol("(%s)",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001341 (unsigned long)*dbg_userword(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001342 printk("\n");
1343 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001344 realobj = (char *)objp + obj_dbghead(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001345 size = obj_reallen(cachep);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001346 for (i = 0; i < size && lines; i += 16, lines--) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001347 int limit;
1348 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001349 if (i + limit > size)
1350 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001351 dump_line(realobj, i, limit);
1352 }
1353}
1354
1355static void check_poison_obj(kmem_cache_t *cachep, void *objp)
1356{
1357 char *realobj;
1358 int size, i;
1359 int lines = 0;
1360
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001361 realobj = (char *)objp + obj_dbghead(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001362 size = obj_reallen(cachep);
1363
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001364 for (i = 0; i < size; i++) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001365 char exp = POISON_FREE;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001366 if (i == size - 1)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001367 exp = POISON_END;
1368 if (realobj[i] != exp) {
1369 int limit;
1370 /* Mismatch ! */
1371 /* Print header */
1372 if (lines == 0) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001373 printk(KERN_ERR
1374 "Slab corruption: start=%p, len=%d\n",
1375 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001376 print_objinfo(cachep, objp, 0);
1377 }
1378 /* Hexdump the affected line */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001379 i = (i / 16) * 16;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001380 limit = 16;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001381 if (i + limit > size)
1382 limit = size - i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001383 dump_line(realobj, i, limit);
1384 i += 16;
1385 lines++;
1386 /* Limit to 5 lines */
1387 if (lines > 5)
1388 break;
1389 }
1390 }
1391 if (lines != 0) {
1392 /* Print some data about the neighboring objects, if they
1393 * exist:
1394 */
Pekka Enberg065d41c2005-11-13 16:06:46 -08001395 struct slab *slabp = page_get_slab(virt_to_page(objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001396 int objnr;
1397
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001398 objnr = (objp - slabp->s_mem) / cachep->objsize;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001399 if (objnr) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001400 objp = slabp->s_mem + (objnr - 1) * cachep->objsize;
1401 realobj = (char *)objp + obj_dbghead(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001402 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001403 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001404 print_objinfo(cachep, objp, 2);
1405 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001406 if (objnr + 1 < cachep->num) {
1407 objp = slabp->s_mem + (objnr + 1) * cachep->objsize;
1408 realobj = (char *)objp + obj_dbghead(cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001409 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001410 realobj, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001411 print_objinfo(cachep, objp, 2);
1412 }
1413 }
1414}
1415#endif
1416
1417/* Destroy all the objs in a slab, and release the mem back to the system.
1418 * Before calling the slab must have been unlinked from the cache.
1419 * The cache-lock is not held/needed.
1420 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001421static void slab_destroy(kmem_cache_t *cachep, struct slab *slabp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001422{
1423 void *addr = slabp->s_mem - slabp->colouroff;
1424
1425#if DEBUG
1426 int i;
1427 for (i = 0; i < cachep->num; i++) {
1428 void *objp = slabp->s_mem + cachep->objsize * i;
1429
1430 if (cachep->flags & SLAB_POISON) {
1431#ifdef CONFIG_DEBUG_PAGEALLOC
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001432 if ((cachep->objsize % PAGE_SIZE) == 0
1433 && OFF_SLAB(cachep))
1434 kernel_map_pages(virt_to_page(objp),
1435 cachep->objsize / PAGE_SIZE,
1436 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001437 else
1438 check_poison_obj(cachep, objp);
1439#else
1440 check_poison_obj(cachep, objp);
1441#endif
1442 }
1443 if (cachep->flags & SLAB_RED_ZONE) {
1444 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1445 slab_error(cachep, "start of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001446 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001447 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1448 slab_error(cachep, "end of a freed object "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001449 "was overwritten");
Linus Torvalds1da177e2005-04-16 15:20:36 -07001450 }
1451 if (cachep->dtor && !(cachep->flags & SLAB_POISON))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001452 (cachep->dtor) (objp + obj_dbghead(cachep), cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001453 }
1454#else
1455 if (cachep->dtor) {
1456 int i;
1457 for (i = 0; i < cachep->num; i++) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001458 void *objp = slabp->s_mem + cachep->objsize * i;
1459 (cachep->dtor) (objp, cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001460 }
1461 }
1462#endif
1463
1464 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1465 struct slab_rcu *slab_rcu;
1466
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001467 slab_rcu = (struct slab_rcu *)slabp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001468 slab_rcu->cachep = cachep;
1469 slab_rcu->addr = addr;
1470 call_rcu(&slab_rcu->head, kmem_rcu_free);
1471 } else {
1472 kmem_freepages(cachep, addr);
1473 if (OFF_SLAB(cachep))
1474 kmem_cache_free(cachep->slabp_cache, slabp);
1475 }
1476}
1477
Christoph Lametere498be72005-09-09 13:03:32 -07001478/* For setting up all the kmem_list3s for cache whose objsize is same
1479 as size of kmem_list3. */
1480static inline void set_up_list3s(kmem_cache_t *cachep, int index)
1481{
1482 int node;
1483
1484 for_each_online_node(node) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001485 cachep->nodelists[node] = &initkmem_list3[index + node];
Christoph Lametere498be72005-09-09 13:03:32 -07001486 cachep->nodelists[node]->next_reap = jiffies +
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001487 REAPTIMEOUT_LIST3 +
1488 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -07001489 }
1490}
1491
Linus Torvalds1da177e2005-04-16 15:20:36 -07001492/**
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001493 * calculate_slab_order - calculate size (page order) of slabs and the number
1494 * of objects per slab.
1495 *
1496 * This could be made much more intelligent. For now, try to avoid using
1497 * high order pages for slabs. When the gfp() functions are more friendly
1498 * towards high-order requests, this should be changed.
1499 */
1500static inline size_t calculate_slab_order(kmem_cache_t *cachep, size_t size,
1501 size_t align, gfp_t flags)
1502{
1503 size_t left_over = 0;
1504
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001505 for (;; cachep->gfporder++) {
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001506 unsigned int num;
1507 size_t remainder;
1508
1509 if (cachep->gfporder > MAX_GFP_ORDER) {
1510 cachep->num = 0;
1511 break;
1512 }
1513
1514 cache_estimate(cachep->gfporder, size, align, flags,
1515 &remainder, &num);
1516 if (!num)
1517 continue;
1518 /* More than offslab_limit objects will cause problems */
1519 if (flags & CFLGS_OFF_SLAB && cachep->num > offslab_limit)
1520 break;
1521
1522 cachep->num = num;
1523 left_over = remainder;
1524
1525 /*
1526 * Large number of objects is good, but very large slabs are
1527 * currently bad for the gfp()s.
1528 */
1529 if (cachep->gfporder >= slab_break_gfp_order)
1530 break;
1531
1532 if ((left_over * 8) <= (PAGE_SIZE << cachep->gfporder))
1533 /* Acceptable internal fragmentation */
1534 break;
1535 }
1536 return left_over;
1537}
1538
1539/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07001540 * kmem_cache_create - Create a cache.
1541 * @name: A string which is used in /proc/slabinfo to identify this cache.
1542 * @size: The size of objects to be created in this cache.
1543 * @align: The required alignment for the objects.
1544 * @flags: SLAB flags
1545 * @ctor: A constructor for the objects.
1546 * @dtor: A destructor for the objects.
1547 *
1548 * Returns a ptr to the cache on success, NULL on failure.
1549 * Cannot be called within a int, but can be interrupted.
1550 * The @ctor is run when new pages are allocated by the cache
1551 * and the @dtor is run before the pages are handed back.
1552 *
1553 * @name must be valid until the cache is destroyed. This implies that
1554 * the module calling this has to destroy the cache before getting
1555 * unloaded.
1556 *
1557 * The flags are
1558 *
1559 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1560 * to catch references to uninitialised memory.
1561 *
1562 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1563 * for buffer overruns.
1564 *
1565 * %SLAB_NO_REAP - Don't automatically reap this cache when we're under
1566 * memory pressure.
1567 *
1568 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1569 * cacheline. This can be beneficial if you're counting cycles as closely
1570 * as davem.
1571 */
1572kmem_cache_t *
1573kmem_cache_create (const char *name, size_t size, size_t align,
1574 unsigned long flags, void (*ctor)(void*, kmem_cache_t *, unsigned long),
1575 void (*dtor)(void*, kmem_cache_t *, unsigned long))
1576{
1577 size_t left_over, slab_size, ralign;
1578 kmem_cache_t *cachep = NULL;
Andrew Morton4f12bb42005-11-07 00:58:00 -08001579 struct list_head *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001580
1581 /*
1582 * Sanity checks... these are all serious usage bugs.
1583 */
1584 if ((!name) ||
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001585 in_interrupt() ||
1586 (size < BYTES_PER_WORD) ||
1587 (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) {
1588 printk(KERN_ERR "%s: Early error in slab %s\n",
1589 __FUNCTION__, name);
1590 BUG();
1591 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001592
Andrew Morton4f12bb42005-11-07 00:58:00 -08001593 down(&cache_chain_sem);
1594
1595 list_for_each(p, &cache_chain) {
1596 kmem_cache_t *pc = list_entry(p, kmem_cache_t, next);
1597 mm_segment_t old_fs = get_fs();
1598 char tmp;
1599 int res;
1600
1601 /*
1602 * This happens when the module gets unloaded and doesn't
1603 * destroy its slab cache and no-one else reuses the vmalloc
1604 * area of the module. Print a warning.
1605 */
1606 set_fs(KERNEL_DS);
1607 res = __get_user(tmp, pc->name);
1608 set_fs(old_fs);
1609 if (res) {
1610 printk("SLAB: cache with size %d has lost its name\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001611 pc->objsize);
Andrew Morton4f12bb42005-11-07 00:58:00 -08001612 continue;
1613 }
1614
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001615 if (!strcmp(pc->name, name)) {
Andrew Morton4f12bb42005-11-07 00:58:00 -08001616 printk("kmem_cache_create: duplicate cache %s\n", name);
1617 dump_stack();
1618 goto oops;
1619 }
1620 }
1621
Linus Torvalds1da177e2005-04-16 15:20:36 -07001622#if DEBUG
1623 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
1624 if ((flags & SLAB_DEBUG_INITIAL) && !ctor) {
1625 /* No constructor, but inital state check requested */
1626 printk(KERN_ERR "%s: No con, but init state check "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001627 "requested - %s\n", __FUNCTION__, name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001628 flags &= ~SLAB_DEBUG_INITIAL;
1629 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001630#if FORCED_DEBUG
1631 /*
1632 * Enable redzoning and last user accounting, except for caches with
1633 * large objects, if the increased size would increase the object size
1634 * above the next power of two: caches with object sizes just above a
1635 * power of two have a significant amount of internal fragmentation.
