USB: Unbreak fsl_usb2_udc
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / mm / slab.c
blob2e338a5f7b14c6d29216f4c54a74ef5259be01fa
1 /*
2 * linux/mm/slab.c
3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
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
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.
28 * This means, that your constructor is used only for newly allocated
29 * slabs and you must pass objects with the same initializations to
30 * kmem_cache_free.
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.
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
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
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.
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.
53 * The c_cpuarray may not be read with enabled local interrupts -
54 * it's changed with a smp_call_function().
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
58 * Several members in struct kmem_cache 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.
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.
68 * Further notes from the original documentation:
70 * 11 April '97. Started multi-threading - markhe
71 * The global cache-chain is protected by the mutex 'cache_chain_mutex'.
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()).
76 * At present, each engine can be growing a cache. This should be blocked.
78 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
89 #include <linux/slab.h>
90 #include <linux/mm.h>
91 #include <linux/poison.h>
92 #include <linux/swap.h>
93 #include <linux/cache.h>
94 #include <linux/interrupt.h>
95 #include <linux/init.h>
96 #include <linux/compiler.h>
97 #include <linux/cpuset.h>
98 #include <linux/seq_file.h>
99 #include <linux/notifier.h>
100 #include <linux/kallsyms.h>
101 #include <linux/cpu.h>
102 #include <linux/sysctl.h>
103 #include <linux/module.h>
104 #include <linux/rcupdate.h>
105 #include <linux/string.h>
106 #include <linux/uaccess.h>
107 #include <linux/nodemask.h>
108 #include <linux/mempolicy.h>
109 #include <linux/mutex.h>
110 #include <linux/fault-inject.h>
111 #include <linux/rtmutex.h>
112 #include <linux/reciprocal_div.h>
114 #include <asm/cacheflush.h>
115 #include <asm/tlbflush.h>
116 #include <asm/page.h>
119 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
120 * 0 for faster, smaller code (especially in the critical paths).
122 * STATS - 1 to collect stats for /proc/slabinfo.
123 * 0 for faster, smaller code (especially in the critical paths).
125 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
128 #ifdef CONFIG_DEBUG_SLAB
129 #define DEBUG 1
130 #define STATS 1
131 #define FORCED_DEBUG 1
132 #else
133 #define DEBUG 0
134 #define STATS 0
135 #define FORCED_DEBUG 0
136 #endif
138 /* Shouldn't this be in a header file somewhere? */
139 #define BYTES_PER_WORD sizeof(void *)
140 #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
142 #ifndef cache_line_size
143 #define cache_line_size() L1_CACHE_BYTES
144 #endif
146 #ifndef ARCH_KMALLOC_MINALIGN
148 * Enforce a minimum alignment for the kmalloc caches.
149 * Usually, the kmalloc caches are cache_line_size() aligned, except when
150 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
151 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
152 * alignment larger than the alignment of a 64-bit integer.
153 * ARCH_KMALLOC_MINALIGN allows that.
154 * Note that increasing this value may disable some debug features.
156 #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
157 #endif
159 #ifndef ARCH_SLAB_MINALIGN
161 * Enforce a minimum alignment for all caches.
162 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
163 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
164 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
165 * some debug features.
167 #define ARCH_SLAB_MINALIGN 0
168 #endif
170 #ifndef ARCH_KMALLOC_FLAGS
171 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
172 #endif
174 /* Legal flag mask for kmem_cache_create(). */
175 #if DEBUG
176 # define CREATE_MASK (SLAB_RED_ZONE | \
177 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
178 SLAB_CACHE_DMA | \
179 SLAB_STORE_USER | \
180 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
181 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
182 #else
183 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
184 SLAB_CACHE_DMA | \
185 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
186 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
187 #endif
190 * kmem_bufctl_t:
192 * Bufctl's are used for linking objs within a slab
193 * linked offsets.
195 * This implementation relies on "struct page" for locating the cache &
196 * slab an object belongs to.
197 * This allows the bufctl structure to be small (one int), but limits
198 * the number of objects a slab (not a cache) can contain when off-slab
199 * bufctls are used. The limit is the size of the largest general cache
200 * that does not use off-slab slabs.
201 * For 32bit archs with 4 kB pages, is this 56.
202 * This is not serious, as it is only for large objects, when it is unwise
203 * to have too many per slab.
204 * Note: This limit can be raised by introducing a general cache whose size
205 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
208 typedef unsigned int kmem_bufctl_t;
209 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
210 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
211 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
212 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
215 * struct slab
217 * Manages the objs in a slab. Placed either at the beginning of mem allocated
218 * for a slab, or allocated from an general cache.
219 * Slabs are chained into three list: fully used, partial, fully free slabs.
221 struct slab {
222 struct list_head list;
223 unsigned long colouroff;
224 void *s_mem; /* including colour offset */
225 unsigned int inuse; /* num of objs active in slab */
226 kmem_bufctl_t free;
227 unsigned short nodeid;
231 * struct slab_rcu
233 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
234 * arrange for kmem_freepages to be called via RCU. This is useful if
235 * we need to approach a kernel structure obliquely, from its address
236 * obtained without the usual locking. We can lock the structure to
237 * stabilize it and check it's still at the given address, only if we
238 * can be sure that the memory has not been meanwhile reused for some
239 * other kind of object (which our subsystem's lock might corrupt).
241 * rcu_read_lock before reading the address, then rcu_read_unlock after
242 * taking the spinlock within the structure expected at that address.
244 * We assume struct slab_rcu can overlay struct slab when destroying.
246 struct slab_rcu {
247 struct rcu_head head;
248 struct kmem_cache *cachep;
249 void *addr;
253 * struct array_cache
255 * Purpose:
256 * - LIFO ordering, to hand out cache-warm objects from _alloc
257 * - reduce the number of linked list operations
258 * - reduce spinlock operations
260 * The limit is stored in the per-cpu structure to reduce the data cache
261 * footprint.
264 struct array_cache {
265 unsigned int avail;
266 unsigned int limit;
267 unsigned int batchcount;
268 unsigned int touched;
269 spinlock_t lock;
270 void *entry[]; /*
271 * Must have this definition in here for the proper
272 * alignment of array_cache. Also simplifies accessing
273 * the entries.
278 * bootstrap: The caches do not work without cpuarrays anymore, but the
279 * cpuarrays are allocated from the generic caches...
281 #define BOOT_CPUCACHE_ENTRIES 1
282 struct arraycache_init {
283 struct array_cache cache;
284 void *entries[BOOT_CPUCACHE_ENTRIES];
288 * The slab lists for all objects.
290 struct kmem_list3 {
291 struct list_head slabs_partial; /* partial list first, better asm code */
292 struct list_head slabs_full;
293 struct list_head slabs_free;
294 unsigned long free_objects;
295 unsigned int free_limit;
296 unsigned int colour_next; /* Per-node cache coloring */
297 spinlock_t list_lock;
298 struct array_cache *shared; /* shared per node */
299 struct array_cache **alien; /* on other nodes */
300 unsigned long next_reap; /* updated without locking */
301 int free_touched; /* updated without locking */
305 * Need this for bootstrapping a per node allocator.
307 #define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
308 struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
309 #define CACHE_CACHE 0
310 #define SIZE_AC 1
311 #define SIZE_L3 (1 + MAX_NUMNODES)
313 static int drain_freelist(struct kmem_cache *cache,
314 struct kmem_list3 *l3, int tofree);
315 static void free_block(struct kmem_cache *cachep, void **objpp, int len,
316 int node);
317 static int enable_cpucache(struct kmem_cache *cachep);
318 static void cache_reap(struct work_struct *unused);
321 * This function must be completely optimized away if a constant is passed to
322 * it. Mostly the same as what is in linux/slab.h except it returns an index.
324 static __always_inline int index_of(const size_t size)
326 extern void __bad_size(void);
328 if (__builtin_constant_p(size)) {
329 int i = 0;
331 #define CACHE(x) \
332 if (size <=x) \
333 return i; \
334 else \
335 i++;
336 #include "linux/kmalloc_sizes.h"
337 #undef CACHE
338 __bad_size();
339 } else
340 __bad_size();
341 return 0;
344 static int slab_early_init = 1;
346 #define INDEX_AC index_of(sizeof(struct arraycache_init))
347 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
349 static void kmem_list3_init(struct kmem_list3 *parent)
351 INIT_LIST_HEAD(&parent->slabs_full);
352 INIT_LIST_HEAD(&parent->slabs_partial);
353 INIT_LIST_HEAD(&parent->slabs_free);
354 parent->shared = NULL;
355 parent->alien = NULL;
356 parent->colour_next = 0;
357 spin_lock_init(&parent->list_lock);
358 parent->free_objects = 0;
359 parent->free_touched = 0;
362 #define MAKE_LIST(cachep, listp, slab, nodeid) \
363 do { \
364 INIT_LIST_HEAD(listp); \
365 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
366 } while (0)
368 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
369 do { \
370 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
371 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
372 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
373 } while (0)
376 * struct kmem_cache
378 * manages a cache.
381 struct kmem_cache {
382 /* 1) per-cpu data, touched during every alloc/free */
383 struct array_cache *array[NR_CPUS];
384 /* 2) Cache tunables. Protected by cache_chain_mutex */
385 unsigned int batchcount;
386 unsigned int limit;
387 unsigned int shared;
389 unsigned int buffer_size;
390 u32 reciprocal_buffer_size;
391 /* 3) touched by every alloc & free from the backend */
393 unsigned int flags; /* constant flags */
394 unsigned int num; /* # of objs per slab */
396 /* 4) cache_grow/shrink */
397 /* order of pgs per slab (2^n) */
398 unsigned int gfporder;
400 /* force GFP flags, e.g. GFP_DMA */
401 gfp_t gfpflags;
403 size_t colour; /* cache colouring range */
404 unsigned int colour_off; /* colour offset */
405 struct kmem_cache *slabp_cache;
406 unsigned int slab_size;
407 unsigned int dflags; /* dynamic flags */
409 /* constructor func */
410 void (*ctor)(struct kmem_cache *, void *);
412 /* 5) cache creation/removal */
413 const char *name;
414 struct list_head next;
416 /* 6) statistics */
417 #if STATS
418 unsigned long num_active;
419 unsigned long num_allocations;
420 unsigned long high_mark;
421 unsigned long grown;
422 unsigned long reaped;
423 unsigned long errors;
424 unsigned long max_freeable;
425 unsigned long node_allocs;
426 unsigned long node_frees;
427 unsigned long node_overflow;
428 atomic_t allochit;
429 atomic_t allocmiss;
430 atomic_t freehit;
431 atomic_t freemiss;
432 #endif
433 #if DEBUG
435 * If debugging is enabled, then the allocator can add additional
436 * fields and/or padding to every object. buffer_size contains the total
437 * object size including these internal fields, the following two
438 * variables contain the offset to the user object and its size.
440 int obj_offset;
441 int obj_size;
442 #endif
444 * We put nodelists[] at the end of kmem_cache, because we want to size
445 * this array to nr_node_ids slots instead of MAX_NUMNODES
446 * (see kmem_cache_init())
447 * We still use [MAX_NUMNODES] and not [1] or [0] because cache_cache
448 * is statically defined, so we reserve the max number of nodes.
450 struct kmem_list3 *nodelists[MAX_NUMNODES];
452 * Do not add fields after nodelists[]
456 #define CFLGS_OFF_SLAB (0x80000000UL)
457 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
459 #define BATCHREFILL_LIMIT 16
461 * Optimization question: fewer reaps means less probability for unnessary
462 * cpucache drain/refill cycles.
464 * OTOH the cpuarrays can contain lots of objects,
465 * which could lock up otherwise freeable slabs.
467 #define REAPTIMEOUT_CPUC (2*HZ)
468 #define REAPTIMEOUT_LIST3 (4*HZ)
470 #if STATS
471 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
472 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
473 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
474 #define STATS_INC_GROWN(x) ((x)->grown++)
475 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
476 #define STATS_SET_HIGH(x) \
477 do { \
478 if ((x)->num_active > (x)->high_mark) \
479 (x)->high_mark = (x)->num_active; \
480 } while (0)
481 #define STATS_INC_ERR(x) ((x)->errors++)
482 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
483 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
484 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
485 #define STATS_SET_FREEABLE(x, i) \
486 do { \
487 if ((x)->max_freeable < i) \
488 (x)->max_freeable = i; \
489 } while (0)
490 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
491 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
492 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
493 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
494 #else
495 #define STATS_INC_ACTIVE(x) do { } while (0)
496 #define STATS_DEC_ACTIVE(x) do { } while (0)
497 #define STATS_INC_ALLOCED(x) do { } while (0)
498 #define STATS_INC_GROWN(x) do { } while (0)
499 #define STATS_ADD_REAPED(x,y) do { } while (0)
500 #define STATS_SET_HIGH(x) do { } while (0)
501 #define STATS_INC_ERR(x) do { } while (0)
502 #define STATS_INC_NODEALLOCS(x) do { } while (0)
503 #define STATS_INC_NODEFREES(x) do { } while (0)
504 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
505 #define STATS_SET_FREEABLE(x, i) do { } while (0)
506 #define STATS_INC_ALLOCHIT(x) do { } while (0)
507 #define STATS_INC_ALLOCMISS(x) do { } while (0)
508 #define STATS_INC_FREEHIT(x) do { } while (0)
509 #define STATS_INC_FREEMISS(x) do { } while (0)
510 #endif
512 #if DEBUG
515 * memory layout of objects:
516 * 0 : objp
517 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
518 * the end of an object is aligned with the end of the real
519 * allocation. Catches writes behind the end of the allocation.
520 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
521 * redzone word.
522 * cachep->obj_offset: The real object.
523 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
524 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
525 * [BYTES_PER_WORD long]
527 static int obj_offset(struct kmem_cache *cachep)
529 return cachep->obj_offset;
532 static int obj_size(struct kmem_cache *cachep)
534 return cachep->obj_size;
537 static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
539 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
540 return (unsigned long long*) (objp + obj_offset(cachep) -
541 sizeof(unsigned long long));
544 static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
546 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
547 if (cachep->flags & SLAB_STORE_USER)
548 return (unsigned long long *)(objp + cachep->buffer_size -
549 sizeof(unsigned long long) -
550 REDZONE_ALIGN);
551 return (unsigned long long *) (objp + cachep->buffer_size -
552 sizeof(unsigned long long));
555 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
557 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
558 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
561 #else
563 #define obj_offset(x) 0
564 #define obj_size(cachep) (cachep->buffer_size)
565 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
566 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
567 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
569 #endif
572 * Do not go above this order unless 0 objects fit into the slab.
574 #define BREAK_GFP_ORDER_HI 1
575 #define BREAK_GFP_ORDER_LO 0
576 static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
579 * Functions for storing/retrieving the cachep and or slab from the page
580 * allocator. These are used to find the slab an obj belongs to. With kfree(),
581 * these are used to find the cache which an obj belongs to.
