Add a 00-INDEX file to Documentation/sysctl/
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / mm / slab.c
blob3ce9bc024d676c9db1f58a298d5aacc5f1d9c6fa
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 intializations 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 int __cpuinit cpuup_callback(struct notifier_block *nfb,
1160 unsigned long action, void *hcpu)
1162 long cpu = (long)hcpu;
1163 struct kmem_cache *cachep;
1164 struct kmem_list3 *l3 = NULL;
1165 int node = cpu_to_node(cpu);
1166 const int memsize = sizeof(struct kmem_list3);
1168 switch (action) {
1169 case CPU_LOCK_ACQUIRE:
1170 mutex_lock(&cache_chain_mutex);
1171 break;
1172 case CPU_UP_PREPARE:
1173 case CPU_UP_PREPARE_FROZEN:
1175 * We need to do this right in the beginning since
1176 * alloc_arraycache's are going to use this list.
1177 * kmalloc_node allows us to add the slab to the right
1178 * kmem_list3 and not this cpu's kmem_list3
1181 list_for_each_entry(cachep, &cache_chain, next) {
1183 * Set up the size64 kmemlist for cpu before we can
1184 * begin anything. Make sure some other cpu on this
1185 * node has not already allocated this
1187 if (!cachep->nodelists[node]) {
1188 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1189 if (!l3)
1190 goto bad;
1191 kmem_list3_init(l3);
1192 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1193 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1196 * The l3s don't come and go as CPUs come and
1197 * go. cache_chain_mutex is sufficient
1198 * protection here.
1200 cachep->nodelists[node] = l3;
1203 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1204 cachep->nodelists[node]->free_limit =
1205 (1 + nr_cpus_node(node)) *
1206 cachep->batchcount + cachep->num;
1207 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1211 * Now we can go ahead with allocating the shared arrays and
1212 * array caches
1214 list_for_each_entry(cachep, &cache_chain, next) {
1215 struct array_cache *nc;
1216 struct array_cache *shared = NULL;
1217 struct array_cache **alien = NULL;
1219 nc = alloc_arraycache(node, cachep->limit,
1220 cachep->batchcount);
1221 if (!nc)
1222 goto bad;
1223 if (cachep->shared) {
1224 shared = alloc_arraycache(node,
1225 cachep->shared * cachep->batchcount,
1226 0xbaadf00d);
1227 if (!shared)
1228 goto bad;
1230 if (use_alien_caches) {
1231 alien = alloc_alien_cache(node, cachep->limit);
1232 if (!alien)
1233 goto bad;
1235 cachep->array[cpu] = nc;
1236 l3 = cachep->nodelists[node];
1237 BUG_ON(!l3);
1239 spin_lock_irq(&l3->list_lock);
1240 if (!l3->shared) {
1242 * We are serialised from CPU_DEAD or
1243 * CPU_UP_CANCELLED by the cpucontrol lock
1245 l3->shared = shared;
1246 shared = NULL;
1248 #ifdef CONFIG_NUMA
1249 if (!l3->alien) {
1250 l3->alien = alien;
1251 alien = NULL;
1253 #endif
1254 spin_unlock_irq(&l3->list_lock);
1255 kfree(shared);
1256 free_alien_cache(alien);
1258 break;
1259 case CPU_ONLINE:
1260 case CPU_ONLINE_FROZEN:
1261 start_cpu_timer(cpu);
1262 break;
1263 #ifdef CONFIG_HOTPLUG_CPU
1264 case CPU_DOWN_PREPARE:
1265 case CPU_DOWN_PREPARE_FROZEN:
1267 * Shutdown cache reaper. Note that the cache_chain_mutex is
1268 * held so that if cache_reap() is invoked it cannot do
1269 * anything expensive but will only modify reap_work
1270 * and reschedule the timer.