1636 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001637 if ((size < 4096
1638 || fls(size - 1) == fls(size - 1 + 3 * BYTES_PER_WORD)))
1639 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001640 if (!(flags & SLAB_DESTROY_BY_RCU))
1641 flags |= SLAB_POISON;
1642#endif
1643 if (flags & SLAB_DESTROY_BY_RCU)
1644 BUG_ON(flags & SLAB_POISON);
1645#endif
1646 if (flags & SLAB_DESTROY_BY_RCU)
1647 BUG_ON(dtor);
1648
1649 /*
1650 * Always checks flags, a caller might be expecting debug
1651 * support which isn't available.
1652 */
1653 if (flags & ~CREATE_MASK)
1654 BUG();
1655
1656 /* Check that size is in terms of words. This is needed to avoid
1657 * unaligned accesses for some archs when redzoning is used, and makes
1658 * sure any on-slab bufctl's are also correctly aligned.
1659 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001660 if (size & (BYTES_PER_WORD - 1)) {
1661 size += (BYTES_PER_WORD - 1);
1662 size &= ~(BYTES_PER_WORD - 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001663 }
1664
1665 /* calculate out the final buffer alignment: */
1666 /* 1) arch recommendation: can be overridden for debug */
1667 if (flags & SLAB_HWCACHE_ALIGN) {
1668 /* Default alignment: as specified by the arch code.
1669 * Except if an object is really small, then squeeze multiple
1670 * objects into one cacheline.
1671 */
1672 ralign = cache_line_size();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001673 while (size <= ralign / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001674 ralign /= 2;
1675 } else {
1676 ralign = BYTES_PER_WORD;
1677 }
1678 /* 2) arch mandated alignment: disables debug if necessary */
1679 if (ralign < ARCH_SLAB_MINALIGN) {
1680 ralign = ARCH_SLAB_MINALIGN;
1681 if (ralign > BYTES_PER_WORD)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001682 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001683 }
1684 /* 3) caller mandated alignment: disables debug if necessary */
1685 if (ralign < align) {
1686 ralign = align;
1687 if (ralign > BYTES_PER_WORD)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001688 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001689 }
1690 /* 4) Store it. Note that the debug code below can reduce
1691 * the alignment to BYTES_PER_WORD.
1692 */
1693 align = ralign;
1694
1695 /* Get cache's description obj. */
1696 cachep = (kmem_cache_t *) kmem_cache_alloc(&cache_cache, SLAB_KERNEL);
1697 if (!cachep)
Andrew Morton4f12bb42005-11-07 00:58:00 -08001698 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001699 memset(cachep, 0, sizeof(kmem_cache_t));
1700
1701#if DEBUG
1702 cachep->reallen = size;
1703
1704 if (flags & SLAB_RED_ZONE) {
1705 /* redzoning only works with word aligned caches */
1706 align = BYTES_PER_WORD;
1707
1708 /* add space for red zone words */
1709 cachep->dbghead += BYTES_PER_WORD;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001710 size += 2 * BYTES_PER_WORD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001711 }
1712 if (flags & SLAB_STORE_USER) {
1713 /* user store requires word alignment and
1714 * one word storage behind the end of the real
1715 * object.
1716 */
1717 align = BYTES_PER_WORD;
1718 size += BYTES_PER_WORD;
1719 }
1720#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001721 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
1722 && cachep->reallen > cache_line_size() && size < PAGE_SIZE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001723 cachep->dbghead += PAGE_SIZE - size;
1724 size = PAGE_SIZE;
1725 }
1726#endif
1727#endif
1728
1729 /* Determine if the slab management is 'on' or 'off' slab. */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001730 if (size >= (PAGE_SIZE >> 3))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001731 /*
1732 * Size is large, assume best to place the slab management obj
1733 * off-slab (should allow better packing of objs).
1734 */
1735 flags |= CFLGS_OFF_SLAB;
1736
1737 size = ALIGN(size, align);
1738
1739 if ((flags & SLAB_RECLAIM_ACCOUNT) && size <= PAGE_SIZE) {
1740 /*
1741 * A VFS-reclaimable slab tends to have most allocations
1742 * as GFP_NOFS and we really don't want to have to be allocating
1743 * higher-order pages when we are unable to shrink dcache.
1744 */
1745 cachep->gfporder = 0;
1746 cache_estimate(cachep->gfporder, size, align, flags,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001747 &left_over, &cachep->num);
Pekka Enberg4d268eb2006-01-08 01:00:36 -08001748 } else
1749 left_over = calculate_slab_order(cachep, size, align, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001750
1751 if (!cachep->num) {
1752 printk("kmem_cache_create: couldn't create cache %s.\n", name);
1753 kmem_cache_free(&cache_cache, cachep);
1754 cachep = NULL;
Andrew Morton4f12bb42005-11-07 00:58:00 -08001755 goto oops;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001756 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001757 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
1758 + sizeof(struct slab), align);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001759
1760 /*
1761 * If the slab has been placed off-slab, and we have enough space then
1762 * move it on-slab. This is at the expense of any extra colouring.
1763 */
1764 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
1765 flags &= ~CFLGS_OFF_SLAB;
1766 left_over -= slab_size;
1767 }
1768
1769 if (flags & CFLGS_OFF_SLAB) {
1770 /* really off slab. No need for manual alignment */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001771 slab_size =
1772 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001773 }
1774
1775 cachep->colour_off = cache_line_size();
1776 /* Offset must be a multiple of the alignment. */
1777 if (cachep->colour_off < align)
1778 cachep->colour_off = align;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001779 cachep->colour = left_over / cachep->colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001780 cachep->slab_size = slab_size;
1781 cachep->flags = flags;
1782 cachep->gfpflags = 0;
1783 if (flags & SLAB_CACHE_DMA)
1784 cachep->gfpflags |= GFP_DMA;
1785 spin_lock_init(&cachep->spinlock);
1786 cachep->objsize = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001787
1788 if (flags & CFLGS_OFF_SLAB)
Victor Fuscob2d55072005-09-10 00:26:36 -07001789 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001790 cachep->ctor = ctor;
1791 cachep->dtor = dtor;
1792 cachep->name = name;
1793
1794 /* Don't let CPUs to come and go */
1795 lock_cpu_hotplug();
1796
1797 if (g_cpucache_up == FULL) {
1798 enable_cpucache(cachep);
1799 } else {
1800 if (g_cpucache_up == NONE) {
1801 /* Note: the first kmem_cache_create must create
1802 * the cache that's used by kmalloc(24), otherwise
1803 * the creation of further caches will BUG().
1804 */
Christoph Lametere498be72005-09-09 13:03:32 -07001805 cachep->array[smp_processor_id()] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001806 &initarray_generic.cache;
Christoph Lametere498be72005-09-09 13:03:32 -07001807
1808 /* If the cache that's used by
1809 * kmalloc(sizeof(kmem_list3)) is the first cache,
1810 * then we need to set up all its list3s, otherwise
1811 * the creation of further caches will BUG().
1812 */
1813 set_up_list3s(cachep, SIZE_AC);
1814 if (INDEX_AC == INDEX_L3)
1815 g_cpucache_up = PARTIAL_L3;
1816 else
1817 g_cpucache_up = PARTIAL_AC;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001818 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07001819 cachep->array[smp_processor_id()] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001820 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001821
1822 if (g_cpucache_up == PARTIAL_AC) {
1823 set_up_list3s(cachep, SIZE_L3);
1824 g_cpucache_up = PARTIAL_L3;
1825 } else {
1826 int node;
1827 for_each_online_node(node) {
1828
1829 cachep->nodelists[node] =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001830 kmalloc_node(sizeof
1831 (struct kmem_list3),
1832 GFP_KERNEL, node);
Christoph Lametere498be72005-09-09 13:03:32 -07001833 BUG_ON(!cachep->nodelists[node]);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001834 kmem_list3_init(cachep->
1835 nodelists[node]);
Christoph Lametere498be72005-09-09 13:03:32 -07001836 }
1837 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001838 }
Christoph Lametere498be72005-09-09 13:03:32 -07001839 cachep->nodelists[numa_node_id()]->next_reap =
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001840 jiffies + REAPTIMEOUT_LIST3 +
1841 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -07001842
Linus Torvalds1da177e2005-04-16 15:20:36 -07001843 BUG_ON(!ac_data(cachep));
1844 ac_data(cachep)->avail = 0;
1845 ac_data(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1846 ac_data(cachep)->batchcount = 1;
1847 ac_data(cachep)->touched = 0;
1848 cachep->batchcount = 1;
1849 cachep->limit = BOOT_CPUCACHE_ENTRIES;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001850 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001851
Linus Torvalds1da177e2005-04-16 15:20:36 -07001852 /* cache setup completed, link it into the list */
1853 list_add(&cachep->next, &cache_chain);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001854 unlock_cpu_hotplug();
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001855 oops:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001856 if (!cachep && (flags & SLAB_PANIC))
1857 panic("kmem_cache_create(): failed to create slab `%s'\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001858 name);
Andrew Morton4f12bb42005-11-07 00:58:00 -08001859 up(&cache_chain_sem);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001860 return cachep;
1861}
1862EXPORT_SYMBOL(kmem_cache_create);
1863
1864#if DEBUG
1865static void check_irq_off(void)
1866{
1867 BUG_ON(!irqs_disabled());
1868}
1869
1870static void check_irq_on(void)
1871{
1872 BUG_ON(irqs_disabled());
1873}
1874
1875static void check_spinlock_acquired(kmem_cache_t *cachep)
1876{
1877#ifdef CONFIG_SMP
1878 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07001879 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001880#endif
1881}
Christoph Lametere498be72005-09-09 13:03:32 -07001882
1883static inline void check_spinlock_acquired_node(kmem_cache_t *cachep, int node)
1884{
1885#ifdef CONFIG_SMP
1886 check_irq_off();
1887 assert_spin_locked(&cachep->nodelists[node]->list_lock);
1888#endif
1889}
1890
Linus Torvalds1da177e2005-04-16 15:20:36 -07001891#else
1892#define check_irq_off() do { } while(0)
1893#define check_irq_on() do { } while(0)
1894#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07001895#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001896#endif
1897
1898/*
1899 * Waits for all CPUs to execute func().
