blob: 9e876d6dfad97f2c4ab60a881edcb91d528c1c3c [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * linux/mm/slab.c
3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
5 *
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
7 *
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
10 *
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
13 *
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
21 *
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
27 *
28 * This means, that your constructor is used only for newly allocated
29 * slabs and you must pass objects with the same intializations to
30 * kmem_cache_free.
31 *
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
35 *
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
38 * partial slabs
39 * empty slabs with no allocated objects
40 *
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
43 *
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
46 *
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
52 *
53 * The c_cpuarray may not be read with enabled local interrupts -
54 * it's changed with a smp_call_function().
55 *
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
58 * Several members in kmem_cache_t and struct slab never change, they
59 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
63 *
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
66 * his patch.
67 *
68 * Further notes from the original documentation:
69 *
70 * 11 April '97. Started multi-threading - markhe
71 * The global cache-chain is protected by the semaphore 'cache_chain_sem'.
72 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
75 *
76 * At present, each engine can be growing a cache. This should be blocked.
77 *
Christoph Lametere498be72005-09-09 13:03:32 -070078 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
83 *
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
Linus Torvalds1da177e2005-04-16 15:20:36 -070087 */
88
89#include <linux/config.h>
90#include <linux/slab.h>
91#include <linux/mm.h>
92#include <linux/swap.h>
93#include <linux/cache.h>
94#include <linux/interrupt.h>
95#include <linux/init.h>
96#include <linux/compiler.h>
97#include <linux/seq_file.h>
98#include <linux/notifier.h>
99#include <linux/kallsyms.h>
100#include <linux/cpu.h>
101#include <linux/sysctl.h>
102#include <linux/module.h>
103#include <linux/rcupdate.h>
Paulo Marques543537b2005-06-23 00:09:02 -0700104#include <linux/string.h>
Christoph Lametere498be72005-09-09 13:03:32 -0700105#include <linux/nodemask.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -0700106
107#include <asm/uaccess.h>
108#include <asm/cacheflush.h>
109#include <asm/tlbflush.h>
110#include <asm/page.h>
111
112/*
113 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL,
114 * SLAB_RED_ZONE & SLAB_POISON.
115 * 0 for faster, smaller code (especially in the critical paths).
116 *
117 * STATS - 1 to collect stats for /proc/slabinfo.
118 * 0 for faster, smaller code (especially in the critical paths).
119 *
120 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
121 */
122
123#ifdef CONFIG_DEBUG_SLAB
124#define DEBUG 1
125#define STATS 1
126#define FORCED_DEBUG 1
127#else
128#define DEBUG 0
129#define STATS 0
130#define FORCED_DEBUG 0
131#endif
132
133
134/* Shouldn't this be in a header file somewhere? */
135#define BYTES_PER_WORD sizeof(void *)
136
137#ifndef cache_line_size
138#define cache_line_size() L1_CACHE_BYTES
139#endif
140
141#ifndef ARCH_KMALLOC_MINALIGN
142/*
143 * Enforce a minimum alignment for the kmalloc caches.
144 * Usually, the kmalloc caches are cache_line_size() aligned, except when
145 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
146 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
147 * alignment larger than BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that.
148 * Note that this flag disables some debug features.
149 */
150#define ARCH_KMALLOC_MINALIGN 0
151#endif
152
153#ifndef ARCH_SLAB_MINALIGN
154/*
155 * Enforce a minimum alignment for all caches.
156 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
157 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
158 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
159 * some debug features.
160 */
161#define ARCH_SLAB_MINALIGN 0
162#endif
163
164#ifndef ARCH_KMALLOC_FLAGS
165#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
166#endif
167
168/* Legal flag mask for kmem_cache_create(). */
169#if DEBUG
170# define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \
171 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
172 SLAB_NO_REAP | SLAB_CACHE_DMA | \
173 SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \
174 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
175 SLAB_DESTROY_BY_RCU)
176#else
177# define CREATE_MASK (SLAB_HWCACHE_ALIGN | SLAB_NO_REAP | \
178 SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \
179 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
180 SLAB_DESTROY_BY_RCU)
181#endif
182
183/*
184 * kmem_bufctl_t:
185 *
186 * Bufctl's are used for linking objs within a slab
187 * linked offsets.
188 *
189 * This implementation relies on "struct page" for locating the cache &
190 * slab an object belongs to.
191 * This allows the bufctl structure to be small (one int), but limits
192 * the number of objects a slab (not a cache) can contain when off-slab
193 * bufctls are used. The limit is the size of the largest general cache
194 * that does not use off-slab slabs.
195 * For 32bit archs with 4 kB pages, is this 56.
196 * This is not serious, as it is only for large objects, when it is unwise
197 * to have too many per slab.
198 * Note: This limit can be raised by introducing a general cache whose size
199 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
200 */
201
Kyle Moffettfa5b08d2005-09-03 15:55:03 -0700202typedef unsigned int kmem_bufctl_t;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700203#define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
204#define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
205#define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-2)
206
207/* Max number of objs-per-slab for caches which use off-slab slabs.
208 * Needed to avoid a possible looping condition in cache_grow().
209 */
210static unsigned long offslab_limit;
211
212/*
213 * struct slab
214 *
215 * Manages the objs in a slab. Placed either at the beginning of mem allocated
216 * for a slab, or allocated from an general cache.
217 * Slabs are chained into three list: fully used, partial, fully free slabs.
218 */
219struct slab {
220 struct list_head list;
221 unsigned long colouroff;
222 void *s_mem; /* including colour offset */
223 unsigned int inuse; /* num of objs active in slab */
224 kmem_bufctl_t free;
Christoph Lametere498be72005-09-09 13:03:32 -0700225 unsigned short nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700226};
227
228/*
229 * struct slab_rcu
230 *
231 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
232 * arrange for kmem_freepages to be called via RCU. This is useful if
233 * we need to approach a kernel structure obliquely, from its address
234 * obtained without the usual locking. We can lock the structure to
235 * stabilize it and check it's still at the given address, only if we
236 * can be sure that the memory has not been meanwhile reused for some
237 * other kind of object (which our subsystem's lock might corrupt).
238 *
239 * rcu_read_lock before reading the address, then rcu_read_unlock after
240 * taking the spinlock within the structure expected at that address.
241 *
242 * We assume struct slab_rcu can overlay struct slab when destroying.
243 */
244struct slab_rcu {
245 struct rcu_head head;
246 kmem_cache_t *cachep;
247 void *addr;
248};
249
250/*
251 * struct array_cache
252 *
Linus Torvalds1da177e2005-04-16 15:20:36 -0700253 * Purpose:
254 * - LIFO ordering, to hand out cache-warm objects from _alloc
255 * - reduce the number of linked list operations
256 * - reduce spinlock operations
257 *
258 * The limit is stored in the per-cpu structure to reduce the data cache
259 * footprint.
260 *
261 */
262struct array_cache {
263 unsigned int avail;
264 unsigned int limit;
265 unsigned int batchcount;
266 unsigned int touched;
Christoph Lametere498be72005-09-09 13:03:32 -0700267 spinlock_t lock;
268 void *entry[0]; /*
269 * Must have this definition in here for the proper
270 * alignment of array_cache. Also simplifies accessing
271 * the entries.
272 * [0] is for gcc 2.95. It should really be [].
273 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700274};
275
276/* bootstrap: The caches do not work without cpuarrays anymore,
277 * but the cpuarrays are allocated from the generic caches...
278 */
279#define BOOT_CPUCACHE_ENTRIES 1
280struct arraycache_init {
281 struct array_cache cache;
282 void * entries[BOOT_CPUCACHE_ENTRIES];
283};
284
285/*
Christoph Lametere498be72005-09-09 13:03:32 -0700286 * The slab lists for all objects.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700287 */
288struct kmem_list3 {
289 struct list_head slabs_partial; /* partial list first, better asm code */
290 struct list_head slabs_full;
291 struct list_head slabs_free;
292 unsigned long free_objects;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700293 unsigned long next_reap;
Christoph Lametere498be72005-09-09 13:03:32 -0700294 int free_touched;
295 unsigned int free_limit;
296 spinlock_t list_lock;
297 struct array_cache *shared; /* shared per node */
298 struct array_cache **alien; /* on other nodes */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700299};
300
Christoph Lametere498be72005-09-09 13:03:32 -0700301/*
302 * Need this for bootstrapping a per node allocator.
303 */
304#define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
305struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
306#define CACHE_CACHE 0
307#define SIZE_AC 1
308#define SIZE_L3 (1 + MAX_NUMNODES)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700309
Christoph Lametere498be72005-09-09 13:03:32 -0700310/*
311 * This function may be completely optimized away if
312 * a constant is passed to it. Mostly the same as
313 * what is in linux/slab.h except it returns an
314 * index.
315 */
316static inline int index_of(const size_t size)
317{
318 if (__builtin_constant_p(size)) {
319 int i = 0;
320
321#define CACHE(x) \
322 if (size <=x) \
323 return i; \
324 else \
325 i++;
326#include "linux/kmalloc_sizes.h"
327#undef CACHE
328 {
329 extern void __bad_size(void);
330 __bad_size();
331 }
332 }
333 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 */
369
370struct kmem_cache_s {
371/* 1) per-cpu data, touched during every alloc/free */
372 struct array_cache *array[NR_CPUS];
373 unsigned int batchcount;
374 unsigned int limit;
Christoph Lametere498be72005-09-09 13:03:32 -0700375 unsigned int shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700376 unsigned int objsize;
Christoph Lametere498be72005-09-09 13:03:32 -0700377/* 2) touched by every alloc & free from the backend */
378 struct kmem_list3 *nodelists[MAX_NUMNODES];
Linus Torvalds1da177e2005-04-16 15:20:36 -0700379 unsigned int flags; /* constant flags */
380 unsigned int num; /* # of objs per slab */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700381 spinlock_t spinlock;
382
383/* 3) cache_grow/shrink */
384 /* order of pgs per slab (2^n) */
385 unsigned int gfporder;
386
387 /* force GFP flags, e.g. GFP_DMA */
388 unsigned int gfpflags;
389
390 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 */
396
397 /* constructor func */
398 void (*ctor)(void *, kmem_cache_t *, unsigned long);
399
400 /* de-constructor func */
401 void (*dtor)(void *, kmem_cache_t *, unsigned long);
402
403/* 4) cache creation/removal */
404 const char *name;
405 struct list_head next;
406
407/* 5) statistics */
408#if STATS
409 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;
Christoph Lametere498be72005-09-09 13:03:32 -0700417 unsigned long node_frees;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700418 atomic_t allochit;
419 atomic_t allocmiss;
420 atomic_t freehit;
421 atomic_t freemiss;
422#endif
423#if DEBUG
424 int dbghead;
425 int reallen;
426#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 *
436 * OTHO the cpuarrays can contain lots of objects,
437 * 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)
525 return (unsigned long*) (objp+cachep->objsize-2*BYTES_PER_WORD);
526 return (unsigned long*) (objp+cachep->objsize-BYTES_PER_WORD);
527}
528
529static void **dbg_userword(kmem_cache_t *cachep, void *objp)
530{
531 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
532 return (void**)(objp+cachep->objsize-BYTES_PER_WORD);
533}
534
535#else
536
537#define obj_dbghead(x) 0
538#define obj_reallen(cachep) (cachep->objsize)
539#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;})
540#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;})
541#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
542
543#endif
544
545/*
546 * Maximum size of an obj (in 2^order pages)
547 * and absolute limit for the gfp order.
548 */
549#if defined(CONFIG_LARGE_ALLOCS)
550#define MAX_OBJ_ORDER 13 /* up to 32Mb */
551#define MAX_GFP_ORDER 13 /* up to 32Mb */
552#elif defined(CONFIG_MMU)
553#define MAX_OBJ_ORDER 5 /* 32 pages */
554#define MAX_GFP_ORDER 5 /* 32 pages */
555#else
556#define MAX_OBJ_ORDER 8 /* up to 1Mb */
557#define MAX_GFP_ORDER 8 /* up to 1Mb */
558#endif
559
560/*
561 * Do not go above this order unless 0 objects fit into the slab.
562 */
563#define BREAK_GFP_ORDER_HI 1
564#define BREAK_GFP_ORDER_LO 0
565static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
566
567/* Macros for storing/retrieving the cachep and or slab from the
568 * global 'mem_map'. These are used to find the slab an obj belongs to.
569 * With kfree(), these are used to find the cache which an obj belongs to.
570 */
571#define SET_PAGE_CACHE(pg,x) ((pg)->lru.next = (struct list_head *)(x))
572#define GET_PAGE_CACHE(pg) ((kmem_cache_t *)(pg)->lru.next)
573#define SET_PAGE_SLAB(pg,x) ((pg)->lru.prev = (struct list_head *)(x))
574#define GET_PAGE_SLAB(pg) ((struct slab *)(pg)->lru.prev)
575
576/* These are the default caches for kmalloc. Custom caches can have other sizes. */
577struct cache_sizes malloc_sizes[] = {
578#define CACHE(x) { .cs_size = (x) },
579#include <linux/kmalloc_sizes.h>
580 CACHE(ULONG_MAX)
581#undef CACHE
582};
583EXPORT_SYMBOL(malloc_sizes);
584
585/* Must match cache_sizes above. Out of line to keep cache footprint low. */
586struct cache_names {
587 char *name;
588 char *name_dma;
589};
590
591static struct cache_names __initdata cache_names[] = {
592#define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
593#include <linux/kmalloc_sizes.h>
594 { NULL, }
595#undef CACHE
596};
597
598static struct arraycache_init initarray_cache __initdata =
599 { { 0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
600static struct arraycache_init initarray_generic =
601 { { 0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
602
603/* internal cache of cache description objs */
604static kmem_cache_t cache_cache = {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700605 .batchcount = 1,
606 .limit = BOOT_CPUCACHE_ENTRIES,
Christoph Lametere498be72005-09-09 13:03:32 -0700607 .shared = 1,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700608 .objsize = sizeof(kmem_cache_t),
609 .flags = SLAB_NO_REAP,
610 .spinlock = SPIN_LOCK_UNLOCKED,
611 .name = "kmem_cache",
612#if DEBUG
613 .reallen = sizeof(kmem_cache_t),
614#endif
615};
616
617/* Guard access to the cache-chain. */
618static struct semaphore cache_chain_sem;
619static struct list_head cache_chain;
620
621/*
622 * vm_enough_memory() looks at this to determine how many
623 * slab-allocated pages are possibly freeable under pressure
624 *
625 * SLAB_RECLAIM_ACCOUNT turns this on per-slab
626 */
627atomic_t slab_reclaim_pages;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700628
629/*
630 * chicken and egg problem: delay the per-cpu array allocation
631 * until the general caches are up.
632 */
633static enum {
634 NONE,
Christoph Lametere498be72005-09-09 13:03:32 -0700635 PARTIAL_AC,
636 PARTIAL_L3,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700637 FULL
638} g_cpucache_up;
639
640static DEFINE_PER_CPU(struct work_struct, reap_work);
641
642static void free_block(kmem_cache_t* cachep, void** objpp, int len);
643static void enable_cpucache (kmem_cache_t *cachep);
644static void cache_reap (void *unused);
Christoph Lametere498be72005-09-09 13:03:32 -0700645static int __node_shrink(kmem_cache_t *cachep, int node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700646
647static inline struct array_cache *ac_data(kmem_cache_t *cachep)
648{
649 return cachep->array[smp_processor_id()];
650}
651
Alexey Dobriyan0db925a2005-07-07 17:56:58 -0700652static inline kmem_cache_t *__find_general_cachep(size_t size,
653 unsigned int __nocast gfpflags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700654{
655 struct cache_sizes *csizep = malloc_sizes;
656
657#if DEBUG
658 /* This happens if someone tries to call
659 * kmem_cache_create(), or __kmalloc(), before
660 * the generic caches are initialized.