1272 cancel_rearming_delayed_work(&per_cpu(reap_work, cpu));
1273 /* Now the cache_reaper is guaranteed to be not running. */
1274 per_cpu(reap_work, cpu).work.func = NULL;
1275 break;
1276 case CPU_DOWN_FAILED:
1277 case CPU_DOWN_FAILED_FROZEN:
1278 start_cpu_timer(cpu);
1279 break;
1280 case CPU_DEAD:
1281 case CPU_DEAD_FROZEN:
1283 * Even if all the cpus of a node are down, we don't free the
1284 * kmem_list3 of any cache. This to avoid a race between
1285 * cpu_down, and a kmalloc allocation from another cpu for
1286 * memory from the node of the cpu going down. The list3
1287 * structure is usually allocated from kmem_cache_create() and
1288 * gets destroyed at kmem_cache_destroy().
1290 /* fall thru */
1291 #endif
1292 case CPU_UP_CANCELED:
1293 case CPU_UP_CANCELED_FROZEN:
1294 list_for_each_entry(cachep, &cache_chain, next) {
1295 struct array_cache *nc;
1296 struct array_cache *shared;
1297 struct array_cache **alien;
1298 cpumask_t mask;
1300 mask = node_to_cpumask(node);
1301 /* cpu is dead; no one can alloc from it. */
1302 nc = cachep->array[cpu];
1303 cachep->array[cpu] = NULL;
1304 l3 = cachep->nodelists[node];
1306 if (!l3)
1307 goto free_array_cache;
1309 spin_lock_irq(&l3->list_lock);
1311 /* Free limit for this kmem_list3 */
1312 l3->free_limit -= cachep->batchcount;
1313 if (nc)
1314 free_block(cachep, nc->entry, nc->avail, node);
1316 if (!cpus_empty(mask)) {
1317 spin_unlock_irq(&l3->list_lock);
1318 goto free_array_cache;
1321 shared = l3->shared;
1322 if (shared) {
1323 free_block(cachep, shared->entry,
1324 shared->avail, node);
1325 l3->shared = NULL;
1328 alien = l3->alien;
1329 l3->alien = NULL;
1331 spin_unlock_irq(&l3->list_lock);
1333 kfree(shared);
1334 if (alien) {
1335 drain_alien_cache(cachep, alien);
1336 free_alien_cache(alien);
1338 free_array_cache:
1339 kfree(nc);
1342 * In the previous loop, all the objects were freed to
1343 * the respective cache's slabs, now we can go ahead and
1344 * shrink each nodelist to its limit.
1346 list_for_each_entry(cachep, &cache_chain, next) {
1347 l3 = cachep->nodelists[node];
1348 if (!l3)
1349 continue;
1350 drain_freelist(cachep, l3, l3->free_objects);
1352 break;
1353 case CPU_LOCK_RELEASE:
1354 mutex_unlock(&cache_chain_mutex);
1355 break;
1357 return NOTIFY_OK;
1358 bad:
1359 return NOTIFY_BAD;
1362 static struct notifier_block __cpuinitdata cpucache_notifier = {
1363 &cpuup_callback, NULL, 0
1367 * swap the static kmem_list3 with kmalloced memory
1369 static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1370 int nodeid)
1372 struct kmem_list3 *ptr;
1374 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1375 BUG_ON(!ptr);
1377 local_irq_disable();
1378 memcpy(ptr, list, sizeof(struct kmem_list3));
1380 * Do not assume that spinlocks can be initialized via memcpy:
1382 spin_lock_init(&ptr->list_lock);
1384 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1385 cachep->nodelists[nodeid] = ptr;
1386 local_irq_enable();
1390 * Initialisation. Called after the page allocator have been initialised and
1391 * before smp_init().