1900 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001901static void smp_call_function_all_cpus(void (*func)(void *arg), void *arg)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001902{
1903 check_irq_on();
1904 preempt_disable();
1905
1906 local_irq_disable();
1907 func(arg);
1908 local_irq_enable();
1909
1910 if (smp_call_function(func, arg, 1, 1))
1911 BUG();
1912
1913 preempt_enable();
1914}
1915
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001916static void drain_array_locked(kmem_cache_t *cachep, struct array_cache *ac,
1917 int force, int node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001918
1919static void do_drain(void *arg)
1920{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001921 kmem_cache_t *cachep = (kmem_cache_t *) arg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001922 struct array_cache *ac;
Christoph Lameterff694162005-09-22 21:44:02 -07001923 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001924
1925 check_irq_off();
1926 ac = ac_data(cachep);
Christoph Lameterff694162005-09-22 21:44:02 -07001927 spin_lock(&cachep->nodelists[node]->list_lock);
1928 free_block(cachep, ac->entry, ac->avail, node);
1929 spin_unlock(&cachep->nodelists[node]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001930 ac->avail = 0;
1931}
1932
1933static void drain_cpu_caches(kmem_cache_t *cachep)
1934{
Christoph Lametere498be72005-09-09 13:03:32 -07001935 struct kmem_list3 *l3;
1936 int node;
1937
Linus Torvalds1da177e2005-04-16 15:20:36 -07001938 smp_call_function_all_cpus(do_drain, cachep);
1939 check_irq_on();
1940 spin_lock_irq(&cachep->spinlock);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001941 for_each_online_node(node) {
Christoph Lametere498be72005-09-09 13:03:32 -07001942 l3 = cachep->nodelists[node];
1943 if (l3) {
1944 spin_lock(&l3->list_lock);
1945 drain_array_locked(cachep, l3->shared, 1, node);
1946 spin_unlock(&l3->list_lock);
1947 if (l3->alien)
1948 drain_alien_cache(cachep, l3);
1949 }
1950 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001951 spin_unlock_irq(&cachep->spinlock);
1952}
1953
Christoph Lametere498be72005-09-09 13:03:32 -07001954static int __node_shrink(kmem_cache_t *cachep, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001955{
1956 struct slab *slabp;
Christoph Lametere498be72005-09-09 13:03:32 -07001957 struct kmem_list3 *l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001958 int ret;
1959
Christoph Lametere498be72005-09-09 13:03:32 -07001960 for (;;) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001961 struct list_head *p;
1962
Christoph Lametere498be72005-09-09 13:03:32 -07001963 p = l3->slabs_free.prev;
1964 if (p == &l3->slabs_free)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001965 break;
1966
Christoph Lametere498be72005-09-09 13:03:32 -07001967 slabp = list_entry(l3->slabs_free.prev, struct slab, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001968#if DEBUG
1969 if (slabp->inuse)
1970 BUG();
1971#endif
1972 list_del(&slabp->list);
1973
Christoph Lametere498be72005-09-09 13:03:32 -07001974 l3->free_objects -= cachep->num;
1975 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001976 slab_destroy(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07001977 spin_lock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001978 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08001979 ret = !list_empty(&l3->slabs_full) || !list_empty(&l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001980 return ret;
1981}
1982
Christoph Lametere498be72005-09-09 13:03:32 -07001983static int __cache_shrink(kmem_cache_t *cachep)
1984{
1985 int ret = 0, i = 0;
1986 struct kmem_list3 *l3;
1987
1988 drain_cpu_caches(cachep);
1989
1990 check_irq_on();
1991 for_each_online_node(i) {
1992 l3 = cachep->nodelists[i];
1993 if (l3) {
1994 spin_lock_irq(&l3->list_lock);
1995 ret += __node_shrink(cachep, i);
1996 spin_unlock_irq(&l3->list_lock);
1997 }
1998 }
1999 return (ret ? 1 : 0);
2000}
2001
Linus Torvalds1da177e2005-04-16 15:20:36 -07002002/**
2003 * kmem_cache_shrink - Shrink a cache.
2004 * @cachep: The cache to shrink.
2005 *
2006 * Releases as many slabs as possible for a cache.
2007 * To help debugging, a zero exit status indicates all slabs were released.
2008 */
2009int kmem_cache_shrink(kmem_cache_t *cachep)
2010{
2011 if (!cachep || in_interrupt())
2012 BUG();
2013
2014 return __cache_shrink(cachep);
2015}
2016EXPORT_SYMBOL(kmem_cache_shrink);
2017
2018/**
2019 * kmem_cache_destroy - delete a cache
2020 * @cachep: the cache to destroy
2021 *
2022 * Remove a kmem_cache_t object from the slab cache.
2023 * Returns 0 on success.
2024 *
2025 * It is expected this function will be called by a module when it is
2026 * unloaded. This will remove the cache completely, and avoid a duplicate
2027 * cache being allocated each time a module is loaded and unloaded, if the
2028 * module doesn't have persistent in-kernel storage across loads and unloads.
2029 *
2030 * The cache must be empty before calling this function.
2031 *
2032 * The caller must guarantee that noone will allocate memory from the cache
2033 * during the kmem_cache_destroy().
2034 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002035int kmem_cache_destroy(kmem_cache_t *cachep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002036{
2037 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07002038 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002039
2040 if (!cachep || in_interrupt())
2041 BUG();
2042
2043 /* Don't let CPUs to come and go */
2044 lock_cpu_hotplug();
2045
2046 /* Find the cache in the chain of caches. */
2047 down(&cache_chain_sem);
2048 /*
2049 * the chain is never empty, cache_cache is never destroyed
2050 */
2051 list_del(&cachep->next);
2052 up(&cache_chain_sem);
2053
2054 if (__cache_shrink(cachep)) {
2055 slab_error(cachep, "Can't free all objects");
2056 down(&cache_chain_sem);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002057 list_add(&cachep->next, &cache_chain);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002058 up(&cache_chain_sem);
2059 unlock_cpu_hotplug();
2060 return 1;
2061 }
2062
2063 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
Paul E. McKenneyfbd568a3e2005-05-01 08:59:04 -07002064 synchronize_rcu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002065
Christoph Lametere498be72005-09-09 13:03:32 -07002066 for_each_online_cpu(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002067 kfree(cachep->array[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002068
2069 /* NUMA: free the list3 structures */
Christoph Lametere498be72005-09-09 13:03:32 -07002070 for_each_online_node(i) {
2071 if ((l3 = cachep->nodelists[i])) {
2072 kfree(l3->shared);
2073 free_alien_cache(l3->alien);
2074 kfree(l3);
2075 }
2076 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002077 kmem_cache_free(&cache_cache, cachep);
2078
2079 unlock_cpu_hotplug();
2080
2081 return 0;
2082}
2083EXPORT_SYMBOL(kmem_cache_destroy);
2084
2085/* Get the memory for a slab management obj. */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002086static struct slab *alloc_slabmgmt(kmem_cache_t *cachep, void *objp,
2087 int colour_off, gfp_t local_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002088{
2089 struct slab *slabp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002090
Linus Torvalds1da177e2005-04-16 15:20:36 -07002091 if (OFF_SLAB(cachep)) {
2092 /* Slab management obj is off-slab. */
2093 slabp = kmem_cache_alloc(cachep->slabp_cache, local_flags);
2094 if (!slabp)
2095 return NULL;
2096 } else {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002097 slabp = objp + colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002098 colour_off += cachep->slab_size;
2099 }
2100 slabp->inuse = 0;
2101 slabp->colouroff = colour_off;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002102 slabp->s_mem = objp + colour_off;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002103
2104 return slabp;
2105}
2106
2107static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2108{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002109 return (kmem_bufctl_t *) (slabp + 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002110}
2111
2112static void cache_init_objs(kmem_cache_t *cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002113 struct slab *slabp, unsigned long ctor_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002114{
2115 int i;
2116
2117 for (i = 0; i < cachep->num; i++) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002118 void *objp = slabp->s_mem + cachep->objsize * i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002119#if DEBUG
2120 /* need to poison the objs? */
2121 if (cachep->flags & SLAB_POISON)
2122 poison_obj(cachep, objp, POISON_FREE);
2123 if (cachep->flags & SLAB_STORE_USER)
2124 *dbg_userword(cachep, objp) = NULL;
2125
2126 if (cachep->flags & SLAB_RED_ZONE) {
2127 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2128 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2129 }
2130 /*
2131 * Constructors are not allowed to allocate memory from
2132 * the same cache which they are a constructor for.
2133 * Otherwise, deadlock. They must also be threaded.
2134 */
2135 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002136 cachep->ctor(objp + obj_dbghead(cachep), cachep,
2137 ctor_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002138
2139 if (cachep->flags & SLAB_RED_ZONE) {
2140 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2141 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002142 " end of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002143 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2144 slab_error(cachep, "constructor overwrote the"
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002145 " start of an object");
Linus Torvalds1da177e2005-04-16 15:20:36 -07002146 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002147 if ((cachep->objsize % PAGE_SIZE) == 0 && OFF_SLAB(cachep)
2148 && cachep->flags & SLAB_POISON)
2149 kernel_map_pages(virt_to_page(objp),
2150 cachep->objsize / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002151#else
2152 if (cachep->ctor)
2153 cachep->ctor(objp, cachep, ctor_flags);
2154#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002155 slab_bufctl(slabp)[i] = i + 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002156 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002157 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002158 slabp->free = 0;
2159}
2160
Al Viro6daa0e22005-10-21 03:18:50 -04002161static void kmem_flagcheck(kmem_cache_t *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002162{
2163 if (flags & SLAB_DMA) {
2164 if (!(cachep->gfpflags & GFP_DMA))
2165 BUG();
2166 } else {
2167 if (cachep->gfpflags & GFP_DMA)
2168 BUG();
2169 }
2170}
2171
2172static void set_slab_attr(kmem_cache_t *cachep, struct slab *slabp, void *objp)
2173{
2174 int i;
2175 struct page *page;
2176
2177 /* Nasty!!!!!! I hope this is OK. */
2178 i = 1 << cachep->gfporder;
2179 page = virt_to_page(objp);
2180 do {
Pekka Enberg065d41c2005-11-13 16:06:46 -08002181 page_set_cache(page, cachep);
2182 page_set_slab(page, slabp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002183 page++;
2184 } while (--i);
2185}
2186
2187/*
2188 * Grow (by 1) the number of slabs within a cache. This is called by
2189 * kmem_cache_alloc() when there are no active objs left in a cache.
2190 */
Al Virodd0fc662005-10-07 07:46:04 +01002191static int cache_grow(kmem_cache_t *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002192{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002193 struct slab *slabp;
2194 void *objp;
2195 size_t offset;
2196 gfp_t local_flags;
2197 unsigned long ctor_flags;
Christoph Lametere498be72005-09-09 13:03:32 -07002198 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002199
2200 /* Be lazy and only check for valid flags here,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002201 * keeping it out of the critical path in kmem_cache_alloc().
Linus Torvalds1da177e2005-04-16 15:20:36 -07002202 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002203 if (flags & ~(SLAB_DMA | SLAB_LEVEL_MASK | SLAB_NO_GROW))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002204 BUG();
2205 if (flags & SLAB_NO_GROW)
2206 return 0;
2207
2208 ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2209 local_flags = (flags & SLAB_LEVEL_MASK);
2210 if (!(local_flags & __GFP_WAIT))
2211 /*
2212 * Not allowed to sleep. Need to tell a constructor about
2213 * this - it might need to know...
2214 */
2215 ctor_flags |= SLAB_CTOR_ATOMIC;
2216
2217 /* About to mess with non-constant members - lock. */
2218 check_irq_off();
2219 spin_lock(&cachep->spinlock);
2220
2221 /* Get colour for the slab, and cal the next value. */
2222 offset = cachep->colour_next;
2223 cachep->colour_next++;
2224 if (cachep->colour_next >= cachep->colour)
2225 cachep->colour_next = 0;
2226 offset *= cachep->colour_off;
2227
2228 spin_unlock(&cachep->spinlock);
2229
Christoph Lametere498be72005-09-09 13:03:32 -07002230 check_irq_off();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002231 if (local_flags & __GFP_WAIT)
2232 local_irq_enable();
2233
2234 /*
2235 * The test for missing atomic flag is performed here, rather than
2236 * the more obvious place, simply to reduce the critical path length
2237 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2238 * will eventually be caught here (where it matters).