661 */
662 BUG_ON(csizep->cs_cachep == NULL);
663#endif
664 while (size > csizep->cs_size)
665 csizep++;
666
667 /*
Martin Hicks0abf40c2005-09-03 15:54:54 -0700668 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
Linus Torvalds1da177e2005-04-16 15:20:36 -0700669 * has cs_{dma,}cachep==NULL. Thus no special case
670 * for large kmalloc calls required.
671 */
672 if (unlikely(gfpflags & GFP_DMA))
673 return csizep->cs_dmacachep;
674 return csizep->cs_cachep;
675}
676
Alexey Dobriyan0db925a2005-07-07 17:56:58 -0700677kmem_cache_t *kmem_find_general_cachep(size_t size,
678 unsigned int __nocast gfpflags)
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700679{
680 return __find_general_cachep(size, gfpflags);
681}
682EXPORT_SYMBOL(kmem_find_general_cachep);
683
Linus Torvalds1da177e2005-04-16 15:20:36 -0700684/* Cal the num objs, wastage, and bytes left over for a given slab size. */
685static void cache_estimate(unsigned long gfporder, size_t size, size_t align,
686 int flags, size_t *left_over, unsigned int *num)
687{
688 int i;
689 size_t wastage = PAGE_SIZE<<gfporder;
690 size_t extra = 0;
691 size_t base = 0;
692
693 if (!(flags & CFLGS_OFF_SLAB)) {
694 base = sizeof(struct slab);
695 extra = sizeof(kmem_bufctl_t);
696 }
697 i = 0;
698 while (i*size + ALIGN(base+i*extra, align) <= wastage)
699 i++;
700 if (i > 0)
701 i--;
702
703 if (i > SLAB_LIMIT)
704 i = SLAB_LIMIT;
705
706 *num = i;
707 wastage -= i*size;
708 wastage -= ALIGN(base+i*extra, align);
709 *left_over = wastage;
710}
711
712#define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
713
714static void __slab_error(const char *function, kmem_cache_t *cachep, char *msg)
715{
716 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
717 function, cachep->name, msg);
718 dump_stack();
719}
720
721/*
722 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
723 * via the workqueue/eventd.
724 * Add the CPU number into the expiration time to minimize the possibility of
725 * the CPUs getting into lockstep and contending for the global cache chain
726 * lock.
727 */
728static void __devinit start_cpu_timer(int cpu)
729{
730 struct work_struct *reap_work = &per_cpu(reap_work, cpu);
731
732 /*
733 * When this gets called from do_initcalls via cpucache_init(),
734 * init_workqueues() has already run, so keventd will be setup
735 * at that time.
736 */
737 if (keventd_up() && reap_work->func == NULL) {
738 INIT_WORK(reap_work, cache_reap, NULL);
739 schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu);
740 }
741}
742
Christoph Lametere498be72005-09-09 13:03:32 -0700743static struct array_cache *alloc_arraycache(int node, int entries,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700744 int batchcount)
745{
746 int memsize = sizeof(void*)*entries+sizeof(struct array_cache);
747 struct array_cache *nc = NULL;
748
Christoph Lametere498be72005-09-09 13:03:32 -0700749 nc = kmalloc_node(memsize, GFP_KERNEL, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700750 if (nc) {
751 nc->avail = 0;
752 nc->limit = entries;
753 nc->batchcount = batchcount;
754 nc->touched = 0;
Christoph Lametere498be72005-09-09 13:03:32 -0700755 spin_lock_init(&nc->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700756 }
757 return nc;
758}
759
Christoph Lametere498be72005-09-09 13:03:32 -0700760#ifdef CONFIG_NUMA
761static inline struct array_cache **alloc_alien_cache(int node, int limit)
762{
763 struct array_cache **ac_ptr;
764 int memsize = sizeof(void*)*MAX_NUMNODES;
765 int i;
766
767 if (limit > 1)
768 limit = 12;
769 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
770 if (ac_ptr) {
771 for_each_node(i) {
772 if (i == node || !node_online(i)) {
773 ac_ptr[i] = NULL;
774 continue;
775 }
776 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
777 if (!ac_ptr[i]) {
778 for (i--; i <=0; i--)
779 kfree(ac_ptr[i]);
780 kfree(ac_ptr);
781 return NULL;
782 }
783 }
784 }
785 return ac_ptr;
786}
787
788static inline void free_alien_cache(struct array_cache **ac_ptr)
789{
790 int i;
791
792 if (!ac_ptr)
793 return;
794
795 for_each_node(i)
796 kfree(ac_ptr[i]);
797
798 kfree(ac_ptr);
799}
800
801static inline void __drain_alien_cache(kmem_cache_t *cachep, struct array_cache *ac, int node)
802{
803 struct kmem_list3 *rl3 = cachep->nodelists[node];
804
805 if (ac->avail) {
806 spin_lock(&rl3->list_lock);
807 free_block(cachep, ac->entry, ac->avail);
808 ac->avail = 0;
809 spin_unlock(&rl3->list_lock);
810 }
811}
812
813static void drain_alien_cache(kmem_cache_t *cachep, struct kmem_list3 *l3)
814{
815 int i=0;
816 struct array_cache *ac;
817 unsigned long flags;
818
819 for_each_online_node(i) {
820 ac = l3->alien[i];
821 if (ac) {
822 spin_lock_irqsave(&ac->lock, flags);
823 __drain_alien_cache(cachep, ac, i);
824 spin_unlock_irqrestore(&ac->lock, flags);
825 }
826 }
827}
828#else
829#define alloc_alien_cache(node, limit) do { } while (0)
830#define free_alien_cache(ac_ptr) do { } while (0)
831#define drain_alien_cache(cachep, l3) do { } while (0)
832#endif
833
Linus Torvalds1da177e2005-04-16 15:20:36 -0700834static int __devinit cpuup_callback(struct notifier_block *nfb,
835 unsigned long action, void *hcpu)
836{
837 long cpu = (long)hcpu;
838 kmem_cache_t* cachep;
Christoph Lametere498be72005-09-09 13:03:32 -0700839 struct kmem_list3 *l3 = NULL;
840 int node = cpu_to_node(cpu);
841 int memsize = sizeof(struct kmem_list3);
842 struct array_cache *nc = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700843
844 switch (action) {
845 case CPU_UP_PREPARE:
846 down(&cache_chain_sem);
Christoph Lametere498be72005-09-09 13:03:32 -0700847 /* we need to do this right in the beginning since
848 * alloc_arraycache's are going to use this list.
849 * kmalloc_node allows us to add the slab to the right
850 * kmem_list3 and not this cpu's kmem_list3
851 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700852
Christoph Lametere498be72005-09-09 13:03:32 -0700853 list_for_each_entry(cachep, &cache_chain, next) {
854 /* setup the size64 kmemlist for cpu before we can
855 * begin anything. Make sure some other cpu on this
856 * node has not already allocated this
857 */
858 if (!cachep->nodelists[node]) {
859 if (!(l3 = kmalloc_node(memsize,
860 GFP_KERNEL, node)))
861 goto bad;
862 kmem_list3_init(l3);
863 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
864 ((unsigned long)cachep)%REAPTIMEOUT_LIST3;
865
866 cachep->nodelists[node] = l3;
867 }
868
869 spin_lock_irq(&cachep->nodelists[node]->list_lock);
870 cachep->nodelists[node]->free_limit =
871 (1 + nr_cpus_node(node)) *
872 cachep->batchcount + cachep->num;
873 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
874 }
875
876 /* Now we can go ahead with allocating the shared array's
877 & array cache's */
878 list_for_each_entry(cachep, &cache_chain, next) {
879 nc = alloc_arraycache(node, cachep->limit,
880 cachep->batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700881 if (!nc)
882 goto bad;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700883 cachep->array[cpu] = nc;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700884
Christoph Lametere498be72005-09-09 13:03:32 -0700885 l3 = cachep->nodelists[node];
886 BUG_ON(!l3);
887 if (!l3->shared) {
888 if (!(nc = alloc_arraycache(node,
889 cachep->shared*cachep->batchcount,
890 0xbaadf00d)))
891 goto bad;
892
893 /* we are serialised from CPU_DEAD or
894 CPU_UP_CANCELLED by the cpucontrol lock */
895 l3->shared = nc;
896 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700897 }
898 up(&cache_chain_sem);
899 break;
900 case CPU_ONLINE:
901 start_cpu_timer(cpu);
902 break;
903#ifdef CONFIG_HOTPLUG_CPU
904 case CPU_DEAD:
905 /* fall thru */
906 case CPU_UP_CANCELED:
907 down(&cache_chain_sem);
908
909 list_for_each_entry(cachep, &cache_chain, next) {
910 struct array_cache *nc;
Christoph Lametere498be72005-09-09 13:03:32 -0700911 cpumask_t mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700912
Christoph Lametere498be72005-09-09 13:03:32 -0700913 mask = node_to_cpumask(node);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700914 spin_lock_irq(&cachep->spinlock);
915 /* cpu is dead; no one can alloc from it. */
916 nc = cachep->array[cpu];
917 cachep->array[cpu] = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -0700918 l3 = cachep->nodelists[node];
919
920 if (!l3)
921 goto unlock_cache;
922
923 spin_lock(&l3->list_lock);
924
925 /* Free limit for this kmem_list3 */
926 l3->free_limit -= cachep->batchcount;
927 if (nc)
928 free_block(cachep, nc->entry, nc->avail);
929
930 if (!cpus_empty(mask)) {
931 spin_unlock(&l3->list_lock);
932 goto unlock_cache;
933 }
934
935 if (l3->shared) {
936 free_block(cachep, l3->shared->entry,
937 l3->shared->avail);
938 kfree(l3->shared);
939 l3->shared = NULL;
940 }
941 if (l3->alien) {
942 drain_alien_cache(cachep, l3);
943 free_alien_cache(l3->alien);
944 l3->alien = NULL;
945 }
946
947 /* free slabs belonging to this node */
948 if (__node_shrink(cachep, node)) {
949 cachep->nodelists[node] = NULL;
950 spin_unlock(&l3->list_lock);
951 kfree(l3);
952 } else {
953 spin_unlock(&l3->list_lock);
954 }
955unlock_cache:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700956 spin_unlock_irq(&cachep->spinlock);
957 kfree(nc);
958 }
959 up(&cache_chain_sem);
960 break;
961#endif
962 }
963 return NOTIFY_OK;
964bad:
965 up(&cache_chain_sem);
966 return NOTIFY_BAD;
967}
968
969static struct notifier_block cpucache_notifier = { &cpuup_callback, NULL, 0 };
970
Christoph Lametere498be72005-09-09 13:03:32 -0700971/*
972 * swap the static kmem_list3 with kmalloced memory
973 */
974static void init_list(kmem_cache_t *cachep, struct kmem_list3 *list,
975 int nodeid)
976{
977 struct kmem_list3 *ptr;
978
979 BUG_ON(cachep->nodelists[nodeid] != list);
980 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
981 BUG_ON(!ptr);
982
983 local_irq_disable();
984 memcpy(ptr, list, sizeof(struct kmem_list3));
985 MAKE_ALL_LISTS(cachep, ptr, nodeid);
986 cachep->nodelists[nodeid] = ptr;
987 local_irq_enable();
988}
989
Linus Torvalds1da177e2005-04-16 15:20:36 -0700990/* Initialisation.
991 * Called after the gfp() functions have been enabled, and before smp_init().
992 */
993void __init kmem_cache_init(void)
994{
995 size_t left_over;
996 struct cache_sizes *sizes;
997 struct cache_names *names;
Christoph Lametere498be72005-09-09 13:03:32 -0700998 int i;
999
1000 for (i = 0; i < NUM_INIT_LISTS; i++) {
1001 kmem_list3_init(&initkmem_list3[i]);
1002 if (i < MAX_NUMNODES)
1003 cache_cache.nodelists[i] = NULL;
1004 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001005
1006 /*
1007 * Fragmentation resistance on low memory - only use bigger
1008 * page orders on machines with more than 32MB of memory.
1009 */
1010 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1011 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1012
Linus Torvalds1da177e2005-04-16 15:20:36 -07001013 /* Bootstrap is tricky, because several objects are allocated
1014 * from caches that do not exist yet:
1015 * 1) initialize the cache_cache cache: it contains the kmem_cache_t
1016 * structures of all caches, except cache_cache itself: cache_cache
1017 * is statically allocated.
Christoph Lametere498be72005-09-09 13:03:32 -07001018 * Initially an __init data area is used for the head array and the
1019 * kmem_list3 structures, it's replaced with a kmalloc allocated
1020 * array at the end of the bootstrap.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001021 * 2) Create the first kmalloc cache.
Christoph Lametere498be72005-09-09 13:03:32 -07001022 * The kmem_cache_t for the new cache is allocated normally.
1023 * An __init data area is used for the head array.
1024 * 3) Create the remaining kmalloc caches, with minimally sized
1025 * head arrays.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001026 * 4) Replace the __init data head arrays for cache_cache and the first
1027 * kmalloc cache with kmalloc allocated arrays.
Christoph Lametere498be72005-09-09 13:03:32 -07001028 * 5) Replace the __init data for kmem_list3 for cache_cache and
1029 * the other cache's with kmalloc allocated memory.
1030 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001031 */
1032
1033 /* 1) create the cache_cache */
1034 init_MUTEX(&cache_chain_sem);
1035 INIT_LIST_HEAD(&cache_chain);
1036 list_add(&cache_cache.next, &cache_chain);
1037 cache_cache.colour_off = cache_line_size();
1038 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
Christoph Lametere498be72005-09-09 13:03:32 -07001039 cache_cache.nodelists[numa_node_id()] = &initkmem_list3[CACHE_CACHE];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001040
1041 cache_cache.objsize = ALIGN(cache_cache.objsize, cache_line_size());
1042
1043 cache_estimate(0, cache_cache.objsize, cache_line_size(), 0,
1044 &left_over, &cache_cache.num);
1045 if (!cache_cache.num)
1046 BUG();
1047
1048 cache_cache.colour = left_over/cache_cache.colour_off;
1049 cache_cache.colour_next = 0;
1050 cache_cache.slab_size = ALIGN(cache_cache.num*sizeof(kmem_bufctl_t) +
1051 sizeof(struct slab), cache_line_size());
1052
1053 /* 2+3) create the kmalloc caches */
1054 sizes = malloc_sizes;
1055 names = cache_names;
1056
Christoph Lametere498be72005-09-09 13:03:32 -07001057 /* Initialize the caches that provide memory for the array cache
1058 * and the kmem_list3 structures first.
1059 * Without this, further allocations will bug
1060 */
1061
1062 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
1063 sizes[INDEX_AC].cs_size, ARCH_KMALLOC_MINALIGN,
1064 (ARCH_KMALLOC_FLAGS | SLAB_PANIC), NULL, NULL);
1065
1066 if (INDEX_AC != INDEX_L3)
1067 sizes[INDEX_L3].cs_cachep =
1068 kmem_cache_create(names[INDEX_L3].name,
1069 sizes[INDEX_L3].cs_size, ARCH_KMALLOC_MINALIGN,
1070 (ARCH_KMALLOC_FLAGS | SLAB_PANIC), NULL, NULL);
1071
Linus Torvalds1da177e2005-04-16 15:20:36 -07001072 while (sizes->cs_size != ULONG_MAX) {
Christoph Lametere498be72005-09-09 13:03:32 -07001073 /*
1074 * For performance, all the general caches are L1 aligned.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001075 * This should be particularly beneficial on SMP boxes, as it
1076 * eliminates "false sharing".
1077 * Note for systems short on memory removing the alignment will
Christoph Lametere498be72005-09-09 13:03:32 -07001078 * allow tighter packing of the smaller caches.