583 static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
585 page->lru.next = (struct list_head *)cache;
588 static inline struct kmem_cache *page_get_cache(struct page *page)
590 page = compound_head(page);
591 BUG_ON(!PageSlab(page));
592 return (struct kmem_cache *)page->lru.next;
595 static inline void page_set_slab(struct page *page, struct slab *slab)
597 page->lru.prev = (struct list_head *)slab;
600 static inline struct slab *page_get_slab(struct page *page)
602 BUG_ON(!PageSlab(page));
603 return (struct slab *)page->lru.prev;
606 static inline struct kmem_cache *virt_to_cache(const void *obj)
608 struct page *page = virt_to_head_page(obj);
609 return page_get_cache(page);
612 static inline struct slab *virt_to_slab(const void *obj)
614 struct page *page = virt_to_head_page(obj);
615 return page_get_slab(page);
618 static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
619 unsigned int idx)
621 return slab->s_mem + cache->buffer_size * idx;
625 * We want to avoid an expensive divide : (offset / cache->buffer_size)
626 * Using the fact that buffer_size is a constant for a particular cache,
627 * we can replace (offset / cache->buffer_size) by
628 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
630 static inline unsigned int obj_to_index(const struct kmem_cache *cache,
631 const struct slab *slab, void *obj)
633 u32 offset = (obj - slab->s_mem);
634 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
638 * These are the default caches for kmalloc. Custom caches can have other sizes.
640 struct cache_sizes malloc_sizes[] = {
641 #define CACHE(x) { .cs_size = (x) },
642 #include <linux/kmalloc_sizes.h>
643 CACHE(ULONG_MAX)
644 #undef CACHE
646 EXPORT_SYMBOL(malloc_sizes);
648 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
649 struct cache_names {
650 char *name;
651 char *name_dma;
654 static struct cache_names __initdata cache_names[] = {
655 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
656 #include <linux/kmalloc_sizes.h>
657 {NULL,}
658 #undef CACHE
661 static struct arraycache_init initarray_cache __initdata =
662 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
663 static struct arraycache_init initarray_generic =
664 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
666 /* internal cache of cache description objs */
667 static struct kmem_cache cache_cache = {
668 .batchcount = 1,
669 .limit = BOOT_CPUCACHE_ENTRIES,
670 .shared = 1,
671 .buffer_size = sizeof(struct kmem_cache),
672 .name = "kmem_cache",
675 #define BAD_ALIEN_MAGIC 0x01020304ul
677 #ifdef CONFIG_LOCKDEP
680 * Slab sometimes uses the kmalloc slabs to store the slab headers
681 * for other slabs "off slab".
682 * The locking for this is tricky in that it nests within the locks
683 * of all other slabs in a few places; to deal with this special
684 * locking we put on-slab caches into a separate lock-class.
686 * We set lock class for alien array caches which are up during init.
687 * The lock annotation will be lost if all cpus of a node goes down and
688 * then comes back up during hotplug
690 static struct lock_class_key on_slab_l3_key;
691 static struct lock_class_key on_slab_alc_key;
693 static inline void init_lock_keys(void)
696 int q;
697 struct cache_sizes *s = malloc_sizes;
699 while (s->cs_size != ULONG_MAX) {
700 for_each_node(q) {
701 struct array_cache **alc;
702 int r;
703 struct kmem_list3 *l3 = s->cs_cachep->nodelists[q];
704 if (!l3 || OFF_SLAB(s->cs_cachep))
705 continue;
706 lockdep_set_class(&l3->list_lock, &on_slab_l3_key);
707 alc = l3->alien;
709 * FIXME: This check for BAD_ALIEN_MAGIC
710 * should go away when common slab code is taught to
711 * work even without alien caches.
712 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
713 * for alloc_alien_cache,
715 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
716 continue;
717 for_each_node(r) {
718 if (alc[r])
719 lockdep_set_class(&alc[r]->lock,
720 &on_slab_alc_key);
723 s++;
726 #else
727 static inline void init_lock_keys(void)
730 #endif
733 * 1. Guard access to the cache-chain.
734 * 2. Protect sanity of cpu_online_map against cpu hotplug events
736 static DEFINE_MUTEX(cache_chain_mutex);
737 static struct list_head cache_chain;
740 * chicken and egg problem: delay the per-cpu array allocation
741 * until the general caches are up.
743 static enum {
744 NONE,
745 PARTIAL_AC,
746 PARTIAL_L3,
747 FULL
748 } g_cpucache_up;
751 * used by boot code to determine if it can use slab based allocator
753 int slab_is_available(void)
755 return g_cpucache_up == FULL;
758 static DEFINE_PER_CPU(struct delayed_work, reap_work);
760 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
762 return cachep->array[smp_processor_id()];
765 static inline struct kmem_cache *__find_general_cachep(size_t size,
766 gfp_t gfpflags)
768 struct cache_sizes *csizep = malloc_sizes;
770 #if DEBUG
771 /* This happens if someone tries to call
772 * kmem_cache_create(), or __kmalloc(), before
773 * the generic caches are initialized.
775 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
776 #endif
777 if (!size)
778 return ZERO_SIZE_PTR;
780 while (size > csizep->cs_size)
781 csizep++;
784 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
785 * has cs_{dma,}cachep==NULL. Thus no special case
786 * for large kmalloc calls required.
788 #ifdef CONFIG_ZONE_DMA
789 if (unlikely(gfpflags & GFP_DMA))
790 return csizep->cs_dmacachep;
791 #endif
792 return csizep->cs_cachep;
795 static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
797 return __find_general_cachep(size, gfpflags);
800 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
802 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
806 * Calculate the number of objects and left-over bytes for a given buffer size.
808 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
809 size_t align, int flags, size_t *left_over,
810 unsigned int *num)
812 int nr_objs;
813 size_t mgmt_size;
814 size_t slab_size = PAGE_SIZE << gfporder;
817 * The slab management structure can be either off the slab or
818 * on it. For the latter case, the memory allocated for a
819 * slab is used for:
821 * - The struct slab
822 * - One kmem_bufctl_t for each object
823 * - Padding to respect alignment of @align
824 * - @buffer_size bytes for each object
826 * If the slab management structure is off the slab, then the
827 * alignment will already be calculated into the size. Because
828 * the slabs are all pages aligned, the objects will be at the
829 * correct alignment when allocated.
831 if (flags & CFLGS_OFF_SLAB) {
832 mgmt_size = 0;
833 nr_objs = slab_size / buffer_size;
835 if (nr_objs > SLAB_LIMIT)
836 nr_objs = SLAB_LIMIT;
837 } else {
839 * Ignore padding for the initial guess. The padding
840 * is at most @align-1 bytes, and @buffer_size is at
841 * least @align. In the worst case, this result will
842 * be one greater than the number of objects that fit
843 * into the memory allocation when taking the padding
844 * into account.
846 nr_objs = (slab_size - sizeof(struct slab)) /
847 (buffer_size + sizeof(kmem_bufctl_t));
850 * This calculated number will be either the right
851 * amount, or one greater than what we want.
853 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
854 > slab_size)
855 nr_objs--;
857 if (nr_objs > SLAB_LIMIT)
858 nr_objs = SLAB_LIMIT;
860 mgmt_size = slab_mgmt_size(nr_objs, align);
862 *num = nr_objs;
863 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
866 #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
868 static void __slab_error(const char *function, struct kmem_cache *cachep,
869 char *msg)
871 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
872 function, cachep->name, msg);
873 dump_stack();
877 * By default on NUMA we use alien caches to stage the freeing of
878 * objects allocated from other nodes. This causes massive memory
879 * inefficiencies when using fake NUMA setup to split memory into a
880 * large number of small nodes, so it can be disabled on the command
881 * line
884 static int use_alien_caches __read_mostly = 1;
885 static int numa_platform __read_mostly = 1;
886 static int __init noaliencache_setup(char *s)
888 use_alien_caches = 0;
889 return 1;
891 __setup("noaliencache", noaliencache_setup);
893 #ifdef CONFIG_NUMA
895 * Special reaping functions for NUMA systems called from cache_reap().
896 * These take care of doing round robin flushing of alien caches (containing
897 * objects freed on different nodes from which they were allocated) and the
898 * flushing of remote pcps by calling drain_node_pages.
900 static DEFINE_PER_CPU(unsigned long, reap_node);
902 static void init_reap_node(int cpu)
904 int node;
906 node = next_node(cpu_to_node(cpu), node_online_map);
907 if (node == MAX_NUMNODES)
908 node = first_node(node_online_map);
910 per_cpu(reap_node, cpu) = node;
913 static void next_reap_node(void)
915 int node = __get_cpu_var(reap_node);
917 node = next_node(node, node_online_map);
918 if (unlikely(node >= MAX_NUMNODES))
919 node = first_node(node_online_map);
920 __get_cpu_var(reap_node) = node;
923 #else
924 #define init_reap_node(cpu) do { } while (0)
925 #define next_reap_node(void) do { } while (0)
926 #endif
929 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
930 * via the workqueue/eventd.
931 * Add the CPU number into the expiration time to minimize the possibility of
932 * the CPUs getting into lockstep and contending for the global cache chain
933 * lock.
935 static void __cpuinit start_cpu_timer(int cpu)
937 struct delayed_work *reap_work = &per_cpu(reap_work, cpu);
940 * When this gets called from do_initcalls via cpucache_init(),
941 * init_workqueues() has already run, so keventd will be setup
942 * at that time.
944 if (keventd_up() && reap_work->work.func == NULL) {
945 init_reap_node(cpu);
946 INIT_DELAYED_WORK(reap_work, cache_reap);
947 schedule_delayed_work_on(cpu, reap_work,
948 __round_jiffies_relative(HZ, cpu));
952 static struct array_cache *alloc_arraycache(int node, int entries,
953 int batchcount)
955 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
956 struct array_cache *nc = NULL;
958 nc = kmalloc_node(memsize, GFP_KERNEL, node);
959 if (nc) {
960 nc->avail = 0;
961 nc->limit = entries;
962 nc->batchcount = batchcount;
963 nc->touched = 0;
964 spin_lock_init(&nc->lock);
966 return nc;
970 * Transfer objects in one arraycache to another.
971 * Locking must be handled by the caller.
973 * Return the number of entries transferred.
975 static int transfer_objects(struct array_cache *to,
976 struct array_cache *from, unsigned int max)
978 /* Figure out how many entries to transfer */
979 int nr = min(min(from->avail, max), to->limit - to->avail);
981 if (!nr)
982 return 0;
984 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
985 sizeof(void *) *nr);
987 from->avail -= nr;
988 to->avail += nr;
989 to->touched = 1;
990 return nr;
993 #ifndef CONFIG_NUMA
995 #define drain_alien_cache(cachep, alien) do { } while (0)
996 #define reap_alien(cachep, l3) do { } while (0)
998 static inline struct array_cache **alloc_alien_cache(int node, int limit)
1000 return (struct array_cache **)BAD_ALIEN_MAGIC;
1003 static inline void free_alien_cache(struct array_cache **ac_ptr)
1007 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1009 return 0;
1012 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
1013 gfp_t flags)
1015 return NULL;
1018 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
1019 gfp_t flags, int nodeid)
1021 return NULL;
1024 #else /* CONFIG_NUMA */
1026 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
1027 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
1029 static struct array_cache **alloc_alien_cache(int node, int limit)
1031 struct array_cache **ac_ptr;
1032 int memsize = sizeof(void *) * nr_node_ids;
1033 int i;
1035 if (limit > 1)
1036 limit = 12;
1037 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
1038 if (ac_ptr) {
1039 for_each_node(i) {
1040 if (i == node || !node_online(i)) {
1041 ac_ptr[i] = NULL;
1042 continue;
1044 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
1045 if (!ac_ptr[i]) {
1046 for (i--; i >= 0; i--)
1047 kfree(ac_ptr[i]);
1048 kfree(ac_ptr);
1049 return NULL;
1053 return ac_ptr;
1056 static void free_alien_cache(struct array_cache **ac_ptr)
1058 int i;
1060 if (!ac_ptr)
1061 return;
1062 for_each_node(i)
1063 kfree(ac_ptr[i]);
1064 kfree(ac_ptr);
1067 static void __drain_alien_cache(struct kmem_cache *cachep,
1068 struct array_cache *ac, int node)
1070 struct kmem_list3 *rl3 = cachep->nodelists[node];
1072 if (ac->avail) {
1073 spin_lock(&rl3->list_lock);
1075 * Stuff objects into the remote nodes shared array first.
1076 * That way we could avoid the overhead of putting the objects
1077 * into the free lists and getting them back later.
1079 if (rl3->shared)
1080 transfer_objects(rl3->shared, ac, ac->limit);
1082 free_block(cachep, ac->entry, ac->avail, node);
1083 ac->avail = 0;
1084 spin_unlock(&rl3->list_lock);
1089 * Called from cache_reap() to regularly drain alien caches round robin.
1091 static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1093 int node = __get_cpu_var(reap_node);
1095 if (l3->alien) {
1096 struct array_cache *ac = l3->alien[node];
1098 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1099 __drain_alien_cache(cachep, ac, node);
1100 spin_unlock_irq(&ac->lock);
1105 static void drain_alien_cache(struct kmem_cache *cachep,
1106 struct array_cache **alien)
1108 int i = 0;
1109 struct array_cache *ac;
1110 unsigned long flags;
1112 for_each_online_node(i) {
1113 ac = alien[i];
1114 if (ac) {
1115 spin_lock_irqsave(&ac->lock, flags);
1116 __drain_alien_cache(cachep, ac, i);
1117 spin_unlock_irqrestore(&ac->lock, flags);
1122 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1124 struct slab *slabp = virt_to_slab(objp);
1125 int nodeid = slabp->nodeid;
1126 struct kmem_list3 *l3;
1127 struct array_cache *alien = NULL;
1128 int node;
1130 node = numa_node_id();
1133 * Make sure we are not freeing a object from another node to the array
1134 * cache on this cpu.
1136 if (likely(slabp->nodeid == node))
1137 return 0;
1139 l3 = cachep->nodelists[node];
1140 STATS_INC_NODEFREES(cachep);
1141 if (l3->alien && l3->alien[nodeid]) {
1142 alien = l3->alien[nodeid];
1143 spin_lock(&alien->lock);
1144 if (unlikely(alien->avail == alien->limit)) {
1145 STATS_INC_ACOVERFLOW(cachep);
1146 __drain_alien_cache(cachep, alien, nodeid);
1148 alien->entry[alien->avail++] = objp;
1149 spin_unlock(&alien->lock);
1150 } else {
1151 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1152 free_block(cachep, &objp, 1, nodeid);
1153 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1155 return 1;
1157 #endif
1159 static void __cpuinit cpuup_canceled(long cpu)
1161 struct kmem_cache *cachep;
1162 struct kmem_list3 *l3 = NULL;
1163 int node = cpu_to_node(cpu);
1165 list_for_each_entry(cachep, &cache_chain, next) {
1166 struct array_cache *nc;
1167 struct array_cache *shared;
1168 struct array_cache **alien;
1169 cpumask_t mask;
1171 mask = node_to_cpumask(node);
1172 /* cpu is dead; no one can alloc from it. */
1173 nc = cachep->array[cpu];
1174 cachep->array[cpu] = NULL;
1175 l3 = cachep->nodelists[node];
1177 if (!l3)
1178 goto free_array_cache;
1180 spin_lock_irq(&l3->list_lock);
1182 /* Free limit for this kmem_list3 */
1183 l3->free_limit -= cachep->batchcount;
1184 if (nc)
1185 free_block(cachep, nc->entry, nc->avail, node);
1187 if (!cpus_empty(mask)) {
1188 spin_unlock_irq(&l3->list_lock);
1189 goto free_array_cache;
1192 shared = l3->shared;
1193 if (shared) {
1194 free_block(cachep, shared->entry,
1195 shared->avail, node);
1196 l3->shared = NULL;
1199 alien = l3->alien;
1200 l3->alien = NULL;
1202 spin_unlock_irq(&l3->list_lock);
1204 kfree(shared);
1205 if (alien) {
1206 drain_alien_cache(cachep, alien);
1207 free_alien_cache(alien);
1209 free_array_cache:
1210 kfree(nc);
1213 * In the previous loop, all the objects were freed to
1214 * the respective cache's slabs, now we can go ahead and
1215 * shrink each nodelist to its limit.