2239 */
2240 kmem_flagcheck(cachep, flags);
2241
Christoph Lametere498be72005-09-09 13:03:32 -07002242 /* Get mem for the objs.
2243 * Attempt to allocate a physical page from 'nodeid',
2244 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002245 if (!(objp = kmem_getpages(cachep, flags, nodeid)))
2246 goto failed;
2247
2248 /* Get slab management. */
2249 if (!(slabp = alloc_slabmgmt(cachep, objp, offset, local_flags)))
2250 goto opps1;
2251
Christoph Lametere498be72005-09-09 13:03:32 -07002252 slabp->nodeid = nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002253 set_slab_attr(cachep, slabp, objp);
2254
2255 cache_init_objs(cachep, slabp, ctor_flags);
2256
2257 if (local_flags & __GFP_WAIT)
2258 local_irq_disable();
2259 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002260 l3 = cachep->nodelists[nodeid];
2261 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002262
2263 /* Make slab active. */
Christoph Lametere498be72005-09-09 13:03:32 -07002264 list_add_tail(&slabp->list, &(l3->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002265 STATS_INC_GROWN(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002266 l3->free_objects += cachep->num;
2267 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002268 return 1;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002269 opps1:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002270 kmem_freepages(cachep, objp);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002271 failed:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002272 if (local_flags & __GFP_WAIT)
2273 local_irq_disable();
2274 return 0;
2275}
2276
2277#if DEBUG
2278
2279/*
2280 * Perform extra freeing checks:
2281 * - detect bad pointers.
2282 * - POISON/RED_ZONE checking
2283 * - destructor calls, for caches with POISON+dtor
2284 */
2285static void kfree_debugcheck(const void *objp)
2286{
2287 struct page *page;
2288
2289 if (!virt_addr_valid(objp)) {
2290 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002291 (unsigned long)objp);
2292 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002293 }
2294 page = virt_to_page(objp);
2295 if (!PageSlab(page)) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002296 printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n",
2297 (unsigned long)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002298 BUG();
2299 }
2300}
2301
2302static void *cache_free_debugcheck(kmem_cache_t *cachep, void *objp,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002303 void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002304{
2305 struct page *page;
2306 unsigned int objnr;
2307 struct slab *slabp;
2308
2309 objp -= obj_dbghead(cachep);
2310 kfree_debugcheck(objp);
2311 page = virt_to_page(objp);
2312
Pekka Enberg065d41c2005-11-13 16:06:46 -08002313 if (page_get_cache(page) != cachep) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002314 printk(KERN_ERR
2315 "mismatch in kmem_cache_free: expected cache %p, got %p\n",
2316 page_get_cache(page), cachep);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002317 printk(KERN_ERR "%p is %s.\n", cachep, cachep->name);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002318 printk(KERN_ERR "%p is %s.\n", page_get_cache(page),
2319 page_get_cache(page)->name);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002320 WARN_ON(1);
2321 }
Pekka Enberg065d41c2005-11-13 16:06:46 -08002322 slabp = page_get_slab(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002323
2324 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002325 if (*dbg_redzone1(cachep, objp) != RED_ACTIVE
2326 || *dbg_redzone2(cachep, objp) != RED_ACTIVE) {
2327 slab_error(cachep,
2328 "double free, or memory outside"
2329 " object was overwritten");
2330 printk(KERN_ERR
2331 "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n",
2332 objp, *dbg_redzone1(cachep, objp),
2333 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002334 }
2335 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2336 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2337 }
2338 if (cachep->flags & SLAB_STORE_USER)
2339 *dbg_userword(cachep, objp) = caller;
2340
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002341 objnr = (objp - slabp->s_mem) / cachep->objsize;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002342
2343 BUG_ON(objnr >= cachep->num);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002344 BUG_ON(objp != slabp->s_mem + objnr * cachep->objsize);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002345
2346 if (cachep->flags & SLAB_DEBUG_INITIAL) {
2347 /* Need to call the slab's constructor so the
2348 * caller can perform a verify of its state (debugging).
2349 * Called without the cache-lock held.
2350 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002351 cachep->ctor(objp + obj_dbghead(cachep),
2352 cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002353 }
2354 if (cachep->flags & SLAB_POISON && cachep->dtor) {
2355 /* we want to cache poison the object,
2356 * call the destruction callback
2357 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002358 cachep->dtor(objp + obj_dbghead(cachep), cachep, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002359 }
2360 if (cachep->flags & SLAB_POISON) {
2361#ifdef CONFIG_DEBUG_PAGEALLOC
2362 if ((cachep->objsize % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) {
2363 store_stackinfo(cachep, objp, (unsigned long)caller);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002364 kernel_map_pages(virt_to_page(objp),
2365 cachep->objsize / PAGE_SIZE, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002366 } else {
2367 poison_obj(cachep, objp, POISON_FREE);
2368 }
2369#else
2370 poison_obj(cachep, objp, POISON_FREE);
2371#endif
2372 }
2373 return objp;
2374}
2375
2376static void check_slabp(kmem_cache_t *cachep, struct slab *slabp)
2377{
2378 kmem_bufctl_t i;
2379 int entries = 0;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002380
Linus Torvalds1da177e2005-04-16 15:20:36 -07002381 /* Check slab's freelist to see if this obj is there. */
2382 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2383 entries++;
2384 if (entries > cachep->num || i >= cachep->num)
2385 goto bad;
2386 }
2387 if (entries != cachep->num - slabp->inuse) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002388 bad:
2389 printk(KERN_ERR
2390 "slab: Internal list corruption detected in cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2391 cachep->name, cachep->num, slabp, slabp->inuse);
2392 for (i = 0;
2393 i < sizeof(slabp) + cachep->num * sizeof(kmem_bufctl_t);
2394 i++) {
2395 if ((i % 16) == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002396 printk("\n%03x:", i);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002397 printk(" %02x", ((unsigned char *)slabp)[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002398 }
2399 printk("\n");
2400 BUG();
2401 }
2402}
2403#else
2404#define kfree_debugcheck(x) do { } while(0)
2405#define cache_free_debugcheck(x,objp,z) (objp)
2406#define check_slabp(x,y) do { } while(0)
2407#endif
2408
Al Virodd0fc662005-10-07 07:46:04 +01002409static void *cache_alloc_refill(kmem_cache_t *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002410{
2411 int batchcount;
2412 struct kmem_list3 *l3;
2413 struct array_cache *ac;
2414
2415 check_irq_off();
2416 ac = ac_data(cachep);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002417 retry:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002418 batchcount = ac->batchcount;
2419 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2420 /* if there was little recent activity on this
2421 * cache, then perform only a partial refill.
2422 * Otherwise we could generate refill bouncing.
2423 */
2424 batchcount = BATCHREFILL_LIMIT;
2425 }
Christoph Lametere498be72005-09-09 13:03:32 -07002426 l3 = cachep->nodelists[numa_node_id()];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002427
Christoph Lametere498be72005-09-09 13:03:32 -07002428 BUG_ON(ac->avail > 0 || !l3);
2429 spin_lock(&l3->list_lock);
2430
Linus Torvalds1da177e2005-04-16 15:20:36 -07002431 if (l3->shared) {
2432 struct array_cache *shared_array = l3->shared;
2433 if (shared_array->avail) {
2434 if (batchcount > shared_array->avail)
2435 batchcount = shared_array->avail;
2436 shared_array->avail -= batchcount;
2437 ac->avail = batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07002438 memcpy(ac->entry,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002439 &(shared_array->entry[shared_array->avail]),
2440 sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002441 shared_array->touched = 1;
2442 goto alloc_done;
2443 }
2444 }
2445 while (batchcount > 0) {
2446 struct list_head *entry;
2447 struct slab *slabp;
2448 /* Get slab alloc is to come from. */
2449 entry = l3->slabs_partial.next;
2450 if (entry == &l3->slabs_partial) {
2451 l3->free_touched = 1;
2452 entry = l3->slabs_free.next;
2453 if (entry == &l3->slabs_free)
2454 goto must_grow;
2455 }
2456
2457 slabp = list_entry(entry, struct slab, list);
2458 check_slabp(cachep, slabp);
2459 check_spinlock_acquired(cachep);
2460 while (slabp->inuse < cachep->num && batchcount--) {
2461 kmem_bufctl_t next;
2462 STATS_INC_ALLOCED(cachep);
2463 STATS_INC_ACTIVE(cachep);
2464 STATS_SET_HIGH(cachep);
2465
2466 /* get obj pointer */
Christoph Lametere498be72005-09-09 13:03:32 -07002467 ac->entry[ac->avail++] = slabp->s_mem +
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002468 slabp->free * cachep->objsize;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002469
2470 slabp->inuse++;
2471 next = slab_bufctl(slabp)[slabp->free];
2472#if DEBUG
2473 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
Christoph Lameter09ad4bb2005-10-29 18:15:52 -07002474 WARN_ON(numa_node_id() != slabp->nodeid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002475#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002476 slabp->free = next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002477 }
2478 check_slabp(cachep, slabp);
2479
2480 /* move slabp to correct slabp list: */
2481 list_del(&slabp->list);
2482 if (slabp->free == BUFCTL_END)
2483 list_add(&slabp->list, &l3->slabs_full);
2484 else
2485 list_add(&slabp->list, &l3->slabs_partial);
2486 }
2487
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002488 must_grow:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002489 l3->free_objects -= ac->avail;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002490 alloc_done:
Christoph Lametere498be72005-09-09 13:03:32 -07002491 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002492
2493 if (unlikely(!ac->avail)) {
2494 int x;
Christoph Lametere498be72005-09-09 13:03:32 -07002495 x = cache_grow(cachep, flags, numa_node_id());
2496
Linus Torvalds1da177e2005-04-16 15:20:36 -07002497 // cache_grow can reenable interrupts, then ac could change.
2498 ac = ac_data(cachep);
2499 if (!x && ac->avail == 0) // no objects in sight? abort
2500 return NULL;
2501
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002502 if (!ac->avail) // objects refilled by interrupt?
