blob: c9e706db46340f81f723e899d45510cdfe78d50c [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 *
78 */
79
80#include <linux/config.h>
81#include <linux/slab.h>
82#include <linux/mm.h>
83#include <linux/swap.h>
84#include <linux/cache.h>
85#include <linux/interrupt.h>
86#include <linux/init.h>
87#include <linux/compiler.h>
88#include <linux/seq_file.h>
89#include <linux/notifier.h>
90#include <linux/kallsyms.h>
91#include <linux/cpu.h>
92#include <linux/sysctl.h>
93#include <linux/module.h>
94#include <linux/rcupdate.h>
Paulo Marques543537b2005-06-23 00:09:02 -070095#include <linux/string.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070096
97#include <asm/uaccess.h>
98#include <asm/cacheflush.h>
99#include <asm/tlbflush.h>
100#include <asm/page.h>
101
102/*
103 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL,
104 * SLAB_RED_ZONE & SLAB_POISON.
105 * 0 for faster, smaller code (especially in the critical paths).
106 *
107 * STATS - 1 to collect stats for /proc/slabinfo.
108 * 0 for faster, smaller code (especially in the critical paths).
109 *
110 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
111 */
112
113#ifdef CONFIG_DEBUG_SLAB
114#define DEBUG 1
115#define STATS 1
116#define FORCED_DEBUG 1
117#else
118#define DEBUG 0
119#define STATS 0
120#define FORCED_DEBUG 0
121#endif
122
123
124/* Shouldn't this be in a header file somewhere? */
125#define BYTES_PER_WORD sizeof(void *)
126
127#ifndef cache_line_size
128#define cache_line_size() L1_CACHE_BYTES
129#endif
130
131#ifndef ARCH_KMALLOC_MINALIGN
132/*
133 * Enforce a minimum alignment for the kmalloc caches.
134 * Usually, the kmalloc caches are cache_line_size() aligned, except when
135 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
136 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
137 * alignment larger than BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that.
138 * Note that this flag disables some debug features.
139 */
140#define ARCH_KMALLOC_MINALIGN 0
141#endif
142
143#ifndef ARCH_SLAB_MINALIGN
144/*
145 * Enforce a minimum alignment for all caches.
146 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
147 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
148 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
149 * some debug features.
150 */
151#define ARCH_SLAB_MINALIGN 0
152#endif
153
154#ifndef ARCH_KMALLOC_FLAGS
155#define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
156#endif
157
158/* Legal flag mask for kmem_cache_create(). */
159#if DEBUG
160# define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \
161 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
162 SLAB_NO_REAP | SLAB_CACHE_DMA | \
163 SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \
164 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
165 SLAB_DESTROY_BY_RCU)
166#else
167# define CREATE_MASK (SLAB_HWCACHE_ALIGN | SLAB_NO_REAP | \
168 SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \
169 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
170 SLAB_DESTROY_BY_RCU)
171#endif
172
173/*
174 * kmem_bufctl_t:
175 *
176 * Bufctl's are used for linking objs within a slab
177 * linked offsets.
178 *
179 * This implementation relies on "struct page" for locating the cache &
180 * slab an object belongs to.
181 * This allows the bufctl structure to be small (one int), but limits
182 * the number of objects a slab (not a cache) can contain when off-slab
183 * bufctls are used. The limit is the size of the largest general cache
184 * that does not use off-slab slabs.
185 * For 32bit archs with 4 kB pages, is this 56.
186 * This is not serious, as it is only for large objects, when it is unwise
187 * to have too many per slab.
188 * Note: This limit can be raised by introducing a general cache whose size
189 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
190 */
191
192#define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
193#define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
194#define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-2)
195
196/* Max number of objs-per-slab for caches which use off-slab slabs.
197 * Needed to avoid a possible looping condition in cache_grow().
198 */
199static unsigned long offslab_limit;
200
201/*
202 * struct slab
203 *
204 * Manages the objs in a slab. Placed either at the beginning of mem allocated
205 * for a slab, or allocated from an general cache.
206 * Slabs are chained into three list: fully used, partial, fully free slabs.
207 */
208struct slab {
209 struct list_head list;
210 unsigned long colouroff;
211 void *s_mem; /* including colour offset */
212 unsigned int inuse; /* num of objs active in slab */
213 kmem_bufctl_t free;
214};
215
216/*
217 * struct slab_rcu
218 *
219 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
220 * arrange for kmem_freepages to be called via RCU. This is useful if
221 * we need to approach a kernel structure obliquely, from its address
222 * obtained without the usual locking. We can lock the structure to
223 * stabilize it and check it's still at the given address, only if we
224 * can be sure that the memory has not been meanwhile reused for some
225 * other kind of object (which our subsystem's lock might corrupt).
226 *
227 * rcu_read_lock before reading the address, then rcu_read_unlock after
228 * taking the spinlock within the structure expected at that address.
229 *
230 * We assume struct slab_rcu can overlay struct slab when destroying.
231 */
232struct slab_rcu {
233 struct rcu_head head;
234 kmem_cache_t *cachep;
235 void *addr;
236};
237
238/*
239 * struct array_cache
240 *
241 * Per cpu structures
242 * Purpose:
243 * - LIFO ordering, to hand out cache-warm objects from _alloc
244 * - reduce the number of linked list operations
245 * - reduce spinlock operations
246 *
247 * The limit is stored in the per-cpu structure to reduce the data cache
248 * footprint.
249 *
250 */
251struct array_cache {
252 unsigned int avail;
253 unsigned int limit;
254 unsigned int batchcount;
255 unsigned int touched;
256};
257
258/* bootstrap: The caches do not work without cpuarrays anymore,
259 * but the cpuarrays are allocated from the generic caches...
260 */
261#define BOOT_CPUCACHE_ENTRIES 1
262struct arraycache_init {
263 struct array_cache cache;
264 void * entries[BOOT_CPUCACHE_ENTRIES];
265};
266
267/*
268 * The slab lists of all objects.
269 * Hopefully reduce the internal fragmentation
270 * NUMA: The spinlock could be moved from the kmem_cache_t
271 * into this structure, too. Figure out what causes
272 * fewer cross-node spinlock operations.
273 */
274struct kmem_list3 {
275 struct list_head slabs_partial; /* partial list first, better asm code */
276 struct list_head slabs_full;
277 struct list_head slabs_free;
278 unsigned long free_objects;
279 int free_touched;
280 unsigned long next_reap;
281 struct array_cache *shared;
282};
283
284#define LIST3_INIT(parent) \
285 { \
286 .slabs_full = LIST_HEAD_INIT(parent.slabs_full), \
287 .slabs_partial = LIST_HEAD_INIT(parent.slabs_partial), \
288 .slabs_free = LIST_HEAD_INIT(parent.slabs_free) \
289 }
290#define list3_data(cachep) \
291 (&(cachep)->lists)
292
293/* NUMA: per-node */
294#define list3_data_ptr(cachep, ptr) \
295 list3_data(cachep)
296
297/*
298 * kmem_cache_t
299 *
300 * manages a cache.
301 */
302
303struct kmem_cache_s {
304/* 1) per-cpu data, touched during every alloc/free */
305 struct array_cache *array[NR_CPUS];
306 unsigned int batchcount;
307 unsigned int limit;
308/* 2) touched by every alloc & free from the backend */
309 struct kmem_list3 lists;
310 /* NUMA: kmem_3list_t *nodelists[MAX_NUMNODES] */
311 unsigned int objsize;
312 unsigned int flags; /* constant flags */
313 unsigned int num; /* # of objs per slab */
314 unsigned int free_limit; /* upper limit of objects in the lists */
315 spinlock_t spinlock;
316
317/* 3) cache_grow/shrink */
318 /* order of pgs per slab (2^n) */
319 unsigned int gfporder;
320
321 /* force GFP flags, e.g. GFP_DMA */
322 unsigned int gfpflags;
323
324 size_t colour; /* cache colouring range */
325 unsigned int colour_off; /* colour offset */
326 unsigned int colour_next; /* cache colouring */
327 kmem_cache_t *slabp_cache;
328 unsigned int slab_size;
329 unsigned int dflags; /* dynamic flags */
330
331 /* constructor func */
332 void (*ctor)(void *, kmem_cache_t *, unsigned long);
333
334 /* de-constructor func */
335 void (*dtor)(void *, kmem_cache_t *, unsigned long);
336
337/* 4) cache creation/removal */
338 const char *name;
339 struct list_head next;
340
341/* 5) statistics */
342#if STATS
343 unsigned long num_active;
344 unsigned long num_allocations;
345 unsigned long high_mark;
346 unsigned long grown;
347 unsigned long reaped;
348 unsigned long errors;
349 unsigned long max_freeable;
350 unsigned long node_allocs;
351 atomic_t allochit;
352 atomic_t allocmiss;
353 atomic_t freehit;
354 atomic_t freemiss;
355#endif
356#if DEBUG
357 int dbghead;
358 int reallen;
359#endif
360};
361
362#define CFLGS_OFF_SLAB (0x80000000UL)
363#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
364
365#define BATCHREFILL_LIMIT 16
366/* Optimization question: fewer reaps means less
367 * probability for unnessary cpucache drain/refill cycles.
368 *
369 * OTHO the cpuarrays can contain lots of objects,
370 * which could lock up otherwise freeable slabs.
371 */
372#define REAPTIMEOUT_CPUC (2*HZ)
373#define REAPTIMEOUT_LIST3 (4*HZ)
374
375#if STATS
376#define STATS_INC_ACTIVE(x) ((x)->num_active++)
377#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
378#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
379#define STATS_INC_GROWN(x) ((x)->grown++)
380#define STATS_INC_REAPED(x) ((x)->reaped++)
381#define STATS_SET_HIGH(x) do { if ((x)->num_active > (x)->high_mark) \
382 (x)->high_mark = (x)->num_active; \
383 } while (0)
384#define STATS_INC_ERR(x) ((x)->errors++)
385#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
386#define STATS_SET_FREEABLE(x, i) \
387 do { if ((x)->max_freeable < i) \
388 (x)->max_freeable = i; \
389 } while (0)
390
391#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
392#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
393#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
394#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
395#else
396#define STATS_INC_ACTIVE(x) do { } while (0)
397#define STATS_DEC_ACTIVE(x) do { } while (0)
398#define STATS_INC_ALLOCED(x) do { } while (0)
399#define STATS_INC_GROWN(x) do { } while (0)
400#define STATS_INC_REAPED(x) do { } while (0)
401#define STATS_SET_HIGH(x) do { } while (0)
402#define STATS_INC_ERR(x) do { } while (0)
403#define STATS_INC_NODEALLOCS(x) do { } while (0)
404#define STATS_SET_FREEABLE(x, i) \
405 do { } while (0)
406
407#define STATS_INC_ALLOCHIT(x) do { } while (0)
408#define STATS_INC_ALLOCMISS(x) do { } while (0)
409#define STATS_INC_FREEHIT(x) do { } while (0)
410#define STATS_INC_FREEMISS(x) do { } while (0)
411#endif
412
413#if DEBUG
414/* Magic nums for obj red zoning.
415 * Placed in the first word before and the first word after an obj.
416 */
417#define RED_INACTIVE 0x5A2CF071UL /* when obj is inactive */
418#define RED_ACTIVE 0x170FC2A5UL /* when obj is active */
419
420/* ...and for poisoning */
421#define POISON_INUSE 0x5a /* for use-uninitialised poisoning */
422#define POISON_FREE 0x6b /* for use-after-free poisoning */
423#define POISON_END 0xa5 /* end-byte of poisoning */
424
425/* memory layout of objects:
426 * 0 : objp
427 * 0 .. cachep->dbghead - BYTES_PER_WORD - 1: padding. This ensures that
428 * the end of an object is aligned with the end of the real
429 * allocation. Catches writes behind the end of the allocation.
430 * cachep->dbghead - BYTES_PER_WORD .. cachep->dbghead - 1:
431 * redzone word.
432 * cachep->dbghead: The real object.
433 * cachep->objsize - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
434 * cachep->objsize - 1* BYTES_PER_WORD: last caller address [BYTES_PER_WORD long]
435 */
436static int obj_dbghead(kmem_cache_t *cachep)
437{
438 return cachep->dbghead;
439}
440
441static int obj_reallen(kmem_cache_t *cachep)
442{
443 return cachep->reallen;
444}
445
446static unsigned long *dbg_redzone1(kmem_cache_t *cachep, void *objp)
447{
448 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
449 return (unsigned long*) (objp+obj_dbghead(cachep)-BYTES_PER_WORD);
450}
451
452static unsigned long *dbg_redzone2(kmem_cache_t *cachep, void *objp)
453{
454 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
455 if (cachep->flags & SLAB_STORE_USER)
456 return (unsigned long*) (objp+cachep->objsize-2*BYTES_PER_WORD);
457 return (unsigned long*) (objp+cachep->objsize-BYTES_PER_WORD);
458}
459
460static void **dbg_userword(kmem_cache_t *cachep, void *objp)
461{
462 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
463 return (void**)(objp+cachep->objsize-BYTES_PER_WORD);
464}
465
466#else
467
468#define obj_dbghead(x) 0
469#define obj_reallen(cachep) (cachep->objsize)
470#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;})
471#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;})
472#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
473
474#endif
475
476/*
477 * Maximum size of an obj (in 2^order pages)
478 * and absolute limit for the gfp order.