1079 */
1080 if(!sizes->cs_cachep)
1081 sizes->cs_cachep = kmem_cache_create(names->name,
1082 sizes->cs_size, ARCH_KMALLOC_MINALIGN,
1083 (ARCH_KMALLOC_FLAGS | SLAB_PANIC), NULL, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001084
1085 /* Inc off-slab bufctl limit until the ceiling is hit. */
1086 if (!(OFF_SLAB(sizes->cs_cachep))) {
1087 offslab_limit = sizes->cs_size-sizeof(struct slab);
1088 offslab_limit /= sizeof(kmem_bufctl_t);
1089 }
1090
1091 sizes->cs_dmacachep = kmem_cache_create(names->name_dma,
1092 sizes->cs_size, ARCH_KMALLOC_MINALIGN,
1093 (ARCH_KMALLOC_FLAGS | SLAB_CACHE_DMA | SLAB_PANIC),
1094 NULL, NULL);
1095
1096 sizes++;
1097 names++;
1098 }
1099 /* 4) Replace the bootstrap head arrays */
1100 {
1101 void * ptr;
Christoph Lametere498be72005-09-09 13:03:32 -07001102
Linus Torvalds1da177e2005-04-16 15:20:36 -07001103 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001104
Linus Torvalds1da177e2005-04-16 15:20:36 -07001105 local_irq_disable();
1106 BUG_ON(ac_data(&cache_cache) != &initarray_cache.cache);
Christoph Lametere498be72005-09-09 13:03:32 -07001107 memcpy(ptr, ac_data(&cache_cache),
1108 sizeof(struct arraycache_init));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001109 cache_cache.array[smp_processor_id()] = ptr;
1110 local_irq_enable();
Christoph Lametere498be72005-09-09 13:03:32 -07001111
Linus Torvalds1da177e2005-04-16 15:20:36 -07001112 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
Christoph Lametere498be72005-09-09 13:03:32 -07001113
Linus Torvalds1da177e2005-04-16 15:20:36 -07001114 local_irq_disable();
Christoph Lametere498be72005-09-09 13:03:32 -07001115 BUG_ON(ac_data(malloc_sizes[INDEX_AC].cs_cachep)
1116 != &initarray_generic.cache);
1117 memcpy(ptr, ac_data(malloc_sizes[INDEX_AC].cs_cachep),
Linus Torvalds1da177e2005-04-16 15:20:36 -07001118 sizeof(struct arraycache_init));
Christoph Lametere498be72005-09-09 13:03:32 -07001119 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1120 ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001121 local_irq_enable();
1122 }
Christoph Lametere498be72005-09-09 13:03:32 -07001123 /* 5) Replace the bootstrap kmem_list3's */
1124 {
1125 int node;
1126 /* Replace the static kmem_list3 structures for the boot cpu */
1127 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE],
1128 numa_node_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -07001129
Christoph Lametere498be72005-09-09 13:03:32 -07001130 for_each_online_node(node) {
1131 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1132 &initkmem_list3[SIZE_AC+node], node);
1133
1134 if (INDEX_AC != INDEX_L3) {
1135 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1136 &initkmem_list3[SIZE_L3+node],
1137 node);
1138 }
1139 }
1140 }
1141
1142 /* 6) resize the head arrays to their final sizes */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001143 {
1144 kmem_cache_t *cachep;
1145 down(&cache_chain_sem);
1146 list_for_each_entry(cachep, &cache_chain, next)
1147 enable_cpucache(cachep);
1148 up(&cache_chain_sem);
1149 }
1150
1151 /* Done! */
1152 g_cpucache_up = FULL;
1153
1154 /* Register a cpu startup notifier callback
1155 * that initializes ac_data for all new cpus
1156 */
1157 register_cpu_notifier(&cpucache_notifier);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001158
1159 /* The reap timers are started later, with a module init call:
1160 * That part of the kernel is not yet operational.
1161 */
1162}
1163
1164static int __init cpucache_init(void)
1165{
1166 int cpu;
1167
1168 /*
1169 * Register the timers that return unneeded
1170 * pages to gfp.
1171 */
Christoph Lametere498be72005-09-09 13:03:32 -07001172 for_each_online_cpu(cpu)
1173 start_cpu_timer(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001174
1175 return 0;
1176}
1177
1178__initcall(cpucache_init);
1179
1180/*
1181 * Interface to system's page allocator. No need to hold the cache-lock.
1182 *
1183 * If we requested dmaable memory, we will get it. Even if we
1184 * did not request dmaable memory, we might get it, but that
1185 * would be relatively rare and ignorable.
1186 */
1187static void *kmem_getpages(kmem_cache_t *cachep, unsigned int __nocast flags, int nodeid)
1188{
1189 struct page *page;
1190 void *addr;
1191 int i;
1192
1193 flags |= cachep->gfpflags;
1194 if (likely(nodeid == -1)) {
1195 page = alloc_pages(flags, cachep->gfporder);
1196 } else {
1197 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
1198 }
1199 if (!page)
1200 return NULL;
1201 addr = page_address(page);
1202
1203 i = (1 << cachep->gfporder);
1204 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1205 atomic_add(i, &slab_reclaim_pages);
1206 add_page_state(nr_slab, i);
1207 while (i--) {
1208 SetPageSlab(page);
1209 page++;
1210 }
1211 return addr;
1212}
1213
1214/*
1215 * Interface to system's page release.
1216 */
1217static void kmem_freepages(kmem_cache_t *cachep, void *addr)
1218{
1219 unsigned long i = (1<<cachep->gfporder);
1220 struct page *page = virt_to_page(addr);
1221 const unsigned long nr_freed = i;
1222
1223 while (i--) {
1224 if (!TestClearPageSlab(page))
1225 BUG();
1226 page++;
1227 }
1228 sub_page_state(nr_slab, nr_freed);
1229 if (current->reclaim_state)
1230 current->reclaim_state->reclaimed_slab += nr_freed;
1231 free_pages((unsigned long)addr, cachep->gfporder);
1232 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1233 atomic_sub(1<<cachep->gfporder, &slab_reclaim_pages);
1234}
1235
1236static void kmem_rcu_free(struct rcu_head *head)
1237{
1238 struct slab_rcu *slab_rcu = (struct slab_rcu *) head;
1239 kmem_cache_t *cachep = slab_rcu->cachep;
1240
1241 kmem_freepages(cachep, slab_rcu->addr);
1242 if (OFF_SLAB(cachep))
1243 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1244}
1245
1246#if DEBUG
1247
1248#ifdef CONFIG_DEBUG_PAGEALLOC
1249static void store_stackinfo(kmem_cache_t *cachep, unsigned long *addr,
1250 unsigned long caller)
1251{
1252 int size = obj_reallen(cachep);
1253
1254 addr = (unsigned long *)&((char*)addr)[obj_dbghead(cachep)];
1255
1256 if (size < 5*sizeof(unsigned long))
1257 return;
1258
1259 *addr++=0x12345678;
1260 *addr++=caller;
1261 *addr++=smp_processor_id();
1262 size -= 3*sizeof(unsigned long);
1263 {
1264 unsigned long *sptr = &caller;
1265 unsigned long svalue;
1266
1267 while (!kstack_end(sptr)) {
1268 svalue = *sptr++;
1269 if (kernel_text_address(svalue)) {
1270 *addr++=svalue;
1271 size -= sizeof(unsigned long);
1272 if (size <= sizeof(unsigned long))
1273 break;
1274 }
1275 }
1276
1277 }
1278 *addr++=0x87654321;
1279}
1280#endif
1281
1282static void poison_obj(kmem_cache_t *cachep, void *addr, unsigned char val)
1283{
1284 int size = obj_reallen(cachep);
1285 addr = &((char*)addr)[obj_dbghead(cachep)];
1286
1287 memset(addr, val, size);
1288 *(unsigned char *)(addr+size-1) = POISON_END;
1289}
1290
1291static void dump_line(char *data, int offset, int limit)
1292{
1293 int i;
1294 printk(KERN_ERR "%03x:", offset);
1295 for (i=0;i<limit;i++) {
1296 printk(" %02x", (unsigned char)data[offset+i]);
1297 }
1298 printk("\n");
1299}
1300#endif
1301
1302#if DEBUG
1303
1304static void print_objinfo(kmem_cache_t *cachep, void *objp, int lines)
1305{
1306 int i, size;
1307 char *realobj;
1308
1309 if (cachep->flags & SLAB_RED_ZONE) {
1310 printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n",
1311 *dbg_redzone1(cachep, objp),
1312 *dbg_redzone2(cachep, objp));
1313 }
1314
1315 if (cachep->flags & SLAB_STORE_USER) {
1316 printk(KERN_ERR "Last user: [<%p>]",
1317 *dbg_userword(cachep, objp));
1318 print_symbol("(%s)",
1319 (unsigned long)*dbg_userword(cachep, objp));
1320 printk("\n");
1321 }
1322 realobj = (char*)objp+obj_dbghead(cachep);
1323 size = obj_reallen(cachep);
1324 for (i=0; i<size && lines;i+=16, lines--) {
1325 int limit;
1326 limit = 16;
1327 if (i+limit > size)
1328 limit = size-i;
1329 dump_line(realobj, i, limit);
1330 }
1331}
1332
1333static void check_poison_obj(kmem_cache_t *cachep, void *objp)
1334{
1335 char *realobj;
1336 int size, i;
1337 int lines = 0;
1338
1339 realobj = (char*)objp+obj_dbghead(cachep);
1340 size = obj_reallen(cachep);
1341
1342 for (i=0;i<size;i++) {
1343 char exp = POISON_FREE;
1344 if (i == size-1)
1345 exp = POISON_END;
1346 if (realobj[i] != exp) {
1347 int limit;
1348 /* Mismatch ! */
1349 /* Print header */
1350 if (lines == 0) {
1351 printk(KERN_ERR "Slab corruption: start=%p, len=%d\n",
1352 realobj, size);
1353 print_objinfo(cachep, objp, 0);
1354 }
1355 /* Hexdump the affected line */
1356 i = (i/16)*16;
1357 limit = 16;
1358 if (i+limit > size)
1359 limit = size-i;
1360 dump_line(realobj, i, limit);
1361 i += 16;
1362 lines++;
1363 /* Limit to 5 lines */
1364 if (lines > 5)
1365 break;
1366 }
1367 }
1368 if (lines != 0) {
1369 /* Print some data about the neighboring objects, if they
1370 * exist:
1371 */
1372 struct slab *slabp = GET_PAGE_SLAB(virt_to_page(objp));
1373 int objnr;
1374
1375 objnr = (objp-slabp->s_mem)/cachep->objsize;
1376 if (objnr) {
1377 objp = slabp->s_mem+(objnr-1)*cachep->objsize;
1378 realobj = (char*)objp+obj_dbghead(cachep);
1379 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1380 realobj, size);
1381 print_objinfo(cachep, objp, 2);
1382 }
1383 if (objnr+1 < cachep->num) {
1384 objp = slabp->s_mem+(objnr+1)*cachep->objsize;
1385 realobj = (char*)objp+obj_dbghead(cachep);
1386 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1387 realobj, size);
1388 print_objinfo(cachep, objp, 2);
1389 }
1390 }
1391}
1392#endif
1393
1394/* Destroy all the objs in a slab, and release the mem back to the system.
1395 * Before calling the slab must have been unlinked from the cache.
1396 * The cache-lock is not held/needed.
1397 */
1398static void slab_destroy (kmem_cache_t *cachep, struct slab *slabp)
1399{
1400 void *addr = slabp->s_mem - slabp->colouroff;
1401
1402#if DEBUG
1403 int i;
1404 for (i = 0; i < cachep->num; i++) {
1405 void *objp = slabp->s_mem + cachep->objsize * i;
1406
1407 if (cachep->flags & SLAB_POISON) {
1408#ifdef CONFIG_DEBUG_PAGEALLOC
1409 if ((cachep->objsize%PAGE_SIZE)==0 && OFF_SLAB(cachep))
1410 kernel_map_pages(virt_to_page(objp), cachep->objsize/PAGE_SIZE,1);
1411 else
1412 check_poison_obj(cachep, objp);
1413#else
1414 check_poison_obj(cachep, objp);
1415#endif
1416 }
1417 if (cachep->flags & SLAB_RED_ZONE) {
1418 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1419 slab_error(cachep, "start of a freed object "
1420 "was overwritten");
1421 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1422 slab_error(cachep, "end of a freed object "
1423 "was overwritten");
1424 }
1425 if (cachep->dtor && !(cachep->flags & SLAB_POISON))
1426 (cachep->dtor)(objp+obj_dbghead(cachep), cachep, 0);
1427 }
1428#else
1429 if (cachep->dtor) {
1430 int i;
1431 for (i = 0; i < cachep->num; i++) {
1432 void* objp = slabp->s_mem+cachep->objsize*i;
1433 (cachep->dtor)(objp, cachep, 0);
1434 }
1435 }
1436#endif
1437
1438 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1439 struct slab_rcu *slab_rcu;
1440
1441 slab_rcu = (struct slab_rcu *) slabp;
1442 slab_rcu->cachep = cachep;
1443 slab_rcu->addr = addr;
1444 call_rcu(&slab_rcu->head, kmem_rcu_free);
1445 } else {
1446 kmem_freepages(cachep, addr);
1447 if (OFF_SLAB(cachep))
1448 kmem_cache_free(cachep->slabp_cache, slabp);
1449 }
1450}
1451
Christoph Lametere498be72005-09-09 13:03:32 -07001452/* For setting up all the kmem_list3s for cache whose objsize is same
1453 as size of kmem_list3. */
1454static inline void set_up_list3s(kmem_cache_t *cachep, int index)
1455{
1456 int node;
1457
1458 for_each_online_node(node) {
1459 cachep->nodelists[node] = &initkmem_list3[index+node];
1460 cachep->nodelists[node]->next_reap = jiffies +
1461 REAPTIMEOUT_LIST3 +
1462 ((unsigned long)cachep)%REAPTIMEOUT_LIST3;
1463 }
1464}
1465
Linus Torvalds1da177e2005-04-16 15:20:36 -07001466/**
1467 * kmem_cache_create - Create a cache.
1468 * @name: A string which is used in /proc/slabinfo to identify this cache.
1469 * @size: The size of objects to be created in this cache.
1470 * @align: The required alignment for the objects.
1471 * @flags: SLAB flags
1472 * @ctor: A constructor for the objects.
1473 * @dtor: A destructor for the objects.
1474 *
1475 * Returns a ptr to the cache on success, NULL on failure.
1476 * Cannot be called within a int, but can be interrupted.
1477 * The @ctor is run when new pages are allocated by the cache
1478 * and the @dtor is run before the pages are handed back.
1479 *
1480 * @name must be valid until the cache is destroyed. This implies that
1481 * the module calling this has to destroy the cache before getting
1482 * unloaded.
1483 *
1484 * The flags are
1485 *
1486 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1487 * to catch references to uninitialised memory.
1488 *
1489 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1490 * for buffer overruns.
1491 *
1492 * %SLAB_NO_REAP - Don't automatically reap this cache when we're under
1493 * memory pressure.
1494 *
1495 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1496 * cacheline. This can be beneficial if you're counting cycles as closely
1497 * as davem.
1498 */
1499kmem_cache_t *
1500kmem_cache_create (const char *name, size_t size, size_t align,
1501 unsigned long flags, void (*ctor)(void*, kmem_cache_t *, unsigned long),
1502 void (*dtor)(void*, kmem_cache_t *, unsigned long))
1503{
1504 size_t left_over, slab_size, ralign;
1505 kmem_cache_t *cachep = NULL;
1506
1507 /*
1508 * Sanity checks... these are all serious usage bugs.
1509 */
1510 if ((!name) ||
1511 in_interrupt() ||
1512 (size < BYTES_PER_WORD) ||
1513 (size > (1<<MAX_OBJ_ORDER)*PAGE_SIZE) ||
1514 (dtor && !ctor)) {
1515 printk(KERN_ERR "%s: Early error in slab %s\n",
1516 __FUNCTION__, name);
1517 BUG();
1518 }
1519
1520#if DEBUG
1521 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
1522 if ((flags & SLAB_DEBUG_INITIAL) && !ctor) {
1523 /* No constructor, but inital state check requested */
1524 printk(KERN_ERR "%s: No con, but init state check "
1525 "requested - %s\n", __FUNCTION__, name);
1526 flags &= ~SLAB_DEBUG_INITIAL;
1527 }
1528
1529#if FORCED_DEBUG
1530 /*
1531 * Enable redzoning and last user accounting, except for caches with
1532 * large objects, if the increased size would increase the object size
1533 * above the next power of two: caches with object sizes just above a
1534 * power of two have a significant amount of internal fragmentation.