Linus Torvalds1da177e2005-04-16 15:20:36 -07002503 goto retry;
2504 }
2505 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07002506 return ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002507}
2508
2509static inline void
Al Virodd0fc662005-10-07 07:46:04 +01002510cache_alloc_debugcheck_before(kmem_cache_t *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002511{
2512 might_sleep_if(flags & __GFP_WAIT);
2513#if DEBUG
2514 kmem_flagcheck(cachep, flags);
2515#endif
2516}
2517
2518#if DEBUG
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002519static void *cache_alloc_debugcheck_after(kmem_cache_t *cachep, gfp_t flags,
2520 void *objp, void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002521{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002522 if (!objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002523 return objp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002524 if (cachep->flags & SLAB_POISON) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002525#ifdef CONFIG_DEBUG_PAGEALLOC
2526 if ((cachep->objsize % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002527 kernel_map_pages(virt_to_page(objp),
2528 cachep->objsize / PAGE_SIZE, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002529 else
2530 check_poison_obj(cachep, objp);
2531#else
2532 check_poison_obj(cachep, objp);
2533#endif
2534 poison_obj(cachep, objp, POISON_INUSE);
2535 }
2536 if (cachep->flags & SLAB_STORE_USER)
2537 *dbg_userword(cachep, objp) = caller;
2538
2539 if (cachep->flags & SLAB_RED_ZONE) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002540 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE
2541 || *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2542 slab_error(cachep,
2543 "double free, or memory outside"
2544 " object was overwritten");
2545 printk(KERN_ERR
2546 "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n",
2547 objp, *dbg_redzone1(cachep, objp),
2548 *dbg_redzone2(cachep, objp));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002549 }
2550 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2551 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2552 }
2553 objp += obj_dbghead(cachep);
2554 if (cachep->ctor && cachep->flags & SLAB_POISON) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002555 unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002556
2557 if (!(flags & __GFP_WAIT))
2558 ctor_flags |= SLAB_CTOR_ATOMIC;
2559
2560 cachep->ctor(objp, cachep, ctor_flags);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002561 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002562 return objp;
2563}
2564#else
2565#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2566#endif
2567
Al Virodd0fc662005-10-07 07:46:04 +01002568static inline void *____cache_alloc(kmem_cache_t *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002569{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002570 void *objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002571 struct array_cache *ac;
2572
Alok N Kataria5c382302005-09-27 21:45:46 -07002573 check_irq_off();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002574 ac = ac_data(cachep);
2575 if (likely(ac->avail)) {
2576 STATS_INC_ALLOCHIT(cachep);
2577 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07002578 objp = ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002579 } else {
2580 STATS_INC_ALLOCMISS(cachep);
2581 objp = cache_alloc_refill(cachep, flags);
2582 }
Alok N Kataria5c382302005-09-27 21:45:46 -07002583 return objp;
2584}
2585
Al Virodd0fc662005-10-07 07:46:04 +01002586static inline void *__cache_alloc(kmem_cache_t *cachep, gfp_t flags)
Alok N Kataria5c382302005-09-27 21:45:46 -07002587{
2588 unsigned long save_flags;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002589 void *objp;
Alok N Kataria5c382302005-09-27 21:45:46 -07002590
2591 cache_alloc_debugcheck_before(cachep, flags);
2592
2593 local_irq_save(save_flags);
2594 objp = ____cache_alloc(cachep, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002595 local_irq_restore(save_flags);
Eric Dumazet34342e82005-09-03 15:55:06 -07002596 objp = cache_alloc_debugcheck_after(cachep, flags, objp,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002597 __builtin_return_address(0));
Eric Dumazet34342e82005-09-03 15:55:06 -07002598 prefetchw(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002599 return objp;
2600}
2601
Christoph Lametere498be72005-09-09 13:03:32 -07002602#ifdef CONFIG_NUMA
2603/*
2604 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002605 */
Al Viro6daa0e22005-10-21 03:18:50 -04002606static void *__cache_alloc_node(kmem_cache_t *cachep, gfp_t flags, int nodeid)
Christoph Lametere498be72005-09-09 13:03:32 -07002607{
2608 struct list_head *entry;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002609 struct slab *slabp;
2610 struct kmem_list3 *l3;
2611 void *obj;
2612 kmem_bufctl_t next;
2613 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002614
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002615 l3 = cachep->nodelists[nodeid];
2616 BUG_ON(!l3);
Christoph Lametere498be72005-09-09 13:03:32 -07002617
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002618 retry:
2619 spin_lock(&l3->list_lock);
2620 entry = l3->slabs_partial.next;
2621 if (entry == &l3->slabs_partial) {
2622 l3->free_touched = 1;
2623 entry = l3->slabs_free.next;
2624 if (entry == &l3->slabs_free)
2625 goto must_grow;
2626 }
Christoph Lametere498be72005-09-09 13:03:32 -07002627
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002628 slabp = list_entry(entry, struct slab, list);
2629 check_spinlock_acquired_node(cachep, nodeid);
2630 check_slabp(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07002631
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002632 STATS_INC_NODEALLOCS(cachep);
2633 STATS_INC_ACTIVE(cachep);
2634 STATS_SET_HIGH(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002635
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002636 BUG_ON(slabp->inuse == cachep->num);
Christoph Lametere498be72005-09-09 13:03:32 -07002637
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002638 /* get obj pointer */
2639 obj = slabp->s_mem + slabp->free * cachep->objsize;
2640 slabp->inuse++;
2641 next = slab_bufctl(slabp)[slabp->free];
Christoph Lametere498be72005-09-09 13:03:32 -07002642#if DEBUG
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002643 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
Christoph Lametere498be72005-09-09 13:03:32 -07002644#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002645 slabp->free = next;
2646 check_slabp(cachep, slabp);
2647 l3->free_objects--;
2648 /* move slabp to correct slabp list: */
2649 list_del(&slabp->list);
Christoph Lametere498be72005-09-09 13:03:32 -07002650
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002651 if (slabp->free == BUFCTL_END) {
2652 list_add(&slabp->list, &l3->slabs_full);
2653 } else {
2654 list_add(&slabp->list, &l3->slabs_partial);
2655 }
Christoph Lametere498be72005-09-09 13:03:32 -07002656
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002657 spin_unlock(&l3->list_lock);
2658 goto done;
Christoph Lametere498be72005-09-09 13:03:32 -07002659
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002660 must_grow:
2661 spin_unlock(&l3->list_lock);
2662 x = cache_grow(cachep, flags, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07002663
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002664 if (!x)
2665 return NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07002666
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002667 goto retry;
2668 done:
2669 return obj;
Christoph Lametere498be72005-09-09 13:03:32 -07002670}
2671#endif
2672
2673/*
2674 * Caller needs to acquire correct kmem_list's list_lock
2675 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002676static void free_block(kmem_cache_t *cachep, void **objpp, int nr_objects,
2677 int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002678{
2679 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07002680 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002681
2682 for (i = 0; i < nr_objects; i++) {
2683 void *objp = objpp[i];
2684 struct slab *slabp;
2685 unsigned int objnr;
2686
Pekka Enberg065d41c2005-11-13 16:06:46 -08002687 slabp = page_get_slab(virt_to_page(objp));
Christoph Lameterff694162005-09-22 21:44:02 -07002688 l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002689 list_del(&slabp->list);
2690 objnr = (objp - slabp->s_mem) / cachep->objsize;
Christoph Lameterff694162005-09-22 21:44:02 -07002691 check_spinlock_acquired_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002692 check_slabp(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07002693
Linus Torvalds1da177e2005-04-16 15:20:36 -07002694#if DEBUG
Christoph Lameter09ad4bb2005-10-29 18:15:52 -07002695 /* Verify that the slab belongs to the intended node */
2696 WARN_ON(slabp->nodeid != node);
2697
Linus Torvalds1da177e2005-04-16 15:20:36 -07002698 if (slab_bufctl(slabp)[objnr] != BUFCTL_FREE) {
Christoph Lametere498be72005-09-09 13:03:32 -07002699 printk(KERN_ERR "slab: double free detected in cache "
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002700 "'%s', objp %p\n", cachep->name, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002701 BUG();
2702 }
2703#endif
2704 slab_bufctl(slabp)[objnr] = slabp->free;
2705 slabp->free = objnr;
2706 STATS_DEC_ACTIVE(cachep);
2707 slabp->inuse--;
Christoph Lametere498be72005-09-09 13:03:32 -07002708 l3->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002709 check_slabp(cachep, slabp);
2710
2711 /* fixup slab chains */
2712 if (slabp->inuse == 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07002713 if (l3->free_objects > l3->free_limit) {
2714 l3->free_objects -= cachep->num;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002715 slab_destroy(cachep, slabp);
2716 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07002717 list_add(&slabp->list, &l3->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002718 }
2719 } else {
2720 /* Unconditionally move a slab to the end of the
2721 * partial list on free - maximum time for the
2722 * other objects to be freed, too.
2723 */
Christoph Lametere498be72005-09-09 13:03:32 -07002724 list_add_tail(&slabp->list, &l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002725 }
2726 }
2727}
2728
2729static void cache_flusharray(kmem_cache_t *cachep, struct array_cache *ac)
2730{
2731 int batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07002732 struct kmem_list3 *l3;
Christoph Lameterff694162005-09-22 21:44:02 -07002733 int node = numa_node_id();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002734
2735 batchcount = ac->batchcount;
2736#if DEBUG
2737 BUG_ON(!batchcount || batchcount > ac->avail);
2738#endif
2739 check_irq_off();
Christoph Lameterff694162005-09-22 21:44:02 -07002740 l3 = cachep->nodelists[node];
Christoph Lametere498be72005-09-09 13:03:32 -07002741 spin_lock(&l3->list_lock);
2742 if (l3->shared) {
2743 struct array_cache *shared_array = l3->shared;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002744 int max = shared_array->limit - shared_array->avail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002745 if (max) {
2746 if (batchcount > max)
2747 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07002748 memcpy(&(shared_array->entry[shared_array->avail]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002749 ac->entry, sizeof(void *) * batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002750 shared_array->avail += batchcount;
2751 goto free_done;
2752 }
2753 }
2754
Christoph Lameterff694162005-09-22 21:44:02 -07002755 free_block(cachep, ac->entry, batchcount, node);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002756 free_done:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002757#if STATS
2758 {
2759 int i = 0;
2760 struct list_head *p;
2761
Christoph Lametere498be72005-09-09 13:03:32 -07002762 p = l3->slabs_free.next;
2763 while (p != &(l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002764 struct slab *slabp;
2765
2766 slabp = list_entry(p, struct slab, list);
2767 BUG_ON(slabp->inuse);
2768
2769 i++;
2770 p = p->next;
2771 }
2772 STATS_SET_FREEABLE(cachep, i);
2773 }
2774#endif
Christoph Lametere498be72005-09-09 13:03:32 -07002775 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002776 ac->avail -= batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07002777 memmove(ac->entry, &(ac->entry[batchcount]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002778 sizeof(void *) * ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002779}
2780
2781/*
2782 * __cache_free
2783 * Release an obj back to its cache. If the obj has a constructed
2784 * state, it must be in this state _before_ it is released.
2785 *
2786 * Called with disabled ints.
2787 */
2788static inline void __cache_free(kmem_cache_t *cachep, void *objp)
2789{
2790 struct array_cache *ac = ac_data(cachep);
2791
2792 check_irq_off();
2793 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
2794
Christoph Lametere498be72005-09-09 13:03:32 -07002795 /* Make sure we are not freeing a object from another
2796 * node to the array cache on this cpu.