1217 list_for_each_entry(cachep, &cache_chain, next) {
1218 l3 = cachep->nodelists[node];
1219 if (!l3)
1220 continue;
1221 drain_freelist(cachep, l3, l3->free_objects);
1225 static int __cpuinit cpuup_prepare(long cpu)
1227 struct kmem_cache *cachep;
1228 struct kmem_list3 *l3 = NULL;
1229 int node = cpu_to_node(cpu);
1230 const int memsize = sizeof(struct kmem_list3);
1233 * We need to do this right in the beginning since
1234 * alloc_arraycache's are going to use this list.
1235 * kmalloc_node allows us to add the slab to the right
1236 * kmem_list3 and not this cpu's kmem_list3
1239 list_for_each_entry(cachep, &cache_chain, next) {
1241 * Set up the size64 kmemlist for cpu before we can
1242 * begin anything. Make sure some other cpu on this
1243 * node has not already allocated this
1245 if (!cachep->nodelists[node]) {
1246 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1247 if (!l3)
1248 goto bad;
1249 kmem_list3_init(l3);
1250 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1251 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1254 * The l3s don't come and go as CPUs come and
1255 * go. cache_chain_mutex is sufficient
1256 * protection here.
1258 cachep->nodelists[node] = l3;
1261 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1262 cachep->nodelists[node]->free_limit =
1263 (1 + nr_cpus_node(node)) *
1264 cachep->batchcount + cachep->num;
1265 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1269 * Now we can go ahead with allocating the shared arrays and
1270 * array caches
1272 list_for_each_entry(cachep, &cache_chain, next) {
1273 struct array_cache *nc;
1274 struct array_cache *shared = NULL;
1275 struct array_cache **alien = NULL;
1277 nc = alloc_arraycache(node, cachep->limit,
1278 cachep->batchcount);
1279 if (!nc)
1280 goto bad;
1281 if (cachep->shared) {
1282 shared = alloc_arraycache(node,
1283 cachep->shared * cachep->batchcount,
1284 0xbaadf00d);
1285 if (!shared) {
1286 kfree(nc);
1287 goto bad;
1290 if (use_alien_caches) {
1291 alien = alloc_alien_cache(node, cachep->limit);
1292 if (!alien) {
1293 kfree(shared);
1294 kfree(nc);
1295 goto bad;
1298 cachep->array[cpu] = nc;
1299 l3 = cachep->nodelists[node];
1300 BUG_ON(!l3);
1302 spin_lock_irq(&l3->list_lock);
1303 if (!l3->shared) {
1305 * We are serialised from CPU_DEAD or
1306 * CPU_UP_CANCELLED by the cpucontrol lock
1308 l3->shared = shared;
1309 shared = NULL;
1311 #ifdef CONFIG_NUMA
1312 if (!l3->alien) {
1313 l3->alien = alien;
1314 alien = NULL;
1316 #endif
1317 spin_unlock_irq(&l3->list_lock);
1318 kfree(shared);
1319 free_alien_cache(alien);
1321 return 0;
1322 bad:
1323 cpuup_canceled(cpu);
1324 return -ENOMEM;
1327 static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1328 unsigned long action, void *hcpu)
1330 long cpu = (long)hcpu;
1331 int err = 0;
1333 switch (action) {
1334 case CPU_LOCK_ACQUIRE:
1335 mutex_lock(&cache_chain_mutex);
1336 break;
1337 case CPU_UP_PREPARE:
1338 case CPU_UP_PREPARE_FROZEN:
1339 err = cpuup_prepare(cpu);
1340 break;
1341 case CPU_ONLINE:
1342 case CPU_ONLINE_FROZEN:
1343 start_cpu_timer(cpu);
1344 break;
1345 #ifdef CONFIG_HOTPLUG_CPU
1346 case CPU_DOWN_PREPARE:
1347 case CPU_DOWN_PREPARE_FROZEN:
1349 * Shutdown cache reaper. Note that the cache_chain_mutex is
1350 * held so that if cache_reap() is invoked it cannot do
1351 * anything expensive but will only modify reap_work
1352 * and reschedule the timer.
1354 cancel_rearming_delayed_work(&per_cpu(reap_work, cpu));
1355 /* Now the cache_reaper is guaranteed to be not running. */
1356 per_cpu(reap_work, cpu).work.func = NULL;
1357 break;
1358 case CPU_DOWN_FAILED:
1359 case CPU_DOWN_FAILED_FROZEN:
1360 start_cpu_timer(cpu);
1361 break;
1362 case CPU_DEAD:
1363 case CPU_DEAD_FROZEN:
1365 * Even if all the cpus of a node are down, we don't free the
1366 * kmem_list3 of any cache. This to avoid a race between
1367 * cpu_down, and a kmalloc allocation from another cpu for
1368 * memory from the node of the cpu going down. The list3
1369 * structure is usually allocated from kmem_cache_create() and
1370 * gets destroyed at kmem_cache_destroy().
1372 /* fall through */
1373 #endif
1374 case CPU_UP_CANCELED:
1375 case CPU_UP_CANCELED_FROZEN:
1376 cpuup_canceled(cpu);
1377 break;
1378 case CPU_LOCK_RELEASE:
1379 mutex_unlock(&cache_chain_mutex);
1380 break;
1382 return err ? NOTIFY_BAD : NOTIFY_OK;
1385 static struct notifier_block __cpuinitdata cpucache_notifier = {
1386 &cpuup_callback, NULL, 0
1390 * swap the static kmem_list3 with kmalloced memory
1392 static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1393 int nodeid)
1395 struct kmem_list3 *ptr;
1397 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1398 BUG_ON(!ptr);
1400 local_irq_disable();
1401 memcpy(ptr, list, sizeof(struct kmem_list3));
1403 * Do not assume that spinlocks can be initialized via memcpy:
1405 spin_lock_init(&ptr->list_lock);
1407 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1408 cachep->nodelists[nodeid] = ptr;
1409 local_irq_enable();
1413 * Initialisation. Called after the page allocator have been initialised and
1414 * before smp_init().
1416 void __init kmem_cache_init(void)
1418 size_t left_over;
1419 struct cache_sizes *sizes;
1420 struct cache_names *names;
1421 int i;
1422 int order;
1423 int node;
1425 if (num_possible_nodes() == 1) {
1426 use_alien_caches = 0;
1427 numa_platform = 0;
1430 for (i = 0; i < NUM_INIT_LISTS; i++) {
1431 kmem_list3_init(&initkmem_list3[i]);
1432 if (i < MAX_NUMNODES)
1433 cache_cache.nodelists[i] = NULL;
1437 * Fragmentation resistance on low memory - only use bigger
1438 * page orders on machines with more than 32MB of memory.
1440 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1441 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1443 /* Bootstrap is tricky, because several objects are allocated
1444 * from caches that do not exist yet:
1445 * 1) initialize the cache_cache cache: it contains the struct
1446 * kmem_cache structures of all caches, except cache_cache itself:
1447 * cache_cache is statically allocated.
1448 * Initially an __init data area is used for the head array and the
1449 * kmem_list3 structures, it's replaced with a kmalloc allocated
1450 * array at the end of the bootstrap.
1451 * 2) Create the first kmalloc cache.
1452 * The struct kmem_cache for the new cache is allocated normally.
1453 * An __init data area is used for the head array.
1454 * 3) Create the remaining kmalloc caches, with minimally sized
1455 * head arrays.
1456 * 4) Replace the __init data head arrays for cache_cache and the first
1457 * kmalloc cache with kmalloc allocated arrays.
1458 * 5) Replace the __init data for kmem_list3 for cache_cache and
1459 * the other cache's with kmalloc allocated memory.
1460 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1463 node = numa_node_id();
1465 /* 1) create the cache_cache */
1466 INIT_LIST_HEAD(&cache_chain);
1467 list_add(&cache_cache.next, &cache_chain);
1468 cache_cache.colour_off = cache_line_size();
1469 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
1470 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE];
1473 * struct kmem_cache size depends on nr_node_ids, which
1474 * can be less than MAX_NUMNODES.
1476 cache_cache.buffer_size = offsetof(struct kmem_cache, nodelists) +
1477 nr_node_ids * sizeof(struct kmem_list3 *);
1478 #if DEBUG
1479 cache_cache.obj_size = cache_cache.buffer_size;
1480 #endif
1481 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1482 cache_line_size());
1483 cache_cache.reciprocal_buffer_size =
1484 reciprocal_value(cache_cache.buffer_size);
1486 for (order = 0; order < MAX_ORDER; order++) {
1487 cache_estimate(order, cache_cache.buffer_size,
1488 cache_line_size(), 0, &left_over, &cache_cache.num);
1489 if (cache_cache.num)
1490 break;
1492 BUG_ON(!cache_cache.num);
1493 cache_cache.gfporder = order;
1494 cache_cache.colour = left_over / cache_cache.colour_off;
1495 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1496 sizeof(struct slab), cache_line_size());
1498 /* 2+3) create the kmalloc caches */
1499 sizes = malloc_sizes;
1500 names = cache_names;
1503 * Initialize the caches that provide memory for the array cache and the
1504 * kmem_list3 structures first. Without this, further allocations will
1505 * bug.
1508 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
1509 sizes[INDEX_AC].cs_size,
1510 ARCH_KMALLOC_MINALIGN,
1511 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1512 NULL);
1514 if (INDEX_AC != INDEX_L3) {
1515 sizes[INDEX_L3].cs_cachep =
1516 kmem_cache_create(names[INDEX_L3].name,
1517 sizes[INDEX_L3].cs_size,
1518 ARCH_KMALLOC_MINALIGN,
1519 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1520 NULL);
1523 slab_early_init = 0;
1525 while (sizes->cs_size != ULONG_MAX) {
1527 * For performance, all the general caches are L1 aligned.
1528 * This should be particularly beneficial on SMP boxes, as it
1529 * eliminates "false sharing".
1530 * Note for systems short on memory removing the alignment will
1531 * allow tighter packing of the smaller caches.
1533 if (!sizes->cs_cachep) {
1534 sizes->cs_cachep = kmem_cache_create(names->name,
1535 sizes->cs_size,
1536 ARCH_KMALLOC_MINALIGN,
1537 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1538 NULL);
1540 #ifdef CONFIG_ZONE_DMA
1541 sizes->cs_dmacachep = kmem_cache_create(
1542 names->name_dma,
1543 sizes->cs_size,
1544 ARCH_KMALLOC_MINALIGN,
1545 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1546 SLAB_PANIC,
1547 NULL);
1548 #endif
1549 sizes++;
1550 names++;
1552 /* 4) Replace the bootstrap head arrays */
1554 struct array_cache *ptr;
1556 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1558 local_irq_disable();
1559 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1560 memcpy(ptr, cpu_cache_get(&cache_cache),
1561 sizeof(struct arraycache_init));
1563 * Do not assume that spinlocks can be initialized via memcpy:
1565 spin_lock_init(&ptr->lock);
1567 cache_cache.array[smp_processor_id()] = ptr;
1568 local_irq_enable();
1570 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1572 local_irq_disable();
1573 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1574 != &initarray_generic.cache);
1575 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1576 sizeof(struct arraycache_init));
1578 * Do not assume that spinlocks can be initialized via memcpy:
1580 spin_lock_init(&ptr->lock);
1582 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1583 ptr;
1584 local_irq_enable();
1586 /* 5) Replace the bootstrap kmem_list3's */
1588 int nid;
1590 /* Replace the static kmem_list3 structures for the boot cpu */
1591 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE], node);
1593 for_each_node_state(nid, N_NORMAL_MEMORY) {
1594 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1595 &initkmem_list3[SIZE_AC + nid], nid);
1597 if (INDEX_AC != INDEX_L3) {
1598 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1599 &initkmem_list3[SIZE_L3 + nid], nid);
1604 /* 6) resize the head arrays to their final sizes */
1606 struct kmem_cache *cachep;
1607 mutex_lock(&cache_chain_mutex);
1608 list_for_each_entry(cachep, &cache_chain, next)
1609 if (enable_cpucache(cachep))
1610 BUG();
1611 mutex_unlock(&cache_chain_mutex);
1614 /* Annotate slab for lockdep -- annotate the malloc caches */
1615 init_lock_keys();
1618 /* Done! */
1619 g_cpucache_up = FULL;
1622 * Register a cpu startup notifier callback that initializes
1623 * cpu_cache_get for all new cpus
1625 register_cpu_notifier(&cpucache_notifier);
1628 * The reap timers are started later, with a module init call: That part
1629 * of the kernel is not yet operational.
1633 static int __init cpucache_init(void)
1635 int cpu;
1638 * Register the timers that return unneeded pages to the page allocator
1640 for_each_online_cpu(cpu)
1641 start_cpu_timer(cpu);
1642 return 0;
1644 __initcall(cpucache_init);
1647 * Interface to system's page allocator. No need to hold the cache-lock.
1649 * If we requested dmaable memory, we will get it. Even if we
1650 * did not request dmaable memory, we might get it, but that
1651 * would be relatively rare and ignorable.
1653 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1655 struct page *page;
1656 int nr_pages;
1657 int i;
1659 #ifndef CONFIG_MMU
1661 * Nommu uses slab's for process anonymous memory allocations, and thus
1662 * requires __GFP_COMP to properly refcount higher order allocations
1664 flags |= __GFP_COMP;
1665 #endif
1667 flags |= cachep->gfpflags;
1668 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1669 flags |= __GFP_RECLAIMABLE;
1671 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
1672 if (!page)
1673 return NULL;
1675 nr_pages = (1 << cachep->gfporder);
1676 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1677 add_zone_page_state(page_zone(page),
1678 NR_SLAB_RECLAIMABLE, nr_pages);
1679 else
1680 add_zone_page_state(page_zone(page),
1681 NR_SLAB_UNRECLAIMABLE, nr_pages);
1682 for (i = 0; i < nr_pages; i++)
1683 __SetPageSlab(page + i);
1684 return page_address(page);
1688 * Interface to system's page release.