1535 */
1536 if ((size < 4096 || fls(size-1) == fls(size-1+3*BYTES_PER_WORD)))
1537 flags |= SLAB_RED_ZONE|SLAB_STORE_USER;
1538 if (!(flags & SLAB_DESTROY_BY_RCU))
1539 flags |= SLAB_POISON;
1540#endif
1541 if (flags & SLAB_DESTROY_BY_RCU)
1542 BUG_ON(flags & SLAB_POISON);
1543#endif
1544 if (flags & SLAB_DESTROY_BY_RCU)
1545 BUG_ON(dtor);
1546
1547 /*
1548 * Always checks flags, a caller might be expecting debug
1549 * support which isn't available.
1550 */
1551 if (flags & ~CREATE_MASK)
1552 BUG();
1553
1554 /* Check that size is in terms of words. This is needed to avoid
1555 * unaligned accesses for some archs when redzoning is used, and makes
1556 * sure any on-slab bufctl's are also correctly aligned.
1557 */
1558 if (size & (BYTES_PER_WORD-1)) {
1559 size += (BYTES_PER_WORD-1);
1560 size &= ~(BYTES_PER_WORD-1);
1561 }
1562
1563 /* calculate out the final buffer alignment: */
1564 /* 1) arch recommendation: can be overridden for debug */
1565 if (flags & SLAB_HWCACHE_ALIGN) {
1566 /* Default alignment: as specified by the arch code.
1567 * Except if an object is really small, then squeeze multiple
1568 * objects into one cacheline.
1569 */
1570 ralign = cache_line_size();
1571 while (size <= ralign/2)
1572 ralign /= 2;
1573 } else {
1574 ralign = BYTES_PER_WORD;
1575 }
1576 /* 2) arch mandated alignment: disables debug if necessary */
1577 if (ralign < ARCH_SLAB_MINALIGN) {
1578 ralign = ARCH_SLAB_MINALIGN;
1579 if (ralign > BYTES_PER_WORD)
1580 flags &= ~(SLAB_RED_ZONE|SLAB_STORE_USER);
1581 }
1582 /* 3) caller mandated alignment: disables debug if necessary */
1583 if (ralign < align) {
1584 ralign = align;
1585 if (ralign > BYTES_PER_WORD)
1586 flags &= ~(SLAB_RED_ZONE|SLAB_STORE_USER);
1587 }
1588 /* 4) Store it. Note that the debug code below can reduce
1589 * the alignment to BYTES_PER_WORD.
1590 */
1591 align = ralign;
1592
1593 /* Get cache's description obj. */
1594 cachep = (kmem_cache_t *) kmem_cache_alloc(&cache_cache, SLAB_KERNEL);
1595 if (!cachep)
1596 goto opps;
1597 memset(cachep, 0, sizeof(kmem_cache_t));
1598
1599#if DEBUG
1600 cachep->reallen = size;
1601
1602 if (flags & SLAB_RED_ZONE) {
1603 /* redzoning only works with word aligned caches */
1604 align = BYTES_PER_WORD;
1605
1606 /* add space for red zone words */
1607 cachep->dbghead += BYTES_PER_WORD;
1608 size += 2*BYTES_PER_WORD;
1609 }
1610 if (flags & SLAB_STORE_USER) {
1611 /* user store requires word alignment and
1612 * one word storage behind the end of the real
1613 * object.
1614 */
1615 align = BYTES_PER_WORD;
1616 size += BYTES_PER_WORD;
1617 }
1618#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
Christoph Lametere498be72005-09-09 13:03:32 -07001619 if (size >= malloc_sizes[INDEX_L3+1].cs_size && cachep->reallen > cache_line_size() && size < PAGE_SIZE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001620 cachep->dbghead += PAGE_SIZE - size;
1621 size = PAGE_SIZE;
1622 }
1623#endif
1624#endif
1625
1626 /* Determine if the slab management is 'on' or 'off' slab. */
1627 if (size >= (PAGE_SIZE>>3))
1628 /*
1629 * Size is large, assume best to place the slab management obj
1630 * off-slab (should allow better packing of objs).
1631 */
1632 flags |= CFLGS_OFF_SLAB;
1633
1634 size = ALIGN(size, align);
1635
1636 if ((flags & SLAB_RECLAIM_ACCOUNT) && size <= PAGE_SIZE) {
1637 /*
1638 * A VFS-reclaimable slab tends to have most allocations
1639 * as GFP_NOFS and we really don't want to have to be allocating
1640 * higher-order pages when we are unable to shrink dcache.
1641 */
1642 cachep->gfporder = 0;
1643 cache_estimate(cachep->gfporder, size, align, flags,
1644 &left_over, &cachep->num);
1645 } else {
1646 /*
1647 * Calculate size (in pages) of slabs, and the num of objs per
1648 * slab. This could be made much more intelligent. For now,
1649 * try to avoid using high page-orders for slabs. When the
1650 * gfp() funcs are more friendly towards high-order requests,
1651 * this should be changed.
1652 */
1653 do {
1654 unsigned int break_flag = 0;
1655cal_wastage:
1656 cache_estimate(cachep->gfporder, size, align, flags,
1657 &left_over, &cachep->num);
1658 if (break_flag)
1659 break;
1660 if (cachep->gfporder >= MAX_GFP_ORDER)
1661 break;
1662 if (!cachep->num)
1663 goto next;
1664 if (flags & CFLGS_OFF_SLAB &&
1665 cachep->num > offslab_limit) {
1666 /* This num of objs will cause problems. */
1667 cachep->gfporder--;
1668 break_flag++;
1669 goto cal_wastage;
1670 }
1671
1672 /*
1673 * Large num of objs is good, but v. large slabs are
1674 * currently bad for the gfp()s.
1675 */
1676 if (cachep->gfporder >= slab_break_gfp_order)
1677 break;
1678
1679 if ((left_over*8) <= (PAGE_SIZE<<cachep->gfporder))
1680 break; /* Acceptable internal fragmentation. */
1681next:
1682 cachep->gfporder++;
1683 } while (1);
1684 }
1685
1686 if (!cachep->num) {
1687 printk("kmem_cache_create: couldn't create cache %s.\n", name);
1688 kmem_cache_free(&cache_cache, cachep);
1689 cachep = NULL;
1690 goto opps;
1691 }
1692 slab_size = ALIGN(cachep->num*sizeof(kmem_bufctl_t)
1693 + sizeof(struct slab), align);
1694
1695 /*
1696 * If the slab has been placed off-slab, and we have enough space then
1697 * move it on-slab. This is at the expense of any extra colouring.
1698 */
1699 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
1700 flags &= ~CFLGS_OFF_SLAB;
1701 left_over -= slab_size;
1702 }
1703
1704 if (flags & CFLGS_OFF_SLAB) {
1705 /* really off slab. No need for manual alignment */
1706 slab_size = cachep->num*sizeof(kmem_bufctl_t)+sizeof(struct slab);
1707 }
1708
1709 cachep->colour_off = cache_line_size();
1710 /* Offset must be a multiple of the alignment. */
1711 if (cachep->colour_off < align)
1712 cachep->colour_off = align;
1713 cachep->colour = left_over/cachep->colour_off;
1714 cachep->slab_size = slab_size;
1715 cachep->flags = flags;
1716 cachep->gfpflags = 0;
1717 if (flags & SLAB_CACHE_DMA)
1718 cachep->gfpflags |= GFP_DMA;
1719 spin_lock_init(&cachep->spinlock);
1720 cachep->objsize = size;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001721
1722 if (flags & CFLGS_OFF_SLAB)
Victor Fuscob2d55072005-09-10 00:26:36 -07001723 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001724 cachep->ctor = ctor;
1725 cachep->dtor = dtor;
1726 cachep->name = name;
1727
1728 /* Don't let CPUs to come and go */
1729 lock_cpu_hotplug();
1730
1731 if (g_cpucache_up == FULL) {
1732 enable_cpucache(cachep);
1733 } else {
1734 if (g_cpucache_up == NONE) {
1735 /* Note: the first kmem_cache_create must create
1736 * the cache that's used by kmalloc(24), otherwise
1737 * the creation of further caches will BUG().
1738 */
Christoph Lametere498be72005-09-09 13:03:32 -07001739 cachep->array[smp_processor_id()] =
1740 &initarray_generic.cache;
1741
1742 /* If the cache that's used by
1743 * kmalloc(sizeof(kmem_list3)) is the first cache,
1744 * then we need to set up all its list3s, otherwise
1745 * the creation of further caches will BUG().
1746 */
1747 set_up_list3s(cachep, SIZE_AC);
1748 if (INDEX_AC == INDEX_L3)
1749 g_cpucache_up = PARTIAL_L3;
1750 else
1751 g_cpucache_up = PARTIAL_AC;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001752 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07001753 cachep->array[smp_processor_id()] =
1754 kmalloc(sizeof(struct arraycache_init),
1755 GFP_KERNEL);
1756
1757 if (g_cpucache_up == PARTIAL_AC) {
1758 set_up_list3s(cachep, SIZE_L3);
1759 g_cpucache_up = PARTIAL_L3;
1760 } else {
1761 int node;
1762 for_each_online_node(node) {
1763
1764 cachep->nodelists[node] =
1765 kmalloc_node(sizeof(struct kmem_list3),
1766 GFP_KERNEL, node);
1767 BUG_ON(!cachep->nodelists[node]);
1768 kmem_list3_init(cachep->nodelists[node]);
1769 }
1770 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001771 }
Christoph Lametere498be72005-09-09 13:03:32 -07001772 cachep->nodelists[numa_node_id()]->next_reap =
1773 jiffies + REAPTIMEOUT_LIST3 +
1774 ((unsigned long)cachep)%REAPTIMEOUT_LIST3;
1775
Linus Torvalds1da177e2005-04-16 15:20:36 -07001776 BUG_ON(!ac_data(cachep));
1777 ac_data(cachep)->avail = 0;
1778 ac_data(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1779 ac_data(cachep)->batchcount = 1;
1780 ac_data(cachep)->touched = 0;
1781 cachep->batchcount = 1;
1782 cachep->limit = BOOT_CPUCACHE_ENTRIES;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001783 }
1784
Linus Torvalds1da177e2005-04-16 15:20:36 -07001785 /* Need the semaphore to access the chain. */
1786 down(&cache_chain_sem);
1787 {
1788 struct list_head *p;
1789 mm_segment_t old_fs;
1790
1791 old_fs = get_fs();
1792 set_fs(KERNEL_DS);
1793 list_for_each(p, &cache_chain) {
1794 kmem_cache_t *pc = list_entry(p, kmem_cache_t, next);
1795 char tmp;
1796 /* This happens when the module gets unloaded and doesn't
1797 destroy its slab cache and noone else reuses the vmalloc
1798 area of the module. Print a warning. */
1799 if (__get_user(tmp,pc->name)) {
1800 printk("SLAB: cache with size %d has lost its name\n",
1801 pc->objsize);
1802 continue;
1803 }
1804 if (!strcmp(pc->name,name)) {
1805 printk("kmem_cache_create: duplicate cache %s\n",name);
1806 up(&cache_chain_sem);
1807 unlock_cpu_hotplug();
1808 BUG();
1809 }
1810 }
1811 set_fs(old_fs);
1812 }
1813
1814 /* cache setup completed, link it into the list */
1815 list_add(&cachep->next, &cache_chain);
1816 up(&cache_chain_sem);
1817 unlock_cpu_hotplug();
1818opps:
1819 if (!cachep && (flags & SLAB_PANIC))
1820 panic("kmem_cache_create(): failed to create slab `%s'\n",
1821 name);
1822 return cachep;
1823}
1824EXPORT_SYMBOL(kmem_cache_create);
1825
1826#if DEBUG
1827static void check_irq_off(void)
1828{
1829 BUG_ON(!irqs_disabled());
1830}
1831
1832static void check_irq_on(void)
1833{
1834 BUG_ON(irqs_disabled());
1835}
1836
1837static void check_spinlock_acquired(kmem_cache_t *cachep)
1838{
1839#ifdef CONFIG_SMP
1840 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07001841 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001842#endif
1843}
Christoph Lametere498be72005-09-09 13:03:32 -07001844
1845static inline void check_spinlock_acquired_node(kmem_cache_t *cachep, int node)
1846{
1847#ifdef CONFIG_SMP
1848 check_irq_off();
1849 assert_spin_locked(&cachep->nodelists[node]->list_lock);
1850#endif
1851}
1852
Linus Torvalds1da177e2005-04-16 15:20:36 -07001853#else
1854#define check_irq_off() do { } while(0)
1855#define check_irq_on() do { } while(0)
1856#define check_spinlock_acquired(x) do { } while(0)
Christoph Lametere498be72005-09-09 13:03:32 -07001857#define check_spinlock_acquired_node(x, y) do { } while(0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001858#endif
1859
1860/*
1861 * Waits for all CPUs to execute func().
1862 */
1863static void smp_call_function_all_cpus(void (*func) (void *arg), void *arg)
1864{
1865 check_irq_on();
1866 preempt_disable();
1867
1868 local_irq_disable();
1869 func(arg);
1870 local_irq_enable();
1871
1872 if (smp_call_function(func, arg, 1, 1))
1873 BUG();
1874
1875 preempt_enable();
1876}
1877
1878static void drain_array_locked(kmem_cache_t* cachep,
Christoph Lametere498be72005-09-09 13:03:32 -07001879 struct array_cache *ac, int force, int node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001880
1881static void do_drain(void *arg)
1882{
1883 kmem_cache_t *cachep = (kmem_cache_t*)arg;
1884 struct array_cache *ac;
1885
1886 check_irq_off();
1887 ac = ac_data(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07001888 spin_lock(&cachep->nodelists[numa_node_id()]->list_lock);
1889 free_block(cachep, ac->entry, ac->avail);
1890 spin_unlock(&cachep->nodelists[numa_node_id()]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001891 ac->avail = 0;
1892}
1893
1894static void drain_cpu_caches(kmem_cache_t *cachep)
1895{
Christoph Lametere498be72005-09-09 13:03:32 -07001896 struct kmem_list3 *l3;
1897 int node;
1898
Linus Torvalds1da177e2005-04-16 15:20:36 -07001899 smp_call_function_all_cpus(do_drain, cachep);
1900 check_irq_on();
1901 spin_lock_irq(&cachep->spinlock);
Christoph Lametere498be72005-09-09 13:03:32 -07001902 for_each_online_node(node) {
1903 l3 = cachep->nodelists[node];
1904 if (l3) {
1905 spin_lock(&l3->list_lock);
1906 drain_array_locked(cachep, l3->shared, 1, node);
1907 spin_unlock(&l3->list_lock);
1908 if (l3->alien)
1909 drain_alien_cache(cachep, l3);
1910 }
1911 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001912 spin_unlock_irq(&cachep->spinlock);
1913}
1914
Christoph Lametere498be72005-09-09 13:03:32 -07001915static int __node_shrink(kmem_cache_t *cachep, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001916{
1917 struct slab *slabp;
Christoph Lametere498be72005-09-09 13:03:32 -07001918 struct kmem_list3 *l3 = cachep->nodelists[node];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001919 int ret;
1920
Christoph Lametere498be72005-09-09 13:03:32 -07001921 for (;;) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001922 struct list_head *p;
1923
Christoph Lametere498be72005-09-09 13:03:32 -07001924 p = l3->slabs_free.prev;
1925 if (p == &l3->slabs_free)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001926 break;
1927
Christoph Lametere498be72005-09-09 13:03:32 -07001928 slabp = list_entry(l3->slabs_free.prev, struct slab, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001929#if DEBUG
1930 if (slabp->inuse)
1931 BUG();
1932#endif
1933 list_del(&slabp->list);
1934
Christoph Lametere498be72005-09-09 13:03:32 -07001935 l3->free_objects -= cachep->num;
1936 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001937 slab_destroy(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07001938 spin_lock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001939 }
Christoph Lametere498be72005-09-09 13:03:32 -07001940 ret = !list_empty(&l3->slabs_full) ||
1941 !list_empty(&l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001942 return ret;
1943}
1944
Christoph Lametere498be72005-09-09 13:03:32 -07001945static int __cache_shrink(kmem_cache_t *cachep)
1946{
1947 int ret = 0, i = 0;
1948 struct kmem_list3 *l3;
1949
1950 drain_cpu_caches(cachep);
1951
1952 check_irq_on();
1953 for_each_online_node(i) {
1954 l3 = cachep->nodelists[i];
1955 if (l3) {
1956 spin_lock_irq(&l3->list_lock);
1957 ret += __node_shrink(cachep, i);
1958 spin_unlock_irq(&l3->list_lock);
1959 }
1960 }
1961 return (ret ? 1 : 0);
1962}
1963
Linus Torvalds1da177e2005-04-16 15:20:36 -07001964/**
1965 * kmem_cache_shrink - Shrink a cache.