479 */
480#if defined(CONFIG_LARGE_ALLOCS)
481#define MAX_OBJ_ORDER 13 /* up to 32Mb */
482#define MAX_GFP_ORDER 13 /* up to 32Mb */
483#elif defined(CONFIG_MMU)
484#define MAX_OBJ_ORDER 5 /* 32 pages */
485#define MAX_GFP_ORDER 5 /* 32 pages */
486#else
487#define MAX_OBJ_ORDER 8 /* up to 1Mb */
488#define MAX_GFP_ORDER 8 /* up to 1Mb */
489#endif
490
491/*
492 * Do not go above this order unless 0 objects fit into the slab.
493 */
494#define BREAK_GFP_ORDER_HI 1
495#define BREAK_GFP_ORDER_LO 0
496static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
497
498/* Macros for storing/retrieving the cachep and or slab from the
499 * global 'mem_map'. These are used to find the slab an obj belongs to.
500 * With kfree(), these are used to find the cache which an obj belongs to.
501 */
502#define SET_PAGE_CACHE(pg,x) ((pg)->lru.next = (struct list_head *)(x))
503#define GET_PAGE_CACHE(pg) ((kmem_cache_t *)(pg)->lru.next)
504#define SET_PAGE_SLAB(pg,x) ((pg)->lru.prev = (struct list_head *)(x))
505#define GET_PAGE_SLAB(pg) ((struct slab *)(pg)->lru.prev)
506
507/* These are the default caches for kmalloc. Custom caches can have other sizes. */
508struct cache_sizes malloc_sizes[] = {
509#define CACHE(x) { .cs_size = (x) },
510#include <linux/kmalloc_sizes.h>
511 CACHE(ULONG_MAX)
512#undef CACHE
513};
514EXPORT_SYMBOL(malloc_sizes);
515
516/* Must match cache_sizes above. Out of line to keep cache footprint low. */
517struct cache_names {
518 char *name;
519 char *name_dma;
520};
521
522static struct cache_names __initdata cache_names[] = {
523#define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
524#include <linux/kmalloc_sizes.h>
525 { NULL, }
526#undef CACHE
527};
528
529static struct arraycache_init initarray_cache __initdata =
530 { { 0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
531static struct arraycache_init initarray_generic =
532 { { 0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
533
534/* internal cache of cache description objs */
535static kmem_cache_t cache_cache = {
536 .lists = LIST3_INIT(cache_cache.lists),
537 .batchcount = 1,
538 .limit = BOOT_CPUCACHE_ENTRIES,
539 .objsize = sizeof(kmem_cache_t),
540 .flags = SLAB_NO_REAP,
541 .spinlock = SPIN_LOCK_UNLOCKED,
542 .name = "kmem_cache",
543#if DEBUG
544 .reallen = sizeof(kmem_cache_t),
545#endif
546};
547
548/* Guard access to the cache-chain. */
549static struct semaphore cache_chain_sem;
550static struct list_head cache_chain;
551
552/*
553 * vm_enough_memory() looks at this to determine how many
554 * slab-allocated pages are possibly freeable under pressure
555 *
556 * SLAB_RECLAIM_ACCOUNT turns this on per-slab
557 */
558atomic_t slab_reclaim_pages;
559EXPORT_SYMBOL(slab_reclaim_pages);
560
561/*
562 * chicken and egg problem: delay the per-cpu array allocation
563 * until the general caches are up.
564 */
565static enum {
566 NONE,
567 PARTIAL,
568 FULL
569} g_cpucache_up;
570
571static DEFINE_PER_CPU(struct work_struct, reap_work);
572
573static void free_block(kmem_cache_t* cachep, void** objpp, int len);
574static void enable_cpucache (kmem_cache_t *cachep);
575static void cache_reap (void *unused);
576
577static inline void **ac_entry(struct array_cache *ac)
578{
579 return (void**)(ac+1);
580}
581
582static inline struct array_cache *ac_data(kmem_cache_t *cachep)
583{
584 return cachep->array[smp_processor_id()];
585}
586
Alexey Dobriyan0db925a2005-07-07 17:56:58 -0700587static inline kmem_cache_t *__find_general_cachep(size_t size,
588 unsigned int __nocast gfpflags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700589{
590 struct cache_sizes *csizep = malloc_sizes;
591
592#if DEBUG
593 /* This happens if someone tries to call
594 * kmem_cache_create(), or __kmalloc(), before
595 * the generic caches are initialized.
596 */
597 BUG_ON(csizep->cs_cachep == NULL);
598#endif
599 while (size > csizep->cs_size)
600 csizep++;
601
602 /*
603 * Really subtile: The last entry with cs->cs_size==ULONG_MAX
604 * has cs_{dma,}cachep==NULL. Thus no special case
605 * for large kmalloc calls required.
606 */
607 if (unlikely(gfpflags & GFP_DMA))
608 return csizep->cs_dmacachep;
609 return csizep->cs_cachep;
610}
611
Alexey Dobriyan0db925a2005-07-07 17:56:58 -0700612kmem_cache_t *kmem_find_general_cachep(size_t size,
613 unsigned int __nocast gfpflags)
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700614{
615 return __find_general_cachep(size, gfpflags);
616}
617EXPORT_SYMBOL(kmem_find_general_cachep);
618
Linus Torvalds1da177e2005-04-16 15:20:36 -0700619/* Cal the num objs, wastage, and bytes left over for a given slab size. */
620static void cache_estimate(unsigned long gfporder, size_t size, size_t align,
621 int flags, size_t *left_over, unsigned int *num)
622{
623 int i;
624 size_t wastage = PAGE_SIZE<<gfporder;
625 size_t extra = 0;
626 size_t base = 0;
627
628 if (!(flags & CFLGS_OFF_SLAB)) {
629 base = sizeof(struct slab);
630 extra = sizeof(kmem_bufctl_t);
631 }
632 i = 0;
633 while (i*size + ALIGN(base+i*extra, align) <= wastage)
634 i++;
635 if (i > 0)
636 i--;
637
638 if (i > SLAB_LIMIT)
639 i = SLAB_LIMIT;
640
641 *num = i;
642 wastage -= i*size;
643 wastage -= ALIGN(base+i*extra, align);
644 *left_over = wastage;
645}
646
647#define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
648
649static void __slab_error(const char *function, kmem_cache_t *cachep, char *msg)
650{
651 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
652 function, cachep->name, msg);
653 dump_stack();
654}
655
656/*
657 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
658 * via the workqueue/eventd.
659 * Add the CPU number into the expiration time to minimize the possibility of
660 * the CPUs getting into lockstep and contending for the global cache chain
661 * lock.
662 */
663static void __devinit start_cpu_timer(int cpu)
664{
665 struct work_struct *reap_work = &per_cpu(reap_work, cpu);
666
667 /*
668 * When this gets called from do_initcalls via cpucache_init(),
669 * init_workqueues() has already run, so keventd will be setup
670 * at that time.
671 */
672 if (keventd_up() && reap_work->func == NULL) {
673 INIT_WORK(reap_work, cache_reap, NULL);
674 schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu);
675 }
676}
677
678static struct array_cache *alloc_arraycache(int cpu, int entries,
679 int batchcount)
680{
681 int memsize = sizeof(void*)*entries+sizeof(struct array_cache);
682 struct array_cache *nc = NULL;
683
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700684 if (cpu == -1)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700685 nc = kmalloc(memsize, GFP_KERNEL);
Manfred Spraul97e2bde2005-05-01 08:58:38 -0700686 else
687 nc = kmalloc_node(memsize, GFP_KERNEL, cpu_to_node(cpu));
688
Linus Torvalds1da177e2005-04-16 15:20:36 -0700689 if (nc) {
690 nc->avail = 0;
691 nc->limit = entries;
692 nc->batchcount = batchcount;
693 nc->touched = 0;
694 }
695 return nc;
696}
697
698static int __devinit cpuup_callback(struct notifier_block *nfb,
699 unsigned long action, void *hcpu)
700{
701 long cpu = (long)hcpu;
702 kmem_cache_t* cachep;
703
704 switch (action) {
705 case CPU_UP_PREPARE:
706 down(&cache_chain_sem);
707 list_for_each_entry(cachep, &cache_chain, next) {
708 struct array_cache *nc;
709
710 nc = alloc_arraycache(cpu, cachep->limit, cachep->batchcount);
711 if (!nc)
712 goto bad;
713
714 spin_lock_irq(&cachep->spinlock);
715 cachep->array[cpu] = nc;
716 cachep->free_limit = (1+num_online_cpus())*cachep->batchcount
717 + cachep->num;
718 spin_unlock_irq(&cachep->spinlock);
719
720 }
721 up(&cache_chain_sem);
722 break;
723 case CPU_ONLINE:
724 start_cpu_timer(cpu);
725 break;
726#ifdef CONFIG_HOTPLUG_CPU
727 case CPU_DEAD:
728 /* fall thru */
729 case CPU_UP_CANCELED:
730 down(&cache_chain_sem);
731
732 list_for_each_entry(cachep, &cache_chain, next) {
733 struct array_cache *nc;
734
735 spin_lock_irq(&cachep->spinlock);
736 /* cpu is dead; no one can alloc from it. */
737 nc = cachep->array[cpu];
738 cachep->array[cpu] = NULL;
739 cachep->free_limit -= cachep->batchcount;
740 free_block(cachep, ac_entry(nc), nc->avail);
741 spin_unlock_irq(&cachep->spinlock);
742 kfree(nc);
743 }
744 up(&cache_chain_sem);
745 break;
746#endif
747 }
748 return NOTIFY_OK;
749bad:
750 up(&cache_chain_sem);
751 return NOTIFY_BAD;
752}
753
754static struct notifier_block cpucache_notifier = { &cpuup_callback, NULL, 0 };
755
756/* Initialisation.
757 * Called after the gfp() functions have been enabled, and before smp_init().
758 */
759void __init kmem_cache_init(void)
760{
761 size_t left_over;
762 struct cache_sizes *sizes;
763 struct cache_names *names;
764
765 /*
766 * Fragmentation resistance on low memory - only use bigger
767 * page orders on machines with more than 32MB of memory.
768 */
769 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
770 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
771
772
773 /* Bootstrap is tricky, because several objects are allocated
774 * from caches that do not exist yet:
775 * 1) initialize the cache_cache cache: it contains the kmem_cache_t
776 * structures of all caches, except cache_cache itself: cache_cache
777 * is statically allocated.
778 * Initially an __init data area is used for the head array, it's
779 * replaced with a kmalloc allocated array at the end of the bootstrap.
780 * 2) Create the first kmalloc cache.
781 * The kmem_cache_t for the new cache is allocated normally. An __init
782 * data area is used for the head array.
783 * 3) Create the remaining kmalloc caches, with minimally sized head arrays.
784 * 4) Replace the __init data head arrays for cache_cache and the first
785 * kmalloc cache with kmalloc allocated arrays.
786 * 5) Resize the head arrays of the kmalloc caches to their final sizes.
787 */
788
789 /* 1) create the cache_cache */
790 init_MUTEX(&cache_chain_sem);
791 INIT_LIST_HEAD(&cache_chain);
792 list_add(&cache_cache.next, &cache_chain);
793 cache_cache.colour_off = cache_line_size();
794 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
795
796 cache_cache.objsize = ALIGN(cache_cache.objsize, cache_line_size());
797
798 cache_estimate(0, cache_cache.objsize, cache_line_size(), 0,
799 &left_over, &cache_cache.num);
800 if (!cache_cache.num)
801 BUG();
802
803 cache_cache.colour = left_over/cache_cache.colour_off;
804 cache_cache.colour_next = 0;
805 cache_cache.slab_size = ALIGN(cache_cache.num*sizeof(kmem_bufctl_t) +
806 sizeof(struct slab), cache_line_size());
807
808 /* 2+3) create the kmalloc caches */
809 sizes = malloc_sizes;
810 names = cache_names;
811
812 while (sizes->cs_size != ULONG_MAX) {
813 /* For performance, all the general caches are L1 aligned.
814 * This should be particularly beneficial on SMP boxes, as it
815 * eliminates "false sharing".
816 * Note for systems short on memory removing the alignment will
817 * allow tighter packing of the smaller caches. */
818 sizes->cs_cachep = kmem_cache_create(names->name,
819 sizes->cs_size, ARCH_KMALLOC_MINALIGN,
820 (ARCH_KMALLOC_FLAGS | SLAB_PANIC), NULL, NULL);
821
822 /* Inc off-slab bufctl limit until the ceiling is hit. */
823 if (!(OFF_SLAB(sizes->cs_cachep))) {
824 offslab_limit = sizes->cs_size-sizeof(struct slab);
825 offslab_limit /= sizeof(kmem_bufctl_t);
826 }
827
828 sizes->cs_dmacachep = kmem_cache_create(names->name_dma,
829 sizes->cs_size, ARCH_KMALLOC_MINALIGN,
830 (ARCH_KMALLOC_FLAGS | SLAB_CACHE_DMA | SLAB_PANIC),
831 NULL, NULL);
832
833 sizes++;
834 names++;
835 }
836 /* 4) Replace the bootstrap head arrays */
837 {
838 void * ptr;
839
840 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
841 local_irq_disable();
842 BUG_ON(ac_data(&cache_cache) != &initarray_cache.cache);
843 memcpy(ptr, ac_data(&cache_cache), sizeof(struct arraycache_init));
844 cache_cache.array[smp_processor_id()] = ptr;
845 local_irq_enable();
846
847 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
848 local_irq_disable();
849 BUG_ON(ac_data(malloc_sizes[0].cs_cachep) != &initarray_generic.cache);
850 memcpy(ptr, ac_data(malloc_sizes[0].cs_cachep),
851 sizeof(struct arraycache_init));
852 malloc_sizes[0].cs_cachep->array[smp_processor_id()] = ptr;
853 local_irq_enable();
854 }
855
856 /* 5) resize the head arrays to their final sizes */
857 {
858 kmem_cache_t *cachep;
859 down(&cache_chain_sem);
860 list_for_each_entry(cachep, &cache_chain, next)
861 enable_cpucache(cachep);
862 up(&cache_chain_sem);
863 }
864
865 /* Done! */
866 g_cpucache_up = FULL;
867
868 /* Register a cpu startup notifier callback
869 * that initializes ac_data for all new cpus
870 */
871 register_cpu_notifier(&cpucache_notifier);
872
873
874 /* The reap timers are started later, with a module init call:
875 * That part of the kernel is not yet operational.