1690 static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1692 unsigned long i = (1 << cachep->gfporder);
1693 struct page *page = virt_to_page(addr);
1694 const unsigned long nr_freed = i;
1696 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1697 sub_zone_page_state(page_zone(page),
1698 NR_SLAB_RECLAIMABLE, nr_freed);
1699 else
1700 sub_zone_page_state(page_zone(page),
1701 NR_SLAB_UNRECLAIMABLE, nr_freed);
1702 while (i--) {
1703 BUG_ON(!PageSlab(page));
1704 __ClearPageSlab(page);
1705 page++;
1707 if (current->reclaim_state)
1708 current->reclaim_state->reclaimed_slab += nr_freed;
1709 free_pages((unsigned long)addr, cachep->gfporder);
1712 static void kmem_rcu_free(struct rcu_head *head)
1714 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1715 struct kmem_cache *cachep = slab_rcu->cachep;
1717 kmem_freepages(cachep, slab_rcu->addr);
1718 if (OFF_SLAB(cachep))
1719 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1722 #if DEBUG
1724 #ifdef CONFIG_DEBUG_PAGEALLOC
1725 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1726 unsigned long caller)
1728 int size = obj_size(cachep);
1730 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1732 if (size < 5 * sizeof(unsigned long))
1733 return;
1735 *addr++ = 0x12345678;
1736 *addr++ = caller;
1737 *addr++ = smp_processor_id();
1738 size -= 3 * sizeof(unsigned long);
1740 unsigned long *sptr = &caller;
1741 unsigned long svalue;
1743 while (!kstack_end(sptr)) {
1744 svalue = *sptr++;
1745 if (kernel_text_address(svalue)) {
1746 *addr++ = svalue;
1747 size -= sizeof(unsigned long);
1748 if (size <= sizeof(unsigned long))
1749 break;
1754 *addr++ = 0x87654321;
1756 #endif
1758 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1760 int size = obj_size(cachep);
1761 addr = &((char *)addr)[obj_offset(cachep)];
1763 memset(addr, val, size);
1764 *(unsigned char *)(addr + size - 1) = POISON_END;
1767 static void dump_line(char *data, int offset, int limit)
1769 int i;
1770 unsigned char error = 0;
1771 int bad_count = 0;
1773 printk(KERN_ERR "%03x:", offset);
1774 for (i = 0; i < limit; i++) {
1775 if (data[offset + i] != POISON_FREE) {
1776 error = data[offset + i];
1777 bad_count++;
1779 printk(" %02x", (unsigned char)data[offset + i]);
1781 printk("\n");
1783 if (bad_count == 1) {
1784 error ^= POISON_FREE;
1785 if (!(error & (error - 1))) {
1786 printk(KERN_ERR "Single bit error detected. Probably "
1787 "bad RAM.\n");
1788 #ifdef CONFIG_X86
1789 printk(KERN_ERR "Run memtest86+ or a similar memory "
1790 "test tool.\n");
1791 #else
1792 printk(KERN_ERR "Run a memory test tool.\n");
1793 #endif
1797 #endif
1799 #if DEBUG
1801 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1803 int i, size;
1804 char *realobj;
1806 if (cachep->flags & SLAB_RED_ZONE) {
1807 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
1808 *dbg_redzone1(cachep, objp),
1809 *dbg_redzone2(cachep, objp));
1812 if (cachep->flags & SLAB_STORE_USER) {
1813 printk(KERN_ERR "Last user: [<%p>]",
1814 *dbg_userword(cachep, objp));
1815 print_symbol("(%s)",
1816 (unsigned long)*dbg_userword(cachep, objp));
1817 printk("\n");
1819 realobj = (char *)objp + obj_offset(cachep);
1820 size = obj_size(cachep);
1821 for (i = 0; i < size && lines; i += 16, lines--) {
1822 int limit;
1823 limit = 16;
1824 if (i + limit > size)
1825 limit = size - i;
1826 dump_line(realobj, i, limit);
1830 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1832 char *realobj;
1833 int size, i;
1834 int lines = 0;
1836 realobj = (char *)objp + obj_offset(cachep);
1837 size = obj_size(cachep);
1839 for (i = 0; i < size; i++) {
1840 char exp = POISON_FREE;
1841 if (i == size - 1)
1842 exp = POISON_END;
1843 if (realobj[i] != exp) {
1844 int limit;
1845 /* Mismatch ! */
1846 /* Print header */
1847 if (lines == 0) {
1848 printk(KERN_ERR
1849 "Slab corruption: %s start=%p, len=%d\n",
1850 cachep->name, realobj, size);
1851 print_objinfo(cachep, objp, 0);
1853 /* Hexdump the affected line */
1854 i = (i / 16) * 16;
1855 limit = 16;
1856 if (i + limit > size)
1857 limit = size - i;
1858 dump_line(realobj, i, limit);
1859 i += 16;
1860 lines++;
1861 /* Limit to 5 lines */
1862 if (lines > 5)
1863 break;
1866 if (lines != 0) {
1867 /* Print some data about the neighboring objects, if they
1868 * exist:
1870 struct slab *slabp = virt_to_slab(objp);
1871 unsigned int objnr;
1873 objnr = obj_to_index(cachep, slabp, objp);
1874 if (objnr) {
1875 objp = index_to_obj(cachep, slabp, objnr - 1);
1876 realobj = (char *)objp + obj_offset(cachep);
1877 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1878 realobj, size);
1879 print_objinfo(cachep, objp, 2);
1881 if (objnr + 1 < cachep->num) {
1882 objp = index_to_obj(cachep, slabp, objnr + 1);
1883 realobj = (char *)objp + obj_offset(cachep);
1884 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1885 realobj, size);
1886 print_objinfo(cachep, objp, 2);
1890 #endif
1892 #if DEBUG
1894 * slab_destroy_objs - destroy a slab and its objects
1895 * @cachep: cache pointer being destroyed
1896 * @slabp: slab pointer being destroyed
1898 * Call the registered destructor for each object in a slab that is being
1899 * destroyed.
1901 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
1903 int i;
1904 for (i = 0; i < cachep->num; i++) {
1905 void *objp = index_to_obj(cachep, slabp, i);
1907 if (cachep->flags & SLAB_POISON) {
1908 #ifdef CONFIG_DEBUG_PAGEALLOC
1909 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1910 OFF_SLAB(cachep))
1911 kernel_map_pages(virt_to_page(objp),
1912 cachep->buffer_size / PAGE_SIZE, 1);
1913 else
1914 check_poison_obj(cachep, objp);
1915 #else
1916 check_poison_obj(cachep, objp);
1917 #endif
1919 if (cachep->flags & SLAB_RED_ZONE) {
1920 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1921 slab_error(cachep, "start of a freed object "
1922 "was overwritten");
1923 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1924 slab_error(cachep, "end of a freed object "
1925 "was overwritten");
1929 #else
1930 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
1933 #endif
1936 * slab_destroy - destroy and release all objects in a slab
1937 * @cachep: cache pointer being destroyed
1938 * @slabp: slab pointer being destroyed
1940 * Destroy all the objs in a slab, and release the mem back to the system.
1941 * Before calling the slab must have been unlinked from the cache. The
1942 * cache-lock is not held/needed.
1944 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
1946 void *addr = slabp->s_mem - slabp->colouroff;
1948 slab_destroy_objs(cachep, slabp);
1949 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1950 struct slab_rcu *slab_rcu;
1952 slab_rcu = (struct slab_rcu *)slabp;
1953 slab_rcu->cachep = cachep;
1954 slab_rcu->addr = addr;
1955 call_rcu(&slab_rcu->head, kmem_rcu_free);
1956 } else {
1957 kmem_freepages(cachep, addr);
1958 if (OFF_SLAB(cachep))
1959 kmem_cache_free(cachep->slabp_cache, slabp);
1964 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1965 * size of kmem_list3.
1967 static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1969 int node;
1971 for_each_node_state(node, N_NORMAL_MEMORY) {
1972 cachep->nodelists[node] = &initkmem_list3[index + node];
1973 cachep->nodelists[node]->next_reap = jiffies +
1974 REAPTIMEOUT_LIST3 +
1975 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1979 static void __kmem_cache_destroy(struct kmem_cache *cachep)
1981 int i;
1982 struct kmem_list3 *l3;
1984 for_each_online_cpu(i)
1985 kfree(cachep->array[i]);
1987 /* NUMA: free the list3 structures */
1988 for_each_online_node(i) {
1989 l3 = cachep->nodelists[i];
1990 if (l3) {
1991 kfree(l3->shared);
1992 free_alien_cache(l3->alien);
1993 kfree(l3);
1996 kmem_cache_free(&cache_cache, cachep);
2001 * calculate_slab_order - calculate size (page order) of slabs
2002 * @cachep: pointer to the cache that is being created
2003 * @size: size of objects to be created in this cache.
2004 * @align: required alignment for the objects.
2005 * @flags: slab allocation flags
2007 * Also calculates the number of objects per slab.
2009 * This could be made much more intelligent. For now, try to avoid using
2010 * high order pages for slabs. When the gfp() functions are more friendly
2011 * towards high-order requests, this should be changed.
2013 static size_t calculate_slab_order(struct kmem_cache *cachep,
2014 size_t size, size_t align, unsigned long flags)
2016 unsigned long offslab_limit;
2017 size_t left_over = 0;
2018 int gfporder;
2020 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
2021 unsigned int num;
2022 size_t remainder;
2024 cache_estimate(gfporder, size, align, flags, &remainder, &num);
2025 if (!num)
2026 continue;
2028 if (flags & CFLGS_OFF_SLAB) {
2030 * Max number of objs-per-slab for caches which
2031 * use off-slab slabs. Needed to avoid a possible
2032 * looping condition in cache_grow().
2034 offslab_limit = size - sizeof(struct slab);
2035 offslab_limit /= sizeof(kmem_bufctl_t);
2037 if (num > offslab_limit)
2038 break;
2041 /* Found something acceptable - save it away */
2042 cachep->num = num;
2043 cachep->gfporder = gfporder;
2044 left_over = remainder;
2047 * A VFS-reclaimable slab tends to have most allocations
2048 * as GFP_NOFS and we really don't want to have to be allocating
2049 * higher-order pages when we are unable to shrink dcache.
2051 if (flags & SLAB_RECLAIM_ACCOUNT)
2052 break;
2055 * Large number of objects is good, but very large slabs are
2056 * currently bad for the gfp()s.
2058 if (gfporder >= slab_break_gfp_order)
2059 break;
2062 * Acceptable internal fragmentation?
2064 if (left_over * 8 <= (PAGE_SIZE << gfporder))
2065 break;
2067 return left_over;
2070 static int __init_refok setup_cpu_cache(struct kmem_cache *cachep)
2072 if (g_cpucache_up == FULL)
2073 return enable_cpucache(cachep);
2075 if (g_cpucache_up == NONE) {
2077 * Note: the first kmem_cache_create must create the cache
2078 * that's used by kmalloc(24), otherwise the creation of
2079 * further caches will BUG().
2081 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2084 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2085 * the first cache, then we need to set up all its list3s,
2086 * otherwise the creation of further caches will BUG().
2088 set_up_list3s(cachep, SIZE_AC);
2089 if (INDEX_AC == INDEX_L3)
2090 g_cpucache_up = PARTIAL_L3;
2091 else
2092 g_cpucache_up = PARTIAL_AC;
2093 } else {
2094 cachep->array[smp_processor_id()] =
2095 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
2097 if (g_cpucache_up == PARTIAL_AC) {
2098 set_up_list3s(cachep, SIZE_L3);
2099 g_cpucache_up = PARTIAL_L3;
2100 } else {
2101 int node;
2102 for_each_node_state(node, N_NORMAL_MEMORY) {
2103 cachep->nodelists[node] =
2104 kmalloc_node(sizeof(struct kmem_list3),
2105 GFP_KERNEL, node);
2106 BUG_ON(!cachep->nodelists[node]);
2107 kmem_list3_init(cachep->nodelists[node]);
2111 cachep->nodelists[numa_node_id()]->next_reap =
2112 jiffies + REAPTIMEOUT_LIST3 +
2113 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2115 cpu_cache_get(cachep)->avail = 0;
2116 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2117 cpu_cache_get(cachep)->batchcount = 1;
2118 cpu_cache_get(cachep)->touched = 0;
2119 cachep->batchcount = 1;
2120 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2121 return 0;
2125 * kmem_cache_create - Create a cache.
2126 * @name: A string which is used in /proc/slabinfo to identify this cache.
2127 * @size: The size of objects to be created in this cache.
2128 * @align: The required alignment for the objects.
2129 * @flags: SLAB flags
2130 * @ctor: A constructor for the objects.
2132 * Returns a ptr to the cache on success, NULL on failure.
2133 * Cannot be called within a int, but can be interrupted.
2134 * The @ctor is run when new pages are allocated by the cache.
2136 * @name must be valid until the cache is destroyed. This implies that
2137 * the module calling this has to destroy the cache before getting unloaded.
2139 * The flags are
2141 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2142 * to catch references to uninitialised memory.
2144 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2145 * for buffer overruns.
2147 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2148 * cacheline. This can be beneficial if you're counting cycles as closely
2149 * as davem.
2151 struct kmem_cache *
2152 kmem_cache_create (const char *name, size_t size, size_t align,
2153 unsigned long flags,
2154 void (*ctor)(struct kmem_cache *, void *))
2156 size_t left_over, slab_size, ralign;
2157 struct kmem_cache *cachep = NULL, *pc;
2160 * Sanity checks... these are all serious usage bugs.
2162 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
2163 size > KMALLOC_MAX_SIZE) {
2164 printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__,
2165 name);
2166 BUG();
2170 * We use cache_chain_mutex to ensure a consistent view of
2171 * cpu_online_map as well. Please see cpuup_callback
2173 mutex_lock(&cache_chain_mutex);
2175 list_for_each_entry(pc, &cache_chain, next) {
2176 char tmp;
2177 int res;
2180 * This happens when the module gets unloaded and doesn't
2181 * destroy its slab cache and no-one else reuses the vmalloc
2182 * area of the module. Print a warning.
2184 res = probe_kernel_address(pc->name, tmp);
2185 if (res) {
2186 printk(KERN_ERR
2187 "SLAB: cache with size %d has lost its name\n",
2188 pc->buffer_size);
2189 continue;
2192 if (!strcmp(pc->name, name)) {
2193 printk(KERN_ERR
2194 "kmem_cache_create: duplicate cache %s\n", name);
2195 dump_stack();
2196 goto oops;
2200 #if DEBUG
2201 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
2202 #if FORCED_DEBUG
2204 * Enable redzoning and last user accounting, except for caches with
2205 * large objects, if the increased size would increase the object size
2206 * above the next power of two: caches with object sizes just above a
2207 * power of two have a significant amount of internal fragmentation.
2209 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2210 2 * sizeof(unsigned long long)))
2211 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2212 if (!(flags & SLAB_DESTROY_BY_RCU))
2213 flags |= SLAB_POISON;
2214 #endif
2215 if (flags & SLAB_DESTROY_BY_RCU)
2216 BUG_ON(flags & SLAB_POISON);
2217 #endif
2219 * Always checks flags, a caller might be expecting debug support which
2220 * isn't available.
2222 BUG_ON(flags & ~CREATE_MASK);
2225 * Check that size is in terms of words. This is needed to avoid
2226 * unaligned accesses for some archs when redzoning is used, and makes
2227 * sure any on-slab bufctl's are also correctly aligned.
2229 if (size & (BYTES_PER_WORD - 1)) {
2230 size += (BYTES_PER_WORD - 1);
2231 size &= ~(BYTES_PER_WORD - 1);
2234 /* calculate the final buffer alignment: */
2236 /* 1) arch recommendation: can be overridden for debug */
2237 if (flags & SLAB_HWCACHE_ALIGN) {
2239 * Default alignment: as specified by the arch code. Except if
2240 * an object is really small, then squeeze multiple objects into
2241 * one cacheline.
2243 ralign = cache_line_size();
2244 while (size <= ralign / 2)
2245 ralign /= 2;
2246 } else {
2247 ralign = BYTES_PER_WORD;
2251 * Redzoning and user store require word alignment or possibly larger.
2252 * Note this will be overridden by architecture or caller mandated
2253 * alignment if either is greater than BYTES_PER_WORD.