1966 * @cachep: The cache to shrink.
1967 *
1968 * Releases as many slabs as possible for a cache.
1969 * To help debugging, a zero exit status indicates all slabs were released.
1970 */
1971int kmem_cache_shrink(kmem_cache_t *cachep)
1972{
1973 if (!cachep || in_interrupt())
1974 BUG();
1975
1976 return __cache_shrink(cachep);
1977}
1978EXPORT_SYMBOL(kmem_cache_shrink);
1979
1980/**
1981 * kmem_cache_destroy - delete a cache
1982 * @cachep: the cache to destroy
1983 *
1984 * Remove a kmem_cache_t object from the slab cache.
1985 * Returns 0 on success.
1986 *
1987 * It is expected this function will be called by a module when it is
1988 * unloaded. This will remove the cache completely, and avoid a duplicate
1989 * cache being allocated each time a module is loaded and unloaded, if the
1990 * module doesn't have persistent in-kernel storage across loads and unloads.
1991 *
1992 * The cache must be empty before calling this function.
1993 *
1994 * The caller must guarantee that noone will allocate memory from the cache
1995 * during the kmem_cache_destroy().
1996 */
1997int kmem_cache_destroy(kmem_cache_t * cachep)
1998{
1999 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07002000 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002001
2002 if (!cachep || in_interrupt())
2003 BUG();
2004
2005 /* Don't let CPUs to come and go */
2006 lock_cpu_hotplug();
2007
2008 /* Find the cache in the chain of caches. */
2009 down(&cache_chain_sem);
2010 /*
2011 * the chain is never empty, cache_cache is never destroyed
2012 */
2013 list_del(&cachep->next);
2014 up(&cache_chain_sem);
2015
2016 if (__cache_shrink(cachep)) {
2017 slab_error(cachep, "Can't free all objects");
2018 down(&cache_chain_sem);
2019 list_add(&cachep->next,&cache_chain);
2020 up(&cache_chain_sem);
2021 unlock_cpu_hotplug();
2022 return 1;
2023 }
2024
2025 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
Paul E. McKenneyfbd568a3e2005-05-01 08:59:04 -07002026 synchronize_rcu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002027
Christoph Lametere498be72005-09-09 13:03:32 -07002028 for_each_online_cpu(i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002029 kfree(cachep->array[i]);
2030
2031 /* NUMA: free the list3 structures */
Christoph Lametere498be72005-09-09 13:03:32 -07002032 for_each_online_node(i) {
2033 if ((l3 = cachep->nodelists[i])) {
2034 kfree(l3->shared);
2035 free_alien_cache(l3->alien);
2036 kfree(l3);
2037 }
2038 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002039 kmem_cache_free(&cache_cache, cachep);
2040
2041 unlock_cpu_hotplug();
2042
2043 return 0;
2044}
2045EXPORT_SYMBOL(kmem_cache_destroy);
2046
2047/* Get the memory for a slab management obj. */
Christoph Lametere498be72005-09-09 13:03:32 -07002048static struct slab* alloc_slabmgmt(kmem_cache_t *cachep, void *objp,
2049 int colour_off, unsigned int __nocast local_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002050{
2051 struct slab *slabp;
2052
2053 if (OFF_SLAB(cachep)) {
2054 /* Slab management obj is off-slab. */
2055 slabp = kmem_cache_alloc(cachep->slabp_cache, local_flags);
2056 if (!slabp)
2057 return NULL;
2058 } else {
2059 slabp = objp+colour_off;
2060 colour_off += cachep->slab_size;
2061 }
2062 slabp->inuse = 0;
2063 slabp->colouroff = colour_off;
2064 slabp->s_mem = objp+colour_off;
2065
2066 return slabp;
2067}
2068
2069static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2070{
2071 return (kmem_bufctl_t *)(slabp+1);
2072}
2073
2074static void cache_init_objs(kmem_cache_t *cachep,
2075 struct slab *slabp, unsigned long ctor_flags)
2076{
2077 int i;
2078
2079 for (i = 0; i < cachep->num; i++) {
Christoph Lametere498be72005-09-09 13:03:32 -07002080 void *objp = slabp->s_mem+cachep->objsize*i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002081#if DEBUG
2082 /* need to poison the objs? */
2083 if (cachep->flags & SLAB_POISON)
2084 poison_obj(cachep, objp, POISON_FREE);
2085 if (cachep->flags & SLAB_STORE_USER)
2086 *dbg_userword(cachep, objp) = NULL;
2087
2088 if (cachep->flags & SLAB_RED_ZONE) {
2089 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2090 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2091 }
2092 /*
2093 * Constructors are not allowed to allocate memory from
2094 * the same cache which they are a constructor for.
2095 * Otherwise, deadlock. They must also be threaded.
2096 */
2097 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2098 cachep->ctor(objp+obj_dbghead(cachep), cachep, ctor_flags);
2099
2100 if (cachep->flags & SLAB_RED_ZONE) {
2101 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2102 slab_error(cachep, "constructor overwrote the"
2103 " end of an object");
2104 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2105 slab_error(cachep, "constructor overwrote the"
2106 " start of an object");
2107 }
2108 if ((cachep->objsize % PAGE_SIZE) == 0 && OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2109 kernel_map_pages(virt_to_page(objp), cachep->objsize/PAGE_SIZE, 0);
2110#else
2111 if (cachep->ctor)
2112 cachep->ctor(objp, cachep, ctor_flags);
2113#endif
2114 slab_bufctl(slabp)[i] = i+1;
2115 }
2116 slab_bufctl(slabp)[i-1] = BUFCTL_END;
2117 slabp->free = 0;
2118}
2119
2120static void kmem_flagcheck(kmem_cache_t *cachep, unsigned int flags)
2121{
2122 if (flags & SLAB_DMA) {
2123 if (!(cachep->gfpflags & GFP_DMA))
2124 BUG();
2125 } else {
2126 if (cachep->gfpflags & GFP_DMA)
2127 BUG();
2128 }
2129}
2130
2131static void set_slab_attr(kmem_cache_t *cachep, struct slab *slabp, void *objp)
2132{
2133 int i;
2134 struct page *page;
2135
2136 /* Nasty!!!!!! I hope this is OK. */
2137 i = 1 << cachep->gfporder;
2138 page = virt_to_page(objp);
2139 do {
2140 SET_PAGE_CACHE(page, cachep);
2141 SET_PAGE_SLAB(page, slabp);
2142 page++;
2143 } while (--i);
2144}
2145
2146/*
2147 * Grow (by 1) the number of slabs within a cache. This is called by
2148 * kmem_cache_alloc() when there are no active objs left in a cache.
2149 */
2150static int cache_grow(kmem_cache_t *cachep, unsigned int __nocast flags, int nodeid)
2151{
2152 struct slab *slabp;
2153 void *objp;
2154 size_t offset;
2155 unsigned int local_flags;
2156 unsigned long ctor_flags;
Christoph Lametere498be72005-09-09 13:03:32 -07002157 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002158
2159 /* Be lazy and only check for valid flags here,
2160 * keeping it out of the critical path in kmem_cache_alloc().
2161 */
2162 if (flags & ~(SLAB_DMA|SLAB_LEVEL_MASK|SLAB_NO_GROW))
2163 BUG();
2164 if (flags & SLAB_NO_GROW)
2165 return 0;
2166
2167 ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2168 local_flags = (flags & SLAB_LEVEL_MASK);
2169 if (!(local_flags & __GFP_WAIT))
2170 /*
2171 * Not allowed to sleep. Need to tell a constructor about
2172 * this - it might need to know...
2173 */
2174 ctor_flags |= SLAB_CTOR_ATOMIC;
2175
2176 /* About to mess with non-constant members - lock. */
2177 check_irq_off();
2178 spin_lock(&cachep->spinlock);
2179
2180 /* Get colour for the slab, and cal the next value. */
2181 offset = cachep->colour_next;
2182 cachep->colour_next++;
2183 if (cachep->colour_next >= cachep->colour)
2184 cachep->colour_next = 0;
2185 offset *= cachep->colour_off;
2186
2187 spin_unlock(&cachep->spinlock);
2188
Christoph Lametere498be72005-09-09 13:03:32 -07002189 check_irq_off();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002190 if (local_flags & __GFP_WAIT)
2191 local_irq_enable();
2192
2193 /*
2194 * The test for missing atomic flag is performed here, rather than
2195 * the more obvious place, simply to reduce the critical path length
2196 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2197 * will eventually be caught here (where it matters).
2198 */
2199 kmem_flagcheck(cachep, flags);
2200
Christoph Lametere498be72005-09-09 13:03:32 -07002201 /* Get mem for the objs.
2202 * Attempt to allocate a physical page from 'nodeid',
2203 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002204 if (!(objp = kmem_getpages(cachep, flags, nodeid)))
2205 goto failed;
2206
2207 /* Get slab management. */
2208 if (!(slabp = alloc_slabmgmt(cachep, objp, offset, local_flags)))
2209 goto opps1;
2210
Christoph Lametere498be72005-09-09 13:03:32 -07002211 slabp->nodeid = nodeid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002212 set_slab_attr(cachep, slabp, objp);
2213
2214 cache_init_objs(cachep, slabp, ctor_flags);
2215
2216 if (local_flags & __GFP_WAIT)
2217 local_irq_disable();
2218 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002219 l3 = cachep->nodelists[nodeid];
2220 spin_lock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002221
2222 /* Make slab active. */
Christoph Lametere498be72005-09-09 13:03:32 -07002223 list_add_tail(&slabp->list, &(l3->slabs_free));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002224 STATS_INC_GROWN(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002225 l3->free_objects += cachep->num;
2226 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002227 return 1;
2228opps1:
2229 kmem_freepages(cachep, objp);
2230failed:
2231 if (local_flags & __GFP_WAIT)
2232 local_irq_disable();
2233 return 0;
2234}
2235
2236#if DEBUG
2237
2238/*
2239 * Perform extra freeing checks:
2240 * - detect bad pointers.
2241 * - POISON/RED_ZONE checking
2242 * - destructor calls, for caches with POISON+dtor
2243 */
2244static void kfree_debugcheck(const void *objp)
2245{
2246 struct page *page;
2247
2248 if (!virt_addr_valid(objp)) {
2249 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2250 (unsigned long)objp);
2251 BUG();
2252 }
2253 page = virt_to_page(objp);
2254 if (!PageSlab(page)) {
2255 printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n", (unsigned long)objp);
2256 BUG();
2257 }
2258}
2259
2260static void *cache_free_debugcheck(kmem_cache_t *cachep, void *objp,
2261 void *caller)
2262{
2263 struct page *page;
2264 unsigned int objnr;
2265 struct slab *slabp;
2266
2267 objp -= obj_dbghead(cachep);
2268 kfree_debugcheck(objp);
2269 page = virt_to_page(objp);
2270
2271 if (GET_PAGE_CACHE(page) != cachep) {
2272 printk(KERN_ERR "mismatch in kmem_cache_free: expected cache %p, got %p\n",
2273 GET_PAGE_CACHE(page),cachep);
2274 printk(KERN_ERR "%p is %s.\n", cachep, cachep->name);
2275 printk(KERN_ERR "%p is %s.\n", GET_PAGE_CACHE(page), GET_PAGE_CACHE(page)->name);
2276 WARN_ON(1);
2277 }
2278 slabp = GET_PAGE_SLAB(page);
2279
2280 if (cachep->flags & SLAB_RED_ZONE) {
2281 if (*dbg_redzone1(cachep, objp) != RED_ACTIVE || *dbg_redzone2(cachep, objp) != RED_ACTIVE) {
2282 slab_error(cachep, "double free, or memory outside"
2283 " object was overwritten");
2284 printk(KERN_ERR "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n",
2285 objp, *dbg_redzone1(cachep, objp), *dbg_redzone2(cachep, objp));
2286 }
2287 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2288 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2289 }
2290 if (cachep->flags & SLAB_STORE_USER)
2291 *dbg_userword(cachep, objp) = caller;
2292
2293 objnr = (objp-slabp->s_mem)/cachep->objsize;
2294
2295 BUG_ON(objnr >= cachep->num);
2296 BUG_ON(objp != slabp->s_mem + objnr*cachep->objsize);
2297
2298 if (cachep->flags & SLAB_DEBUG_INITIAL) {
2299 /* Need to call the slab's constructor so the
2300 * caller can perform a verify of its state (debugging).
2301 * Called without the cache-lock held.
2302 */
2303 cachep->ctor(objp+obj_dbghead(cachep),
2304 cachep, SLAB_CTOR_CONSTRUCTOR|SLAB_CTOR_VERIFY);
2305 }
2306 if (cachep->flags & SLAB_POISON && cachep->dtor) {
2307 /* we want to cache poison the object,
2308 * call the destruction callback
2309 */
2310 cachep->dtor(objp+obj_dbghead(cachep), cachep, 0);
2311 }
2312 if (cachep->flags & SLAB_POISON) {
2313#ifdef CONFIG_DEBUG_PAGEALLOC
2314 if ((cachep->objsize % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) {
2315 store_stackinfo(cachep, objp, (unsigned long)caller);
2316 kernel_map_pages(virt_to_page(objp), cachep->objsize/PAGE_SIZE, 0);
2317 } else {
2318 poison_obj(cachep, objp, POISON_FREE);
2319 }
2320#else
2321 poison_obj(cachep, objp, POISON_FREE);
2322#endif
2323 }
2324 return objp;
2325}
2326
2327static void check_slabp(kmem_cache_t *cachep, struct slab *slabp)
2328{
2329 kmem_bufctl_t i;
2330 int entries = 0;
2331
Linus Torvalds1da177e2005-04-16 15:20:36 -07002332 /* Check slab's freelist to see if this obj is there. */
2333 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2334 entries++;
2335 if (entries > cachep->num || i >= cachep->num)
2336 goto bad;
2337 }
2338 if (entries != cachep->num - slabp->inuse) {
2339bad:
2340 printk(KERN_ERR "slab: Internal list corruption detected in cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2341 cachep->name, cachep->num, slabp, slabp->inuse);
2342 for (i=0;i<sizeof(slabp)+cachep->num*sizeof(kmem_bufctl_t);i++) {
2343 if ((i%16)==0)
2344 printk("\n%03x:", i);
2345 printk(" %02x", ((unsigned char*)slabp)[i]);
2346 }
2347 printk("\n");
2348 BUG();
2349 }
2350}
2351#else
2352#define kfree_debugcheck(x) do { } while(0)
2353#define cache_free_debugcheck(x,objp,z) (objp)
2354#define check_slabp(x,y) do { } while(0)
2355#endif
2356
2357static void *cache_alloc_refill(kmem_cache_t *cachep, unsigned int __nocast flags)
2358{
2359 int batchcount;
2360 struct kmem_list3 *l3;
2361 struct array_cache *ac;
2362
2363 check_irq_off();
2364 ac = ac_data(cachep);
2365retry:
2366 batchcount = ac->batchcount;
2367 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2368 /* if there was little recent activity on this
2369 * cache, then perform only a partial refill.