876 */
877}
878
879static int __init cpucache_init(void)
880{
881 int cpu;
882
883 /*
884 * Register the timers that return unneeded
885 * pages to gfp.
886 */
887 for (cpu = 0; cpu < NR_CPUS; cpu++) {
888 if (cpu_online(cpu))
889 start_cpu_timer(cpu);
890 }
891
892 return 0;
893}
894
895__initcall(cpucache_init);
896
897/*
898 * Interface to system's page allocator. No need to hold the cache-lock.
899 *
900 * If we requested dmaable memory, we will get it. Even if we
901 * did not request dmaable memory, we might get it, but that
902 * would be relatively rare and ignorable.
903 */
904static void *kmem_getpages(kmem_cache_t *cachep, unsigned int __nocast flags, int nodeid)
905{
906 struct page *page;
907 void *addr;
908 int i;
909
910 flags |= cachep->gfpflags;
911 if (likely(nodeid == -1)) {
912 page = alloc_pages(flags, cachep->gfporder);
913 } else {
914 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
915 }
916 if (!page)
917 return NULL;
918 addr = page_address(page);
919
920 i = (1 << cachep->gfporder);
921 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
922 atomic_add(i, &slab_reclaim_pages);
923 add_page_state(nr_slab, i);
924 while (i--) {
925 SetPageSlab(page);
926 page++;
927 }
928 return addr;
929}
930
931/*
932 * Interface to system's page release.
933 */
934static void kmem_freepages(kmem_cache_t *cachep, void *addr)
935{
936 unsigned long i = (1<<cachep->gfporder);
937 struct page *page = virt_to_page(addr);
938 const unsigned long nr_freed = i;
939
940 while (i--) {
941 if (!TestClearPageSlab(page))
942 BUG();
943 page++;
944 }
945 sub_page_state(nr_slab, nr_freed);
946 if (current->reclaim_state)
947 current->reclaim_state->reclaimed_slab += nr_freed;
948 free_pages((unsigned long)addr, cachep->gfporder);
949 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
950 atomic_sub(1<<cachep->gfporder, &slab_reclaim_pages);
951}
952
953static void kmem_rcu_free(struct rcu_head *head)
954{
955 struct slab_rcu *slab_rcu = (struct slab_rcu *) head;
956 kmem_cache_t *cachep = slab_rcu->cachep;
957
958 kmem_freepages(cachep, slab_rcu->addr);
959 if (OFF_SLAB(cachep))
960 kmem_cache_free(cachep->slabp_cache, slab_rcu);
961}
962
963#if DEBUG
964
965#ifdef CONFIG_DEBUG_PAGEALLOC
966static void store_stackinfo(kmem_cache_t *cachep, unsigned long *addr,
967 unsigned long caller)
968{
969 int size = obj_reallen(cachep);
970
971 addr = (unsigned long *)&((char*)addr)[obj_dbghead(cachep)];
972
973 if (size < 5*sizeof(unsigned long))
974 return;
975
976 *addr++=0x12345678;
977 *addr++=caller;
978 *addr++=smp_processor_id();
979 size -= 3*sizeof(unsigned long);
980 {
981 unsigned long *sptr = &caller;
982 unsigned long svalue;
983
984 while (!kstack_end(sptr)) {
985 svalue = *sptr++;
986 if (kernel_text_address(svalue)) {
987 *addr++=svalue;
988 size -= sizeof(unsigned long);
989 if (size <= sizeof(unsigned long))
990 break;
991 }
992 }
993
994 }
995 *addr++=0x87654321;
996}
997#endif
998
999static void poison_obj(kmem_cache_t *cachep, void *addr, unsigned char val)
1000{
1001 int size = obj_reallen(cachep);
1002 addr = &((char*)addr)[obj_dbghead(cachep)];
1003
1004 memset(addr, val, size);
1005 *(unsigned char *)(addr+size-1) = POISON_END;
1006}
1007
1008static void dump_line(char *data, int offset, int limit)
1009{
1010 int i;
1011 printk(KERN_ERR "%03x:", offset);
1012 for (i=0;i<limit;i++) {
1013 printk(" %02x", (unsigned char)data[offset+i]);
1014 }
1015 printk("\n");
1016}
1017#endif
1018
1019#if DEBUG
1020
1021static void print_objinfo(kmem_cache_t *cachep, void *objp, int lines)
1022{
1023 int i, size;
1024 char *realobj;
1025
1026 if (cachep->flags & SLAB_RED_ZONE) {
1027 printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n",
1028 *dbg_redzone1(cachep, objp),
1029 *dbg_redzone2(cachep, objp));
1030 }
1031
1032 if (cachep->flags & SLAB_STORE_USER) {
1033 printk(KERN_ERR "Last user: [<%p>]",
1034 *dbg_userword(cachep, objp));
1035 print_symbol("(%s)",
1036 (unsigned long)*dbg_userword(cachep, objp));
1037 printk("\n");
1038 }
1039 realobj = (char*)objp+obj_dbghead(cachep);
1040 size = obj_reallen(cachep);
1041 for (i=0; i<size && lines;i+=16, lines--) {
1042 int limit;
1043 limit = 16;
1044 if (i+limit > size)
1045 limit = size-i;
1046 dump_line(realobj, i, limit);
1047 }
1048}
1049
1050static void check_poison_obj(kmem_cache_t *cachep, void *objp)
1051{
1052 char *realobj;
1053 int size, i;
1054 int lines = 0;
1055
1056 realobj = (char*)objp+obj_dbghead(cachep);
1057 size = obj_reallen(cachep);
1058
1059 for (i=0;i<size;i++) {
1060 char exp = POISON_FREE;
1061 if (i == size-1)
1062 exp = POISON_END;
1063 if (realobj[i] != exp) {
1064 int limit;
1065 /* Mismatch ! */
1066 /* Print header */
1067 if (lines == 0) {
1068 printk(KERN_ERR "Slab corruption: start=%p, len=%d\n",
1069 realobj, size);
1070 print_objinfo(cachep, objp, 0);
1071 }
1072 /* Hexdump the affected line */
1073 i = (i/16)*16;
1074 limit = 16;
1075 if (i+limit > size)
1076 limit = size-i;
1077 dump_line(realobj, i, limit);
1078 i += 16;
1079 lines++;
1080 /* Limit to 5 lines */
1081 if (lines > 5)
1082 break;
1083 }
1084 }
1085 if (lines != 0) {
1086 /* Print some data about the neighboring objects, if they
1087 * exist:
1088 */
1089 struct slab *slabp = GET_PAGE_SLAB(virt_to_page(objp));
1090 int objnr;
1091
1092 objnr = (objp-slabp->s_mem)/cachep->objsize;
1093 if (objnr) {
1094 objp = slabp->s_mem+(objnr-1)*cachep->objsize;
1095 realobj = (char*)objp+obj_dbghead(cachep);
1096 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1097 realobj, size);
1098 print_objinfo(cachep, objp, 2);
1099 }
1100 if (objnr+1 < cachep->num) {
1101 objp = slabp->s_mem+(objnr+1)*cachep->objsize;
1102 realobj = (char*)objp+obj_dbghead(cachep);
1103 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1104 realobj, size);
1105 print_objinfo(cachep, objp, 2);
1106 }
1107 }
1108}
1109#endif
1110
1111/* Destroy all the objs in a slab, and release the mem back to the system.
1112 * Before calling the slab must have been unlinked from the cache.
1113 * The cache-lock is not held/needed.
1114 */
1115static void slab_destroy (kmem_cache_t *cachep, struct slab *slabp)
1116{
1117 void *addr = slabp->s_mem - slabp->colouroff;
1118
1119#if DEBUG
1120 int i;
1121 for (i = 0; i < cachep->num; i++) {
1122 void *objp = slabp->s_mem + cachep->objsize * i;
1123
1124 if (cachep->flags & SLAB_POISON) {
1125#ifdef CONFIG_DEBUG_PAGEALLOC
1126 if ((cachep->objsize%PAGE_SIZE)==0 && OFF_SLAB(cachep))
1127 kernel_map_pages(virt_to_page(objp), cachep->objsize/PAGE_SIZE,1);
1128 else
1129 check_poison_obj(cachep, objp);
1130#else
1131 check_poison_obj(cachep, objp);
1132#endif
1133 }
1134 if (cachep->flags & SLAB_RED_ZONE) {
1135 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1136 slab_error(cachep, "start of a freed object "
1137 "was overwritten");
1138 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1139 slab_error(cachep, "end of a freed object "
1140 "was overwritten");
1141 }
1142 if (cachep->dtor && !(cachep->flags & SLAB_POISON))
1143 (cachep->dtor)(objp+obj_dbghead(cachep), cachep, 0);
1144 }
1145#else
1146 if (cachep->dtor) {
1147 int i;
1148 for (i = 0; i < cachep->num; i++) {
1149 void* objp = slabp->s_mem+cachep->objsize*i;
1150 (cachep->dtor)(objp, cachep, 0);
1151 }
1152 }
1153#endif
1154
1155 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1156 struct slab_rcu *slab_rcu;
1157
1158 slab_rcu = (struct slab_rcu *) slabp;
1159 slab_rcu->cachep = cachep;
1160 slab_rcu->addr = addr;
1161 call_rcu(&slab_rcu->head, kmem_rcu_free);
1162 } else {
1163 kmem_freepages(cachep, addr);
1164 if (OFF_SLAB(cachep))
1165 kmem_cache_free(cachep->slabp_cache, slabp);
1166 }
1167}
1168
1169/**
1170 * kmem_cache_create - Create a cache.
1171 * @name: A string which is used in /proc/slabinfo to identify this cache.
1172 * @size: The size of objects to be created in this cache.
1173 * @align: The required alignment for the objects.
1174 * @flags: SLAB flags
1175 * @ctor: A constructor for the objects.
1176 * @dtor: A destructor for the objects.
1177 *
1178 * Returns a ptr to the cache on success, NULL on failure.
1179 * Cannot be called within a int, but can be interrupted.
1180 * The @ctor is run when new pages are allocated by the cache
1181 * and the @dtor is run before the pages are handed back.
1182 *
1183 * @name must be valid until the cache is destroyed. This implies that
1184 * the module calling this has to destroy the cache before getting
1185 * unloaded.
1186 *
1187 * The flags are
1188 *
1189 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1190 * to catch references to uninitialised memory.
1191 *
1192 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1193 * for buffer overruns.
1194 *
1195 * %SLAB_NO_REAP - Don't automatically reap this cache when we're under
1196 * memory pressure.
1197 *
1198 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
1199 * cacheline. This can be beneficial if you're counting cycles as closely
1200 * as davem.
1201 */
1202kmem_cache_t *
1203kmem_cache_create (const char *name, size_t size, size_t align,
1204 unsigned long flags, void (*ctor)(void*, kmem_cache_t *, unsigned long),
1205 void (*dtor)(void*, kmem_cache_t *, unsigned long))
1206{
1207 size_t left_over, slab_size, ralign;
1208 kmem_cache_t *cachep = NULL;
1209
1210 /*
1211 * Sanity checks... these are all serious usage bugs.
1212 */
1213 if ((!name) ||
1214 in_interrupt() ||
1215 (size < BYTES_PER_WORD) ||
1216 (size > (1<<MAX_OBJ_ORDER)*PAGE_SIZE) ||
1217 (dtor && !ctor)) {
1218 printk(KERN_ERR "%s: Early error in slab %s\n",
1219 __FUNCTION__, name);
1220 BUG();
1221 }
1222
1223#if DEBUG
1224 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
1225 if ((flags & SLAB_DEBUG_INITIAL) && !ctor) {
1226 /* No constructor, but inital state check requested */
1227 printk(KERN_ERR "%s: No con, but init state check "
1228 "requested - %s\n", __FUNCTION__, name);
1229 flags &= ~SLAB_DEBUG_INITIAL;
1230 }
1231
1232#if FORCED_DEBUG
1233 /*
1234 * Enable redzoning and last user accounting, except for caches with
1235 * large objects, if the increased size would increase the object size
1236 * above the next power of two: caches with object sizes just above a
1237 * power of two have a significant amount of internal fragmentation.
1238 */
1239 if ((size < 4096 || fls(size-1) == fls(size-1+3*BYTES_PER_WORD)))
1240 flags |= SLAB_RED_ZONE|SLAB_STORE_USER;
1241 if (!(flags & SLAB_DESTROY_BY_RCU))
1242 flags |= SLAB_POISON;
1243#endif
1244 if (flags & SLAB_DESTROY_BY_RCU)
1245 BUG_ON(flags & SLAB_POISON);
1246#endif
1247 if (flags & SLAB_DESTROY_BY_RCU)
1248 BUG_ON(dtor);
1249
1250 /*
1251 * Always checks flags, a caller might be expecting debug
1252 * support which isn't available.