2797 */
2798#ifdef CONFIG_NUMA
2799 {
2800 struct slab *slabp;
Pekka Enberg065d41c2005-11-13 16:06:46 -08002801 slabp = page_get_slab(virt_to_page(objp));
Christoph Lametere498be72005-09-09 13:03:32 -07002802 if (unlikely(slabp->nodeid != numa_node_id())) {
2803 struct array_cache *alien = NULL;
2804 int nodeid = slabp->nodeid;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002805 struct kmem_list3 *l3 =
2806 cachep->nodelists[numa_node_id()];
Christoph Lametere498be72005-09-09 13:03:32 -07002807
2808 STATS_INC_NODEFREES(cachep);
2809 if (l3->alien && l3->alien[nodeid]) {
2810 alien = l3->alien[nodeid];
2811 spin_lock(&alien->lock);
2812 if (unlikely(alien->avail == alien->limit))
2813 __drain_alien_cache(cachep,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002814 alien, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07002815 alien->entry[alien->avail++] = objp;
2816 spin_unlock(&alien->lock);
2817 } else {
2818 spin_lock(&(cachep->nodelists[nodeid])->
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002819 list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -07002820 free_block(cachep, &objp, 1, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07002821 spin_unlock(&(cachep->nodelists[nodeid])->
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002822 list_lock);
Christoph Lametere498be72005-09-09 13:03:32 -07002823 }
2824 return;
2825 }
2826 }
2827#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002828 if (likely(ac->avail < ac->limit)) {
2829 STATS_INC_FREEHIT(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002830 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002831 return;
2832 } else {
2833 STATS_INC_FREEMISS(cachep);
2834 cache_flusharray(cachep, ac);
Christoph Lametere498be72005-09-09 13:03:32 -07002835 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002836 }
2837}
2838
2839/**
2840 * kmem_cache_alloc - Allocate an object
2841 * @cachep: The cache to allocate from.
2842 * @flags: See kmalloc().
2843 *
2844 * Allocate an object from this cache. The flags are only relevant
2845 * if the cache has no available objects.
2846 */
Al Virodd0fc662005-10-07 07:46:04 +01002847void *kmem_cache_alloc(kmem_cache_t *cachep, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002848{
2849 return __cache_alloc(cachep, flags);
2850}
2851EXPORT_SYMBOL(kmem_cache_alloc);
2852
2853/**
2854 * kmem_ptr_validate - check if an untrusted pointer might
2855 * be a slab entry.
2856 * @cachep: the cache we're checking against
2857 * @ptr: pointer to validate
2858 *
2859 * This verifies that the untrusted pointer looks sane:
2860 * it is _not_ a guarantee that the pointer is actually
2861 * part of the slab cache in question, but it at least
2862 * validates that the pointer can be dereferenced and
2863 * looks half-way sane.
2864 *
2865 * Currently only used for dentry validation.
2866 */
2867int fastcall kmem_ptr_validate(kmem_cache_t *cachep, void *ptr)
2868{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002869 unsigned long addr = (unsigned long)ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002870 unsigned long min_addr = PAGE_OFFSET;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002871 unsigned long align_mask = BYTES_PER_WORD - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002872 unsigned long size = cachep->objsize;
2873 struct page *page;
2874
2875 if (unlikely(addr < min_addr))
2876 goto out;
2877 if (unlikely(addr > (unsigned long)high_memory - size))
2878 goto out;
2879 if (unlikely(addr & align_mask))
2880 goto out;
2881 if (unlikely(!kern_addr_valid(addr)))
2882 goto out;
2883 if (unlikely(!kern_addr_valid(addr + size - 1)))
2884 goto out;
2885 page = virt_to_page(ptr);
2886 if (unlikely(!PageSlab(page)))
2887 goto out;
Pekka Enberg065d41c2005-11-13 16:06:46 -08002888 if (unlikely(page_get_cache(page) != cachep))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002889 goto out;
2890 return 1;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002891 out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002892 return 0;
2893}
2894
2895#ifdef CONFIG_NUMA
2896/**
2897 * kmem_cache_alloc_node - Allocate an object on the specified node
2898 * @cachep: The cache to allocate from.
2899 * @flags: See kmalloc().
2900 * @nodeid: node number of the target node.
2901 *
2902 * Identical to kmem_cache_alloc, except that this function is slow
2903 * and can sleep. And it will allocate memory on the given node, which
2904 * can improve the performance for cpu bound structures.
Christoph Lametere498be72005-09-09 13:03:32 -07002905 * New and improved: it will now make sure that the object gets
2906 * put on the correct node list so that there is no false sharing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002907 */
Al Virodd0fc662005-10-07 07:46:04 +01002908void *kmem_cache_alloc_node(kmem_cache_t *cachep, gfp_t flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002909{
Christoph Lametere498be72005-09-09 13:03:32 -07002910 unsigned long save_flags;
2911 void *ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002912
Christoph Lameterff694162005-09-22 21:44:02 -07002913 if (nodeid == -1)
Christoph Lametere498be72005-09-09 13:03:32 -07002914 return __cache_alloc(cachep, flags);
Christoph Lameter83b78bd2005-07-06 10:47:07 -07002915
Christoph Lametere498be72005-09-09 13:03:32 -07002916 if (unlikely(!cachep->nodelists[nodeid])) {
2917 /* Fall back to __cache_alloc if we run into trouble */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002918 printk(KERN_WARNING
2919 "slab: not allocating in inactive node %d for cache %s\n",
2920 nodeid, cachep->name);
2921 return __cache_alloc(cachep, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002922 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002923
Christoph Lametere498be72005-09-09 13:03:32 -07002924 cache_alloc_debugcheck_before(cachep, flags);
2925 local_irq_save(save_flags);
Alok N Kataria5c382302005-09-27 21:45:46 -07002926 if (nodeid == numa_node_id())
2927 ptr = ____cache_alloc(cachep, flags);
2928 else
2929 ptr = __cache_alloc_node(cachep, flags, nodeid);
Christoph Lametere498be72005-09-09 13:03:32 -07002930 local_irq_restore(save_flags);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002931 ptr =
2932 cache_alloc_debugcheck_after(cachep, flags, ptr,
2933 __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002934
Christoph Lametere498be72005-09-09 13:03:32 -07002935 return ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002936}
2937EXPORT_SYMBOL(kmem_cache_alloc_node);
2938
Al Virodd0fc662005-10-07 07:46:04 +01002939void *kmalloc_node(size_t size, gfp_t flags, int node)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07002940{
2941 kmem_cache_t *cachep;
2942
2943 cachep = kmem_find_general_cachep(size, flags);
2944 if (unlikely(cachep == NULL))
2945 return NULL;
2946 return kmem_cache_alloc_node(cachep, flags, node);
2947}
2948EXPORT_SYMBOL(kmalloc_node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002949#endif
2950
2951/**
2952 * kmalloc - allocate memory
2953 * @size: how many bytes of memory are required.
2954 * @flags: the type of memory to allocate.
2955 *
2956 * kmalloc is the normal method of allocating memory
2957 * in the kernel.
2958 *
2959 * The @flags argument may be one of:
2960 *
2961 * %GFP_USER - Allocate memory on behalf of user. May sleep.
2962 *
2963 * %GFP_KERNEL - Allocate normal kernel ram. May sleep.
2964 *
2965 * %GFP_ATOMIC - Allocation will not sleep. Use inside interrupt handlers.
2966 *
2967 * Additionally, the %GFP_DMA flag may be set to indicate the memory
2968 * must be suitable for DMA. This can mean different things on different
2969 * platforms. For example, on i386, it means that the memory must come
2970 * from the first 16MB.
2971 */
Al Virodd0fc662005-10-07 07:46:04 +01002972void *__kmalloc(size_t size, gfp_t flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002973{
2974 kmem_cache_t *cachep;
2975
Manfred Spraul97e2bde2005-05-01 08:58:38 -07002976 /* If you want to save a few bytes .text space: replace
2977 * __ with kmem_.
2978 * Then kmalloc uses the uninlined functions instead of the inline
2979 * functions.
2980 */
2981 cachep = __find_general_cachep(size, flags);
Andrew Mortondbdb9042005-09-23 13:24:10 -07002982 if (unlikely(cachep == NULL))
2983 return NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002984 return __cache_alloc(cachep, flags);
2985}
2986EXPORT_SYMBOL(__kmalloc);
2987
2988#ifdef CONFIG_SMP
2989/**
2990 * __alloc_percpu - allocate one copy of the object for every present
2991 * cpu in the system, zeroing them.
2992 * Objects should be dereferenced using the per_cpu_ptr macro only.
2993 *
2994 * @size: how many bytes of memory are required.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002995 */
Pekka Enbergf9f75002006-01-08 01:00:33 -08002996void *__alloc_percpu(size_t size)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002997{
2998 int i;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08002999 struct percpu_data *pdata = kmalloc(sizeof(*pdata), GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003000
3001 if (!pdata)
3002 return NULL;
3003
Christoph Lametere498be72005-09-09 13:03:32 -07003004 /*
3005 * Cannot use for_each_online_cpu since a cpu may come online
3006 * and we have no way of figuring out how to fix the array
3007 * that we have allocated then....
3008 */
3009 for_each_cpu(i) {
3010 int node = cpu_to_node(i);
3011
3012 if (node_online(node))
3013 pdata->ptrs[i] = kmalloc_node(size, GFP_KERNEL, node);
3014 else
3015 pdata->ptrs[i] = kmalloc(size, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003016
3017 if (!pdata->ptrs[i])
3018 goto unwind_oom;
3019 memset(pdata->ptrs[i], 0, size);
3020 }
3021
3022 /* Catch derefs w/o wrappers */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003023 return (void *)(~(unsigned long)pdata);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003024
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003025 unwind_oom:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003026 while (--i >= 0) {
3027 if (!cpu_possible(i))
3028 continue;
3029 kfree(pdata->ptrs[i]);
3030 }
3031 kfree(pdata);
3032 return NULL;
3033}
3034EXPORT_SYMBOL(__alloc_percpu);
3035#endif
3036
3037/**
3038 * kmem_cache_free - Deallocate an object
3039 * @cachep: The cache the allocation was from.
3040 * @objp: The previously allocated object.
3041 *
3042 * Free an object which was previously allocated from this
3043 * cache.
3044 */
3045void kmem_cache_free(kmem_cache_t *cachep, void *objp)
3046{
3047 unsigned long flags;
3048
3049 local_irq_save(flags);
3050 __cache_free(cachep, objp);
3051 local_irq_restore(flags);
3052}
3053EXPORT_SYMBOL(kmem_cache_free);
3054
3055/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07003056 * kfree - free previously allocated memory
3057 * @objp: pointer returned by kmalloc.
3058 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003059 * If @objp is NULL, no operation is performed.
3060 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003061 * Don't free memory not originally allocated by kmalloc()
3062 * or you will run into trouble.