2370 * Otherwise we could generate refill bouncing.
2371 */
2372 batchcount = BATCHREFILL_LIMIT;
2373 }
Christoph Lametere498be72005-09-09 13:03:32 -07002374 l3 = cachep->nodelists[numa_node_id()];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002375
Christoph Lametere498be72005-09-09 13:03:32 -07002376 BUG_ON(ac->avail > 0 || !l3);
2377 spin_lock(&l3->list_lock);
2378
Linus Torvalds1da177e2005-04-16 15:20:36 -07002379 if (l3->shared) {
2380 struct array_cache *shared_array = l3->shared;
2381 if (shared_array->avail) {
2382 if (batchcount > shared_array->avail)
2383 batchcount = shared_array->avail;
2384 shared_array->avail -= batchcount;
2385 ac->avail = batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07002386 memcpy(ac->entry,
2387 &(shared_array->entry[shared_array->avail]),
2388 sizeof(void*)*batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002389 shared_array->touched = 1;
2390 goto alloc_done;
2391 }
2392 }
2393 while (batchcount > 0) {
2394 struct list_head *entry;
2395 struct slab *slabp;
2396 /* Get slab alloc is to come from. */
2397 entry = l3->slabs_partial.next;
2398 if (entry == &l3->slabs_partial) {
2399 l3->free_touched = 1;
2400 entry = l3->slabs_free.next;
2401 if (entry == &l3->slabs_free)
2402 goto must_grow;
2403 }
2404
2405 slabp = list_entry(entry, struct slab, list);
2406 check_slabp(cachep, slabp);
2407 check_spinlock_acquired(cachep);
2408 while (slabp->inuse < cachep->num && batchcount--) {
2409 kmem_bufctl_t next;
2410 STATS_INC_ALLOCED(cachep);
2411 STATS_INC_ACTIVE(cachep);
2412 STATS_SET_HIGH(cachep);
2413
2414 /* get obj pointer */
Christoph Lametere498be72005-09-09 13:03:32 -07002415 ac->entry[ac->avail++] = slabp->s_mem +
2416 slabp->free*cachep->objsize;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002417
2418 slabp->inuse++;
2419 next = slab_bufctl(slabp)[slabp->free];
2420#if DEBUG
2421 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2422#endif
2423 slabp->free = next;
2424 }
2425 check_slabp(cachep, slabp);
2426
2427 /* move slabp to correct slabp list: */
2428 list_del(&slabp->list);
2429 if (slabp->free == BUFCTL_END)
2430 list_add(&slabp->list, &l3->slabs_full);
2431 else
2432 list_add(&slabp->list, &l3->slabs_partial);
2433 }
2434
2435must_grow:
2436 l3->free_objects -= ac->avail;
2437alloc_done:
Christoph Lametere498be72005-09-09 13:03:32 -07002438 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002439
2440 if (unlikely(!ac->avail)) {
2441 int x;
Christoph Lametere498be72005-09-09 13:03:32 -07002442 x = cache_grow(cachep, flags, numa_node_id());
2443
Linus Torvalds1da177e2005-04-16 15:20:36 -07002444 // cache_grow can reenable interrupts, then ac could change.
2445 ac = ac_data(cachep);
2446 if (!x && ac->avail == 0) // no objects in sight? abort
2447 return NULL;
2448
2449 if (!ac->avail) // objects refilled by interrupt?
2450 goto retry;
2451 }
2452 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07002453 return ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002454}
2455
2456static inline void
2457cache_alloc_debugcheck_before(kmem_cache_t *cachep, unsigned int __nocast flags)
2458{
2459 might_sleep_if(flags & __GFP_WAIT);
2460#if DEBUG
2461 kmem_flagcheck(cachep, flags);
2462#endif
2463}
2464
2465#if DEBUG
2466static void *
2467cache_alloc_debugcheck_after(kmem_cache_t *cachep,
Alexey Dobriyan0db925a2005-07-07 17:56:58 -07002468 unsigned int __nocast flags, void *objp, void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002469{
2470 if (!objp)
2471 return objp;
2472 if (cachep->flags & SLAB_POISON) {
2473#ifdef CONFIG_DEBUG_PAGEALLOC
2474 if ((cachep->objsize % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
2475 kernel_map_pages(virt_to_page(objp), cachep->objsize/PAGE_SIZE, 1);
2476 else
2477 check_poison_obj(cachep, objp);
2478#else
2479 check_poison_obj(cachep, objp);
2480#endif
2481 poison_obj(cachep, objp, POISON_INUSE);
2482 }
2483 if (cachep->flags & SLAB_STORE_USER)
2484 *dbg_userword(cachep, objp) = caller;
2485
2486 if (cachep->flags & SLAB_RED_ZONE) {
2487 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE || *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2488 slab_error(cachep, "double free, or memory outside"
2489 " object was overwritten");
2490 printk(KERN_ERR "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n",
2491 objp, *dbg_redzone1(cachep, objp), *dbg_redzone2(cachep, objp));
2492 }
2493 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2494 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2495 }
2496 objp += obj_dbghead(cachep);
2497 if (cachep->ctor && cachep->flags & SLAB_POISON) {
2498 unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2499
2500 if (!(flags & __GFP_WAIT))
2501 ctor_flags |= SLAB_CTOR_ATOMIC;
2502
2503 cachep->ctor(objp, cachep, ctor_flags);
2504 }
2505 return objp;
2506}
2507#else
2508#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2509#endif
2510
2511
2512static inline void *__cache_alloc(kmem_cache_t *cachep, unsigned int __nocast flags)
2513{
2514 unsigned long save_flags;
2515 void* objp;
2516 struct array_cache *ac;
2517
2518 cache_alloc_debugcheck_before(cachep, flags);
2519
2520 local_irq_save(save_flags);
2521 ac = ac_data(cachep);
2522 if (likely(ac->avail)) {
2523 STATS_INC_ALLOCHIT(cachep);
2524 ac->touched = 1;
Christoph Lametere498be72005-09-09 13:03:32 -07002525 objp = ac->entry[--ac->avail];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002526 } else {
2527 STATS_INC_ALLOCMISS(cachep);
2528 objp = cache_alloc_refill(cachep, flags);
2529 }
2530 local_irq_restore(save_flags);
Eric Dumazet34342e82005-09-03 15:55:06 -07002531 objp = cache_alloc_debugcheck_after(cachep, flags, objp,
2532 __builtin_return_address(0));
2533 prefetchw(objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002534 return objp;
2535}
2536
Christoph Lametere498be72005-09-09 13:03:32 -07002537#ifdef CONFIG_NUMA
2538/*
2539 * A interface to enable slab creation on nodeid
Linus Torvalds1da177e2005-04-16 15:20:36 -07002540 */
Christoph Lametere498be72005-09-09 13:03:32 -07002541static void *__cache_alloc_node(kmem_cache_t *cachep, int flags, int nodeid)
2542{
2543 struct list_head *entry;
2544 struct slab *slabp;
2545 struct kmem_list3 *l3;
2546 void *obj;
2547 kmem_bufctl_t next;
2548 int x;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002549
Christoph Lametere498be72005-09-09 13:03:32 -07002550 l3 = cachep->nodelists[nodeid];
2551 BUG_ON(!l3);
2552
2553retry:
2554 spin_lock(&l3->list_lock);
2555 entry = l3->slabs_partial.next;
2556 if (entry == &l3->slabs_partial) {
2557 l3->free_touched = 1;
2558 entry = l3->slabs_free.next;
2559 if (entry == &l3->slabs_free)
2560 goto must_grow;
2561 }
2562
2563 slabp = list_entry(entry, struct slab, list);
2564 check_spinlock_acquired_node(cachep, nodeid);
2565 check_slabp(cachep, slabp);
2566
2567 STATS_INC_NODEALLOCS(cachep);
2568 STATS_INC_ACTIVE(cachep);
2569 STATS_SET_HIGH(cachep);
2570
2571 BUG_ON(slabp->inuse == cachep->num);
2572
2573 /* get obj pointer */
2574 obj = slabp->s_mem + slabp->free*cachep->objsize;
2575 slabp->inuse++;
2576 next = slab_bufctl(slabp)[slabp->free];
2577#if DEBUG
2578 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2579#endif
2580 slabp->free = next;
2581 check_slabp(cachep, slabp);
2582 l3->free_objects--;
2583 /* move slabp to correct slabp list: */
2584 list_del(&slabp->list);
2585
2586 if (slabp->free == BUFCTL_END) {
2587 list_add(&slabp->list, &l3->slabs_full);
2588 } else {
2589 list_add(&slabp->list, &l3->slabs_partial);
2590 }
2591
2592 spin_unlock(&l3->list_lock);
2593 goto done;
2594
2595must_grow:
2596 spin_unlock(&l3->list_lock);
2597 x = cache_grow(cachep, flags, nodeid);
2598
2599 if (!x)
2600 return NULL;
2601
2602 goto retry;
2603done:
2604 return obj;
2605}
2606#endif
2607
2608/*
2609 * Caller needs to acquire correct kmem_list's list_lock
2610 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002611static void free_block(kmem_cache_t *cachep, void **objpp, int nr_objects)
2612{
2613 int i;
Christoph Lametere498be72005-09-09 13:03:32 -07002614 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002615
2616 for (i = 0; i < nr_objects; i++) {
2617 void *objp = objpp[i];
2618 struct slab *slabp;
2619 unsigned int objnr;
Christoph Lametere498be72005-09-09 13:03:32 -07002620 int nodeid = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002621
2622 slabp = GET_PAGE_SLAB(virt_to_page(objp));
Christoph Lametere498be72005-09-09 13:03:32 -07002623 nodeid = slabp->nodeid;
2624 l3 = cachep->nodelists[nodeid];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002625 list_del(&slabp->list);
2626 objnr = (objp - slabp->s_mem) / cachep->objsize;
Christoph Lametere498be72005-09-09 13:03:32 -07002627 check_spinlock_acquired_node(cachep, nodeid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002628 check_slabp(cachep, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07002629
2630
Linus Torvalds1da177e2005-04-16 15:20:36 -07002631#if DEBUG
2632 if (slab_bufctl(slabp)[objnr] != BUFCTL_FREE) {
Christoph Lametere498be72005-09-09 13:03:32 -07002633 printk(KERN_ERR "slab: double free detected in cache "
2634 "'%s', objp %p\n", cachep->name, objp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002635 BUG();
2636 }
2637#endif
2638 slab_bufctl(slabp)[objnr] = slabp->free;
2639 slabp->free = objnr;
2640 STATS_DEC_ACTIVE(cachep);
2641 slabp->inuse--;
Christoph Lametere498be72005-09-09 13:03:32 -07002642 l3->free_objects++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002643 check_slabp(cachep, slabp);
2644
2645 /* fixup slab chains */
2646 if (slabp->inuse == 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07002647 if (l3->free_objects > l3->free_limit) {
2648 l3->free_objects -= cachep->num;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002649 slab_destroy(cachep, slabp);
2650 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07002651 list_add(&slabp->list, &l3->slabs_free);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002652 }
2653 } else {
2654 /* Unconditionally move a slab to the end of the
2655 * partial list on free - maximum time for the
2656 * other objects to be freed, too.
2657 */
Christoph Lametere498be72005-09-09 13:03:32 -07002658 list_add_tail(&slabp->list, &l3->slabs_partial);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002659 }
2660 }
2661}
2662
2663static void cache_flusharray(kmem_cache_t *cachep, struct array_cache *ac)
2664{
2665 int batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07002666 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002667
2668 batchcount = ac->batchcount;
2669#if DEBUG
2670 BUG_ON(!batchcount || batchcount > ac->avail);
2671#endif
2672 check_irq_off();
Christoph Lametere498be72005-09-09 13:03:32 -07002673 l3 = cachep->nodelists[numa_node_id()];
2674 spin_lock(&l3->list_lock);
2675 if (l3->shared) {
2676 struct array_cache *shared_array = l3->shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002677 int max = shared_array->limit-shared_array->avail;
2678 if (max) {
2679 if (batchcount > max)
2680 batchcount = max;
Christoph Lametere498be72005-09-09 13:03:32 -07002681 memcpy(&(shared_array->entry[shared_array->avail]),
2682 ac->entry,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002683 sizeof(void*)*batchcount);
2684 shared_array->avail += batchcount;
2685 goto free_done;
2686 }
2687 }
2688
Christoph Lametere498be72005-09-09 13:03:32 -07002689 free_block(cachep, ac->entry, batchcount);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002690free_done:
2691#if STATS
2692 {
2693 int i = 0;
2694 struct list_head *p;
2695
Christoph Lametere498be72005-09-09 13:03:32 -07002696 p = l3->slabs_free.next;
2697 while (p != &(l3->slabs_free)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002698 struct slab *slabp;
2699
2700 slabp = list_entry(p, struct slab, list);
2701 BUG_ON(slabp->inuse);
2702
2703 i++;
2704 p = p->next;
2705 }
2706 STATS_SET_FREEABLE(cachep, i);
2707 }
2708#endif
Christoph Lametere498be72005-09-09 13:03:32 -07002709 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002710 ac->avail -= batchcount;
Christoph Lametere498be72005-09-09 13:03:32 -07002711 memmove(ac->entry, &(ac->entry[batchcount]),
Linus Torvalds1da177e2005-04-16 15:20:36 -07002712 sizeof(void*)*ac->avail);
2713}
2714
Christoph Lametere498be72005-09-09 13:03:32 -07002715
Linus Torvalds1da177e2005-04-16 15:20:36 -07002716/*
2717 * __cache_free
2718 * Release an obj back to its cache. If the obj has a constructed
2719 * state, it must be in this state _before_ it is released.
2720 *
2721 * Called with disabled ints.
2722 */
2723static inline void __cache_free(kmem_cache_t *cachep, void *objp)
2724{
2725 struct array_cache *ac = ac_data(cachep);
2726
2727 check_irq_off();
2728 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
2729
Christoph Lametere498be72005-09-09 13:03:32 -07002730 /* Make sure we are not freeing a object from another
2731 * node to the array cache on this cpu.
2732 */
2733#ifdef CONFIG_NUMA
2734 {
2735 struct slab *slabp;
2736 slabp = GET_PAGE_SLAB(virt_to_page(objp));
2737 if (unlikely(slabp->nodeid != numa_node_id())) {
2738 struct array_cache *alien = NULL;
2739 int nodeid = slabp->nodeid;
2740 struct kmem_list3 *l3 = cachep->nodelists[numa_node_id()];
2741
2742 STATS_INC_NODEFREES(cachep);
2743 if (l3->alien && l3->alien[nodeid]) {
2744 alien = l3->alien[nodeid];
2745 spin_lock(&alien->lock);
2746 if (unlikely(alien->avail == alien->limit))
2747 __drain_alien_cache(cachep,
2748 alien, nodeid);
2749 alien->entry[alien->avail++] = objp;
2750 spin_unlock(&alien->lock);
2751 } else {
2752 spin_lock(&(cachep->nodelists[nodeid])->
2753 list_lock);
2754 free_block(cachep, &objp, 1);
2755 spin_unlock(&(cachep->nodelists[nodeid])->
2756 list_lock);
2757 }
2758 return;
2759 }
2760 }
2761#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002762 if (likely(ac->avail < ac->limit)) {
2763 STATS_INC_FREEHIT(cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07002764 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002765 return;
2766 } else {
2767 STATS_INC_FREEMISS(cachep);
2768 cache_flusharray(cachep, ac);
Christoph Lametere498be72005-09-09 13:03:32 -07002769 ac->entry[ac->avail++] = objp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002770 }
2771}
2772
2773/**
2774 * kmem_cache_alloc - Allocate an object
2775 * @cachep: The cache to allocate from.
2776 * @flags: See kmalloc().
2777 *
2778 * Allocate an object from this cache. The flags are only relevant
2779 * if the cache has no available objects.