1253 */
1254 if (flags & ~CREATE_MASK)
1255 BUG();
1256
1257 /* Check that size is in terms of words. This is needed to avoid
1258 * unaligned accesses for some archs when redzoning is used, and makes
1259 * sure any on-slab bufctl's are also correctly aligned.
1260 */
1261 if (size & (BYTES_PER_WORD-1)) {
1262 size += (BYTES_PER_WORD-1);
1263 size &= ~(BYTES_PER_WORD-1);
1264 }
1265
1266 /* calculate out the final buffer alignment: */
1267 /* 1) arch recommendation: can be overridden for debug */
1268 if (flags & SLAB_HWCACHE_ALIGN) {
1269 /* Default alignment: as specified by the arch code.
1270 * Except if an object is really small, then squeeze multiple
1271 * objects into one cacheline.
1272 */
1273 ralign = cache_line_size();
1274 while (size <= ralign/2)
1275 ralign /= 2;
1276 } else {
1277 ralign = BYTES_PER_WORD;
1278 }
1279 /* 2) arch mandated alignment: disables debug if necessary */
1280 if (ralign < ARCH_SLAB_MINALIGN) {
1281 ralign = ARCH_SLAB_MINALIGN;
1282 if (ralign > BYTES_PER_WORD)
1283 flags &= ~(SLAB_RED_ZONE|SLAB_STORE_USER);
1284 }
1285 /* 3) caller mandated alignment: disables debug if necessary */
1286 if (ralign < align) {
1287 ralign = align;
1288 if (ralign > BYTES_PER_WORD)
1289 flags &= ~(SLAB_RED_ZONE|SLAB_STORE_USER);
1290 }
1291 /* 4) Store it. Note that the debug code below can reduce
1292 * the alignment to BYTES_PER_WORD.
1293 */
1294 align = ralign;
1295
1296 /* Get cache's description obj. */
1297 cachep = (kmem_cache_t *) kmem_cache_alloc(&cache_cache, SLAB_KERNEL);
1298 if (!cachep)
1299 goto opps;
1300 memset(cachep, 0, sizeof(kmem_cache_t));
1301
1302#if DEBUG
1303 cachep->reallen = size;
1304
1305 if (flags & SLAB_RED_ZONE) {
1306 /* redzoning only works with word aligned caches */
1307 align = BYTES_PER_WORD;
1308
1309 /* add space for red zone words */
1310 cachep->dbghead += BYTES_PER_WORD;
1311 size += 2*BYTES_PER_WORD;
1312 }
1313 if (flags & SLAB_STORE_USER) {
1314 /* user store requires word alignment and
1315 * one word storage behind the end of the real
1316 * object.
1317 */
1318 align = BYTES_PER_WORD;
1319 size += BYTES_PER_WORD;
1320 }
1321#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
1322 if (size > 128 && cachep->reallen > cache_line_size() && size < PAGE_SIZE) {
1323 cachep->dbghead += PAGE_SIZE - size;
1324 size = PAGE_SIZE;
1325 }
1326#endif
1327#endif
1328
1329 /* Determine if the slab management is 'on' or 'off' slab. */
1330 if (size >= (PAGE_SIZE>>3))
1331 /*
1332 * Size is large, assume best to place the slab management obj
1333 * off-slab (should allow better packing of objs).
1334 */
1335 flags |= CFLGS_OFF_SLAB;
1336
1337 size = ALIGN(size, align);
1338
1339 if ((flags & SLAB_RECLAIM_ACCOUNT) && size <= PAGE_SIZE) {
1340 /*
1341 * A VFS-reclaimable slab tends to have most allocations
1342 * as GFP_NOFS and we really don't want to have to be allocating
1343 * higher-order pages when we are unable to shrink dcache.
1344 */
1345 cachep->gfporder = 0;
1346 cache_estimate(cachep->gfporder, size, align, flags,
1347 &left_over, &cachep->num);
1348 } else {
1349 /*
1350 * Calculate size (in pages) of slabs, and the num of objs per
1351 * slab. This could be made much more intelligent. For now,
1352 * try to avoid using high page-orders for slabs. When the
1353 * gfp() funcs are more friendly towards high-order requests,
1354 * this should be changed.
1355 */
1356 do {
1357 unsigned int break_flag = 0;
1358cal_wastage:
1359 cache_estimate(cachep->gfporder, size, align, flags,
1360 &left_over, &cachep->num);
1361 if (break_flag)
1362 break;
1363 if (cachep->gfporder >= MAX_GFP_ORDER)
1364 break;
1365 if (!cachep->num)
1366 goto next;
1367 if (flags & CFLGS_OFF_SLAB &&
1368 cachep->num > offslab_limit) {
1369 /* This num of objs will cause problems. */
1370 cachep->gfporder--;
1371 break_flag++;
1372 goto cal_wastage;
1373 }
1374
1375 /*
1376 * Large num of objs is good, but v. large slabs are
1377 * currently bad for the gfp()s.
1378 */
1379 if (cachep->gfporder >= slab_break_gfp_order)
1380 break;
1381
1382 if ((left_over*8) <= (PAGE_SIZE<<cachep->gfporder))
1383 break; /* Acceptable internal fragmentation. */
1384next:
1385 cachep->gfporder++;
1386 } while (1);
1387 }
1388
1389 if (!cachep->num) {
1390 printk("kmem_cache_create: couldn't create cache %s.\n", name);
1391 kmem_cache_free(&cache_cache, cachep);
1392 cachep = NULL;
1393 goto opps;
1394 }
1395 slab_size = ALIGN(cachep->num*sizeof(kmem_bufctl_t)
1396 + sizeof(struct slab), align);
1397
1398 /*
1399 * If the slab has been placed off-slab, and we have enough space then
1400 * move it on-slab. This is at the expense of any extra colouring.
1401 */
1402 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
1403 flags &= ~CFLGS_OFF_SLAB;
1404 left_over -= slab_size;
1405 }
1406
1407 if (flags & CFLGS_OFF_SLAB) {
1408 /* really off slab. No need for manual alignment */
1409 slab_size = cachep->num*sizeof(kmem_bufctl_t)+sizeof(struct slab);
1410 }
1411
1412 cachep->colour_off = cache_line_size();
1413 /* Offset must be a multiple of the alignment. */
1414 if (cachep->colour_off < align)
1415 cachep->colour_off = align;
1416 cachep->colour = left_over/cachep->colour_off;
1417 cachep->slab_size = slab_size;
1418 cachep->flags = flags;
1419 cachep->gfpflags = 0;
1420 if (flags & SLAB_CACHE_DMA)
1421 cachep->gfpflags |= GFP_DMA;
1422 spin_lock_init(&cachep->spinlock);
1423 cachep->objsize = size;
1424 /* NUMA */
1425 INIT_LIST_HEAD(&cachep->lists.slabs_full);
1426 INIT_LIST_HEAD(&cachep->lists.slabs_partial);
1427 INIT_LIST_HEAD(&cachep->lists.slabs_free);
1428
1429 if (flags & CFLGS_OFF_SLAB)
1430 cachep->slabp_cache = kmem_find_general_cachep(slab_size,0);
1431 cachep->ctor = ctor;
1432 cachep->dtor = dtor;
1433 cachep->name = name;
1434
1435 /* Don't let CPUs to come and go */
1436 lock_cpu_hotplug();
1437
1438 if (g_cpucache_up == FULL) {
1439 enable_cpucache(cachep);
1440 } else {
1441 if (g_cpucache_up == NONE) {
1442 /* Note: the first kmem_cache_create must create
1443 * the cache that's used by kmalloc(24), otherwise
1444 * the creation of further caches will BUG().
1445 */
1446 cachep->array[smp_processor_id()] = &initarray_generic.cache;
1447 g_cpucache_up = PARTIAL;
1448 } else {
1449 cachep->array[smp_processor_id()] = kmalloc(sizeof(struct arraycache_init),GFP_KERNEL);
1450 }
1451 BUG_ON(!ac_data(cachep));
1452 ac_data(cachep)->avail = 0;
1453 ac_data(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1454 ac_data(cachep)->batchcount = 1;
1455 ac_data(cachep)->touched = 0;
1456 cachep->batchcount = 1;
1457 cachep->limit = BOOT_CPUCACHE_ENTRIES;
1458 cachep->free_limit = (1+num_online_cpus())*cachep->batchcount
1459 + cachep->num;
1460 }
1461
1462 cachep->lists.next_reap = jiffies + REAPTIMEOUT_LIST3 +
1463 ((unsigned long)cachep)%REAPTIMEOUT_LIST3;
1464
1465 /* Need the semaphore to access the chain. */
1466 down(&cache_chain_sem);
1467 {
1468 struct list_head *p;
1469 mm_segment_t old_fs;
1470
1471 old_fs = get_fs();
1472 set_fs(KERNEL_DS);
1473 list_for_each(p, &cache_chain) {
1474 kmem_cache_t *pc = list_entry(p, kmem_cache_t, next);
1475 char tmp;
1476 /* This happens when the module gets unloaded and doesn't
1477 destroy its slab cache and noone else reuses the vmalloc
1478 area of the module. Print a warning. */
1479 if (__get_user(tmp,pc->name)) {
1480 printk("SLAB: cache with size %d has lost its name\n",
1481 pc->objsize);
1482 continue;
1483 }
1484 if (!strcmp(pc->name,name)) {
1485 printk("kmem_cache_create: duplicate cache %s\n",name);
1486 up(&cache_chain_sem);
1487 unlock_cpu_hotplug();
1488 BUG();
1489 }
1490 }
1491 set_fs(old_fs);
1492 }
1493
1494 /* cache setup completed, link it into the list */
1495 list_add(&cachep->next, &cache_chain);
1496 up(&cache_chain_sem);
1497 unlock_cpu_hotplug();
1498opps:
1499 if (!cachep && (flags & SLAB_PANIC))
1500 panic("kmem_cache_create(): failed to create slab `%s'\n",
1501 name);
1502 return cachep;
1503}
1504EXPORT_SYMBOL(kmem_cache_create);
1505
1506#if DEBUG
1507static void check_irq_off(void)
1508{
1509 BUG_ON(!irqs_disabled());
1510}
1511
1512static void check_irq_on(void)
1513{
1514 BUG_ON(irqs_disabled());
1515}
1516
1517static void check_spinlock_acquired(kmem_cache_t *cachep)
1518{
1519#ifdef CONFIG_SMP
1520 check_irq_off();
1521 BUG_ON(spin_trylock(&cachep->spinlock));
1522#endif
1523}
1524#else
1525#define check_irq_off() do { } while(0)
1526#define check_irq_on() do { } while(0)
1527#define check_spinlock_acquired(x) do { } while(0)
1528#endif
1529
1530/*
1531 * Waits for all CPUs to execute func().
1532 */
1533static void smp_call_function_all_cpus(void (*func) (void *arg), void *arg)
1534{
1535 check_irq_on();
1536 preempt_disable();
1537
1538 local_irq_disable();
1539 func(arg);
1540 local_irq_enable();
1541
1542 if (smp_call_function(func, arg, 1, 1))
1543 BUG();
1544
1545 preempt_enable();
1546}
1547
1548static void drain_array_locked(kmem_cache_t* cachep,
1549 struct array_cache *ac, int force);
1550
1551static void do_drain(void *arg)
1552{
1553 kmem_cache_t *cachep = (kmem_cache_t*)arg;
1554 struct array_cache *ac;
1555
1556 check_irq_off();
1557 ac = ac_data(cachep);
1558 spin_lock(&cachep->spinlock);
1559 free_block(cachep, &ac_entry(ac)[0], ac->avail);
1560 spin_unlock(&cachep->spinlock);
1561 ac->avail = 0;
1562}
1563
1564static void drain_cpu_caches(kmem_cache_t *cachep)
1565{
1566 smp_call_function_all_cpus(do_drain, cachep);
1567 check_irq_on();
1568 spin_lock_irq(&cachep->spinlock);
1569 if (cachep->lists.shared)
1570 drain_array_locked(cachep, cachep->lists.shared, 1);
1571 spin_unlock_irq(&cachep->spinlock);
1572}
1573
1574
1575/* NUMA shrink all list3s */
1576static int __cache_shrink(kmem_cache_t *cachep)
1577{
1578 struct slab *slabp;
1579 int ret;
1580
1581 drain_cpu_caches(cachep);
1582
1583 check_irq_on();
1584 spin_lock_irq(&cachep->spinlock);
1585
1586 for(;;) {
1587 struct list_head *p;
1588
1589 p = cachep->lists.slabs_free.prev;
1590 if (p == &cachep->lists.slabs_free)
1591 break;
1592
1593 slabp = list_entry(cachep->lists.slabs_free.prev, struct slab, list);
1594#if DEBUG
1595 if (slabp->inuse)
1596 BUG();
1597#endif
1598 list_del(&slabp->list);
1599
1600 cachep->lists.free_objects -= cachep->num;
1601 spin_unlock_irq(&cachep->spinlock);
1602 slab_destroy(cachep, slabp);
1603 spin_lock_irq(&cachep->spinlock);
1604 }
1605 ret = !list_empty(&cachep->lists.slabs_full) ||
1606 !list_empty(&cachep->lists.slabs_partial);
1607 spin_unlock_irq(&cachep->spinlock);
1608 return ret;
1609}
1610
1611/**
1612 * kmem_cache_shrink - Shrink a cache.