3063 */
3064void kfree(const void *objp)
3065{
3066 kmem_cache_t *c;
3067 unsigned long flags;
3068
3069 if (unlikely(!objp))
3070 return;
3071 local_irq_save(flags);
3072 kfree_debugcheck(objp);
Pekka Enberg065d41c2005-11-13 16:06:46 -08003073 c = page_get_cache(virt_to_page(objp));
David Woodhousea4fc7ab2006-01-11 14:41:26 +00003074 mutex_debug_check_no_locks_freed(objp, obj_reallen(c));
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003075 __cache_free(c, (void *)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003076 local_irq_restore(flags);
3077}
3078EXPORT_SYMBOL(kfree);
3079
3080#ifdef CONFIG_SMP
3081/**
3082 * free_percpu - free previously allocated percpu memory
3083 * @objp: pointer returned by alloc_percpu.
3084 *
3085 * Don't free memory not originally allocated by alloc_percpu()
3086 * The complemented objp is to check for that.
3087 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003088void free_percpu(const void *objp)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003089{
3090 int i;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003091 struct percpu_data *p = (struct percpu_data *)(~(unsigned long)objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003092
Christoph Lametere498be72005-09-09 13:03:32 -07003093 /*
3094 * We allocate for all cpus so we cannot use for online cpu here.
3095 */
3096 for_each_cpu(i)
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003097 kfree(p->ptrs[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003098 kfree(p);
3099}
3100EXPORT_SYMBOL(free_percpu);
3101#endif
3102
3103unsigned int kmem_cache_size(kmem_cache_t *cachep)
3104{
3105 return obj_reallen(cachep);
3106}
3107EXPORT_SYMBOL(kmem_cache_size);
3108
Arnaldo Carvalho de Melo19449722005-06-18 22:46:19 -07003109const char *kmem_cache_name(kmem_cache_t *cachep)
3110{
3111 return cachep->name;
3112}
3113EXPORT_SYMBOL_GPL(kmem_cache_name);
3114
Christoph Lametere498be72005-09-09 13:03:32 -07003115/*
3116 * This initializes kmem_list3 for all nodes.
3117 */
3118static int alloc_kmemlist(kmem_cache_t *cachep)
3119{
3120 int node;
3121 struct kmem_list3 *l3;
3122 int err = 0;
3123
3124 for_each_online_node(node) {
3125 struct array_cache *nc = NULL, *new;
3126 struct array_cache **new_alien = NULL;
3127#ifdef CONFIG_NUMA
3128 if (!(new_alien = alloc_alien_cache(node, cachep->limit)))
3129 goto fail;
3130#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003131 if (!(new = alloc_arraycache(node, (cachep->shared *
3132 cachep->batchcount),
3133 0xbaadf00d)))
Christoph Lametere498be72005-09-09 13:03:32 -07003134 goto fail;
3135 if ((l3 = cachep->nodelists[node])) {
3136
3137 spin_lock_irq(&l3->list_lock);
3138
3139 if ((nc = cachep->nodelists[node]->shared))
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003140 free_block(cachep, nc->entry, nc->avail, node);
Christoph Lametere498be72005-09-09 13:03:32 -07003141
3142 l3->shared = new;
3143 if (!cachep->nodelists[node]->alien) {
3144 l3->alien = new_alien;
3145 new_alien = NULL;
3146 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003147 l3->free_limit = (1 + nr_cpus_node(node)) *
3148 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003149 spin_unlock_irq(&l3->list_lock);
3150 kfree(nc);
3151 free_alien_cache(new_alien);
3152 continue;
3153 }
3154 if (!(l3 = kmalloc_node(sizeof(struct kmem_list3),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003155 GFP_KERNEL, node)))
Christoph Lametere498be72005-09-09 13:03:32 -07003156 goto fail;
3157
3158 kmem_list3_init(l3);
3159 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003160 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
Christoph Lametere498be72005-09-09 13:03:32 -07003161 l3->shared = new;
3162 l3->alien = new_alien;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003163 l3->free_limit = (1 + nr_cpus_node(node)) *
3164 cachep->batchcount + cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003165 cachep->nodelists[node] = l3;
3166 }
3167 return err;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003168 fail:
Christoph Lametere498be72005-09-09 13:03:32 -07003169 err = -ENOMEM;
3170 return err;
3171}
3172
Linus Torvalds1da177e2005-04-16 15:20:36 -07003173struct ccupdate_struct {
3174 kmem_cache_t *cachep;
3175 struct array_cache *new[NR_CPUS];
3176};
3177
3178static void do_ccupdate_local(void *info)
3179{
3180 struct ccupdate_struct *new = (struct ccupdate_struct *)info;
3181 struct array_cache *old;
3182
3183 check_irq_off();
3184 old = ac_data(new->cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003185
Linus Torvalds1da177e2005-04-16 15:20:36 -07003186 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3187 new->new[smp_processor_id()] = old;
3188}
3189
Linus Torvalds1da177e2005-04-16 15:20:36 -07003190static int do_tune_cpucache(kmem_cache_t *cachep, int limit, int batchcount,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003191 int shared)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003192{
3193 struct ccupdate_struct new;
Christoph Lametere498be72005-09-09 13:03:32 -07003194 int i, err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003195
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003196 memset(&new.new, 0, sizeof(new.new));
Christoph Lametere498be72005-09-09 13:03:32 -07003197 for_each_online_cpu(i) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003198 new.new[i] =
3199 alloc_arraycache(cpu_to_node(i), limit, batchcount);
Christoph Lametere498be72005-09-09 13:03:32 -07003200 if (!new.new[i]) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003201 for (i--; i >= 0; i--)
3202 kfree(new.new[i]);
Christoph Lametere498be72005-09-09 13:03:32 -07003203 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003204 }
3205 }
3206 new.cachep = cachep;
3207
3208 smp_call_function_all_cpus(do_ccupdate_local, (void *)&new);
Christoph Lametere498be72005-09-09 13:03:32 -07003209
Linus Torvalds1da177e2005-04-16 15:20:36 -07003210 check_irq_on();
3211 spin_lock_irq(&cachep->spinlock);
3212 cachep->batchcount = batchcount;
3213 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07003214 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003215 spin_unlock_irq(&cachep->spinlock);
3216
Christoph Lametere498be72005-09-09 13:03:32 -07003217 for_each_online_cpu(i) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003218 struct array_cache *ccold = new.new[i];
3219 if (!ccold)
3220 continue;
Christoph Lametere498be72005-09-09 13:03:32 -07003221 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Christoph Lameterff694162005-09-22 21:44:02 -07003222 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
Christoph Lametere498be72005-09-09 13:03:32 -07003223 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003224 kfree(ccold);
3225 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003226
Christoph Lametere498be72005-09-09 13:03:32 -07003227 err = alloc_kmemlist(cachep);
3228 if (err) {
3229 printk(KERN_ERR "alloc_kmemlist failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003230 cachep->name, -err);
Christoph Lametere498be72005-09-09 13:03:32 -07003231 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003232 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003233 return 0;
3234}
3235
Linus Torvalds1da177e2005-04-16 15:20:36 -07003236static void enable_cpucache(kmem_cache_t *cachep)
3237{
3238 int err;
3239 int limit, shared;
3240
3241 /* The head array serves three purposes:
3242 * - create a LIFO ordering, i.e. return objects that are cache-warm
3243 * - reduce the number of spinlock operations.
3244 * - reduce the number of linked list operations on the slab and
3245 * bufctl chains: array operations are cheaper.
3246 * The numbers are guessed, we should auto-tune as described by
3247 * Bonwick.
3248 */
3249 if (cachep->objsize > 131072)
3250 limit = 1;
3251 else if (cachep->objsize > PAGE_SIZE)
3252 limit = 8;
3253 else if (cachep->objsize > 1024)
3254 limit = 24;
3255 else if (cachep->objsize > 256)
3256 limit = 54;
3257 else
3258 limit = 120;
3259
3260 /* Cpu bound tasks (e.g. network routing) can exhibit cpu bound
3261 * allocation behaviour: Most allocs on one cpu, most free operations
3262 * on another cpu. For these cases, an efficient object passing between
3263 * cpus is necessary. This is provided by a shared array. The array
3264 * replaces Bonwick's magazine layer.
3265 * On uniprocessor, it's functionally equivalent (but less efficient)
3266 * to a larger limit. Thus disabled by default.
3267 */
3268 shared = 0;
3269#ifdef CONFIG_SMP
3270 if (cachep->objsize <= PAGE_SIZE)
3271 shared = 8;
3272#endif
3273
3274#if DEBUG
3275 /* With debugging enabled, large batchcount lead to excessively
3276 * long periods with disabled local interrupts. Limit the
3277 * batchcount
3278 */
3279 if (limit > 32)
3280 limit = 32;
3281#endif
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003282 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003283 if (err)
3284 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003285 cachep->name, -err);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003286}
3287
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003288static void drain_array_locked(kmem_cache_t *cachep, struct array_cache *ac,
3289 int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003290{
3291 int tofree;
3292
Christoph Lametere498be72005-09-09 13:03:32 -07003293 check_spinlock_acquired_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003294 if (ac->touched && !force) {
3295 ac->touched = 0;
3296 } else if (ac->avail) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003297 tofree = force ? ac->avail : (ac->limit + 4) / 5;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003298 if (tofree > ac->avail) {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003299 tofree = (ac->avail + 1) / 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003300 }
Christoph Lameterff694162005-09-22 21:44:02 -07003301 free_block(cachep, ac->entry, tofree, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003302 ac->avail -= tofree;
Christoph Lametere498be72005-09-09 13:03:32 -07003303 memmove(ac->entry, &(ac->entry[tofree]),
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003304 sizeof(void *) * ac->avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003305 }
3306}
3307
3308/**
3309 * cache_reap - Reclaim memory from caches.
Randy Dunlap1e5d5332005-11-07 01:01:06 -08003310 * @unused: unused parameter
Linus Torvalds1da177e2005-04-16 15:20:36 -07003311 *
3312 * Called from workqueue/eventd every few seconds.
3313 * Purpose:
3314 * - clear the per-cpu caches for this CPU.
3315 * - return freeable pages to the main free memory pool.
3316 *
3317 * If we cannot acquire the cache chain semaphore then just give up - we'll
3318 * try again on the next iteration.