2255 if (flags & SLAB_STORE_USER)
2256 ralign = BYTES_PER_WORD;
2258 if (flags & SLAB_RED_ZONE) {
2259 ralign = REDZONE_ALIGN;
2260 /* If redzoning, ensure that the second redzone is suitably
2261 * aligned, by adjusting the object size accordingly. */
2262 size += REDZONE_ALIGN - 1;
2263 size &= ~(REDZONE_ALIGN - 1);
2266 /* 2) arch mandated alignment */
2267 if (ralign < ARCH_SLAB_MINALIGN) {
2268 ralign = ARCH_SLAB_MINALIGN;
2270 /* 3) caller mandated alignment */
2271 if (ralign < align) {
2272 ralign = align;
2274 /* disable debug if necessary */
2275 if (ralign > __alignof__(unsigned long long))
2276 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2278 * 4) Store it.
2280 align = ralign;
2282 /* Get cache's description obj. */
2283 cachep = kmem_cache_zalloc(&cache_cache, GFP_KERNEL);
2284 if (!cachep)
2285 goto oops;
2287 #if DEBUG
2288 cachep->obj_size = size;
2291 * Both debugging options require word-alignment which is calculated
2292 * into align above.
2294 if (flags & SLAB_RED_ZONE) {
2295 /* add space for red zone words */
2296 cachep->obj_offset += sizeof(unsigned long long);
2297 size += 2 * sizeof(unsigned long long);
2299 if (flags & SLAB_STORE_USER) {
2300 /* user store requires one word storage behind the end of
2301 * the real object. But if the second red zone needs to be
2302 * aligned to 64 bits, we must allow that much space.
2304 if (flags & SLAB_RED_ZONE)
2305 size += REDZONE_ALIGN;
2306 else
2307 size += BYTES_PER_WORD;
2309 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2310 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
2311 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2312 cachep->obj_offset += PAGE_SIZE - size;
2313 size = PAGE_SIZE;
2315 #endif
2316 #endif
2319 * Determine if the slab management is 'on' or 'off' slab.
2320 * (bootstrapping cannot cope with offslab caches so don't do
2321 * it too early on.)
2323 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init)
2325 * Size is large, assume best to place the slab management obj
2326 * off-slab (should allow better packing of objs).
2328 flags |= CFLGS_OFF_SLAB;
2330 size = ALIGN(size, align);
2332 left_over = calculate_slab_order(cachep, size, align, flags);
2334 if (!cachep->num) {
2335 printk(KERN_ERR
2336 "kmem_cache_create: couldn't create cache %s.\n", name);
2337 kmem_cache_free(&cache_cache, cachep);
2338 cachep = NULL;
2339 goto oops;
2341 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2342 + sizeof(struct slab), align);
2345 * If the slab has been placed off-slab, and we have enough space then
2346 * move it on-slab. This is at the expense of any extra colouring.
2348 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2349 flags &= ~CFLGS_OFF_SLAB;
2350 left_over -= slab_size;
2353 if (flags & CFLGS_OFF_SLAB) {
2354 /* really off slab. No need for manual alignment */
2355 slab_size =
2356 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2359 cachep->colour_off = cache_line_size();
2360 /* Offset must be a multiple of the alignment. */
2361 if (cachep->colour_off < align)
2362 cachep->colour_off = align;
2363 cachep->colour = left_over / cachep->colour_off;
2364 cachep->slab_size = slab_size;
2365 cachep->flags = flags;
2366 cachep->gfpflags = 0;
2367 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2368 cachep->gfpflags |= GFP_DMA;
2369 cachep->buffer_size = size;
2370 cachep->reciprocal_buffer_size = reciprocal_value(size);
2372 if (flags & CFLGS_OFF_SLAB) {
2373 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2375 * This is a possibility for one of the malloc_sizes caches.
2376 * But since we go off slab only for object size greater than
2377 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2378 * this should not happen at all.
2379 * But leave a BUG_ON for some lucky dude.
2381 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
2383 cachep->ctor = ctor;
2384 cachep->name = name;
2386 if (setup_cpu_cache(cachep)) {
2387 __kmem_cache_destroy(cachep);
2388 cachep = NULL;
2389 goto oops;
2392 /* cache setup completed, link it into the list */
2393 list_add(&cachep->next, &cache_chain);
2394 oops:
2395 if (!cachep && (flags & SLAB_PANIC))
2396 panic("kmem_cache_create(): failed to create slab `%s'\n",
2397 name);
2398 mutex_unlock(&cache_chain_mutex);
2399 return cachep;
2401 EXPORT_SYMBOL(kmem_cache_create);
2403 #if DEBUG
2404 static void check_irq_off(void)
2406 BUG_ON(!irqs_disabled());
2409 static void check_irq_on(void)
2411 BUG_ON(irqs_disabled());
2414 static void check_spinlock_acquired(struct kmem_cache *cachep)
2416 #ifdef CONFIG_SMP
2417 check_irq_off();
2418 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
2419 #endif
2422 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2424 #ifdef CONFIG_SMP
2425 check_irq_off();
2426 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2427 #endif
2430 #else
2431 #define check_irq_off() do { } while(0)
2432 #define check_irq_on() do { } while(0)
2433 #define check_spinlock_acquired(x) do { } while(0)
2434 #define check_spinlock_acquired_node(x, y) do { } while(0)
2435 #endif
2437 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2438 struct array_cache *ac,
2439 int force, int node);
2441 static void do_drain(void *arg)
2443 struct kmem_cache *cachep = arg;
2444 struct array_cache *ac;
2445 int node = numa_node_id();
2447 check_irq_off();
2448 ac = cpu_cache_get(cachep);
2449 spin_lock(&cachep->nodelists[node]->list_lock);
2450 free_block(cachep, ac->entry, ac->avail, node);
2451 spin_unlock(&cachep->nodelists[node]->list_lock);
2452 ac->avail = 0;
2455 static void drain_cpu_caches(struct kmem_cache *cachep)
2457 struct kmem_list3 *l3;
2458 int node;
2460 on_each_cpu(do_drain, cachep, 1, 1);
2461 check_irq_on();
2462 for_each_online_node(node) {
2463 l3 = cachep->nodelists[node];
2464 if (l3 && l3->alien)
2465 drain_alien_cache(cachep, l3->alien);
2468 for_each_online_node(node) {
2469 l3 = cachep->nodelists[node];
2470 if (l3)
2471 drain_array(cachep, l3, l3->shared, 1, node);
2476 * Remove slabs from the list of free slabs.
2477 * Specify the number of slabs to drain in tofree.
2479 * Returns the actual number of slabs released.
2481 static int drain_freelist(struct kmem_cache *cache,
2482 struct kmem_list3 *l3, int tofree)
2484 struct list_head *p;
2485 int nr_freed;
2486 struct slab *slabp;
2488 nr_freed = 0;
2489 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
2491 spin_lock_irq(&l3->list_lock);
2492 p = l3->slabs_free.prev;
2493 if (p == &l3->slabs_free) {
2494 spin_unlock_irq(&l3->list_lock);
2495 goto out;
2498 slabp = list_entry(p, struct slab, list);
2499 #if DEBUG
2500 BUG_ON(slabp->inuse);
2501 #endif
2502 list_del(&slabp->list);
2504 * Safe to drop the lock. The slab is no longer linked
2505 * to the cache.
2507 l3->free_objects -= cache->num;
2508 spin_unlock_irq(&l3->list_lock);
2509 slab_destroy(cache, slabp);
2510 nr_freed++;
2512 out:
2513 return nr_freed;
2516 /* Called with cache_chain_mutex held to protect against cpu hotplug */
2517 static int __cache_shrink(struct kmem_cache *cachep)
2519 int ret = 0, i = 0;
2520 struct kmem_list3 *l3;
2522 drain_cpu_caches(cachep);
2524 check_irq_on();
2525 for_each_online_node(i) {
2526 l3 = cachep->nodelists[i];
2527 if (!l3)
2528 continue;
2530 drain_freelist(cachep, l3, l3->free_objects);
2532 ret += !list_empty(&l3->slabs_full) ||
2533 !list_empty(&l3->slabs_partial);
2535 return (ret ? 1 : 0);
2539 * kmem_cache_shrink - Shrink a cache.
2540 * @cachep: The cache to shrink.
2542 * Releases as many slabs as possible for a cache.
2543 * To help debugging, a zero exit status indicates all slabs were released.
2545 int kmem_cache_shrink(struct kmem_cache *cachep)
2547 int ret;
2548 BUG_ON(!cachep || in_interrupt());
2550 mutex_lock(&cache_chain_mutex);
2551 ret = __cache_shrink(cachep);
2552 mutex_unlock(&cache_chain_mutex);
2553 return ret;
2555 EXPORT_SYMBOL(kmem_cache_shrink);
2558 * kmem_cache_destroy - delete a cache
2559 * @cachep: the cache to destroy
2561 * Remove a &struct kmem_cache object from the slab cache.
2563 * It is expected this function will be called by a module when it is
2564 * unloaded. This will remove the cache completely, and avoid a duplicate
2565 * cache being allocated each time a module is loaded and unloaded, if the
2566 * module doesn't have persistent in-kernel storage across loads and unloads.
2568 * The cache must be empty before calling this function.
2570 * The caller must guarantee that noone will allocate memory from the cache
2571 * during the kmem_cache_destroy().
2573 void kmem_cache_destroy(struct kmem_cache *cachep)
2575 BUG_ON(!cachep || in_interrupt());
2577 /* Find the cache in the chain of caches. */
2578 mutex_lock(&cache_chain_mutex);
2580 * the chain is never empty, cache_cache is never destroyed
2582 list_del(&cachep->next);
2583 if (__cache_shrink(cachep)) {
2584 slab_error(cachep, "Can't free all objects");
2585 list_add(&cachep->next, &cache_chain);
2586 mutex_unlock(&cache_chain_mutex);
2587 return;
2590 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
2591 synchronize_rcu();
2593 __kmem_cache_destroy(cachep);
2594 mutex_unlock(&cache_chain_mutex);
2596 EXPORT_SYMBOL(kmem_cache_destroy);
2599 * Get the memory for a slab management obj.
2600 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2601 * always come from malloc_sizes caches. The slab descriptor cannot
2602 * come from the same cache which is getting created because,
2603 * when we are searching for an appropriate cache for these
2604 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2605 * If we are creating a malloc_sizes cache here it would not be visible to
2606 * kmem_find_general_cachep till the initialization is complete.
2607 * Hence we cannot have slabp_cache same as the original cache.
2609 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2610 int colour_off, gfp_t local_flags,
2611 int nodeid)
2613 struct slab *slabp;
2615 if (OFF_SLAB(cachep)) {
2616 /* Slab management obj is off-slab. */
2617 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2618 local_flags & ~GFP_THISNODE, nodeid);
2619 if (!slabp)
2620 return NULL;
2621 } else {
2622 slabp = objp + colour_off;
2623 colour_off += cachep->slab_size;
2625 slabp->inuse = 0;
2626 slabp->colouroff = colour_off;
2627 slabp->s_mem = objp + colour_off;
2628 slabp->nodeid = nodeid;
2629 return slabp;
2632 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2634 return (kmem_bufctl_t *) (slabp + 1);
2637 static void cache_init_objs(struct kmem_cache *cachep,
2638 struct slab *slabp)
2640 int i;
2642 for (i = 0; i < cachep->num; i++) {
2643 void *objp = index_to_obj(cachep, slabp, i);
2644 #if DEBUG
2645 /* need to poison the objs? */
2646 if (cachep->flags & SLAB_POISON)
2647 poison_obj(cachep, objp, POISON_FREE);
2648 if (cachep->flags & SLAB_STORE_USER)
2649 *dbg_userword(cachep, objp) = NULL;
2651 if (cachep->flags & SLAB_RED_ZONE) {
2652 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2653 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2656 * Constructors are not allowed to allocate memory from the same
2657 * cache which they are a constructor for. Otherwise, deadlock.
2658 * They must also be threaded.
2660 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2661 cachep->ctor(cachep, objp + obj_offset(cachep));
2663 if (cachep->flags & SLAB_RED_ZONE) {
2664 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2665 slab_error(cachep, "constructor overwrote the"
2666 " end of an object");
2667 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2668 slab_error(cachep, "constructor overwrote the"
2669 " start of an object");
2671 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2672 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2673 kernel_map_pages(virt_to_page(objp),
2674 cachep->buffer_size / PAGE_SIZE, 0);
2675 #else
2676 if (cachep->ctor)
2677 cachep->ctor(cachep, objp);
2678 #endif
2679 slab_bufctl(slabp)[i] = i + 1;
2681 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2682 slabp->free = 0;
2685 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2687 if (CONFIG_ZONE_DMA_FLAG) {
2688 if (flags & GFP_DMA)
2689 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2690 else
2691 BUG_ON(cachep->gfpflags & GFP_DMA);
2695 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2696 int nodeid)
2698 void *objp = index_to_obj(cachep, slabp, slabp->free);
2699 kmem_bufctl_t next;
2701 slabp->inuse++;
2702 next = slab_bufctl(slabp)[slabp->free];
2703 #if DEBUG
2704 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2705 WARN_ON(slabp->nodeid != nodeid);
2706 #endif
2707 slabp->free = next;
2709 return objp;
2712 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2713 void *objp, int nodeid)
2715 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2717 #if DEBUG
2718 /* Verify that the slab belongs to the intended node */
2719 WARN_ON(slabp->nodeid != nodeid);
2721 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2722 printk(KERN_ERR "slab: double free detected in cache "
2723 "'%s', objp %p\n", cachep->name, objp);
2724 BUG();
2726 #endif
2727 slab_bufctl(slabp)[objnr] = slabp->free;
2728 slabp->free = objnr;
2729 slabp->inuse--;
2733 * Map pages beginning at addr to the given cache and slab. This is required
2734 * for the slab allocator to be able to lookup the cache and slab of a
2735 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2737 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2738 void *addr)
2740 int nr_pages;
2741 struct page *page;
2743 page = virt_to_page(addr);
2745 nr_pages = 1;
2746 if (likely(!PageCompound(page)))
2747 nr_pages <<= cache->gfporder;
2749 do {
2750 page_set_cache(page, cache);
2751 page_set_slab(page, slab);
2752 page++;
2753 } while (--nr_pages);
2757 * Grow (by 1) the number of slabs within a cache. This is called by
2758 * kmem_cache_alloc() when there are no active objs left in a cache.
2760 static int cache_grow(struct kmem_cache *cachep,
2761 gfp_t flags, int nodeid, void *objp)
2763 struct slab *slabp;
2764 size_t offset;
2765 gfp_t local_flags;
2766 struct kmem_list3 *l3;
2769 * Be lazy and only check for valid flags here, keeping it out of the
2770 * critical path in kmem_cache_alloc().
2772 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2773 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2775 /* Take the l3 list lock to change the colour_next on this node */
2776 check_irq_off();
2777 l3 = cachep->nodelists[nodeid];
2778 spin_lock(&l3->list_lock);
2780 /* Get colour for the slab, and cal the next value. */
2781 offset = l3->colour_next;
2782 l3->colour_next++;
2783 if (l3->colour_next >= cachep->colour)
2784 l3->colour_next = 0;
2785 spin_unlock(&l3->list_lock);
2787 offset *= cachep->colour_off;
2789 if (local_flags & __GFP_WAIT)
2790 local_irq_enable();
2793 * The test for missing atomic flag is performed here, rather than
2794 * the more obvious place, simply to reduce the critical path length
2795 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2796 * will eventually be caught here (where it matters).
2798 kmem_flagcheck(cachep, flags);
2801 * Get mem for the objs. Attempt to allocate a physical page from
2802 * 'nodeid'.