2780 */
2781void *kmem_cache_alloc(kmem_cache_t *cachep, unsigned int __nocast flags)
2782{
2783 return __cache_alloc(cachep, flags);
2784}
2785EXPORT_SYMBOL(kmem_cache_alloc);
2786
2787/**
2788 * kmem_ptr_validate - check if an untrusted pointer might
2789 * be a slab entry.
2790 * @cachep: the cache we're checking against
2791 * @ptr: pointer to validate
2792 *
2793 * This verifies that the untrusted pointer looks sane:
2794 * it is _not_ a guarantee that the pointer is actually
2795 * part of the slab cache in question, but it at least
2796 * validates that the pointer can be dereferenced and
2797 * looks half-way sane.
2798 *
2799 * Currently only used for dentry validation.
2800 */
2801int fastcall kmem_ptr_validate(kmem_cache_t *cachep, void *ptr)
2802{
2803 unsigned long addr = (unsigned long) ptr;
2804 unsigned long min_addr = PAGE_OFFSET;
2805 unsigned long align_mask = BYTES_PER_WORD-1;
2806 unsigned long size = cachep->objsize;
2807 struct page *page;
2808
2809 if (unlikely(addr < min_addr))
2810 goto out;
2811 if (unlikely(addr > (unsigned long)high_memory - size))
2812 goto out;
2813 if (unlikely(addr & align_mask))
2814 goto out;
2815 if (unlikely(!kern_addr_valid(addr)))
2816 goto out;
2817 if (unlikely(!kern_addr_valid(addr + size - 1)))
2818 goto out;
2819 page = virt_to_page(ptr);
2820 if (unlikely(!PageSlab(page)))
2821 goto out;
2822 if (unlikely(GET_PAGE_CACHE(page) != cachep))
2823 goto out;
2824 return 1;
2825out:
2826 return 0;
2827}
2828
2829#ifdef CONFIG_NUMA
2830/**
2831 * kmem_cache_alloc_node - Allocate an object on the specified node
2832 * @cachep: The cache to allocate from.
2833 * @flags: See kmalloc().
2834 * @nodeid: node number of the target node.
2835 *
2836 * Identical to kmem_cache_alloc, except that this function is slow
2837 * and can sleep. And it will allocate memory on the given node, which
2838 * can improve the performance for cpu bound structures.
Christoph Lametere498be72005-09-09 13:03:32 -07002839 * New and improved: it will now make sure that the object gets
2840 * put on the correct node list so that there is no false sharing.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002841 */
Victor Fuscob2d55072005-09-10 00:26:36 -07002842void *kmem_cache_alloc_node(kmem_cache_t *cachep, unsigned int __nocast flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002843{
Christoph Lametere498be72005-09-09 13:03:32 -07002844 unsigned long save_flags;
2845 void *ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002846
Christoph Lametere498be72005-09-09 13:03:32 -07002847 if (nodeid == numa_node_id() || nodeid == -1)
2848 return __cache_alloc(cachep, flags);
Christoph Lameter83b78bd2005-07-06 10:47:07 -07002849
Christoph Lametere498be72005-09-09 13:03:32 -07002850 if (unlikely(!cachep->nodelists[nodeid])) {
2851 /* Fall back to __cache_alloc if we run into trouble */
2852 printk(KERN_WARNING "slab: not allocating in inactive node %d for cache %s\n", nodeid, cachep->name);
2853 return __cache_alloc(cachep,flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002854 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002855
Christoph Lametere498be72005-09-09 13:03:32 -07002856 cache_alloc_debugcheck_before(cachep, flags);
2857 local_irq_save(save_flags);
2858 ptr = __cache_alloc_node(cachep, flags, nodeid);
2859 local_irq_restore(save_flags);
2860 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, __builtin_return_address(0));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002861
Christoph Lametere498be72005-09-09 13:03:32 -07002862 return ptr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002863}
2864EXPORT_SYMBOL(kmem_cache_alloc_node);
2865
Alexey Dobriyan0db925a2005-07-07 17:56:58 -07002866void *kmalloc_node(size_t size, unsigned int __nocast flags, int node)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07002867{
2868 kmem_cache_t *cachep;
2869
2870 cachep = kmem_find_general_cachep(size, flags);
2871 if (unlikely(cachep == NULL))
2872 return NULL;
2873 return kmem_cache_alloc_node(cachep, flags, node);
2874}
2875EXPORT_SYMBOL(kmalloc_node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002876#endif
2877
2878/**
2879 * kmalloc - allocate memory
2880 * @size: how many bytes of memory are required.
2881 * @flags: the type of memory to allocate.
2882 *
2883 * kmalloc is the normal method of allocating memory
2884 * in the kernel.
2885 *
2886 * The @flags argument may be one of:
2887 *
2888 * %GFP_USER - Allocate memory on behalf of user. May sleep.
2889 *
2890 * %GFP_KERNEL - Allocate normal kernel ram. May sleep.
2891 *
2892 * %GFP_ATOMIC - Allocation will not sleep. Use inside interrupt handlers.
2893 *
2894 * Additionally, the %GFP_DMA flag may be set to indicate the memory
2895 * must be suitable for DMA. This can mean different things on different
2896 * platforms. For example, on i386, it means that the memory must come
2897 * from the first 16MB.
2898 */
2899void *__kmalloc(size_t size, unsigned int __nocast flags)
2900{
2901 kmem_cache_t *cachep;
2902
Manfred Spraul97e2bde2005-05-01 08:58:38 -07002903 /* If you want to save a few bytes .text space: replace
2904 * __ with kmem_.
2905 * Then kmalloc uses the uninlined functions instead of the inline
2906 * functions.
2907 */
2908 cachep = __find_general_cachep(size, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002909 if (unlikely(cachep == NULL))
2910 return NULL;
2911 return __cache_alloc(cachep, flags);
2912}
2913EXPORT_SYMBOL(__kmalloc);
2914
2915#ifdef CONFIG_SMP
2916/**
2917 * __alloc_percpu - allocate one copy of the object for every present
2918 * cpu in the system, zeroing them.
2919 * Objects should be dereferenced using the per_cpu_ptr macro only.
2920 *
2921 * @size: how many bytes of memory are required.
2922 * @align: the alignment, which can't be greater than SMP_CACHE_BYTES.
2923 */
2924void *__alloc_percpu(size_t size, size_t align)
2925{
2926 int i;
2927 struct percpu_data *pdata = kmalloc(sizeof (*pdata), GFP_KERNEL);
2928
2929 if (!pdata)
2930 return NULL;
2931
Christoph Lametere498be72005-09-09 13:03:32 -07002932 /*
2933 * Cannot use for_each_online_cpu since a cpu may come online
2934 * and we have no way of figuring out how to fix the array
2935 * that we have allocated then....
2936 */
2937 for_each_cpu(i) {
2938 int node = cpu_to_node(i);
2939
2940 if (node_online(node))
2941 pdata->ptrs[i] = kmalloc_node(size, GFP_KERNEL, node);
2942 else
2943 pdata->ptrs[i] = kmalloc(size, GFP_KERNEL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002944
2945 if (!pdata->ptrs[i])
2946 goto unwind_oom;
2947 memset(pdata->ptrs[i], 0, size);
2948 }
2949
2950 /* Catch derefs w/o wrappers */
2951 return (void *) (~(unsigned long) pdata);
2952
2953unwind_oom:
2954 while (--i >= 0) {
2955 if (!cpu_possible(i))
2956 continue;
2957 kfree(pdata->ptrs[i]);
2958 }
2959 kfree(pdata);
2960 return NULL;
2961}
2962EXPORT_SYMBOL(__alloc_percpu);
2963#endif
2964
2965/**
2966 * kmem_cache_free - Deallocate an object
2967 * @cachep: The cache the allocation was from.
2968 * @objp: The previously allocated object.
2969 *
2970 * Free an object which was previously allocated from this
2971 * cache.
2972 */
2973void kmem_cache_free(kmem_cache_t *cachep, void *objp)
2974{
2975 unsigned long flags;
2976
2977 local_irq_save(flags);
2978 __cache_free(cachep, objp);
2979 local_irq_restore(flags);
2980}
2981EXPORT_SYMBOL(kmem_cache_free);
2982
2983/**
Pekka J Enbergdd392712005-09-06 15:18:31 -07002984 * kzalloc - allocate memory. The memory is set to zero.
2985 * @size: how many bytes of memory are required.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002986 * @flags: the type of memory to allocate.
2987 */
Pekka J Enbergdd392712005-09-06 15:18:31 -07002988void *kzalloc(size_t size, unsigned int __nocast flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002989{
Pekka J Enbergdd392712005-09-06 15:18:31 -07002990 void *ret = kmalloc(size, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002991 if (ret)
Pekka J Enbergdd392712005-09-06 15:18:31 -07002992 memset(ret, 0, size);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002993 return ret;
2994}
Pekka J Enbergdd392712005-09-06 15:18:31 -07002995EXPORT_SYMBOL(kzalloc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002996
2997/**
2998 * kfree - free previously allocated memory
2999 * @objp: pointer returned by kmalloc.
3000 *
Pekka Enberg80e93ef2005-09-09 13:10:16 -07003001 * If @objp is NULL, no operation is performed.
3002 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07003003 * Don't free memory not originally allocated by kmalloc()
3004 * or you will run into trouble.
3005 */
3006void kfree(const void *objp)
3007{
3008 kmem_cache_t *c;
3009 unsigned long flags;
3010
3011 if (unlikely(!objp))
3012 return;
3013 local_irq_save(flags);
3014 kfree_debugcheck(objp);
3015 c = GET_PAGE_CACHE(virt_to_page(objp));
3016 __cache_free(c, (void*)objp);
3017 local_irq_restore(flags);
3018}
3019EXPORT_SYMBOL(kfree);
3020
3021#ifdef CONFIG_SMP
3022/**
3023 * free_percpu - free previously allocated percpu memory
3024 * @objp: pointer returned by alloc_percpu.
3025 *
3026 * Don't free memory not originally allocated by alloc_percpu()
3027 * The complemented objp is to check for that.
3028 */
3029void
3030free_percpu(const void *objp)
3031{
3032 int i;
3033 struct percpu_data *p = (struct percpu_data *) (~(unsigned long) objp);
3034
Christoph Lametere498be72005-09-09 13:03:32 -07003035 /*
3036 * We allocate for all cpus so we cannot use for online cpu here.
3037 */
3038 for_each_cpu(i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003039 kfree(p->ptrs[i]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003040 kfree(p);
3041}
3042EXPORT_SYMBOL(free_percpu);
3043#endif
3044
3045unsigned int kmem_cache_size(kmem_cache_t *cachep)
3046{
3047 return obj_reallen(cachep);
3048}
3049EXPORT_SYMBOL(kmem_cache_size);
3050
Arnaldo Carvalho de Melo19449722005-06-18 22:46:19 -07003051const char *kmem_cache_name(kmem_cache_t *cachep)
3052{
3053 return cachep->name;
3054}
3055EXPORT_SYMBOL_GPL(kmem_cache_name);
3056
Christoph Lametere498be72005-09-09 13:03:32 -07003057/*
3058 * This initializes kmem_list3 for all nodes.
3059 */
3060static int alloc_kmemlist(kmem_cache_t *cachep)
3061{
3062 int node;
3063 struct kmem_list3 *l3;
3064 int err = 0;
3065
3066 for_each_online_node(node) {
3067 struct array_cache *nc = NULL, *new;
3068 struct array_cache **new_alien = NULL;
3069#ifdef CONFIG_NUMA
3070 if (!(new_alien = alloc_alien_cache(node, cachep->limit)))
3071 goto fail;
3072#endif
3073 if (!(new = alloc_arraycache(node, (cachep->shared*
3074 cachep->batchcount), 0xbaadf00d)))
3075 goto fail;
3076 if ((l3 = cachep->nodelists[node])) {
3077
3078 spin_lock_irq(&l3->list_lock);
3079
3080 if ((nc = cachep->nodelists[node]->shared))
3081 free_block(cachep, nc->entry,
3082 nc->avail);
3083
3084 l3->shared = new;
3085 if (!cachep->nodelists[node]->alien) {
3086 l3->alien = new_alien;
3087 new_alien = NULL;
3088 }
3089 l3->free_limit = (1 + nr_cpus_node(node))*
3090 cachep->batchcount + cachep->num;
3091 spin_unlock_irq(&l3->list_lock);
3092 kfree(nc);
3093 free_alien_cache(new_alien);
3094 continue;
3095 }
3096 if (!(l3 = kmalloc_node(sizeof(struct kmem_list3),
3097 GFP_KERNEL, node)))
3098 goto fail;
3099
3100 kmem_list3_init(l3);
3101 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3102 ((unsigned long)cachep)%REAPTIMEOUT_LIST3;
3103 l3->shared = new;
3104 l3->alien = new_alien;
3105 l3->free_limit = (1 + nr_cpus_node(node))*
3106 cachep->batchcount + cachep->num;
3107 cachep->nodelists[node] = l3;
3108 }
3109 return err;
3110fail:
3111 err = -ENOMEM;
3112 return err;
3113}
3114
Linus Torvalds1da177e2005-04-16 15:20:36 -07003115struct ccupdate_struct {
3116 kmem_cache_t *cachep;
3117 struct array_cache *new[NR_CPUS];
3118};
3119
3120static void do_ccupdate_local(void *info)
3121{
3122 struct ccupdate_struct *new = (struct ccupdate_struct *)info;
3123 struct array_cache *old;
3124
3125 check_irq_off();
3126 old = ac_data(new->cachep);
Christoph Lametere498be72005-09-09 13:03:32 -07003127
Linus Torvalds1da177e2005-04-16 15:20:36 -07003128 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3129 new->new[smp_processor_id()] = old;
3130}
3131
3132
3133static int do_tune_cpucache(kmem_cache_t *cachep, int limit, int batchcount,
3134 int shared)
3135{
3136 struct ccupdate_struct new;
Christoph Lametere498be72005-09-09 13:03:32 -07003137 int i, err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003138
3139 memset(&new.new,0,sizeof(new.new));
Christoph Lametere498be72005-09-09 13:03:32 -07003140 for_each_online_cpu(i) {
3141 new.new[i] = alloc_arraycache(cpu_to_node(i), limit, batchcount);
3142 if (!new.new[i]) {
3143 for (i--; i >= 0; i--) kfree(new.new[i]);
3144 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003145 }
3146 }
3147 new.cachep = cachep;
3148
3149 smp_call_function_all_cpus(do_ccupdate_local, (void *)&new);
Christoph Lametere498be72005-09-09 13:03:32 -07003150
Linus Torvalds1da177e2005-04-16 15:20:36 -07003151 check_irq_on();
3152 spin_lock_irq(&cachep->spinlock);
3153 cachep->batchcount = batchcount;
3154 cachep->limit = limit;
Christoph Lametere498be72005-09-09 13:03:32 -07003155 cachep->shared = shared;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003156 spin_unlock_irq(&cachep->spinlock);
3157
Christoph Lametere498be72005-09-09 13:03:32 -07003158 for_each_online_cpu(i) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003159 struct array_cache *ccold = new.new[i];
3160 if (!ccold)
3161 continue;
Christoph Lametere498be72005-09-09 13:03:32 -07003162 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3163 free_block(cachep, ccold->entry, ccold->avail);
3164 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003165 kfree(ccold);
3166 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003167
Christoph Lametere498be72005-09-09 13:03:32 -07003168 err = alloc_kmemlist(cachep);
3169 if (err) {
3170 printk(KERN_ERR "alloc_kmemlist failed for %s, error %d.\n",
3171 cachep->name, -err);
3172 BUG();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003173 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003174 return 0;
3175}
3176
3177
3178static void enable_cpucache(kmem_cache_t *cachep)
3179{
3180 int err;
3181 int limit, shared;
3182
3183 /* The head array serves three purposes:
3184 * - create a LIFO ordering, i.e. return objects that are cache-warm
3185 * - reduce the number of spinlock operations.