3319 */
3320static void cache_reap(void *unused)
3321{
3322 struct list_head *walk;
Christoph Lametere498be72005-09-09 13:03:32 -07003323 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003324
3325 if (down_trylock(&cache_chain_sem)) {
3326 /* Give up. Setup the next iteration. */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003327 schedule_delayed_work(&__get_cpu_var(reap_work),
3328 REAPTIMEOUT_CPUC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003329 return;
3330 }
3331
3332 list_for_each(walk, &cache_chain) {
3333 kmem_cache_t *searchp;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003334 struct list_head *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003335 int tofree;
3336 struct slab *slabp;
3337
3338 searchp = list_entry(walk, kmem_cache_t, next);
3339
3340 if (searchp->flags & SLAB_NO_REAP)
3341 goto next;
3342
3343 check_irq_on();
3344
Christoph Lametere498be72005-09-09 13:03:32 -07003345 l3 = searchp->nodelists[numa_node_id()];
3346 if (l3->alien)
3347 drain_alien_cache(searchp, l3);
3348 spin_lock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003349
Christoph Lametere498be72005-09-09 13:03:32 -07003350 drain_array_locked(searchp, ac_data(searchp), 0,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003351 numa_node_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -07003352
Christoph Lametere498be72005-09-09 13:03:32 -07003353 if (time_after(l3->next_reap, jiffies))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003354 goto next_unlock;
3355
Christoph Lametere498be72005-09-09 13:03:32 -07003356 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003357
Christoph Lametere498be72005-09-09 13:03:32 -07003358 if (l3->shared)
3359 drain_array_locked(searchp, l3->shared, 0,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003360 numa_node_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -07003361
Christoph Lametere498be72005-09-09 13:03:32 -07003362 if (l3->free_touched) {
3363 l3->free_touched = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003364 goto next_unlock;
3365 }
3366
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003367 tofree =
3368 (l3->free_limit + 5 * searchp->num -
3369 1) / (5 * searchp->num);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003370 do {
Christoph Lametere498be72005-09-09 13:03:32 -07003371 p = l3->slabs_free.next;
3372 if (p == &(l3->slabs_free))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003373 break;
3374
3375 slabp = list_entry(p, struct slab, list);
3376 BUG_ON(slabp->inuse);
3377 list_del(&slabp->list);
3378 STATS_INC_REAPED(searchp);
3379
3380 /* Safe to drop the lock. The slab is no longer
3381 * linked to the cache.
3382 * searchp cannot disappear, we hold
3383 * cache_chain_lock
3384 */
Christoph Lametere498be72005-09-09 13:03:32 -07003385 l3->free_objects -= searchp->num;
3386 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003387 slab_destroy(searchp, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07003388 spin_lock_irq(&l3->list_lock);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003389 } while (--tofree > 0);
3390 next_unlock:
Christoph Lametere498be72005-09-09 13:03:32 -07003391 spin_unlock_irq(&l3->list_lock);
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003392 next:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003393 cond_resched();
3394 }
3395 check_irq_on();
3396 up(&cache_chain_sem);
Christoph Lameter4ae7c032005-06-21 17:14:57 -07003397 drain_remote_pages();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003398 /* Setup the next iteration */
Manfred Spraulcd61ef62005-11-07 00:58:02 -08003399 schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003400}
3401
3402#ifdef CONFIG_PROC_FS
3403
Pekka Enberg85289f92006-01-08 01:00:36 -08003404static void print_slabinfo_header(struct seq_file *m)
3405{
3406 /*
3407 * Output format version, so at least we can change it
3408 * without _too_ many complaints.
3409 */
3410#if STATS
3411 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
3412#else
3413 seq_puts(m, "slabinfo - version: 2.1\n");
3414#endif
3415 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
3416 "<objperslab> <pagesperslab>");
3417 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
3418 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
3419#if STATS
3420 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
3421 "<error> <maxfreeable> <nodeallocs> <remotefrees>");
3422 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
3423#endif
3424 seq_putc(m, '\n');
3425}
3426
Linus Torvalds1da177e2005-04-16 15:20:36 -07003427static void *s_start(struct seq_file *m, loff_t *pos)
3428{
3429 loff_t n = *pos;
3430 struct list_head *p;
3431
3432 down(&cache_chain_sem);
Pekka Enberg85289f92006-01-08 01:00:36 -08003433 if (!n)
3434 print_slabinfo_header(m);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003435 p = cache_chain.next;
3436 while (n--) {
3437 p = p->next;
3438 if (p == &cache_chain)
3439 return NULL;
3440 }
3441 return list_entry(p, kmem_cache_t, next);
3442}
3443
3444static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3445{
3446 kmem_cache_t *cachep = p;
3447 ++*pos;
3448 return cachep->next.next == &cache_chain ? NULL
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003449 : list_entry(cachep->next.next, kmem_cache_t, next);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003450}
3451
3452static void s_stop(struct seq_file *m, void *p)
3453{
3454 up(&cache_chain_sem);
3455}
3456
3457static int s_show(struct seq_file *m, void *p)
3458{
3459 kmem_cache_t *cachep = p;
3460 struct list_head *q;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003461 struct slab *slabp;
3462 unsigned long active_objs;
3463 unsigned long num_objs;
3464 unsigned long active_slabs = 0;
3465 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07003466 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003467 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003468 int node;
3469 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003470
3471 check_irq_on();
3472 spin_lock_irq(&cachep->spinlock);
3473 active_objs = 0;
3474 num_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07003475 for_each_online_node(node) {
3476 l3 = cachep->nodelists[node];
3477 if (!l3)
3478 continue;
3479
3480 spin_lock(&l3->list_lock);
3481
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003482 list_for_each(q, &l3->slabs_full) {
Christoph Lametere498be72005-09-09 13:03:32 -07003483 slabp = list_entry(q, struct slab, list);
3484 if (slabp->inuse != cachep->num && !error)
3485 error = "slabs_full accounting error";
3486 active_objs += cachep->num;
3487 active_slabs++;
3488 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003489 list_for_each(q, &l3->slabs_partial) {
Christoph Lametere498be72005-09-09 13:03:32 -07003490 slabp = list_entry(q, struct slab, list);
3491 if (slabp->inuse == cachep->num && !error)
3492 error = "slabs_partial inuse accounting error";
3493 if (!slabp->inuse && !error)
3494 error = "slabs_partial/inuse accounting error";
3495 active_objs += slabp->inuse;
3496 active_slabs++;
3497 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003498 list_for_each(q, &l3->slabs_free) {
Christoph Lametere498be72005-09-09 13:03:32 -07003499 slabp = list_entry(q, struct slab, list);
3500 if (slabp->inuse && !error)
3501 error = "slabs_free/inuse accounting error";
3502 num_slabs++;
3503 }
3504 free_objects += l3->free_objects;
3505 shared_avail += l3->shared->avail;
3506
3507 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003508 }
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003509 num_slabs += active_slabs;
3510 num_objs = num_slabs * cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003511 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003512 error = "free_objects accounting error";
3513
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003514 name = cachep->name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003515 if (error)
3516 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
3517
3518 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003519 name, active_objs, num_objs, cachep->objsize,
3520 cachep->num, (1 << cachep->gfporder));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003521 seq_printf(m, " : tunables %4u %4u %4u",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003522 cachep->limit, cachep->batchcount, cachep->shared);
Christoph Lametere498be72005-09-09 13:03:32 -07003523 seq_printf(m, " : slabdata %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003524 active_slabs, num_slabs, shared_avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003525#if STATS
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003526 { /* list3 stats */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003527 unsigned long high = cachep->high_mark;
3528 unsigned long allocs = cachep->num_allocations;
3529 unsigned long grown = cachep->grown;
3530 unsigned long reaped = cachep->reaped;
3531 unsigned long errors = cachep->errors;
3532 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003533 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07003534 unsigned long node_frees = cachep->node_frees;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003535
Christoph Lametere498be72005-09-09 13:03:32 -07003536 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003537 %4lu %4lu %4lu %4lu", allocs, high, grown, reaped, errors, max_freeable, node_allocs, node_frees);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003538 }
3539 /* cpu stats */
3540 {
3541 unsigned long allochit = atomic_read(&cachep->allochit);
3542 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
3543 unsigned long freehit = atomic_read(&cachep->freehit);
3544 unsigned long freemiss = atomic_read(&cachep->freemiss);
3545
3546 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003547 allochit, allocmiss, freehit, freemiss);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003548 }
3549#endif
3550 seq_putc(m, '\n');
3551 spin_unlock_irq(&cachep->spinlock);
3552 return 0;
3553}
3554
3555/*
3556 * slabinfo_op - iterator that generates /proc/slabinfo
3557 *
3558 * Output layout:
3559 * cache-name
3560 * num-active-objs
3561 * total-objs
3562 * object size
3563 * num-active-slabs
3564 * total-slabs
3565 * num-pages-per-slab
3566 * + further values on SMP and with statistics enabled
3567 */
3568
3569struct seq_operations slabinfo_op = {
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003570 .start = s_start,
3571 .next = s_next,
3572 .stop = s_stop,
3573 .show = s_show,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003574};
3575
3576#define MAX_SLABINFO_WRITE 128
3577/**
3578 * slabinfo_write - Tuning for the slab allocator
3579 * @file: unused
3580 * @buffer: user buffer
3581 * @count: data length
3582 * @ppos: unused
3583 */
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003584ssize_t slabinfo_write(struct file *file, const char __user * buffer,
3585 size_t count, loff_t *ppos)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003586{
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003587 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003588 int limit, batchcount, shared, res;
3589 struct list_head *p;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003590
Linus Torvalds1da177e2005-04-16 15:20:36 -07003591 if (count > MAX_SLABINFO_WRITE)
3592 return -EINVAL;
3593 if (copy_from_user(&kbuf, buffer, count))
3594 return -EFAULT;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003595 kbuf[MAX_SLABINFO_WRITE] = '\0';
Linus Torvalds1da177e2005-04-16 15:20:36 -07003596
3597 tmp = strchr(kbuf, ' ');
3598 if (!tmp)
3599 return -EINVAL;
3600 *tmp = '\0';
3601 tmp++;
3602 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
3603 return -EINVAL;
3604
3605 /* Find the cache in the chain of caches. */
3606 down(&cache_chain_sem);
3607 res = -EINVAL;
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003608 list_for_each(p, &cache_chain) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003609 kmem_cache_t *cachep = list_entry(p, kmem_cache_t, next);
3610
3611 if (!strcmp(cachep->name, kbuf)) {
3612 if (limit < 1 ||
3613 batchcount < 1 ||
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003614 batchcount > limit || shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07003615 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003616 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07003617 res = do_tune_cpucache(cachep, limit,
Pekka Enbergb28a02d2006-01-08 01:00:37 -08003618 batchcount, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003619 }
3620 break;
3621 }
3622 }
3623 up(&cache_chain_sem);
3624 if (res >= 0)
3625 res = count;
3626 return res;
3627}
3628#endif
3629
Manfred Spraul00e145b2005-09-03 15:55:07 -07003630/**
3631 * ksize - get the actual amount of memory allocated for a given object
3632 * @objp: Pointer to the object
3633 *
3634 * kmalloc may internally round up allocations and return more memory
3635 * than requested. ksize() can be used to determine the actual amount of
3636 * memory allocated. The caller may use this additional memory, even though
3637 * a smaller amount of memory was initially specified with the kmalloc call.
3638 * The caller must guarantee that objp points to a valid object previously
3639 * allocated with either kmalloc() or kmem_cache_alloc(). The object
3640 * must not be freed during the duration of the call.
3641 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003642unsigned int ksize(const void *objp)
3643{
Manfred Spraul00e145b2005-09-03 15:55:07 -07003644 if (unlikely(objp == NULL))
3645 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003646
Pekka Enberg065d41c2005-11-13 16:06:46 -08003647 return obj_reallen(page_get_cache(virt_to_page(objp)));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003648}