1393 void __init kmem_cache_init(void)
1395 size_t left_over;
1396 struct cache_sizes *sizes;
1397 struct cache_names *names;
1398 int i;
1399 int order;
1400 int node;
1402 if (num_possible_nodes() == 1) {
1403 use_alien_caches = 0;
1404 numa_platform = 0;
1407 for (i = 0; i < NUM_INIT_LISTS; i++) {
1408 kmem_list3_init(&initkmem_list3[i]);
1409 if (i < MAX_NUMNODES)
1410 cache_cache.nodelists[i] = NULL;
1414 * Fragmentation resistance on low memory - only use bigger
1415 * page orders on machines with more than 32MB of memory.
1417 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1418 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1420 /* Bootstrap is tricky, because several objects are allocated
1421 * from caches that do not exist yet:
1422 * 1) initialize the cache_cache cache: it contains the struct
1423 * kmem_cache structures of all caches, except cache_cache itself:
1424 * cache_cache is statically allocated.
1425 * Initially an __init data area is used for the head array and the
1426 * kmem_list3 structures, it's replaced with a kmalloc allocated
1427 * array at the end of the bootstrap.
1428 * 2) Create the first kmalloc cache.
1429 * The struct kmem_cache for the new cache is allocated normally.
1430 * An __init data area is used for the head array.
1431 * 3) Create the remaining kmalloc caches, with minimally sized
1432 * head arrays.
1433 * 4) Replace the __init data head arrays for cache_cache and the first
1434 * kmalloc cache with kmalloc allocated arrays.
1435 * 5) Replace the __init data for kmem_list3 for cache_cache and
1436 * the other cache's with kmalloc allocated memory.
1437 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1440 node = numa_node_id();
1442 /* 1) create the cache_cache */
1443 INIT_LIST_HEAD(&cache_chain);
1444 list_add(&cache_cache.next, &cache_chain);
1445 cache_cache.colour_off = cache_line_size();
1446 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
1447 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE];
1450 * struct kmem_cache size depends on nr_node_ids, which
1451 * can be less than MAX_NUMNODES.
1453 cache_cache.buffer_size = offsetof(struct kmem_cache, nodelists) +
1454 nr_node_ids * sizeof(struct kmem_list3 *);
1455 #if DEBUG
1456 cache_cache.obj_size = cache_cache.buffer_size;
1457 #endif
1458 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1459 cache_line_size());
1460 cache_cache.reciprocal_buffer_size =
1461 reciprocal_value(cache_cache.buffer_size);
1463 for (order = 0; order < MAX_ORDER; order++) {
1464 cache_estimate(order, cache_cache.buffer_size,
1465 cache_line_size(), 0, &left_over, &cache_cache.num);
1466 if (cache_cache.num)
1467 break;
1469 BUG_ON(!cache_cache.num);
1470 cache_cache.gfporder = order;
1471 cache_cache.colour = left_over / cache_cache.colour_off;
1472 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1473 sizeof(struct slab), cache_line_size());
1475 /* 2+3) create the kmalloc caches */
1476 sizes = malloc_sizes;
1477 names = cache_names;
1480 * Initialize the caches that provide memory for the array cache and the
1481 * kmem_list3 structures first. Without this, further allocations will
1482 * bug.
1485 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
1486 sizes[INDEX_AC].cs_size,
1487 ARCH_KMALLOC_MINALIGN,
1488 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1489 NULL);
1491 if (INDEX_AC != INDEX_L3) {
1492 sizes[INDEX_L3].cs_cachep =
1493 kmem_cache_create(names[INDEX_L3].name,
1494 sizes[INDEX_L3].cs_size,
1495 ARCH_KMALLOC_MINALIGN,
1496 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1497 NULL);
1500 slab_early_init = 0;
1502 while (sizes->cs_size != ULONG_MAX) {
1504 * For performance, all the general caches are L1 aligned.
1505 * This should be particularly beneficial on SMP boxes, as it
1506 * eliminates "false sharing".
1507 * Note for systems short on memory removing the alignment will
1508 * allow tighter packing of the smaller caches.