1613 * @cachep: The cache to shrink.
1614 *
1615 * Releases as many slabs as possible for a cache.
1616 * To help debugging, a zero exit status indicates all slabs were released.
1617 */
1618int kmem_cache_shrink(kmem_cache_t *cachep)
1619{
1620 if (!cachep || in_interrupt())
1621 BUG();
1622
1623 return __cache_shrink(cachep);
1624}
1625EXPORT_SYMBOL(kmem_cache_shrink);
1626
1627/**
1628 * kmem_cache_destroy - delete a cache
1629 * @cachep: the cache to destroy
1630 *
1631 * Remove a kmem_cache_t object from the slab cache.
1632 * Returns 0 on success.
1633 *
1634 * It is expected this function will be called by a module when it is
1635 * unloaded. This will remove the cache completely, and avoid a duplicate
1636 * cache being allocated each time a module is loaded and unloaded, if the
1637 * module doesn't have persistent in-kernel storage across loads and unloads.
1638 *
1639 * The cache must be empty before calling this function.
1640 *
1641 * The caller must guarantee that noone will allocate memory from the cache
1642 * during the kmem_cache_destroy().
1643 */
1644int kmem_cache_destroy(kmem_cache_t * cachep)
1645{
1646 int i;
1647
1648 if (!cachep || in_interrupt())
1649 BUG();
1650
1651 /* Don't let CPUs to come and go */
1652 lock_cpu_hotplug();
1653
1654 /* Find the cache in the chain of caches. */
1655 down(&cache_chain_sem);
1656 /*
1657 * the chain is never empty, cache_cache is never destroyed
1658 */
1659 list_del(&cachep->next);
1660 up(&cache_chain_sem);
1661
1662 if (__cache_shrink(cachep)) {
1663 slab_error(cachep, "Can't free all objects");
1664 down(&cache_chain_sem);
1665 list_add(&cachep->next,&cache_chain);
1666 up(&cache_chain_sem);
1667 unlock_cpu_hotplug();
1668 return 1;
1669 }
1670
1671 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
Paul E. McKenneyfbd568a3e2005-05-01 08:59:04 -07001672 synchronize_rcu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001673
1674 /* no cpu_online check required here since we clear the percpu
1675 * array on cpu offline and set this to NULL.
1676 */
1677 for (i = 0; i < NR_CPUS; i++)
1678 kfree(cachep->array[i]);
1679
1680 /* NUMA: free the list3 structures */
1681 kfree(cachep->lists.shared);
1682 cachep->lists.shared = NULL;
1683 kmem_cache_free(&cache_cache, cachep);
1684
1685 unlock_cpu_hotplug();
1686
1687 return 0;
1688}
1689EXPORT_SYMBOL(kmem_cache_destroy);
1690
1691/* Get the memory for a slab management obj. */
1692static struct slab* alloc_slabmgmt(kmem_cache_t *cachep,
1693 void *objp, int colour_off, unsigned int __nocast local_flags)
1694{
1695 struct slab *slabp;
1696
1697 if (OFF_SLAB(cachep)) {
1698 /* Slab management obj is off-slab. */
1699 slabp = kmem_cache_alloc(cachep->slabp_cache, local_flags);
1700 if (!slabp)
1701 return NULL;
1702 } else {
1703 slabp = objp+colour_off;
1704 colour_off += cachep->slab_size;
1705 }
1706 slabp->inuse = 0;
1707 slabp->colouroff = colour_off;
1708 slabp->s_mem = objp+colour_off;
1709
1710 return slabp;
1711}
1712
1713static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
1714{
1715 return (kmem_bufctl_t *)(slabp+1);
1716}
1717
1718static void cache_init_objs(kmem_cache_t *cachep,
1719 struct slab *slabp, unsigned long ctor_flags)
1720{
1721 int i;
1722
1723 for (i = 0; i < cachep->num; i++) {
1724 void* objp = slabp->s_mem+cachep->objsize*i;
1725#if DEBUG
1726 /* need to poison the objs? */
1727 if (cachep->flags & SLAB_POISON)
1728 poison_obj(cachep, objp, POISON_FREE);
1729 if (cachep->flags & SLAB_STORE_USER)
1730 *dbg_userword(cachep, objp) = NULL;
1731
1732 if (cachep->flags & SLAB_RED_ZONE) {
1733 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
1734 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
1735 }
1736 /*
1737 * Constructors are not allowed to allocate memory from
1738 * the same cache which they are a constructor for.
1739 * Otherwise, deadlock. They must also be threaded.
1740 */
1741 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
1742 cachep->ctor(objp+obj_dbghead(cachep), cachep, ctor_flags);
1743
1744 if (cachep->flags & SLAB_RED_ZONE) {
1745 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1746 slab_error(cachep, "constructor overwrote the"
1747 " end of an object");
1748 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1749 slab_error(cachep, "constructor overwrote the"
1750 " start of an object");
1751 }
1752 if ((cachep->objsize % PAGE_SIZE) == 0 && OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
1753 kernel_map_pages(virt_to_page(objp), cachep->objsize/PAGE_SIZE, 0);
1754#else
1755 if (cachep->ctor)
1756 cachep->ctor(objp, cachep, ctor_flags);
1757#endif
1758 slab_bufctl(slabp)[i] = i+1;
1759 }
1760 slab_bufctl(slabp)[i-1] = BUFCTL_END;
1761 slabp->free = 0;
1762}
1763
1764static void kmem_flagcheck(kmem_cache_t *cachep, unsigned int flags)
1765{
1766 if (flags & SLAB_DMA) {
1767 if (!(cachep->gfpflags & GFP_DMA))
1768 BUG();
1769 } else {
1770 if (cachep->gfpflags & GFP_DMA)
1771 BUG();
1772 }
1773}
1774
1775static void set_slab_attr(kmem_cache_t *cachep, struct slab *slabp, void *objp)
1776{
1777 int i;
1778 struct page *page;
1779
1780 /* Nasty!!!!!! I hope this is OK. */
1781 i = 1 << cachep->gfporder;
1782 page = virt_to_page(objp);
1783 do {
1784 SET_PAGE_CACHE(page, cachep);
1785 SET_PAGE_SLAB(page, slabp);
1786 page++;
1787 } while (--i);
1788}
1789
1790/*
1791 * Grow (by 1) the number of slabs within a cache. This is called by
1792 * kmem_cache_alloc() when there are no active objs left in a cache.
1793 */
1794static int cache_grow(kmem_cache_t *cachep, unsigned int __nocast flags, int nodeid)
1795{
1796 struct slab *slabp;
1797 void *objp;
1798 size_t offset;
1799 unsigned int local_flags;
1800 unsigned long ctor_flags;
1801
1802 /* Be lazy and only check for valid flags here,
1803 * keeping it out of the critical path in kmem_cache_alloc().
1804 */
1805 if (flags & ~(SLAB_DMA|SLAB_LEVEL_MASK|SLAB_NO_GROW))
1806 BUG();
1807 if (flags & SLAB_NO_GROW)
1808 return 0;
1809
1810 ctor_flags = SLAB_CTOR_CONSTRUCTOR;
1811 local_flags = (flags & SLAB_LEVEL_MASK);
1812 if (!(local_flags & __GFP_WAIT))
1813 /*
1814 * Not allowed to sleep. Need to tell a constructor about
1815 * this - it might need to know...
1816 */
1817 ctor_flags |= SLAB_CTOR_ATOMIC;
1818
1819 /* About to mess with non-constant members - lock. */
1820 check_irq_off();
1821 spin_lock(&cachep->spinlock);
1822
1823 /* Get colour for the slab, and cal the next value. */
1824 offset = cachep->colour_next;
1825 cachep->colour_next++;
1826 if (cachep->colour_next >= cachep->colour)
1827 cachep->colour_next = 0;
1828 offset *= cachep->colour_off;
1829
1830 spin_unlock(&cachep->spinlock);
1831
1832 if (local_flags & __GFP_WAIT)
1833 local_irq_enable();
1834
1835 /*
1836 * The test for missing atomic flag is performed here, rather than
1837 * the more obvious place, simply to reduce the critical path length
1838 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
1839 * will eventually be caught here (where it matters).
1840 */
1841 kmem_flagcheck(cachep, flags);
1842
1843
1844 /* Get mem for the objs. */
1845 if (!(objp = kmem_getpages(cachep, flags, nodeid)))
1846 goto failed;
1847
1848 /* Get slab management. */
1849 if (!(slabp = alloc_slabmgmt(cachep, objp, offset, local_flags)))
1850 goto opps1;
1851
1852 set_slab_attr(cachep, slabp, objp);
1853
1854 cache_init_objs(cachep, slabp, ctor_flags);
1855
1856 if (local_flags & __GFP_WAIT)
1857 local_irq_disable();
1858 check_irq_off();
1859 spin_lock(&cachep->spinlock);
1860
1861 /* Make slab active. */
1862 list_add_tail(&slabp->list, &(list3_data(cachep)->slabs_free));
1863 STATS_INC_GROWN(cachep);
1864 list3_data(cachep)->free_objects += cachep->num;
1865 spin_unlock(&cachep->spinlock);
1866 return 1;
1867opps1:
1868 kmem_freepages(cachep, objp);
1869failed:
1870 if (local_flags & __GFP_WAIT)
1871 local_irq_disable();
1872 return 0;
1873}
1874
1875#if DEBUG
1876
1877/*
1878 * Perform extra freeing checks:
1879 * - detect bad pointers.
1880 * - POISON/RED_ZONE checking
1881 * - destructor calls, for caches with POISON+dtor
1882 */
1883static void kfree_debugcheck(const void *objp)
1884{
1885 struct page *page;
1886
1887 if (!virt_addr_valid(objp)) {
1888 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
1889 (unsigned long)objp);
1890 BUG();
1891 }
1892 page = virt_to_page(objp);
1893 if (!PageSlab(page)) {
1894 printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n", (unsigned long)objp);
1895 BUG();
1896 }
1897}
1898
1899static void *cache_free_debugcheck(kmem_cache_t *cachep, void *objp,
1900 void *caller)
1901{
1902 struct page *page;
1903 unsigned int objnr;
1904 struct slab *slabp;
1905
1906 objp -= obj_dbghead(cachep);
1907 kfree_debugcheck(objp);
1908 page = virt_to_page(objp);
1909
1910 if (GET_PAGE_CACHE(page) != cachep) {
1911 printk(KERN_ERR "mismatch in kmem_cache_free: expected cache %p, got %p\n",
1912 GET_PAGE_CACHE(page),cachep);
1913 printk(KERN_ERR "%p is %s.\n", cachep, cachep->name);
1914 printk(KERN_ERR "%p is %s.\n", GET_PAGE_CACHE(page), GET_PAGE_CACHE(page)->name);
1915 WARN_ON(1);
1916 }
1917 slabp = GET_PAGE_SLAB(page);
1918
1919 if (cachep->flags & SLAB_RED_ZONE) {
1920 if (*dbg_redzone1(cachep, objp) != RED_ACTIVE || *dbg_redzone2(cachep, objp) != RED_ACTIVE) {
1921 slab_error(cachep, "double free, or memory outside"
1922 " object was overwritten");
1923 printk(KERN_ERR "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n",
1924 objp, *dbg_redzone1(cachep, objp), *dbg_redzone2(cachep, objp));
1925 }
1926 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
1927 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
1928 }
1929 if (cachep->flags & SLAB_STORE_USER)
1930 *dbg_userword(cachep, objp) = caller;
1931
1932 objnr = (objp-slabp->s_mem)/cachep->objsize;
1933
1934 BUG_ON(objnr >= cachep->num);
1935 BUG_ON(objp != slabp->s_mem + objnr*cachep->objsize);
1936
1937 if (cachep->flags & SLAB_DEBUG_INITIAL) {
1938 /* Need to call the slab's constructor so the
1939 * caller can perform a verify of its state (debugging).
1940 * Called without the cache-lock held.
1941 */
1942 cachep->ctor(objp+obj_dbghead(cachep),
1943 cachep, SLAB_CTOR_CONSTRUCTOR|SLAB_CTOR_VERIFY);
1944 }
1945 if (cachep->flags & SLAB_POISON && cachep->dtor) {
1946 /* we want to cache poison the object,
1947 * call the destruction callback
1948 */
1949 cachep->dtor(objp+obj_dbghead(cachep), cachep, 0);
1950 }
1951 if (cachep->flags & SLAB_POISON) {
1952#ifdef CONFIG_DEBUG_PAGEALLOC
1953 if ((cachep->objsize % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) {
1954 store_stackinfo(cachep, objp, (unsigned long)caller);
1955 kernel_map_pages(virt_to_page(objp), cachep->objsize/PAGE_SIZE, 0);
1956 } else {
1957 poison_obj(cachep, objp, POISON_FREE);
1958 }
1959#else
1960 poison_obj(cachep, objp, POISON_FREE);
1961#endif
1962 }
1963 return objp;
1964}
1965
1966static void check_slabp(kmem_cache_t *cachep, struct slab *slabp)
1967{
1968 kmem_bufctl_t i;
1969 int entries = 0;
1970
1971 check_spinlock_acquired(cachep);
1972 /* Check slab's freelist to see if this obj is there. */
1973 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
1974 entries++;
1975 if (entries > cachep->num || i >= cachep->num)
1976 goto bad;
1977 }
1978 if (entries != cachep->num - slabp->inuse) {
1979bad:
1980 printk(KERN_ERR "slab: Internal list corruption detected in cache '%s'(%d), slabp %p(%d). Hexdump:\n",
1981 cachep->name, cachep->num, slabp, slabp->inuse);
1982 for (i=0;i<sizeof(slabp)+cachep->num*sizeof(kmem_bufctl_t);i++) {
1983 if ((i%16)==0)
1984 printk("\n%03x:", i);
1985 printk(" %02x", ((unsigned char*)slabp)[i]);
1986 }
1987 printk("\n");
1988 BUG();
1989 }
1990}
1991#else
1992#define kfree_debugcheck(x) do { } while(0)
1993#define cache_free_debugcheck(x,objp,z) (objp)
1994#define check_slabp(x,y) do { } while(0)
1995#endif
1996
1997static void *cache_alloc_refill(kmem_cache_t *cachep, unsigned int __nocast flags)
1998{
1999 int batchcount;
2000 struct kmem_list3 *l3;
2001 struct array_cache *ac;
2002
2003 check_irq_off();
2004 ac = ac_data(cachep);
2005retry:
2006 batchcount = ac->batchcount;
2007 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2008 /* if there was little recent activity on this
2009 * cache, then perform only a partial refill.