3186 * - reduce the number of linked list operations on the slab and
3187 * bufctl chains: array operations are cheaper.
3188 * The numbers are guessed, we should auto-tune as described by
3189 * Bonwick.
3190 */
3191 if (cachep->objsize > 131072)
3192 limit = 1;
3193 else if (cachep->objsize > PAGE_SIZE)
3194 limit = 8;
3195 else if (cachep->objsize > 1024)
3196 limit = 24;
3197 else if (cachep->objsize > 256)
3198 limit = 54;
3199 else
3200 limit = 120;
3201
3202 /* Cpu bound tasks (e.g. network routing) can exhibit cpu bound
3203 * allocation behaviour: Most allocs on one cpu, most free operations
3204 * on another cpu. For these cases, an efficient object passing between
3205 * cpus is necessary. This is provided by a shared array. The array
3206 * replaces Bonwick's magazine layer.
3207 * On uniprocessor, it's functionally equivalent (but less efficient)
3208 * to a larger limit. Thus disabled by default.
3209 */
3210 shared = 0;
3211#ifdef CONFIG_SMP
3212 if (cachep->objsize <= PAGE_SIZE)
3213 shared = 8;
3214#endif
3215
3216#if DEBUG
3217 /* With debugging enabled, large batchcount lead to excessively
3218 * long periods with disabled local interrupts. Limit the
3219 * batchcount
3220 */
3221 if (limit > 32)
3222 limit = 32;
3223#endif
3224 err = do_tune_cpucache(cachep, limit, (limit+1)/2, shared);
3225 if (err)
3226 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
3227 cachep->name, -err);
3228}
3229
3230static void drain_array_locked(kmem_cache_t *cachep,
Christoph Lametere498be72005-09-09 13:03:32 -07003231 struct array_cache *ac, int force, int node)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003232{
3233 int tofree;
3234
Christoph Lametere498be72005-09-09 13:03:32 -07003235 check_spinlock_acquired_node(cachep, node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003236 if (ac->touched && !force) {
3237 ac->touched = 0;
3238 } else if (ac->avail) {
3239 tofree = force ? ac->avail : (ac->limit+4)/5;
3240 if (tofree > ac->avail) {
3241 tofree = (ac->avail+1)/2;
3242 }
Christoph Lametere498be72005-09-09 13:03:32 -07003243 free_block(cachep, ac->entry, tofree);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003244 ac->avail -= tofree;
Christoph Lametere498be72005-09-09 13:03:32 -07003245 memmove(ac->entry, &(ac->entry[tofree]),
Linus Torvalds1da177e2005-04-16 15:20:36 -07003246 sizeof(void*)*ac->avail);
3247 }
3248}
3249
3250/**
3251 * cache_reap - Reclaim memory from caches.
3252 *
3253 * Called from workqueue/eventd every few seconds.
3254 * Purpose:
3255 * - clear the per-cpu caches for this CPU.
3256 * - return freeable pages to the main free memory pool.
3257 *
3258 * If we cannot acquire the cache chain semaphore then just give up - we'll
3259 * try again on the next iteration.
3260 */
3261static void cache_reap(void *unused)
3262{
3263 struct list_head *walk;
Christoph Lametere498be72005-09-09 13:03:32 -07003264 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003265
3266 if (down_trylock(&cache_chain_sem)) {
3267 /* Give up. Setup the next iteration. */
3268 schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC + smp_processor_id());
3269 return;
3270 }
3271
3272 list_for_each(walk, &cache_chain) {
3273 kmem_cache_t *searchp;
3274 struct list_head* p;
3275 int tofree;
3276 struct slab *slabp;
3277
3278 searchp = list_entry(walk, kmem_cache_t, next);
3279
3280 if (searchp->flags & SLAB_NO_REAP)
3281 goto next;
3282
3283 check_irq_on();
3284
Christoph Lametere498be72005-09-09 13:03:32 -07003285 l3 = searchp->nodelists[numa_node_id()];
3286 if (l3->alien)
3287 drain_alien_cache(searchp, l3);
3288 spin_lock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003289
Christoph Lametere498be72005-09-09 13:03:32 -07003290 drain_array_locked(searchp, ac_data(searchp), 0,
3291 numa_node_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -07003292
Christoph Lametere498be72005-09-09 13:03:32 -07003293 if (time_after(l3->next_reap, jiffies))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003294 goto next_unlock;
3295
Christoph Lametere498be72005-09-09 13:03:32 -07003296 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003297
Christoph Lametere498be72005-09-09 13:03:32 -07003298 if (l3->shared)
3299 drain_array_locked(searchp, l3->shared, 0,
3300 numa_node_id());
Linus Torvalds1da177e2005-04-16 15:20:36 -07003301
Christoph Lametere498be72005-09-09 13:03:32 -07003302 if (l3->free_touched) {
3303 l3->free_touched = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003304 goto next_unlock;
3305 }
3306
Christoph Lametere498be72005-09-09 13:03:32 -07003307 tofree = (l3->free_limit+5*searchp->num-1)/(5*searchp->num);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003308 do {
Christoph Lametere498be72005-09-09 13:03:32 -07003309 p = l3->slabs_free.next;
3310 if (p == &(l3->slabs_free))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003311 break;
3312
3313 slabp = list_entry(p, struct slab, list);
3314 BUG_ON(slabp->inuse);
3315 list_del(&slabp->list);
3316 STATS_INC_REAPED(searchp);
3317
3318 /* Safe to drop the lock. The slab is no longer
3319 * linked to the cache.
3320 * searchp cannot disappear, we hold
3321 * cache_chain_lock
3322 */
Christoph Lametere498be72005-09-09 13:03:32 -07003323 l3->free_objects -= searchp->num;
3324 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003325 slab_destroy(searchp, slabp);
Christoph Lametere498be72005-09-09 13:03:32 -07003326 spin_lock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003327 } while(--tofree > 0);
3328next_unlock:
Christoph Lametere498be72005-09-09 13:03:32 -07003329 spin_unlock_irq(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003330next:
3331 cond_resched();
3332 }
3333 check_irq_on();
3334 up(&cache_chain_sem);
Christoph Lameter4ae7c032005-06-21 17:14:57 -07003335 drain_remote_pages();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003336 /* Setup the next iteration */
3337 schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC + smp_processor_id());
3338}
3339
3340#ifdef CONFIG_PROC_FS
3341
3342static void *s_start(struct seq_file *m, loff_t *pos)
3343{
3344 loff_t n = *pos;
3345 struct list_head *p;
3346
3347 down(&cache_chain_sem);
3348 if (!n) {
3349 /*
3350 * Output format version, so at least we can change it
3351 * without _too_ many complaints.
3352 */
3353#if STATS
3354 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
3355#else
3356 seq_puts(m, "slabinfo - version: 2.1\n");
3357#endif
3358 seq_puts(m, "# name <active_objs> <num_objs> <objsize> <objperslab> <pagesperslab>");
3359 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
3360 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
3361#if STATS
3362 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped>"
Christoph Lametere498be72005-09-09 13:03:32 -07003363 " <error> <maxfreeable> <nodeallocs> <remotefrees>");
Linus Torvalds1da177e2005-04-16 15:20:36 -07003364 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
3365#endif
3366 seq_putc(m, '\n');
3367 }
3368 p = cache_chain.next;
3369 while (n--) {
3370 p = p->next;
3371 if (p == &cache_chain)
3372 return NULL;
3373 }
3374 return list_entry(p, kmem_cache_t, next);
3375}
3376
3377static void *s_next(struct seq_file *m, void *p, loff_t *pos)
3378{
3379 kmem_cache_t *cachep = p;
3380 ++*pos;
3381 return cachep->next.next == &cache_chain ? NULL
3382 : list_entry(cachep->next.next, kmem_cache_t, next);
3383}
3384
3385static void s_stop(struct seq_file *m, void *p)
3386{
3387 up(&cache_chain_sem);
3388}
3389
3390static int s_show(struct seq_file *m, void *p)
3391{
3392 kmem_cache_t *cachep = p;
3393 struct list_head *q;
3394 struct slab *slabp;
3395 unsigned long active_objs;
3396 unsigned long num_objs;
3397 unsigned long active_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07003398 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
3399 const char *name;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003400 char *error = NULL;
Christoph Lametere498be72005-09-09 13:03:32 -07003401 int node;
3402 struct kmem_list3 *l3;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003403
3404 check_irq_on();
3405 spin_lock_irq(&cachep->spinlock);
3406 active_objs = 0;
3407 num_slabs = 0;
Christoph Lametere498be72005-09-09 13:03:32 -07003408 for_each_online_node(node) {
3409 l3 = cachep->nodelists[node];
3410 if (!l3)
3411 continue;
3412
3413 spin_lock(&l3->list_lock);
3414
3415 list_for_each(q,&l3->slabs_full) {
3416 slabp = list_entry(q, struct slab, list);
3417 if (slabp->inuse != cachep->num && !error)
3418 error = "slabs_full accounting error";
3419 active_objs += cachep->num;
3420 active_slabs++;
3421 }
3422 list_for_each(q,&l3->slabs_partial) {
3423 slabp = list_entry(q, struct slab, list);
3424 if (slabp->inuse == cachep->num && !error)
3425 error = "slabs_partial inuse accounting error";
3426 if (!slabp->inuse && !error)
3427 error = "slabs_partial/inuse accounting error";
3428 active_objs += slabp->inuse;
3429 active_slabs++;
3430 }
3431 list_for_each(q,&l3->slabs_free) {
3432 slabp = list_entry(q, struct slab, list);
3433 if (slabp->inuse && !error)
3434 error = "slabs_free/inuse accounting error";
3435 num_slabs++;
3436 }
3437 free_objects += l3->free_objects;
3438 shared_avail += l3->shared->avail;
3439
3440 spin_unlock(&l3->list_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003441 }
3442 num_slabs+=active_slabs;
3443 num_objs = num_slabs*cachep->num;
Christoph Lametere498be72005-09-09 13:03:32 -07003444 if (num_objs - active_objs != free_objects && !error)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003445 error = "free_objects accounting error";
3446
3447 name = cachep->name;
3448 if (error)
3449 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
3450
3451 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
3452 name, active_objs, num_objs, cachep->objsize,
3453 cachep->num, (1<<cachep->gfporder));
3454 seq_printf(m, " : tunables %4u %4u %4u",
3455 cachep->limit, cachep->batchcount,
Christoph Lametere498be72005-09-09 13:03:32 -07003456 cachep->shared);
3457 seq_printf(m, " : slabdata %6lu %6lu %6lu",
3458 active_slabs, num_slabs, shared_avail);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003459#if STATS
3460 { /* list3 stats */
3461 unsigned long high = cachep->high_mark;
3462 unsigned long allocs = cachep->num_allocations;
3463 unsigned long grown = cachep->grown;
3464 unsigned long reaped = cachep->reaped;
3465 unsigned long errors = cachep->errors;
3466 unsigned long max_freeable = cachep->max_freeable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003467 unsigned long node_allocs = cachep->node_allocs;
Christoph Lametere498be72005-09-09 13:03:32 -07003468 unsigned long node_frees = cachep->node_frees;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003469
Christoph Lametere498be72005-09-09 13:03:32 -07003470 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
3471 %4lu %4lu %4lu %4lu",
3472 allocs, high, grown, reaped, errors,
3473 max_freeable, node_allocs, node_frees);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003474 }
3475 /* cpu stats */
3476 {
3477 unsigned long allochit = atomic_read(&cachep->allochit);
3478 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
3479 unsigned long freehit = atomic_read(&cachep->freehit);
3480 unsigned long freemiss = atomic_read(&cachep->freemiss);
3481
3482 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
3483 allochit, allocmiss, freehit, freemiss);
3484 }
3485#endif
3486 seq_putc(m, '\n');
3487 spin_unlock_irq(&cachep->spinlock);
3488 return 0;
3489}
3490
3491/*
3492 * slabinfo_op - iterator that generates /proc/slabinfo
3493 *
3494 * Output layout:
3495 * cache-name
3496 * num-active-objs
3497 * total-objs
3498 * object size
3499 * num-active-slabs
3500 * total-slabs
3501 * num-pages-per-slab
3502 * + further values on SMP and with statistics enabled
3503 */
3504
3505struct seq_operations slabinfo_op = {
3506 .start = s_start,
3507 .next = s_next,
3508 .stop = s_stop,
3509 .show = s_show,
3510};
3511
3512#define MAX_SLABINFO_WRITE 128
3513/**
3514 * slabinfo_write - Tuning for the slab allocator
3515 * @file: unused
3516 * @buffer: user buffer
3517 * @count: data length
3518 * @ppos: unused
3519 */
3520ssize_t slabinfo_write(struct file *file, const char __user *buffer,
3521 size_t count, loff_t *ppos)
3522{
3523 char kbuf[MAX_SLABINFO_WRITE+1], *tmp;
3524 int limit, batchcount, shared, res;
3525 struct list_head *p;
3526
3527 if (count > MAX_SLABINFO_WRITE)
3528 return -EINVAL;
3529 if (copy_from_user(&kbuf, buffer, count))
3530 return -EFAULT;
3531 kbuf[MAX_SLABINFO_WRITE] = '\0';
3532
3533 tmp = strchr(kbuf, ' ');
3534 if (!tmp)
3535 return -EINVAL;
3536 *tmp = '\0';
3537 tmp++;
3538 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
3539 return -EINVAL;
3540
3541 /* Find the cache in the chain of caches. */
3542 down(&cache_chain_sem);
3543 res = -EINVAL;
3544 list_for_each(p,&cache_chain) {
3545 kmem_cache_t *cachep = list_entry(p, kmem_cache_t, next);
3546
3547 if (!strcmp(cachep->name, kbuf)) {
3548 if (limit < 1 ||
3549 batchcount < 1 ||
3550 batchcount > limit ||
3551 shared < 0) {
Christoph Lametere498be72005-09-09 13:03:32 -07003552 res = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003553 } else {
Christoph Lametere498be72005-09-09 13:03:32 -07003554 res = do_tune_cpucache(cachep, limit,
3555 batchcount, shared);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003556 }
3557 break;
3558 }
3559 }
3560 up(&cache_chain_sem);
3561 if (res >= 0)
3562 res = count;
3563 return res;
3564}
3565#endif
3566
Manfred Spraul00e145b2005-09-03 15:55:07 -07003567/**
3568 * ksize - get the actual amount of memory allocated for a given object
3569 * @objp: Pointer to the object
3570 *
3571 * kmalloc may internally round up allocations and return more memory
3572 * than requested. ksize() can be used to determine the actual amount of
3573 * memory allocated. The caller may use this additional memory, even though
3574 * a smaller amount of memory was initially specified with the kmalloc call.
3575 * The caller must guarantee that objp points to a valid object previously
3576 * allocated with either kmalloc() or kmem_cache_alloc(). The object
3577 * must not be freed during the duration of the call.
3578 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003579unsigned int ksize(const void *objp)
3580{
Manfred Spraul00e145b2005-09-03 15:55:07 -07003581 if (unlikely(objp == NULL))
3582 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003583
Manfred Spraul00e145b2005-09-03 15:55:07 -07003584 return obj_reallen(GET_PAGE_CACHE(virt_to_page(objp)));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003585}
Paulo Marques543537b2005-06-23 00:09:02 -07003586
3587
3588/*
3589 * kstrdup - allocate space for and copy an existing string
3590 *
3591 * @s: the string to duplicate
3592 * @gfp: the GFP mask used in the kmalloc() call when allocating memory
3593 */
Alexey Dobriyan0db925a2005-07-07 17:56:58 -07003594char *kstrdup(const char *s, unsigned int __nocast gfp)
Paulo Marques543537b2005-06-23 00:09:02 -07003595{
3596 size_t len;
3597 char *buf;
3598
3599 if (!s)
3600 return NULL;
3601
3602 len = strlen(s) + 1;
3603 buf = kmalloc(len, gfp);
3604 if (buf)
3605 memcpy(buf, s, len);
3606 return buf;
3607}
3608EXPORT_SYMBOL(kstrdup);