1510 if (!sizes->cs_cachep) {
1511 sizes->cs_cachep = kmem_cache_create(names->name,
1512 sizes->cs_size,
1513 ARCH_KMALLOC_MINALIGN,
1514 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1515 NULL);
1517 #ifdef CONFIG_ZONE_DMA
1518 sizes->cs_dmacachep = kmem_cache_create(
1519 names->name_dma,
1520 sizes->cs_size,
1521 ARCH_KMALLOC_MINALIGN,
1522 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1523 SLAB_PANIC,
1524 NULL);
1525 #endif
1526 sizes++;
1527 names++;
1529 /* 4) Replace the bootstrap head arrays */
1531 struct array_cache *ptr;
1533 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1535 local_irq_disable();
1536 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1537 memcpy(ptr, cpu_cache_get(&cache_cache),
1538 sizeof(struct arraycache_init));
1540 * Do not assume that spinlocks can be initialized via memcpy:
1542 spin_lock_init(&ptr->lock);
1544 cache_cache.array[smp_processor_id()] = ptr;
1545 local_irq_enable();
1547 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1549 local_irq_disable();
1550 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1551 != &initarray_generic.cache);
1552 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1553 sizeof(struct arraycache_init));
1555 * Do not assume that spinlocks can be initialized via memcpy:
1557 spin_lock_init(&ptr->lock);
1559 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1560 ptr;
1561 local_irq_enable();
1563 /* 5) Replace the bootstrap kmem_list3's */
1565 int nid;
1567 /* Replace the static kmem_list3 structures for the boot cpu */
1568 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE], node);
1570 for_each_node_state(nid, N_NORMAL_MEMORY) {
1571 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1572 &initkmem_list3[SIZE_AC + nid], nid);
1574 if (INDEX_AC != INDEX_L3) {
1575 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1576 &initkmem_list3[SIZE_L3 + nid], nid);
1581 /* 6) resize the head arrays to their final sizes */
1583 struct kmem_cache *cachep;
1584 mutex_lock(&cache_chain_mutex);
1585 list_for_each_entry(cachep, &cache_chain, next)
1586 if (enable_cpucache(cachep))
1587 BUG();
1588 mutex_unlock(&cache_chain_mutex);
1591 /* Annotate slab for lockdep -- annotate the malloc caches */
1592 init_lock_keys();
1595 /* Done! */
1596 g_cpucache_up = FULL;
1599 * Register a cpu startup notifier callback that initializes
1600 * cpu_cache_get for all new cpus
1602 register_cpu_notifier(&cpucache_notifier);
1605 * The reap timers are started later, with a module init call: That part
1606 * of the kernel is not yet operational.
1610 static int __init cpucache_init(void)
1612 int cpu;
1615 * Register the timers that return unneeded pages to the page allocator
1617 for_each_online_cpu(cpu)
1618 start_cpu_timer(cpu);
1619 return 0;
1621 __initcall(cpucache_init);
1624 * Interface to system's page allocator. No need to hold the cache-lock.
1626 * If we requested dmaable memory, we will get it. Even if we
1627 * did not request dmaable memory, we might get it, but that
1628 * would be relatively rare and ignorable.
1630 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1632 struct page *page;
1633 int nr_pages;
1634 int i;
1636 #ifndef CONFIG_MMU
1638 * Nommu uses slab's for process anonymous memory allocations, and thus
1639 * requires __GFP_COMP to properly refcount higher order allocations
1641 flags |= __GFP_COMP;
1642 #endif
1644 flags |= cachep->gfpflags;
1645 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1646 flags |= __GFP_RECLAIMABLE;
1648 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
1649 if (!page)
1650 return NULL;
1652 nr_pages = (1 << cachep->gfporder);
1653 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1654 add_zone_page_state(page_zone(page),
1655 NR_SLAB_RECLAIMABLE, nr_pages);
1656 else
1657 add_zone_page_state(page_zone(page),
1658 NR_SLAB_UNRECLAIMABLE, nr_pages);
1659 for (i = 0; i < nr_pages; i++)
1660 __SetPageSlab(page + i);
1661 return page_address(page);
1665 * Interface to system's page release.