2010 * Otherwise we could generate refill bouncing.
2011 */
2012 batchcount = BATCHREFILL_LIMIT;
2013 }
2014 l3 = list3_data(cachep);
2015
2016 BUG_ON(ac->avail > 0);
2017 spin_lock(&cachep->spinlock);
2018 if (l3->shared) {
2019 struct array_cache *shared_array = l3->shared;
2020 if (shared_array->avail) {
2021 if (batchcount > shared_array->avail)
2022 batchcount = shared_array->avail;
2023 shared_array->avail -= batchcount;
2024 ac->avail = batchcount;
2025 memcpy(ac_entry(ac), &ac_entry(shared_array)[shared_array->avail],
2026 sizeof(void*)*batchcount);
2027 shared_array->touched = 1;
2028 goto alloc_done;
2029 }
2030 }
2031 while (batchcount > 0) {
2032 struct list_head *entry;
2033 struct slab *slabp;
2034 /* Get slab alloc is to come from. */
2035 entry = l3->slabs_partial.next;
2036 if (entry == &l3->slabs_partial) {
2037 l3->free_touched = 1;
2038 entry = l3->slabs_free.next;
2039 if (entry == &l3->slabs_free)
2040 goto must_grow;
2041 }
2042
2043 slabp = list_entry(entry, struct slab, list);
2044 check_slabp(cachep, slabp);
2045 check_spinlock_acquired(cachep);
2046 while (slabp->inuse < cachep->num && batchcount--) {
2047 kmem_bufctl_t next;
2048 STATS_INC_ALLOCED(cachep);
2049 STATS_INC_ACTIVE(cachep);
2050 STATS_SET_HIGH(cachep);
2051
2052 /* get obj pointer */
2053 ac_entry(ac)[ac->avail++] = slabp->s_mem + slabp->free*cachep->objsize;
2054
2055 slabp->inuse++;
2056 next = slab_bufctl(slabp)[slabp->free];
2057#if DEBUG
2058 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2059#endif
2060 slabp->free = next;
2061 }
2062 check_slabp(cachep, slabp);
2063
2064 /* move slabp to correct slabp list: */
2065 list_del(&slabp->list);
2066 if (slabp->free == BUFCTL_END)
2067 list_add(&slabp->list, &l3->slabs_full);
2068 else
2069 list_add(&slabp->list, &l3->slabs_partial);
2070 }
2071
2072must_grow:
2073 l3->free_objects -= ac->avail;
2074alloc_done:
2075 spin_unlock(&cachep->spinlock);
2076
2077 if (unlikely(!ac->avail)) {
2078 int x;
2079 x = cache_grow(cachep, flags, -1);
2080
2081 // cache_grow can reenable interrupts, then ac could change.
2082 ac = ac_data(cachep);
2083 if (!x && ac->avail == 0) // no objects in sight? abort
2084 return NULL;
2085
2086 if (!ac->avail) // objects refilled by interrupt?
2087 goto retry;
2088 }
2089 ac->touched = 1;
2090 return ac_entry(ac)[--ac->avail];
2091}
2092
2093static inline void
2094cache_alloc_debugcheck_before(kmem_cache_t *cachep, unsigned int __nocast flags)
2095{
2096 might_sleep_if(flags & __GFP_WAIT);
2097#if DEBUG
2098 kmem_flagcheck(cachep, flags);
2099#endif
2100}
2101
2102#if DEBUG
2103static void *
2104cache_alloc_debugcheck_after(kmem_cache_t *cachep,
Alexey Dobriyan0db925a2005-07-07 17:56:58 -07002105 unsigned int __nocast flags, void *objp, void *caller)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002106{
2107 if (!objp)
2108 return objp;
2109 if (cachep->flags & SLAB_POISON) {
2110#ifdef CONFIG_DEBUG_PAGEALLOC
2111 if ((cachep->objsize % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
2112 kernel_map_pages(virt_to_page(objp), cachep->objsize/PAGE_SIZE, 1);
2113 else
2114 check_poison_obj(cachep, objp);
2115#else
2116 check_poison_obj(cachep, objp);
2117#endif
2118 poison_obj(cachep, objp, POISON_INUSE);
2119 }
2120 if (cachep->flags & SLAB_STORE_USER)
2121 *dbg_userword(cachep, objp) = caller;
2122
2123 if (cachep->flags & SLAB_RED_ZONE) {
2124 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE || *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2125 slab_error(cachep, "double free, or memory outside"
2126 " object was overwritten");
2127 printk(KERN_ERR "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n",
2128 objp, *dbg_redzone1(cachep, objp), *dbg_redzone2(cachep, objp));
2129 }
2130 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2131 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2132 }
2133 objp += obj_dbghead(cachep);
2134 if (cachep->ctor && cachep->flags & SLAB_POISON) {
2135 unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2136
2137 if (!(flags & __GFP_WAIT))
2138 ctor_flags |= SLAB_CTOR_ATOMIC;
2139
2140 cachep->ctor(objp, cachep, ctor_flags);
2141 }
2142 return objp;
2143}
2144#else
2145#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2146#endif
2147
2148
2149static inline void *__cache_alloc(kmem_cache_t *cachep, unsigned int __nocast flags)
2150{
2151 unsigned long save_flags;
2152 void* objp;
2153 struct array_cache *ac;
2154
2155 cache_alloc_debugcheck_before(cachep, flags);
2156
2157 local_irq_save(save_flags);
2158 ac = ac_data(cachep);
2159 if (likely(ac->avail)) {
2160 STATS_INC_ALLOCHIT(cachep);
2161 ac->touched = 1;
2162 objp = ac_entry(ac)[--ac->avail];
2163 } else {
2164 STATS_INC_ALLOCMISS(cachep);
2165 objp = cache_alloc_refill(cachep, flags);
2166 }
2167 local_irq_restore(save_flags);
2168 objp = cache_alloc_debugcheck_after(cachep, flags, objp, __builtin_return_address(0));
2169 return objp;
2170}
2171
2172/*
2173 * NUMA: different approach needed if the spinlock is moved into
2174 * the l3 structure
2175 */
2176
2177static void free_block(kmem_cache_t *cachep, void **objpp, int nr_objects)
2178{
2179 int i;
2180
2181 check_spinlock_acquired(cachep);
2182
2183 /* NUMA: move add into loop */
2184 cachep->lists.free_objects += nr_objects;
2185
2186 for (i = 0; i < nr_objects; i++) {
2187 void *objp = objpp[i];
2188 struct slab *slabp;
2189 unsigned int objnr;
2190
2191 slabp = GET_PAGE_SLAB(virt_to_page(objp));
2192 list_del(&slabp->list);
2193 objnr = (objp - slabp->s_mem) / cachep->objsize;
2194 check_slabp(cachep, slabp);
2195#if DEBUG
2196 if (slab_bufctl(slabp)[objnr] != BUFCTL_FREE) {
2197 printk(KERN_ERR "slab: double free detected in cache '%s', objp %p.\n",
2198 cachep->name, objp);
2199 BUG();
2200 }
2201#endif
2202 slab_bufctl(slabp)[objnr] = slabp->free;
2203 slabp->free = objnr;
2204 STATS_DEC_ACTIVE(cachep);
2205 slabp->inuse--;
2206 check_slabp(cachep, slabp);
2207
2208 /* fixup slab chains */
2209 if (slabp->inuse == 0) {
2210 if (cachep->lists.free_objects > cachep->free_limit) {
2211 cachep->lists.free_objects -= cachep->num;
2212 slab_destroy(cachep, slabp);
2213 } else {
2214 list_add(&slabp->list,
2215 &list3_data_ptr(cachep, objp)->slabs_free);
2216 }
2217 } else {
2218 /* Unconditionally move a slab to the end of the
2219 * partial list on free - maximum time for the
2220 * other objects to be freed, too.
2221 */
2222 list_add_tail(&slabp->list,
2223 &list3_data_ptr(cachep, objp)->slabs_partial);
2224 }
2225 }
2226}
2227
2228static void cache_flusharray(kmem_cache_t *cachep, struct array_cache *ac)
2229{
2230 int batchcount;
2231
2232 batchcount = ac->batchcount;
2233#if DEBUG
2234 BUG_ON(!batchcount || batchcount > ac->avail);
2235#endif
2236 check_irq_off();
2237 spin_lock(&cachep->spinlock);
2238 if (cachep->lists.shared) {
2239 struct array_cache *shared_array = cachep->lists.shared;
2240 int max = shared_array->limit-shared_array->avail;
2241 if (max) {
2242 if (batchcount > max)
2243 batchcount = max;
2244 memcpy(&ac_entry(shared_array)[shared_array->avail],
2245 &ac_entry(ac)[0],
2246 sizeof(void*)*batchcount);
2247 shared_array->avail += batchcount;
2248 goto free_done;
2249 }
2250 }
2251
2252 free_block(cachep, &ac_entry(ac)[0], batchcount);
2253free_done:
2254#if STATS
2255 {
2256 int i = 0;
2257 struct list_head *p;
2258
2259 p = list3_data(cachep)->slabs_free.next;
2260 while (p != &(list3_data(cachep)->slabs_free)) {
2261 struct slab *slabp;
2262
2263 slabp = list_entry(p, struct slab, list);
2264 BUG_ON(slabp->inuse);
2265
2266 i++;
2267 p = p->next;
2268 }
2269 STATS_SET_FREEABLE(cachep, i);
2270 }
2271#endif
2272 spin_unlock(&cachep->spinlock);
2273 ac->avail -= batchcount;
2274 memmove(&ac_entry(ac)[0], &ac_entry(ac)[batchcount],
2275 sizeof(void*)*ac->avail);
2276}
2277
2278/*
2279 * __cache_free
2280 * Release an obj back to its cache. If the obj has a constructed
2281 * state, it must be in this state _before_ it is released.
2282 *
2283 * Called with disabled ints.
2284 */
2285static inline void __cache_free(kmem_cache_t *cachep, void *objp)
2286{
2287 struct array_cache *ac = ac_data(cachep);
2288
2289 check_irq_off();
2290 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
2291
2292 if (likely(ac->avail < ac->limit)) {
2293 STATS_INC_FREEHIT(cachep);
2294 ac_entry(ac)[ac->avail++] = objp;
2295 return;
2296 } else {
2297 STATS_INC_FREEMISS(cachep);
2298 cache_flusharray(cachep, ac);
2299 ac_entry(ac)[ac->avail++] = objp;
2300 }
2301}
2302
2303/**
2304 * kmem_cache_alloc - Allocate an object
2305 * @cachep: The cache to allocate from.
2306 * @flags: See kmalloc().
2307 *
2308 * Allocate an object from this cache. The flags are only relevant
2309 * if the cache has no available objects.
2310 */
2311void *kmem_cache_alloc(kmem_cache_t *cachep, unsigned int __nocast flags)
2312{
2313 return __cache_alloc(cachep, flags);
2314}
2315EXPORT_SYMBOL(kmem_cache_alloc);
2316
2317/**
2318 * kmem_ptr_validate - check if an untrusted pointer might
2319 * be a slab entry.
2320 * @cachep: the cache we're checking against
2321 * @ptr: pointer to validate
2322 *
2323 * This verifies that the untrusted pointer looks sane:
2324 * it is _not_ a guarantee that the pointer is actually
2325 * part of the slab cache in question, but it at least
2326 * validates that the pointer can be dereferenced and
2327 * looks half-way sane.
2328 *
2329 * Currently only used for dentry validation.
2330 */
2331int fastcall kmem_ptr_validate(kmem_cache_t *cachep, void *ptr)
2332{
2333 unsigned long addr = (unsigned long) ptr;
2334 unsigned long min_addr = PAGE_OFFSET;
2335 unsigned long align_mask = BYTES_PER_WORD-1;
2336 unsigned long size = cachep->objsize;
2337 struct page *page;
2338
2339 if (unlikely(addr < min_addr))
2340 goto out;
2341 if (unlikely(addr > (unsigned long)high_memory - size))
2342 goto out;
2343 if (unlikely(addr & align_mask))
2344 goto out;
2345 if (unlikely(!kern_addr_valid(addr)))
2346 goto out;
2347 if (unlikely(!kern_addr_valid(addr + size - 1)))
2348 goto out;
2349 page = virt_to_page(ptr);
2350 if (unlikely(!PageSlab(page)))
2351 goto out;
2352 if (unlikely(GET_PAGE_CACHE(page) != cachep))
2353 goto out;
2354 return 1;
2355out:
2356 return 0;
2357}
2358
2359#ifdef CONFIG_NUMA
2360/**
2361 * kmem_cache_alloc_node - Allocate an object on the specified node
2362 * @cachep: The cache to allocate from.