2804 if (!objp)
2805 objp = kmem_getpages(cachep, local_flags, nodeid);
2806 if (!objp)
2807 goto failed;
2809 /* Get slab management. */
2810 slabp = alloc_slabmgmt(cachep, objp, offset,
2811 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2812 if (!slabp)
2813 goto opps1;
2815 slabp->nodeid = nodeid;
2816 slab_map_pages(cachep, slabp, objp);
2818 cache_init_objs(cachep, slabp);
2820 if (local_flags & __GFP_WAIT)
2821 local_irq_disable();
2822 check_irq_off();
2823 spin_lock(&l3->list_lock);
2825 /* Make slab active. */
2826 list_add_tail(&slabp->list, &(l3->slabs_free));
2827 STATS_INC_GROWN(cachep);
2828 l3->free_objects += cachep->num;
2829 spin_unlock(&l3->list_lock);
2830 return 1;
2831 opps1:
2832 kmem_freepages(cachep, objp);
2833 failed:
2834 if (local_flags & __GFP_WAIT)
2835 local_irq_disable();
2836 return 0;
2839 #if DEBUG
2842 * Perform extra freeing checks:
2843 * - detect bad pointers.
2844 * - POISON/RED_ZONE checking
2846 static void kfree_debugcheck(const void *objp)
2848 if (!virt_addr_valid(objp)) {
2849 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2850 (unsigned long)objp);
2851 BUG();
2855 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2857 unsigned long long redzone1, redzone2;
2859 redzone1 = *dbg_redzone1(cache, obj);
2860 redzone2 = *dbg_redzone2(cache, obj);
2863 * Redzone is ok.
2865 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2866 return;
2868 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2869 slab_error(cache, "double free detected");
2870 else
2871 slab_error(cache, "memory outside object was overwritten");
2873 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2874 obj, redzone1, redzone2);
2877 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2878 void *caller)
2880 struct page *page;
2881 unsigned int objnr;
2882 struct slab *slabp;
2884 BUG_ON(virt_to_cache(objp) != cachep);
2886 objp -= obj_offset(cachep);
2887 kfree_debugcheck(objp);
2888 page = virt_to_head_page(objp);
2890 slabp = page_get_slab(page);
2892 if (cachep->flags & SLAB_RED_ZONE) {
2893 verify_redzone_free(cachep, objp);
2894 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2895 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2897 if (cachep->flags & SLAB_STORE_USER)
2898 *dbg_userword(cachep, objp) = caller;
2900 objnr = obj_to_index(cachep, slabp, objp);
2902 BUG_ON(objnr >= cachep->num);
2903 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
2905 #ifdef CONFIG_DEBUG_SLAB_LEAK
2906 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2907 #endif
2908 if (cachep->flags & SLAB_POISON) {
2909 #ifdef CONFIG_DEBUG_PAGEALLOC
2910 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
2911 store_stackinfo(cachep, objp, (unsigned long)caller);
2912 kernel_map_pages(virt_to_page(objp),
2913 cachep->buffer_size / PAGE_SIZE, 0);
2914 } else {
2915 poison_obj(cachep, objp, POISON_FREE);
2917 #else
2918 poison_obj(cachep, objp, POISON_FREE);
2919 #endif
2921 return objp;
2924 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
2926 kmem_bufctl_t i;
2927 int entries = 0;
2929 /* Check slab's freelist to see if this obj is there. */
2930 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2931 entries++;
2932 if (entries > cachep->num || i >= cachep->num)
2933 goto bad;
2935 if (entries != cachep->num - slabp->inuse) {
2936 bad:
2937 printk(KERN_ERR "slab: Internal list corruption detected in "
2938 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2939 cachep->name, cachep->num, slabp, slabp->inuse);
2940 for (i = 0;
2941 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
2942 i++) {
2943 if (i % 16 == 0)
2944 printk("\n%03x:", i);
2945 printk(" %02x", ((unsigned char *)slabp)[i]);
2947 printk("\n");
2948 BUG();
2951 #else
2952 #define kfree_debugcheck(x) do { } while(0)
2953 #define cache_free_debugcheck(x,objp,z) (objp)
2954 #define check_slabp(x,y) do { } while(0)
2955 #endif
2957 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
2959 int batchcount;
2960 struct kmem_list3 *l3;
2961 struct array_cache *ac;
2962 int node;
2964 node = numa_node_id();
2966 check_irq_off();
2967 ac = cpu_cache_get(cachep);
2968 retry:
2969 batchcount = ac->batchcount;
2970 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2972 * If there was little recent activity on this cache, then
2973 * perform only a partial refill. Otherwise we could generate
2974 * refill bouncing.
2976 batchcount = BATCHREFILL_LIMIT;
2978 l3 = cachep->nodelists[node];
2980 BUG_ON(ac->avail > 0 || !l3);
2981 spin_lock(&l3->list_lock);
2983 /* See if we can refill from the shared array */
2984 if (l3->shared && transfer_objects(ac, l3->shared, batchcount))
2985 goto alloc_done;
2987 while (batchcount > 0) {
2988 struct list_head *entry;
2989 struct slab *slabp;
2990 /* Get slab alloc is to come from. */
2991 entry = l3->slabs_partial.next;
2992 if (entry == &l3->slabs_partial) {
2993 l3->free_touched = 1;
2994 entry = l3->slabs_free.next;
2995 if (entry == &l3->slabs_free)
2996 goto must_grow;
2999 slabp = list_entry(entry, struct slab, list);
3000 check_slabp(cachep, slabp);
3001 check_spinlock_acquired(cachep);
3004 * The slab was either on partial or free list so
3005 * there must be at least one object available for
3006 * allocation.
3008 BUG_ON(slabp->inuse < 0 || slabp->inuse >= cachep->num);
3010 while (slabp->inuse < cachep->num && batchcount--) {
3011 STATS_INC_ALLOCED(cachep);
3012 STATS_INC_ACTIVE(cachep);
3013 STATS_SET_HIGH(cachep);
3015 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
3016 node);
3018 check_slabp(cachep, slabp);
3020 /* move slabp to correct slabp list: */
3021 list_del(&slabp->list);
3022 if (slabp->free == BUFCTL_END)
3023 list_add(&slabp->list, &l3->slabs_full);
3024 else
3025 list_add(&slabp->list, &l3->slabs_partial);
3028 must_grow:
3029 l3->free_objects -= ac->avail;
3030 alloc_done:
3031 spin_unlock(&l3->list_lock);
3033 if (unlikely(!ac->avail)) {
3034 int x;
3035 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
3037 /* cache_grow can reenable interrupts, then ac could change. */
3038 ac = cpu_cache_get(cachep);
3039 if (!x && ac->avail == 0) /* no objects in sight? abort */
3040 return NULL;
3042 if (!ac->avail) /* objects refilled by interrupt? */
3043 goto retry;
3045 ac->touched = 1;
3046 return ac->entry[--ac->avail];
3049 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3050 gfp_t flags)
3052 might_sleep_if(flags & __GFP_WAIT);
3053 #if DEBUG
3054 kmem_flagcheck(cachep, flags);
3055 #endif
3058 #if DEBUG
3059 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3060 gfp_t flags, void *objp, void *caller)
3062 if (!objp)
3063 return objp;
3064 if (cachep->flags & SLAB_POISON) {
3065 #ifdef CONFIG_DEBUG_PAGEALLOC
3066 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
3067 kernel_map_pages(virt_to_page(objp),
3068 cachep->buffer_size / PAGE_SIZE, 1);
3069 else
3070 check_poison_obj(cachep, objp);
3071 #else
3072 check_poison_obj(cachep, objp);
3073 #endif
3074 poison_obj(cachep, objp, POISON_INUSE);
3076 if (cachep->flags & SLAB_STORE_USER)
3077 *dbg_userword(cachep, objp) = caller;
3079 if (cachep->flags & SLAB_RED_ZONE) {
3080 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3081 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3082 slab_error(cachep, "double free, or memory outside"
3083 " object was overwritten");
3084 printk(KERN_ERR
3085 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3086 objp, *dbg_redzone1(cachep, objp),
3087 *dbg_redzone2(cachep, objp));
3089 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3090 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3092 #ifdef CONFIG_DEBUG_SLAB_LEAK
3094 struct slab *slabp;
3095 unsigned objnr;
3097 slabp = page_get_slab(virt_to_head_page(objp));
3098 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3099 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3101 #endif
3102 objp += obj_offset(cachep);
3103 if (cachep->ctor && cachep->flags & SLAB_POISON)
3104 cachep->ctor(cachep, objp);
3105 #if ARCH_SLAB_MINALIGN
3106 if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) {
3107 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3108 objp, ARCH_SLAB_MINALIGN);
3110 #endif
3111 return objp;
3113 #else
3114 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3115 #endif
3117 #ifdef CONFIG_FAILSLAB
3119 static struct failslab_attr {
3121 struct fault_attr attr;
3123 u32 ignore_gfp_wait;
3124 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3125 struct dentry *ignore_gfp_wait_file;
3126 #endif
3128 } failslab = {
3129 .attr = FAULT_ATTR_INITIALIZER,
3130 .ignore_gfp_wait = 1,
3133 static int __init setup_failslab(char *str)
3135 return setup_fault_attr(&failslab.attr, str);
3137 __setup("failslab=", setup_failslab);
3139 static int should_failslab(struct kmem_cache *cachep, gfp_t flags)
3141 if (cachep == &cache_cache)
3142 return 0;
3143 if (flags & __GFP_NOFAIL)
3144 return 0;
3145 if (failslab.ignore_gfp_wait && (flags & __GFP_WAIT))
3146 return 0;
3148 return should_fail(&failslab.attr, obj_size(cachep));
3151 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3153 static int __init failslab_debugfs(void)
3155 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
3156 struct dentry *dir;
3157 int err;
3159 err = init_fault_attr_dentries(&failslab.attr, "failslab");
3160 if (err)
3161 return err;
3162 dir = failslab.attr.dentries.dir;
3164 failslab.ignore_gfp_wait_file =
3165 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3166 &failslab.ignore_gfp_wait);
3168 if (!failslab.ignore_gfp_wait_file) {
3169 err = -ENOMEM;
3170 debugfs_remove(failslab.ignore_gfp_wait_file);
3171 cleanup_fault_attr_dentries(&failslab.attr);
3174 return err;
3177 late_initcall(failslab_debugfs);
3179 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3181 #else /* CONFIG_FAILSLAB */
3183 static inline int should_failslab(struct kmem_cache *cachep, gfp_t flags)
3185 return 0;
3188 #endif /* CONFIG_FAILSLAB */
3190 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3192 void *objp;
3193 struct array_cache *ac;
3195 check_irq_off();
3197 ac = cpu_cache_get(cachep);
3198 if (likely(ac->avail)) {
3199 STATS_INC_ALLOCHIT(cachep);
3200 ac->touched = 1;
3201 objp = ac->entry[--ac->avail];
3202 } else {
3203 STATS_INC_ALLOCMISS(cachep);
3204 objp = cache_alloc_refill(cachep, flags);
3206 return objp;
3209 #ifdef CONFIG_NUMA
3211 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3213 * If we are in_interrupt, then process context, including cpusets and
3214 * mempolicy, may not apply and should not be used for allocation policy.
3216 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3218 int nid_alloc, nid_here;
3220 if (in_interrupt() || (flags & __GFP_THISNODE))
3221 return NULL;
3222 nid_alloc = nid_here = numa_node_id();
3223 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3224 nid_alloc = cpuset_mem_spread_node();
3225 else if (current->mempolicy)
3226 nid_alloc = slab_node(current->mempolicy);
3227 if (nid_alloc != nid_here)
3228 return ____cache_alloc_node(cachep, flags, nid_alloc);
3229 return NULL;
3233 * Fallback function if there was no memory available and no objects on a
3234 * certain node and fall back is permitted. First we scan all the
3235 * available nodelists for available objects. If that fails then we
3236 * perform an allocation without specifying a node. This allows the page
3237 * allocator to do its reclaim / fallback magic. We then insert the
3238 * slab into the proper nodelist and then allocate from it.
3240 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3242 struct zonelist *zonelist;
3243 gfp_t local_flags;
3244 struct zone **z;
3245 void *obj = NULL;
3246 int nid;
3248 if (flags & __GFP_THISNODE)
3249 return NULL;
3251 zonelist = &NODE_DATA(slab_node(current->mempolicy))
3252 ->node_zonelists[gfp_zone(flags)];
3253 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3255 retry:
3257 * Look through allowed nodes for objects available
3258 * from existing per node queues.
3260 for (z = zonelist->zones; *z && !obj; z++) {
3261 nid = zone_to_nid(*z);
3263 if (cpuset_zone_allowed_hardwall(*z, flags) &&
3264 cache->nodelists[nid] &&
3265 cache->nodelists[nid]->free_objects)
3266 obj = ____cache_alloc_node(cache,
3267 flags | GFP_THISNODE, nid);
3270 if (!obj) {
3272 * This allocation will be performed within the constraints
3273 * of the current cpuset / memory policy requirements.
3274 * We may trigger various forms of reclaim on the allowed
3275 * set and go into memory reserves if necessary.
3277 if (local_flags & __GFP_WAIT)
3278 local_irq_enable();
3279 kmem_flagcheck(cache, flags);
3280 obj = kmem_getpages(cache, flags, -1);
3281 if (local_flags & __GFP_WAIT)
3282 local_irq_disable();
3283 if (obj) {
3285 * Insert into the appropriate per node queues
3287 nid = page_to_nid(virt_to_page(obj));
3288 if (cache_grow(cache, flags, nid, obj)) {
3289 obj = ____cache_alloc_node(cache,
3290 flags | GFP_THISNODE, nid);
3291 if (!obj)
3293 * Another processor may allocate the
3294 * objects in the slab since we are
3295 * not holding any locks.
3297 goto retry;
3298 } else {
3299 /* cache_grow already freed obj */
3300 obj = NULL;
3304 return obj;
3308 * A interface to enable slab creation on nodeid
3310 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3311 int nodeid)
3313 struct list_head *entry;
3314 struct slab *slabp;
3315 struct kmem_list3 *l3;
3316 void *obj;
3317 int x;
3319 l3 = cachep->nodelists[nodeid];
3320 BUG_ON(!l3);
3322 retry:
3323 check_irq_off();
3324 spin_lock(&l3->list_lock);
3325 entry = l3->slabs_partial.next;
3326 if (entry == &l3->slabs_partial) {
3327 l3->free_touched = 1;
3328 entry = l3->slabs_free.next;
3329 if (entry == &l3->slabs_free)
3330 goto must_grow;
3333 slabp = list_entry(entry, struct slab, list);
3334 check_spinlock_acquired_node(cachep, nodeid);
3335 check_slabp(cachep, slabp);
3337 STATS_INC_NODEALLOCS(cachep);
3338 STATS_INC_ACTIVE(cachep);
3339 STATS_SET_HIGH(cachep);
3341 BUG_ON(slabp->inuse == cachep->num);
3343 obj = slab_get_obj(cachep, slabp, nodeid);
3344 check_slabp(cachep, slabp);
3345 l3->free_objects--;
3346 /* move slabp to correct slabp list: */
3347 list_del(&slabp->list);
3349 if (slabp->free == BUFCTL_END)
3350 list_add(&slabp->list, &l3->slabs_full);
3351 else
3352 list_add(&slabp->list, &l3->slabs_partial);
3354 spin_unlock(&l3->list_lock);
3355 goto done;
3357 must_grow:
3358 spin_unlock(&l3->list_lock);
3359 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
3360 if (x)
3361 goto retry;
3363 return fallback_alloc(cachep, flags);
3365 done:
3366 return obj;
3370 * kmem_cache_alloc_node - Allocate an object on the specified node
3371 * @cachep: The cache to allocate from.