1667 static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1669 unsigned long i = (1 << cachep->gfporder);
1670 struct page *page = virt_to_page(addr);
1671 const unsigned long nr_freed = i;
1673 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1674 sub_zone_page_state(page_zone(page),
1675 NR_SLAB_RECLAIMABLE, nr_freed);
1676 else
1677 sub_zone_page_state(page_zone(page),
1678 NR_SLAB_UNRECLAIMABLE, nr_freed);
1679 while (i--) {
1680 BUG_ON(!PageSlab(page));
1681 __ClearPageSlab(page);
1682 page++;
1684 if (current->reclaim_state)
1685 current->reclaim_state->reclaimed_slab += nr_freed;
1686 free_pages((unsigned long)addr, cachep->gfporder);
1689 static void kmem_rcu_free(struct rcu_head *head)
1691 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1692 struct kmem_cache *cachep = slab_rcu->cachep;
1694 kmem_freepages(cachep, slab_rcu->addr);
1695 if (OFF_SLAB(cachep))
1696 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1699 #if DEBUG
1701 #ifdef CONFIG_DEBUG_PAGEALLOC
1702 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1703 unsigned long caller)
1705 int size = obj_size(cachep);
1707 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1709 if (size < 5 * sizeof(unsigned long))
1710 return;
1712 *addr++ = 0x12345678;
1713 *addr++ = caller;
1714 *addr++ = smp_processor_id();
1715 size -= 3 * sizeof(unsigned long);
1717 unsigned long *sptr = &caller;
1718 unsigned long svalue;
1720 while (!kstack_end(sptr)) {
1721 svalue = *sptr++;
1722 if (kernel_text_address(svalue)) {
1723 *addr++ = svalue;
1724 size -= sizeof(unsigned long);
1725 if (size <= sizeof(unsigned long))
1726 break;
1731 *addr++ = 0x87654321;
1733 #endif
1735 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1737 int size = obj_size(cachep);
1738 addr = &((char *)addr)[obj_offset(cachep)];
1740 memset(addr, val, size);
1741 *(unsigned char *)(addr + size - 1) = POISON_END;
1744 static void dump_line(char *data, int offset, int limit)
1746 int i;
1747 unsigned char error = 0;
1748 int bad_count = 0;
1750 printk(KERN_ERR "%03x:", offset);
1751 for (i = 0; i < limit; i++) {
1752 if (data[offset + i] != POISON_FREE) {
1753 error = data[offset + i];
1754 bad_count++;
1756 printk(" %02x", (unsigned char)data[offset + i]);
1758 printk("\n");
1760 if (bad_count == 1) {
1761 error ^= POISON_FREE;
1762 if (!(error & (error - 1))) {
1763 printk(KERN_ERR "Single bit error detected. Probably "
1764 "bad RAM.\n");
1765 #ifdef CONFIG_X86
1766 printk(KERN_ERR "Run memtest86+ or a similar memory "
1767 "test tool.\n");
1768 #else
1769 printk(KERN_ERR "Run a memory test tool.\n");
1770 #endif
1774 #endif
1776 #if DEBUG
1778 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1780 int i, size;
1781 char *realobj;
1783 if (cachep->flags & SLAB_RED_ZONE) {
1784 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
1785 *dbg_redzone1(cachep, objp),
1786 *dbg_redzone2(cachep, objp));
1789 if (cachep->flags & SLAB_STORE_USER) {
1790 printk(KERN_ERR "Last user: [<%p>]",
1791 *dbg_userword(cachep, objp));
1792 print_symbol("(%s)",
1793 (unsigned long)*dbg_userword(cachep, objp));
1794 printk("\n");
1796 realobj = (char *)objp + obj_offset(cachep);
1797 size = obj_size(cachep);
1798 for (i = 0; i < size && lines; i += 16, lines--) {
1799 int limit;
1800 limit = 16;
1801 if (i + limit > size)
1802 limit = size - i;
1803 dump_line(realobj, i, limit);
1807 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1809 char *realobj;
1810 int size, i;
1811 int lines = 0;