2363 * @flags: See kmalloc().
2364 * @nodeid: node number of the target node.
2365 *
2366 * Identical to kmem_cache_alloc, except that this function is slow
2367 * and can sleep. And it will allocate memory on the given node, which
2368 * can improve the performance for cpu bound structures.
2369 */
Manfred Spraul97e2bde2005-05-01 08:58:38 -07002370void *kmem_cache_alloc_node(kmem_cache_t *cachep, int flags, int nodeid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002371{
2372 int loop;
2373 void *objp;
2374 struct slab *slabp;
2375 kmem_bufctl_t next;
2376
Christoph Lameter83b78bd2005-07-06 10:47:07 -07002377 if (nodeid == -1)
2378 return kmem_cache_alloc(cachep, flags);
2379
Linus Torvalds1da177e2005-04-16 15:20:36 -07002380 for (loop = 0;;loop++) {
2381 struct list_head *q;
2382
2383 objp = NULL;
2384 check_irq_on();
2385 spin_lock_irq(&cachep->spinlock);
2386 /* walk through all partial and empty slab and find one
2387 * from the right node */
2388 list_for_each(q,&cachep->lists.slabs_partial) {
2389 slabp = list_entry(q, struct slab, list);
2390
2391 if (page_to_nid(virt_to_page(slabp->s_mem)) == nodeid ||
2392 loop > 2)
2393 goto got_slabp;
2394 }
2395 list_for_each(q, &cachep->lists.slabs_free) {
2396 slabp = list_entry(q, struct slab, list);
2397
2398 if (page_to_nid(virt_to_page(slabp->s_mem)) == nodeid ||
2399 loop > 2)
2400 goto got_slabp;
2401 }
2402 spin_unlock_irq(&cachep->spinlock);
2403
2404 local_irq_disable();
Manfred Spraul97e2bde2005-05-01 08:58:38 -07002405 if (!cache_grow(cachep, flags, nodeid)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002406 local_irq_enable();
2407 return NULL;
2408 }
2409 local_irq_enable();
2410 }
2411got_slabp:
2412 /* found one: allocate object */
2413 check_slabp(cachep, slabp);
2414 check_spinlock_acquired(cachep);
2415
2416 STATS_INC_ALLOCED(cachep);
2417 STATS_INC_ACTIVE(cachep);
2418 STATS_SET_HIGH(cachep);
2419 STATS_INC_NODEALLOCS(cachep);
2420
2421 objp = slabp->s_mem + slabp->free*cachep->objsize;
2422
2423 slabp->inuse++;
2424 next = slab_bufctl(slabp)[slabp->free];
2425#if DEBUG
2426 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2427#endif
2428 slabp->free = next;
2429 check_slabp(cachep, slabp);
2430
2431 /* move slabp to correct slabp list: */
2432 list_del(&slabp->list);
2433 if (slabp->free == BUFCTL_END)
2434 list_add(&slabp->list, &cachep->lists.slabs_full);
2435 else
2436 list_add(&slabp->list, &cachep->lists.slabs_partial);
2437
2438 list3_data(cachep)->free_objects--;
2439 spin_unlock_irq(&cachep->spinlock);
2440
2441 objp = cache_alloc_debugcheck_after(cachep, GFP_KERNEL, objp,
2442 __builtin_return_address(0));
2443 return objp;
2444}
2445EXPORT_SYMBOL(kmem_cache_alloc_node);
2446
Alexey Dobriyan0db925a2005-07-07 17:56:58 -07002447void *kmalloc_node(size_t size, unsigned int __nocast flags, int node)
Manfred Spraul97e2bde2005-05-01 08:58:38 -07002448{
2449 kmem_cache_t *cachep;
2450
2451 cachep = kmem_find_general_cachep(size, flags);
2452 if (unlikely(cachep == NULL))
2453 return NULL;
2454 return kmem_cache_alloc_node(cachep, flags, node);
2455}
2456EXPORT_SYMBOL(kmalloc_node);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002457#endif
2458
2459/**
2460 * kmalloc - allocate memory
2461 * @size: how many bytes of memory are required.
2462 * @flags: the type of memory to allocate.
2463 *
2464 * kmalloc is the normal method of allocating memory
2465 * in the kernel.
2466 *
2467 * The @flags argument may be one of:
2468 *
2469 * %GFP_USER - Allocate memory on behalf of user. May sleep.
2470 *
2471 * %GFP_KERNEL - Allocate normal kernel ram. May sleep.
2472 *
2473 * %GFP_ATOMIC - Allocation will not sleep. Use inside interrupt handlers.
2474 *
2475 * Additionally, the %GFP_DMA flag may be set to indicate the memory
2476 * must be suitable for DMA. This can mean different things on different
2477 * platforms. For example, on i386, it means that the memory must come
2478 * from the first 16MB.
2479 */
2480void *__kmalloc(size_t size, unsigned int __nocast flags)
2481{
2482 kmem_cache_t *cachep;
2483
Manfred Spraul97e2bde2005-05-01 08:58:38 -07002484 /* If you want to save a few bytes .text space: replace
2485 * __ with kmem_.
2486 * Then kmalloc uses the uninlined functions instead of the inline
2487 * functions.
2488 */
2489 cachep = __find_general_cachep(size, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002490 if (unlikely(cachep == NULL))
2491 return NULL;
2492 return __cache_alloc(cachep, flags);
2493}
2494EXPORT_SYMBOL(__kmalloc);
2495
2496#ifdef CONFIG_SMP
2497/**
2498 * __alloc_percpu - allocate one copy of the object for every present
2499 * cpu in the system, zeroing them.
2500 * Objects should be dereferenced using the per_cpu_ptr macro only.
2501 *
2502 * @size: how many bytes of memory are required.
2503 * @align: the alignment, which can't be greater than SMP_CACHE_BYTES.
2504 */
2505void *__alloc_percpu(size_t size, size_t align)
2506{
2507 int i;
2508 struct percpu_data *pdata = kmalloc(sizeof (*pdata), GFP_KERNEL);
2509
2510 if (!pdata)
2511 return NULL;
2512
2513 for (i = 0; i < NR_CPUS; i++) {
2514 if (!cpu_possible(i))
2515 continue;
Manfred Spraul97e2bde2005-05-01 08:58:38 -07002516 pdata->ptrs[i] = kmalloc_node(size, GFP_KERNEL,
2517 cpu_to_node(i));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002518
2519 if (!pdata->ptrs[i])
2520 goto unwind_oom;
2521 memset(pdata->ptrs[i], 0, size);
2522 }
2523
2524 /* Catch derefs w/o wrappers */
2525 return (void *) (~(unsigned long) pdata);
2526
2527unwind_oom:
2528 while (--i >= 0) {
2529 if (!cpu_possible(i))
2530 continue;
2531 kfree(pdata->ptrs[i]);
2532 }
2533 kfree(pdata);
2534 return NULL;
2535}
2536EXPORT_SYMBOL(__alloc_percpu);
2537#endif
2538
2539/**
2540 * kmem_cache_free - Deallocate an object
2541 * @cachep: The cache the allocation was from.
2542 * @objp: The previously allocated object.
2543 *
2544 * Free an object which was previously allocated from this
2545 * cache.
2546 */
2547void kmem_cache_free(kmem_cache_t *cachep, void *objp)
2548{
2549 unsigned long flags;
2550
2551 local_irq_save(flags);
2552 __cache_free(cachep, objp);
2553 local_irq_restore(flags);
2554}
2555EXPORT_SYMBOL(kmem_cache_free);
2556
2557/**
2558 * kcalloc - allocate memory for an array. The memory is set to zero.
2559 * @n: number of elements.
2560 * @size: element size.
2561 * @flags: the type of memory to allocate.
2562 */
2563void *kcalloc(size_t n, size_t size, unsigned int __nocast flags)
2564{
2565 void *ret = NULL;
2566
2567 if (n != 0 && size > INT_MAX / n)
2568 return ret;
2569
2570 ret = kmalloc(n * size, flags);
2571 if (ret)
2572 memset(ret, 0, n * size);
2573 return ret;
2574}
2575EXPORT_SYMBOL(kcalloc);
2576
2577/**
2578 * kfree - free previously allocated memory
2579 * @objp: pointer returned by kmalloc.
2580 *
2581 * Don't free memory not originally allocated by kmalloc()
2582 * or you will run into trouble.
2583 */
2584void kfree(const void *objp)
2585{
2586 kmem_cache_t *c;
2587 unsigned long flags;
2588
2589 if (unlikely(!objp))
2590 return;
2591 local_irq_save(flags);
2592 kfree_debugcheck(objp);
2593 c = GET_PAGE_CACHE(virt_to_page(objp));
2594 __cache_free(c, (void*)objp);
2595 local_irq_restore(flags);
2596}
2597EXPORT_SYMBOL(kfree);
2598
2599#ifdef CONFIG_SMP
2600/**
2601 * free_percpu - free previously allocated percpu memory
2602 * @objp: pointer returned by alloc_percpu.
2603 *
2604 * Don't free memory not originally allocated by alloc_percpu()
2605 * The complemented objp is to check for that.
2606 */
2607void
2608free_percpu(const void *objp)
2609{
2610 int i;
2611 struct percpu_data *p = (struct percpu_data *) (~(unsigned long) objp);
2612
2613 for (i = 0; i < NR_CPUS; i++) {
2614 if (!cpu_possible(i))
2615 continue;
2616 kfree(p->ptrs[i]);
2617 }
2618 kfree(p);
2619}
2620EXPORT_SYMBOL(free_percpu);
2621#endif
2622
2623unsigned int kmem_cache_size(kmem_cache_t *cachep)
2624{
2625 return obj_reallen(cachep);
2626}
2627EXPORT_SYMBOL(kmem_cache_size);
2628
Arnaldo Carvalho de Melo19449722005-06-18 22:46:19 -07002629const char *kmem_cache_name(kmem_cache_t *cachep)
2630{
2631 return cachep->name;
2632}
2633EXPORT_SYMBOL_GPL(kmem_cache_name);
2634
Linus Torvalds1da177e2005-04-16 15:20:36 -07002635struct ccupdate_struct {
2636 kmem_cache_t *cachep;
2637 struct array_cache *new[NR_CPUS];
2638};
2639
2640static void do_ccupdate_local(void *info)
2641{
2642 struct ccupdate_struct *new = (struct ccupdate_struct *)info;
2643 struct array_cache *old;
2644
2645 check_irq_off();
2646 old = ac_data(new->cachep);
2647
2648 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
2649 new->new[smp_processor_id()] = old;
2650}
2651
2652
2653static int do_tune_cpucache(kmem_cache_t *cachep, int limit, int batchcount,
2654 int shared)
2655{
2656 struct ccupdate_struct new;
2657 struct array_cache *new_shared;
2658 int i;
2659
2660 memset(&new.new,0,sizeof(new.new));
2661 for (i = 0; i < NR_CPUS; i++) {
2662 if (cpu_online(i)) {
2663 new.new[i] = alloc_arraycache(i, limit, batchcount);
2664 if (!new.new[i]) {
2665 for (i--; i >= 0; i--) kfree(new.new[i]);
2666 return -ENOMEM;
2667 }
2668 } else {
2669 new.new[i] = NULL;
2670 }
2671 }
2672 new.cachep = cachep;
2673
2674 smp_call_function_all_cpus(do_ccupdate_local, (void *)&new);
2675
2676 check_irq_on();
2677 spin_lock_irq(&cachep->spinlock);
2678 cachep->batchcount = batchcount;
2679 cachep->limit = limit;
2680 cachep->free_limit = (1+num_online_cpus())*cachep->batchcount + cachep->num;
2681 spin_unlock_irq(&cachep->spinlock);
2682
2683 for (i = 0; i < NR_CPUS; i++) {
2684 struct array_cache *ccold = new.new[i];
2685 if (!ccold)
2686 continue;
2687 spin_lock_irq(&cachep->spinlock);
2688 free_block(cachep, ac_entry(ccold), ccold->avail);
2689 spin_unlock_irq(&cachep->spinlock);
2690 kfree(ccold);
2691 }
2692 new_shared = alloc_arraycache(-1, batchcount*shared, 0xbaadf00d);
2693 if (new_shared) {
2694 struct array_cache *old;
2695
2696 spin_lock_irq(&cachep->spinlock);
2697 old = cachep->lists.shared;
2698 cachep->lists.shared = new_shared;
2699 if (old)
2700 free_block(cachep, ac_entry(old), old->avail);
2701 spin_unlock_irq(&cachep->spinlock);
2702 kfree(old);
2703 }
2704
2705 return 0;
2706}
2707
2708
2709static void enable_cpucache(kmem_cache_t *cachep)
2710{
2711 int err;
2712 int limit, shared;
2713
2714 /* The head array serves three purposes:
2715 * - create a LIFO ordering, i.e. return objects that are cache-warm
2716 * - reduce the number of spinlock operations.
2717 * - reduce the number of linked list operations on the slab and
2718 * bufctl chains: array operations are cheaper.
2719 * The numbers are guessed, we should auto-tune as described by
2720 * Bonwick.