3372 * @flags: See kmalloc().
3373 * @nodeid: node number of the target node.
3374 * @caller: return address of caller, used for debug information
3376 * Identical to kmem_cache_alloc but it will allocate memory on the given
3377 * node, which can improve the performance for cpu bound structures.
3379 * Fallback to other node is possible if __GFP_THISNODE is not set.
3381 static __always_inline void *
3382 __cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3383 void *caller)
3385 unsigned long save_flags;
3386 void *ptr;
3388 if (should_failslab(cachep, flags))
3389 return NULL;
3391 cache_alloc_debugcheck_before(cachep, flags);
3392 local_irq_save(save_flags);
3394 if (unlikely(nodeid == -1))
3395 nodeid = numa_node_id();
3397 if (unlikely(!cachep->nodelists[nodeid])) {
3398 /* Node not bootstrapped yet */
3399 ptr = fallback_alloc(cachep, flags);
3400 goto out;
3403 if (nodeid == numa_node_id()) {
3405 * Use the locally cached objects if possible.
3406 * However ____cache_alloc does not allow fallback
3407 * to other nodes. It may fail while we still have
3408 * objects on other nodes available.
3410 ptr = ____cache_alloc(cachep, flags);
3411 if (ptr)
3412 goto out;
3414 /* ___cache_alloc_node can fall back to other nodes */
3415 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3416 out:
3417 local_irq_restore(save_flags);
3418 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3420 if (unlikely((flags & __GFP_ZERO) && ptr))
3421 memset(ptr, 0, obj_size(cachep));
3423 return ptr;
3426 static __always_inline void *
3427 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3429 void *objp;
3431 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3432 objp = alternate_node_alloc(cache, flags);
3433 if (objp)
3434 goto out;
3436 objp = ____cache_alloc(cache, flags);
3439 * We may just have run out of memory on the local node.
3440 * ____cache_alloc_node() knows how to locate memory on other nodes
3442 if (!objp)
3443 objp = ____cache_alloc_node(cache, flags, numa_node_id());
3445 out:
3446 return objp;
3448 #else
3450 static __always_inline void *
3451 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3453 return ____cache_alloc(cachep, flags);
3456 #endif /* CONFIG_NUMA */
3458 static __always_inline void *
3459 __cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3461 unsigned long save_flags;
3462 void *objp;
3464 if (should_failslab(cachep, flags))
3465 return NULL;
3467 cache_alloc_debugcheck_before(cachep, flags);
3468 local_irq_save(save_flags);
3469 objp = __do_cache_alloc(cachep, flags);
3470 local_irq_restore(save_flags);
3471 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3472 prefetchw(objp);
3474 if (unlikely((flags & __GFP_ZERO) && objp))
3475 memset(objp, 0, obj_size(cachep));
3477 return objp;
3481 * Caller needs to acquire correct kmem_list's list_lock
3483 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3484 int node)
3486 int i;
3487 struct kmem_list3 *l3;
3489 for (i = 0; i < nr_objects; i++) {
3490 void *objp = objpp[i];
3491 struct slab *slabp;
3493 slabp = virt_to_slab(objp);
3494 l3 = cachep->nodelists[node];
3495 list_del(&slabp->list);
3496 check_spinlock_acquired_node(cachep, node);
3497 check_slabp(cachep, slabp);
3498 slab_put_obj(cachep, slabp, objp, node);
3499 STATS_DEC_ACTIVE(cachep);
3500 l3->free_objects++;
3501 check_slabp(cachep, slabp);
3503 /* fixup slab chains */
3504 if (slabp->inuse == 0) {
3505 if (l3->free_objects > l3->free_limit) {
3506 l3->free_objects -= cachep->num;
3507 /* No need to drop any previously held
3508 * lock here, even if we have a off-slab slab
3509 * descriptor it is guaranteed to come from
3510 * a different cache, refer to comments before
3511 * alloc_slabmgmt.
3513 slab_destroy(cachep, slabp);
3514 } else {
3515 list_add(&slabp->list, &l3->slabs_free);
3517 } else {
3518 /* Unconditionally move a slab to the end of the
3519 * partial list on free - maximum time for the
3520 * other objects to be freed, too.
3522 list_add_tail(&slabp->list, &l3->slabs_partial);
3527 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3529 int batchcount;
3530 struct kmem_list3 *l3;
3531 int node = numa_node_id();
3533 batchcount = ac->batchcount;
3534 #if DEBUG
3535 BUG_ON(!batchcount || batchcount > ac->avail);
3536 #endif
3537 check_irq_off();
3538 l3 = cachep->nodelists[node];
3539 spin_lock(&l3->list_lock);
3540 if (l3->shared) {
3541 struct array_cache *shared_array = l3->shared;
3542 int max = shared_array->limit - shared_array->avail;
3543 if (max) {
3544 if (batchcount > max)
3545 batchcount = max;
3546 memcpy(&(shared_array->entry[shared_array->avail]),
3547 ac->entry, sizeof(void *) * batchcount);
3548 shared_array->avail += batchcount;
3549 goto free_done;
3553 free_block(cachep, ac->entry, batchcount, node);
3554 free_done:
3555 #if STATS
3557 int i = 0;
3558 struct list_head *p;
3560 p = l3->slabs_free.next;
3561 while (p != &(l3->slabs_free)) {
3562 struct slab *slabp;
3564 slabp = list_entry(p, struct slab, list);
3565 BUG_ON(slabp->inuse);
3567 i++;
3568 p = p->next;
3570 STATS_SET_FREEABLE(cachep, i);
3572 #endif
3573 spin_unlock(&l3->list_lock);
3574 ac->avail -= batchcount;
3575 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3579 * Release an obj back to its cache. If the obj has a constructed state, it must
3580 * be in this state _before_ it is released. Called with disabled ints.
3582 static inline void __cache_free(struct kmem_cache *cachep, void *objp)
3584 struct array_cache *ac = cpu_cache_get(cachep);
3586 check_irq_off();
3587 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3590 * Skip calling cache_free_alien() when the platform is not numa.
3591 * This will avoid cache misses that happen while accessing slabp (which
3592 * is per page memory reference) to get nodeid. Instead use a global
3593 * variable to skip the call, which is mostly likely to be present in
3594 * the cache.
3596 if (numa_platform && cache_free_alien(cachep, objp))
3597 return;
3599 if (likely(ac->avail < ac->limit)) {
3600 STATS_INC_FREEHIT(cachep);
3601 ac->entry[ac->avail++] = objp;
3602 return;
3603 } else {
3604 STATS_INC_FREEMISS(cachep);
3605 cache_flusharray(cachep, ac);
3606 ac->entry[ac->avail++] = objp;
3611 * kmem_cache_alloc - Allocate an object
3612 * @cachep: The cache to allocate from.
3613 * @flags: See kmalloc().
3615 * Allocate an object from this cache. The flags are only relevant
3616 * if the cache has no available objects.
3618 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3620 return __cache_alloc(cachep, flags, __builtin_return_address(0));
3622 EXPORT_SYMBOL(kmem_cache_alloc);
3625 * kmem_ptr_validate - check if an untrusted pointer might
3626 * be a slab entry.
3627 * @cachep: the cache we're checking against
3628 * @ptr: pointer to validate
3630 * This verifies that the untrusted pointer looks sane:
3631 * it is _not_ a guarantee that the pointer is actually
3632 * part of the slab cache in question, but it at least
3633 * validates that the pointer can be dereferenced and
3634 * looks half-way sane.
3636 * Currently only used for dentry validation.
3638 int kmem_ptr_validate(struct kmem_cache *cachep, const void *ptr)
3640 unsigned long addr = (unsigned long)ptr;
3641 unsigned long min_addr = PAGE_OFFSET;
3642 unsigned long align_mask = BYTES_PER_WORD - 1;
3643 unsigned long size = cachep->buffer_size;
3644 struct page *page;
3646 if (unlikely(addr < min_addr))
3647 goto out;
3648 if (unlikely(addr > (unsigned long)high_memory - size))
3649 goto out;
3650 if (unlikely(addr & align_mask))
3651 goto out;
3652 if (unlikely(!kern_addr_valid(addr)))
3653 goto out;
3654 if (unlikely(!kern_addr_valid(addr + size - 1)))
3655 goto out;
3656 page = virt_to_page(ptr);
3657 if (unlikely(!PageSlab(page)))
3658 goto out;
3659 if (unlikely(page_get_cache(page) != cachep))
3660 goto out;
3661 return 1;
3662 out:
3663 return 0;
3666 #ifdef CONFIG_NUMA
3667 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3669 return __cache_alloc_node(cachep, flags, nodeid,
3670 __builtin_return_address(0));
3672 EXPORT_SYMBOL(kmem_cache_alloc_node);
3674 static __always_inline void *
3675 __do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
3677 struct kmem_cache *cachep;
3679 cachep = kmem_find_general_cachep(size, flags);
3680 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3681 return cachep;
3682 return kmem_cache_alloc_node(cachep, flags, node);
3685 #ifdef CONFIG_DEBUG_SLAB
3686 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3688 return __do_kmalloc_node(size, flags, node,
3689 __builtin_return_address(0));
3691 EXPORT_SYMBOL(__kmalloc_node);
3693 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3694 int node, void *caller)
3696 return __do_kmalloc_node(size, flags, node, caller);
3698 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3699 #else
3700 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3702 return __do_kmalloc_node(size, flags, node, NULL);
3704 EXPORT_SYMBOL(__kmalloc_node);
3705 #endif /* CONFIG_DEBUG_SLAB */
3706 #endif /* CONFIG_NUMA */
3709 * __do_kmalloc - allocate memory
3710 * @size: how many bytes of memory are required.
3711 * @flags: the type of memory to allocate (see kmalloc).
3712 * @caller: function caller for debug tracking of the caller
3714 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3715 void *caller)
3717 struct kmem_cache *cachep;
3719 /* If you want to save a few bytes .text space: replace
3720 * __ with kmem_.
3721 * Then kmalloc uses the uninlined functions instead of the inline
3722 * functions.
3724 cachep = __find_general_cachep(size, flags);
3725 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3726 return cachep;
3727 return __cache_alloc(cachep, flags, caller);
3731 #ifdef CONFIG_DEBUG_SLAB
3732 void *__kmalloc(size_t size, gfp_t flags)
3734 return __do_kmalloc(size, flags, __builtin_return_address(0));
3736 EXPORT_SYMBOL(__kmalloc);
3738 void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller)
3740 return __do_kmalloc(size, flags, caller);
3742 EXPORT_SYMBOL(__kmalloc_track_caller);
3744 #else
3745 void *__kmalloc(size_t size, gfp_t flags)
3747 return __do_kmalloc(size, flags, NULL);
3749 EXPORT_SYMBOL(__kmalloc);
3750 #endif
3753 * kmem_cache_free - Deallocate an object
3754 * @cachep: The cache the allocation was from.
3755 * @objp: The previously allocated object.
3757 * Free an object which was previously allocated from this
3758 * cache.
3760 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3762 unsigned long flags;
3764 local_irq_save(flags);
3765 debug_check_no_locks_freed(objp, obj_size(cachep));
3766 __cache_free(cachep, objp);
3767 local_irq_restore(flags);
3769 EXPORT_SYMBOL(kmem_cache_free);
3772 * kfree - free previously allocated memory
3773 * @objp: pointer returned by kmalloc.
3775 * If @objp is NULL, no operation is performed.
3777 * Don't free memory not originally allocated by kmalloc()
3778 * or you will run into trouble.
3780 void kfree(const void *objp)
3782 struct kmem_cache *c;
3783 unsigned long flags;
3785 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3786 return;
3787 local_irq_save(flags);
3788 kfree_debugcheck(objp);
3789 c = virt_to_cache(objp);
3790 debug_check_no_locks_freed(objp, obj_size(c));
3791 __cache_free(c, (void *)objp);
3792 local_irq_restore(flags);
3794 EXPORT_SYMBOL(kfree);
3796 unsigned int kmem_cache_size(struct kmem_cache *cachep)
3798 return obj_size(cachep);
3800 EXPORT_SYMBOL(kmem_cache_size);
3802 const char *kmem_cache_name(struct kmem_cache *cachep)
3804 return cachep->name;
3806 EXPORT_SYMBOL_GPL(kmem_cache_name);
3809 * This initializes kmem_list3 or resizes various caches for all nodes.
3811 static int alloc_kmemlist(struct kmem_cache *cachep)
3813 int node;
3814 struct kmem_list3 *l3;
3815 struct array_cache *new_shared;
3816 struct array_cache **new_alien = NULL;
3818 for_each_node_state(node, N_NORMAL_MEMORY) {
3820 if (use_alien_caches) {
3821 new_alien = alloc_alien_cache(node, cachep->limit);
3822 if (!new_alien)
3823 goto fail;
3826 new_shared = NULL;
3827 if (cachep->shared) {
3828 new_shared = alloc_arraycache(node,
3829 cachep->shared*cachep->batchcount,
3830 0xbaadf00d);
3831 if (!new_shared) {
3832 free_alien_cache(new_alien);
3833 goto fail;
3837 l3 = cachep->nodelists[node];
3838 if (l3) {
3839 struct array_cache *shared = l3->shared;
3841 spin_lock_irq(&l3->list_lock);
3843 if (shared)
3844 free_block(cachep, shared->entry,
3845 shared->avail, node);
3847 l3->shared = new_shared;
3848 if (!l3->alien) {
3849 l3->alien = new_alien;
3850 new_alien = NULL;
3852 l3->free_limit = (1 + nr_cpus_node(node)) *
3853 cachep->batchcount + cachep->num;
3854 spin_unlock_irq(&l3->list_lock);
3855 kfree(shared);
3856 free_alien_cache(new_alien);
3857 continue;
3859 l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node);
3860 if (!l3) {
3861 free_alien_cache(new_alien);
3862 kfree(new_shared);
3863 goto fail;
3866 kmem_list3_init(l3);
3867 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3868 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3869 l3->shared = new_shared;
3870 l3->alien = new_alien;
3871 l3->free_limit = (1 + nr_cpus_node(node)) *
3872 cachep->batchcount + cachep->num;
3873 cachep->nodelists[node] = l3;
3875 return 0;
3877 fail:
3878 if (!cachep->next.next) {
3879 /* Cache is not active yet. Roll back what we did */
3880 node--;
3881 while (node >= 0) {
3882 if (cachep->nodelists[node]) {
3883 l3 = cachep->nodelists[node];
3885 kfree(l3->shared);
3886 free_alien_cache(l3->alien);
3887 kfree(l3);
3888 cachep->nodelists[node] = NULL;
3890 node--;
3893 return -ENOMEM;
3896 struct ccupdate_struct {
3897 struct kmem_cache *cachep;
3898 struct array_cache *new[NR_CPUS];
3901 static void do_ccupdate_local(void *info)
3903 struct ccupdate_struct *new = info;
3904 struct array_cache *old;
3906 check_irq_off();
3907 old = cpu_cache_get(new->cachep);
3909 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3910 new->new[smp_processor_id()] = old;
3913 /* Always called with the cache_chain_mutex held */
3914 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3915 int batchcount, int shared)
3917 struct ccupdate_struct *new;
3918 int i;
3920 new = kzalloc(sizeof(*new), GFP_KERNEL);
3921 if (!new)
3922 return -ENOMEM;
3924 for_each_online_cpu(i) {
3925 new->new[i] = alloc_arraycache(cpu_to_node(i), limit,
3926 batchcount);
3927 if (!new->new[i]) {
3928 for (i--; i >= 0; i--)
3929 kfree(new->new[i]);
3930 kfree(new);
3931 return -ENOMEM;
3934 new->cachep = cachep;
3936 on_each_cpu(do_ccupdate_local, (void *)new, 1, 1);
3938 check_irq_on();
3939 cachep->batchcount = batchcount;
3940 cachep->limit = limit;
3941 cachep->shared = shared;
3943 for_each_online_cpu(i) {
3944 struct array_cache *ccold = new->new[i];
3945 if (!ccold)
3946 continue;
3947 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3948 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
3949 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3950 kfree(ccold);
3952 kfree(new);
3953 return alloc_kmemlist(cachep);
3956 /* Called with cache_chain_mutex held always */
3957 static int enable_cpucache(struct kmem_cache *cachep)
3959 int err;
3960 int limit, shared;
3963 * The head array serves three purposes:
3964 * - create a LIFO ordering, i.e. return objects that are cache-warm
3965 * - reduce the number of spinlock operations.