1813 realobj = (char *)objp + obj_offset(cachep);
1814 size = obj_size(cachep);
1816 for (i = 0; i < size; i++) {
1817 char exp = POISON_FREE;
1818 if (i == size - 1)
1819 exp = POISON_END;
1820 if (realobj[i] != exp) {
1821 int limit;
1822 /* Mismatch ! */
1823 /* Print header */
1824 if (lines == 0) {
1825 printk(KERN_ERR
1826 "Slab corruption: %s start=%p, len=%d\n",
1827 cachep->name, realobj, size);
1828 print_objinfo(cachep, objp, 0);
1830 /* Hexdump the affected line */
1831 i = (i / 16) * 16;
1832 limit = 16;
1833 if (i + limit > size)
1834 limit = size - i;
1835 dump_line(realobj, i, limit);
1836 i += 16;
1837 lines++;
1838 /* Limit to 5 lines */
1839 if (lines > 5)
1840 break;
1843 if (lines != 0) {
1844 /* Print some data about the neighboring objects, if they
1845 * exist:
1847 struct slab *slabp = virt_to_slab(objp);
1848 unsigned int objnr;
1850 objnr = obj_to_index(cachep, slabp, objp);
1851 if (objnr) {
1852 objp = index_to_obj(cachep, slabp, objnr - 1);
1853 realobj = (char *)objp + obj_offset(cachep);
1854 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1855 realobj, size);
1856 print_objinfo(cachep, objp, 2);
1858 if (objnr + 1 < cachep->num) {
1859 objp = index_to_obj(cachep, slabp, objnr + 1);
1860 realobj = (char *)objp + obj_offset(cachep);
1861 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1862 realobj, size);
1863 print_objinfo(cachep, objp, 2);
1867 #endif
1869 #if DEBUG
1871 * slab_destroy_objs - destroy a slab and its objects
1872 * @cachep: cache pointer being destroyed
1873 * @slabp: slab pointer being destroyed
1875 * Call the registered destructor for each object in a slab that is being
1876 * destroyed.
1878 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
1880 int i;
1881 for (i = 0; i < cachep->num; i++) {
1882 void *objp = index_to_obj(cachep, slabp, i);
1884 if (cachep->flags & SLAB_POISON) {
1885 #ifdef CONFIG_DEBUG_PAGEALLOC
1886 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1887 OFF_SLAB(cachep))
1888 kernel_map_pages(virt_to_page(objp),
1889 cachep->buffer_size / PAGE_SIZE, 1);
1890 else
1891 check_poison_obj(cachep, objp);
1892 #else
1893 check_poison_obj(cachep, objp);
1894 #endif
1896 if (cachep->flags & SLAB_RED_ZONE) {
1897 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1898 slab_error(cachep, "start of a freed object "
1899 "was overwritten");
1900 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1901 slab_error(cachep, "end of a freed object "
1902 "was overwritten");
1906 #else
1907 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
1910 #endif
1913 * slab_destroy - destroy and release all objects in a slab
1914 * @cachep: cache pointer being destroyed
1915 * @slabp: slab pointer being destroyed
1917 * Destroy all the objs in a slab, and release the mem back to the system.
1918 * Before calling the slab must have been unlinked from the cache. The
1919 * cache-lock is not held/needed.
1921 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
1923 void *addr = slabp->s_mem - slabp->colouroff;
1925 slab_destroy_objs(cachep, slabp);
1926 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1927 struct slab_rcu *slab_rcu;
1929 slab_rcu = (struct slab_rcu *)slabp;
1930 slab_rcu->cachep = cachep;
1931 slab_rcu->addr = addr;
1932 call_rcu(&slab_rcu->head, kmem_rcu_free);
1933 } else {
1934 kmem_freepages(cachep, addr);
1935 if (OFF_SLAB(cachep))
1936 kmem_cache_free(cachep->slabp_cache, slabp);
1941 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1942 * size of kmem_list3.