2721 */
2722 if (cachep->objsize > 131072)
2723 limit = 1;
2724 else if (cachep->objsize > PAGE_SIZE)
2725 limit = 8;
2726 else if (cachep->objsize > 1024)
2727 limit = 24;
2728 else if (cachep->objsize > 256)
2729 limit = 54;
2730 else
2731 limit = 120;
2732
2733 /* Cpu bound tasks (e.g. network routing) can exhibit cpu bound
2734 * allocation behaviour: Most allocs on one cpu, most free operations
2735 * on another cpu. For these cases, an efficient object passing between
2736 * cpus is necessary. This is provided by a shared array. The array
2737 * replaces Bonwick's magazine layer.
2738 * On uniprocessor, it's functionally equivalent (but less efficient)
2739 * to a larger limit. Thus disabled by default.
2740 */
2741 shared = 0;
2742#ifdef CONFIG_SMP
2743 if (cachep->objsize <= PAGE_SIZE)
2744 shared = 8;
2745#endif
2746
2747#if DEBUG
2748 /* With debugging enabled, large batchcount lead to excessively
2749 * long periods with disabled local interrupts. Limit the
2750 * batchcount
2751 */
2752 if (limit > 32)
2753 limit = 32;
2754#endif
2755 err = do_tune_cpucache(cachep, limit, (limit+1)/2, shared);
2756 if (err)
2757 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
2758 cachep->name, -err);
2759}
2760
2761static void drain_array_locked(kmem_cache_t *cachep,
2762 struct array_cache *ac, int force)
2763{
2764 int tofree;
2765
2766 check_spinlock_acquired(cachep);
2767 if (ac->touched && !force) {
2768 ac->touched = 0;
2769 } else if (ac->avail) {
2770 tofree = force ? ac->avail : (ac->limit+4)/5;
2771 if (tofree > ac->avail) {
2772 tofree = (ac->avail+1)/2;
2773 }
2774 free_block(cachep, ac_entry(ac), tofree);
2775 ac->avail -= tofree;
2776 memmove(&ac_entry(ac)[0], &ac_entry(ac)[tofree],
2777 sizeof(void*)*ac->avail);
2778 }
2779}
2780
2781/**
2782 * cache_reap - Reclaim memory from caches.
2783 *
2784 * Called from workqueue/eventd every few seconds.
2785 * Purpose:
2786 * - clear the per-cpu caches for this CPU.
2787 * - return freeable pages to the main free memory pool.
2788 *
2789 * If we cannot acquire the cache chain semaphore then just give up - we'll
2790 * try again on the next iteration.
2791 */
2792static void cache_reap(void *unused)
2793{
2794 struct list_head *walk;
2795
2796 if (down_trylock(&cache_chain_sem)) {
2797 /* Give up. Setup the next iteration. */
2798 schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC + smp_processor_id());
2799 return;
2800 }
2801
2802 list_for_each(walk, &cache_chain) {
2803 kmem_cache_t *searchp;
2804 struct list_head* p;
2805 int tofree;
2806 struct slab *slabp;
2807
2808 searchp = list_entry(walk, kmem_cache_t, next);
2809
2810 if (searchp->flags & SLAB_NO_REAP)
2811 goto next;
2812
2813 check_irq_on();
2814
2815 spin_lock_irq(&searchp->spinlock);
2816
2817 drain_array_locked(searchp, ac_data(searchp), 0);
2818
2819 if(time_after(searchp->lists.next_reap, jiffies))
2820 goto next_unlock;
2821
2822 searchp->lists.next_reap = jiffies + REAPTIMEOUT_LIST3;
2823
2824 if (searchp->lists.shared)
2825 drain_array_locked(searchp, searchp->lists.shared, 0);
2826
2827 if (searchp->lists.free_touched) {
2828 searchp->lists.free_touched = 0;
2829 goto next_unlock;
2830 }
2831
2832 tofree = (searchp->free_limit+5*searchp->num-1)/(5*searchp->num);
2833 do {
2834 p = list3_data(searchp)->slabs_free.next;
2835 if (p == &(list3_data(searchp)->slabs_free))
2836 break;
2837
2838 slabp = list_entry(p, struct slab, list);
2839 BUG_ON(slabp->inuse);
2840 list_del(&slabp->list);
2841 STATS_INC_REAPED(searchp);
2842
2843 /* Safe to drop the lock. The slab is no longer
2844 * linked to the cache.
2845 * searchp cannot disappear, we hold
2846 * cache_chain_lock
2847 */
2848 searchp->lists.free_objects -= searchp->num;
2849 spin_unlock_irq(&searchp->spinlock);
2850 slab_destroy(searchp, slabp);
2851 spin_lock_irq(&searchp->spinlock);
2852 } while(--tofree > 0);
2853next_unlock:
2854 spin_unlock_irq(&searchp->spinlock);
2855next:
2856 cond_resched();
2857 }
2858 check_irq_on();
2859 up(&cache_chain_sem);
Christoph Lameter4ae7c032005-06-21 17:14:57 -07002860 drain_remote_pages();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002861 /* Setup the next iteration */
2862 schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC + smp_processor_id());
2863}
2864
2865#ifdef CONFIG_PROC_FS
2866
2867static void *s_start(struct seq_file *m, loff_t *pos)
2868{
2869 loff_t n = *pos;
2870 struct list_head *p;
2871
2872 down(&cache_chain_sem);
2873 if (!n) {
2874 /*
2875 * Output format version, so at least we can change it
2876 * without _too_ many complaints.
2877 */
2878#if STATS
2879 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
2880#else
2881 seq_puts(m, "slabinfo - version: 2.1\n");
2882#endif
2883 seq_puts(m, "# name <active_objs> <num_objs> <objsize> <objperslab> <pagesperslab>");
2884 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
2885 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
2886#if STATS
2887 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped>"
2888 " <error> <maxfreeable> <freelimit> <nodeallocs>");
2889 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
2890#endif
2891 seq_putc(m, '\n');
2892 }
2893 p = cache_chain.next;
2894 while (n--) {
2895 p = p->next;
2896 if (p == &cache_chain)
2897 return NULL;
2898 }
2899 return list_entry(p, kmem_cache_t, next);
2900}
2901
2902static void *s_next(struct seq_file *m, void *p, loff_t *pos)
2903{
2904 kmem_cache_t *cachep = p;
2905 ++*pos;
2906 return cachep->next.next == &cache_chain ? NULL
2907 : list_entry(cachep->next.next, kmem_cache_t, next);
2908}
2909
2910static void s_stop(struct seq_file *m, void *p)
2911{
2912 up(&cache_chain_sem);
2913}
2914
2915static int s_show(struct seq_file *m, void *p)
2916{
2917 kmem_cache_t *cachep = p;
2918 struct list_head *q;
2919 struct slab *slabp;
2920 unsigned long active_objs;
2921 unsigned long num_objs;
2922 unsigned long active_slabs = 0;
2923 unsigned long num_slabs;
2924 const char *name;
2925 char *error = NULL;
2926
2927 check_irq_on();
2928 spin_lock_irq(&cachep->spinlock);
2929 active_objs = 0;
2930 num_slabs = 0;
2931 list_for_each(q,&cachep->lists.slabs_full) {
2932 slabp = list_entry(q, struct slab, list);
2933 if (slabp->inuse != cachep->num && !error)
2934 error = "slabs_full accounting error";
2935 active_objs += cachep->num;
2936 active_slabs++;
2937 }
2938 list_for_each(q,&cachep->lists.slabs_partial) {
2939 slabp = list_entry(q, struct slab, list);
2940 if (slabp->inuse == cachep->num && !error)
2941 error = "slabs_partial inuse accounting error";
2942 if (!slabp->inuse && !error)
2943 error = "slabs_partial/inuse accounting error";
2944 active_objs += slabp->inuse;
2945 active_slabs++;
2946 }
2947 list_for_each(q,&cachep->lists.slabs_free) {
2948 slabp = list_entry(q, struct slab, list);
2949 if (slabp->inuse && !error)
2950 error = "slabs_free/inuse accounting error";
2951 num_slabs++;
2952 }
2953 num_slabs+=active_slabs;
2954 num_objs = num_slabs*cachep->num;
2955 if (num_objs - active_objs != cachep->lists.free_objects && !error)
2956 error = "free_objects accounting error";
2957
2958 name = cachep->name;
2959 if (error)
2960 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
2961
2962 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
2963 name, active_objs, num_objs, cachep->objsize,
2964 cachep->num, (1<<cachep->gfporder));
2965 seq_printf(m, " : tunables %4u %4u %4u",
2966 cachep->limit, cachep->batchcount,
2967 cachep->lists.shared->limit/cachep->batchcount);
2968 seq_printf(m, " : slabdata %6lu %6lu %6u",
2969 active_slabs, num_slabs, cachep->lists.shared->avail);
2970#if STATS
2971 { /* list3 stats */
2972 unsigned long high = cachep->high_mark;
2973 unsigned long allocs = cachep->num_allocations;
2974 unsigned long grown = cachep->grown;
2975 unsigned long reaped = cachep->reaped;
2976 unsigned long errors = cachep->errors;
2977 unsigned long max_freeable = cachep->max_freeable;
2978 unsigned long free_limit = cachep->free_limit;
2979 unsigned long node_allocs = cachep->node_allocs;
2980
2981 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu %4lu %4lu %4lu %4lu",
2982 allocs, high, grown, reaped, errors,
2983 max_freeable, free_limit, node_allocs);
2984 }
2985 /* cpu stats */
2986 {
2987 unsigned long allochit = atomic_read(&cachep->allochit);
2988 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
2989 unsigned long freehit = atomic_read(&cachep->freehit);
2990 unsigned long freemiss = atomic_read(&cachep->freemiss);
2991
2992 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
2993 allochit, allocmiss, freehit, freemiss);
2994 }
2995#endif
2996 seq_putc(m, '\n');
2997 spin_unlock_irq(&cachep->spinlock);
2998 return 0;
2999}
3000
3001/*
3002 * slabinfo_op - iterator that generates /proc/slabinfo
3003 *
3004 * Output layout:
3005 * cache-name
3006 * num-active-objs
3007 * total-objs
3008 * object size
3009 * num-active-slabs
3010 * total-slabs
3011 * num-pages-per-slab
3012 * + further values on SMP and with statistics enabled
3013 */
3014
3015struct seq_operations slabinfo_op = {
3016 .start = s_start,
3017 .next = s_next,
3018 .stop = s_stop,
3019 .show = s_show,
3020};
3021
3022#define MAX_SLABINFO_WRITE 128
3023/**
3024 * slabinfo_write - Tuning for the slab allocator
3025 * @file: unused
3026 * @buffer: user buffer
3027 * @count: data length
3028 * @ppos: unused
3029 */
3030ssize_t slabinfo_write(struct file *file, const char __user *buffer,
3031 size_t count, loff_t *ppos)
3032{
3033 char kbuf[MAX_SLABINFO_WRITE+1], *tmp;
3034 int limit, batchcount, shared, res;
3035 struct list_head *p;
3036
3037 if (count > MAX_SLABINFO_WRITE)
3038 return -EINVAL;
3039 if (copy_from_user(&kbuf, buffer, count))
3040 return -EFAULT;
3041 kbuf[MAX_SLABINFO_WRITE] = '\0';
3042
3043 tmp = strchr(kbuf, ' ');
3044 if (!tmp)
3045 return -EINVAL;
3046 *tmp = '\0';
3047 tmp++;
3048 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
3049 return -EINVAL;
3050
3051 /* Find the cache in the chain of caches. */
3052 down(&cache_chain_sem);
3053 res = -EINVAL;
3054 list_for_each(p,&cache_chain) {
3055 kmem_cache_t *cachep = list_entry(p, kmem_cache_t, next);
3056
3057 if (!strcmp(cachep->name, kbuf)) {
3058 if (limit < 1 ||
3059 batchcount < 1 ||
3060 batchcount > limit ||
3061 shared < 0) {
3062 res = -EINVAL;
3063 } else {
3064 res = do_tune_cpucache(cachep, limit, batchcount, shared);
3065 }
3066 break;
3067 }
3068 }
3069 up(&cache_chain_sem);
3070 if (res >= 0)
3071 res = count;
3072 return res;
3073}
3074#endif
3075
3076unsigned int ksize(const void *objp)
3077{
3078 kmem_cache_t *c;
3079 unsigned long flags;
3080 unsigned int size = 0;
3081
3082 if (likely(objp != NULL)) {
3083 local_irq_save(flags);
3084 c = GET_PAGE_CACHE(virt_to_page(objp));
3085 size = kmem_cache_size(c);
3086 local_irq_restore(flags);
3087 }
3088
3089 return size;
3090}
Paulo Marques543537b2005-06-23 00:09:02 -07003091
3092
3093/*
3094 * kstrdup - allocate space for and copy an existing string
3095 *
3096 * @s: the string to duplicate
3097 * @gfp: the GFP mask used in the kmalloc() call when allocating memory
3098 */
Alexey Dobriyan0db925a2005-07-07 17:56:58 -07003099char *kstrdup(const char *s, unsigned int __nocast gfp)
Paulo Marques543537b2005-06-23 00:09:02 -07003100{
3101 size_t len;
3102 char *buf;
3103
3104 if (!s)
3105 return NULL;
3106
3107 len = strlen(s) + 1;
3108 buf = kmalloc(len, gfp);
3109 if (buf)
3110 memcpy(buf, s, len);
3111 return buf;
3112}
3113EXPORT_SYMBOL(kstrdup);