3966 * - reduce the number of linked list operations on the slab and
3967 * bufctl chains: array operations are cheaper.
3968 * The numbers are guessed, we should auto-tune as described by
3969 * Bonwick.
3971 if (cachep->buffer_size > 131072)
3972 limit = 1;
3973 else if (cachep->buffer_size > PAGE_SIZE)
3974 limit = 8;
3975 else if (cachep->buffer_size > 1024)
3976 limit = 24;
3977 else if (cachep->buffer_size > 256)
3978 limit = 54;
3979 else
3980 limit = 120;
3983 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3984 * allocation behaviour: Most allocs on one cpu, most free operations
3985 * on another cpu. For these cases, an efficient object passing between
3986 * cpus is necessary. This is provided by a shared array. The array
3987 * replaces Bonwick's magazine layer.
3988 * On uniprocessor, it's functionally equivalent (but less efficient)
3989 * to a larger limit. Thus disabled by default.
3991 shared = 0;
3992 if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1)
3993 shared = 8;
3995 #if DEBUG
3997 * With debugging enabled, large batchcount lead to excessively long
3998 * periods with disabled local interrupts. Limit the batchcount
4000 if (limit > 32)
4001 limit = 32;
4002 #endif
4003 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
4004 if (err)
4005 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
4006 cachep->name, -err);
4007 return err;
4011 * Drain an array if it contains any elements taking the l3 lock only if
4012 * necessary. Note that the l3 listlock also protects the array_cache
4013 * if drain_array() is used on the shared array.
4015 void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
4016 struct array_cache *ac, int force, int node)
4018 int tofree;
4020 if (!ac || !ac->avail)
4021 return;
4022 if (ac->touched && !force) {
4023 ac->touched = 0;
4024 } else {
4025 spin_lock_irq(&l3->list_lock);
4026 if (ac->avail) {
4027 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4028 if (tofree > ac->avail)
4029 tofree = (ac->avail + 1) / 2;
4030 free_block(cachep, ac->entry, tofree, node);
4031 ac->avail -= tofree;
4032 memmove(ac->entry, &(ac->entry[tofree]),
4033 sizeof(void *) * ac->avail);
4035 spin_unlock_irq(&l3->list_lock);
4040 * cache_reap - Reclaim memory from caches.
4041 * @w: work descriptor
4043 * Called from workqueue/eventd every few seconds.
4044 * Purpose:
4045 * - clear the per-cpu caches for this CPU.
4046 * - return freeable pages to the main free memory pool.
4048 * If we cannot acquire the cache chain mutex then just give up - we'll try
4049 * again on the next iteration.
4051 static void cache_reap(struct work_struct *w)
4053 struct kmem_cache *searchp;
4054 struct kmem_list3 *l3;
4055 int node = numa_node_id();
4056 struct delayed_work *work =
4057 container_of(w, struct delayed_work, work);
4059 if (!mutex_trylock(&cache_chain_mutex))
4060 /* Give up. Setup the next iteration. */
4061 goto out;
4063 list_for_each_entry(searchp, &cache_chain, next) {
4064 check_irq_on();
4067 * We only take the l3 lock if absolutely necessary and we
4068 * have established with reasonable certainty that
4069 * we can do some work if the lock was obtained.
4071 l3 = searchp->nodelists[node];
4073 reap_alien(searchp, l3);
4075 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
4078 * These are racy checks but it does not matter
4079 * if we skip one check or scan twice.
4081 if (time_after(l3->next_reap, jiffies))
4082 goto next;
4084 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
4086 drain_array(searchp, l3, l3->shared, 0, node);
4088 if (l3->free_touched)
4089 l3->free_touched = 0;
4090 else {
4091 int freed;
4093 freed = drain_freelist(searchp, l3, (l3->free_limit +
4094 5 * searchp->num - 1) / (5 * searchp->num));
4095 STATS_ADD_REAPED(searchp, freed);
4097 next:
4098 cond_resched();
4100 check_irq_on();
4101 mutex_unlock(&cache_chain_mutex);
4102 next_reap_node();
4103 out:
4104 /* Set up the next iteration */
4105 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
4108 #ifdef CONFIG_PROC_FS
4110 static void print_slabinfo_header(struct seq_file *m)
4113 * Output format version, so at least we can change it
4114 * without _too_ many complaints.
4116 #if STATS
4117 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4118 #else
4119 seq_puts(m, "slabinfo - version: 2.1\n");
4120 #endif
4121 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4122 "<objperslab> <pagesperslab>");
4123 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4124 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4125 #if STATS
4126 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
4127 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
4128 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4129 #endif
4130 seq_putc(m, '\n');
4133 static void *s_start(struct seq_file *m, loff_t *pos)
4135 loff_t n = *pos;
4137 mutex_lock(&cache_chain_mutex);
4138 if (!n)
4139 print_slabinfo_header(m);
4141 return seq_list_start(&cache_chain, *pos);
4144 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4146 return seq_list_next(p, &cache_chain, pos);
4149 static void s_stop(struct seq_file *m, void *p)
4151 mutex_unlock(&cache_chain_mutex);
4154 static int s_show(struct seq_file *m, void *p)
4156 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
4157 struct slab *slabp;
4158 unsigned long active_objs;
4159 unsigned long num_objs;
4160 unsigned long active_slabs = 0;
4161 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4162 const char *name;
4163 char *error = NULL;
4164 int node;
4165 struct kmem_list3 *l3;
4167 active_objs = 0;
4168 num_slabs = 0;
4169 for_each_online_node(node) {
4170 l3 = cachep->nodelists[node];
4171 if (!l3)
4172 continue;
4174 check_irq_on();
4175 spin_lock_irq(&l3->list_lock);
4177 list_for_each_entry(slabp, &l3->slabs_full, list) {
4178 if (slabp->inuse != cachep->num && !error)
4179 error = "slabs_full accounting error";
4180 active_objs += cachep->num;
4181 active_slabs++;
4183 list_for_each_entry(slabp, &l3->slabs_partial, list) {
4184 if (slabp->inuse == cachep->num && !error)
4185 error = "slabs_partial inuse accounting error";
4186 if (!slabp->inuse && !error)
4187 error = "slabs_partial/inuse accounting error";
4188 active_objs += slabp->inuse;
4189 active_slabs++;
4191 list_for_each_entry(slabp, &l3->slabs_free, list) {
4192 if (slabp->inuse && !error)
4193 error = "slabs_free/inuse accounting error";
4194 num_slabs++;
4196 free_objects += l3->free_objects;
4197 if (l3->shared)
4198 shared_avail += l3->shared->avail;
4200 spin_unlock_irq(&l3->list_lock);
4202 num_slabs += active_slabs;
4203 num_objs = num_slabs * cachep->num;
4204 if (num_objs - active_objs != free_objects && !error)
4205 error = "free_objects accounting error";
4207 name = cachep->name;
4208 if (error)
4209 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4211 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
4212 name, active_objs, num_objs, cachep->buffer_size,
4213 cachep->num, (1 << cachep->gfporder));
4214 seq_printf(m, " : tunables %4u %4u %4u",
4215 cachep->limit, cachep->batchcount, cachep->shared);
4216 seq_printf(m, " : slabdata %6lu %6lu %6lu",
4217 active_slabs, num_slabs, shared_avail);
4218 #if STATS
4219 { /* list3 stats */
4220 unsigned long high = cachep->high_mark;
4221 unsigned long allocs = cachep->num_allocations;
4222 unsigned long grown = cachep->grown;
4223 unsigned long reaped = cachep->reaped;
4224 unsigned long errors = cachep->errors;
4225 unsigned long max_freeable = cachep->max_freeable;
4226 unsigned long node_allocs = cachep->node_allocs;
4227 unsigned long node_frees = cachep->node_frees;
4228 unsigned long overflows = cachep->node_overflow;
4230 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
4231 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown,
4232 reaped, errors, max_freeable, node_allocs,
4233 node_frees, overflows);
4235 /* cpu stats */
4237 unsigned long allochit = atomic_read(&cachep->allochit);
4238 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4239 unsigned long freehit = atomic_read(&cachep->freehit);
4240 unsigned long freemiss = atomic_read(&cachep->freemiss);
4242 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4243 allochit, allocmiss, freehit, freemiss);
4245 #endif
4246 seq_putc(m, '\n');
4247 return 0;
4251 * slabinfo_op - iterator that generates /proc/slabinfo
4253 * Output layout:
4254 * cache-name
4255 * num-active-objs
4256 * total-objs
4257 * object size
4258 * num-active-slabs
4259 * total-slabs
4260 * num-pages-per-slab
4261 * + further values on SMP and with statistics enabled
4264 const struct seq_operations slabinfo_op = {
4265 .start = s_start,
4266 .next = s_next,
4267 .stop = s_stop,
4268 .show = s_show,
4271 #define MAX_SLABINFO_WRITE 128
4273 * slabinfo_write - Tuning for the slab allocator
4274 * @file: unused
4275 * @buffer: user buffer
4276 * @count: data length
4277 * @ppos: unused
4279 ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4280 size_t count, loff_t *ppos)
4282 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4283 int limit, batchcount, shared, res;
4284 struct kmem_cache *cachep;
4286 if (count > MAX_SLABINFO_WRITE)
4287 return -EINVAL;
4288 if (copy_from_user(&kbuf, buffer, count))
4289 return -EFAULT;
4290 kbuf[MAX_SLABINFO_WRITE] = '\0';
4292 tmp = strchr(kbuf, ' ');
4293 if (!tmp)
4294 return -EINVAL;
4295 *tmp = '\0';
4296 tmp++;
4297 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4298 return -EINVAL;
4300 /* Find the cache in the chain of caches. */
4301 mutex_lock(&cache_chain_mutex);
4302 res = -EINVAL;
4303 list_for_each_entry(cachep, &cache_chain, next) {
4304 if (!strcmp(cachep->name, kbuf)) {
4305 if (limit < 1 || batchcount < 1 ||
4306 batchcount > limit || shared < 0) {
4307 res = 0;
4308 } else {
4309 res = do_tune_cpucache(cachep, limit,
4310 batchcount, shared);
4312 break;
4315 mutex_unlock(&cache_chain_mutex);
4316 if (res >= 0)
4317 res = count;
4318 return res;
4321 #ifdef CONFIG_DEBUG_SLAB_LEAK
4323 static void *leaks_start(struct seq_file *m, loff_t *pos)
4325 mutex_lock(&cache_chain_mutex);
4326 return seq_list_start(&cache_chain, *pos);
4329 static inline int add_caller(unsigned long *n, unsigned long v)
4331 unsigned long *p;
4332 int l;
4333 if (!v)
4334 return 1;
4335 l = n[1];
4336 p = n + 2;
4337 while (l) {
4338 int i = l/2;
4339 unsigned long *q = p + 2 * i;
4340 if (*q == v) {
4341 q[1]++;
4342 return 1;
4344 if (*q > v) {
4345 l = i;
4346 } else {
4347 p = q + 2;
4348 l -= i + 1;
4351 if (++n[1] == n[0])
4352 return 0;
4353 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4354 p[0] = v;
4355 p[1] = 1;
4356 return 1;
4359 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4361 void *p;
4362 int i;
4363 if (n[0] == n[1])
4364 return;
4365 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4366 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4367 continue;
4368 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4369 return;
4373 static void show_symbol(struct seq_file *m, unsigned long address)
4375 #ifdef CONFIG_KALLSYMS
4376 unsigned long offset, size;
4377 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4379 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4380 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4381 if (modname[0])
4382 seq_printf(m, " [%s]", modname);
4383 return;
4385 #endif
4386 seq_printf(m, "%p", (void *)address);
4389 static int leaks_show(struct seq_file *m, void *p)
4391 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
4392 struct slab *slabp;
4393 struct kmem_list3 *l3;
4394 const char *name;
4395 unsigned long *n = m->private;
4396 int node;
4397 int i;
4399 if (!(cachep->flags & SLAB_STORE_USER))
4400 return 0;
4401 if (!(cachep->flags & SLAB_RED_ZONE))
4402 return 0;
4404 /* OK, we can do it */
4406 n[1] = 0;
4408 for_each_online_node(node) {
4409 l3 = cachep->nodelists[node];
4410 if (!l3)
4411 continue;
4413 check_irq_on();
4414 spin_lock_irq(&l3->list_lock);
4416 list_for_each_entry(slabp, &l3->slabs_full, list)
4417 handle_slab(n, cachep, slabp);
4418 list_for_each_entry(slabp, &l3->slabs_partial, list)
4419 handle_slab(n, cachep, slabp);
4420 spin_unlock_irq(&l3->list_lock);
4422 name = cachep->name;
4423 if (n[0] == n[1]) {
4424 /* Increase the buffer size */
4425 mutex_unlock(&cache_chain_mutex);
4426 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4427 if (!m->private) {
4428 /* Too bad, we are really out */
4429 m->private = n;
4430 mutex_lock(&cache_chain_mutex);
4431 return -ENOMEM;
4433 *(unsigned long *)m->private = n[0] * 2;
4434 kfree(n);
4435 mutex_lock(&cache_chain_mutex);
4436 /* Now make sure this entry will be retried */
4437 m->count = m->size;
4438 return 0;
4440 for (i = 0; i < n[1]; i++) {
4441 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4442 show_symbol(m, n[2*i+2]);
4443 seq_putc(m, '\n');
4446 return 0;
4449 const struct seq_operations slabstats_op = {
4450 .start = leaks_start,
4451 .next = s_next,
4452 .stop = s_stop,
4453 .show = leaks_show,
4455 #endif
4456 #endif
4459 * ksize - get the actual amount of memory allocated for a given object
4460 * @objp: Pointer to the object
4462 * kmalloc may internally round up allocations and return more memory
4463 * than requested. ksize() can be used to determine the actual amount of
4464 * memory allocated. The caller may use this additional memory, even though
4465 * a smaller amount of memory was initially specified with the kmalloc call.
4466 * The caller must guarantee that objp points to a valid object previously
4467 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4468 * must not be freed during the duration of the call.
4470 size_t ksize(const void *objp)
4472 BUG_ON(!objp);
4473 if (unlikely(objp == ZERO_SIZE_PTR))
4474 return 0;
4476 return obj_size(virt_to_cache(objp));
4478 EXPORT_SYMBOL(ksize);