1944 static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1946 int node;
1948 for_each_node_state(node, N_NORMAL_MEMORY) {
1949 cachep->nodelists[node] = &initkmem_list3[index + node];
1950 cachep->nodelists[node]->next_reap = jiffies +
1951 REAPTIMEOUT_LIST3 +
1952 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1956 static void __kmem_cache_destroy(struct kmem_cache *cachep)
1958 int i;
1959 struct kmem_list3 *l3;
1961 for_each_online_cpu(i)
1962 kfree(cachep->array[i]);
1964 /* NUMA: free the list3 structures */
1965 for_each_online_node(i) {
1966 l3 = cachep->nodelists[i];
1967 if (l3) {
1968 kfree(l3->shared);
1969 free_alien_cache(l3->alien);
1970 kfree(l3);
1973 kmem_cache_free(&cache_cache, cachep);
1978 * calculate_slab_order - calculate size (page order) of slabs
1979 * @cachep: pointer to the cache that is being created
1980 * @size: size of objects to be created in this cache.
1981 * @align: required alignment for the objects.
1982 * @flags: slab allocation flags
1984 * Also calculates the number of objects per slab.
1986 * This could be made much more intelligent. For now, try to avoid using
1987 * high order pages for slabs. When the gfp() functions are more friendly
1988 * towards high-order requests, this should be changed.
1990 static size_t calculate_slab_order(struct kmem_cache *cachep,
1991 size_t size, size_t align, unsigned long flags)
1993 unsigned long offslab_limit;
1994 size_t left_over = 0;
1995 int gfporder;
1997 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
1998 unsigned int num;
1999 size_t remainder;
2001 cache_estimate(gfporder, size, align, flags, &remainder, &num);
2002 if (!num)
2003 continue;
2005 if (flags & CFLGS_OFF_SLAB) {
2007 * Max number of objs-per-slab for caches which
2008 * use off-slab slabs. Needed to avoid a possible
2009 * looping condition in cache_grow().
2011 offslab_limit = size - sizeof(struct slab);
2012 offslab_limit /= sizeof(kmem_bufctl_t);
2014 if (num > offslab_limit)
2015 break;
2018 /* Found something acceptable - save it away */
2019 cachep->num = num;
2020 cachep->gfporder = gfporder;
2021 left_over = remainder;
2024 * A VFS-reclaimable slab tends to have most allocations
2025 * as GFP_NOFS and we really don't want to have to be allocating
2026 * higher-order pages when we are unable to shrink dcache.
2028 if (flags & SLAB_RECLAIM_ACCOUNT)
2029 break;
2032 * Large number of objects is good, but very large slabs are
2033 * currently bad for the gfp()s.
2035 if (gfporder >= slab_break_gfp_order)
2036 break;
2039 * Acceptable internal fragmentation?
2041 if (left_over * 8 <= (PAGE_SIZE << gfporder))
2042 break;
2044 return left_over;
2047 static int __init_refok setup_cpu_cache(struct kmem_cache *cachep)
2049 if (g_cpucache_up == FULL)
2050 return enable_cpucache(cachep);
2052 if (g_cpucache_up == NONE) {
2054 * Note: the first kmem_cache_create must create the cache
2055 * that's used by kmalloc(24), otherwise the creation of
2056 * further caches will BUG().
2058 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2061 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2062 * the first cache, then we need to set up all its list3s,
2063 * otherwise the creation of further caches will BUG().
2065 set_up_list3s(cachep, SIZE_AC);
2066 if (INDEX_AC == INDEX_L3)
2067 g_cpucache_up = PARTIAL_L3;
2068 else
2069 g_cpucache_up = PARTIAL_AC;
2070 } else {
2071 cachep->array[smp_processor_id()] =
2072 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
2074 if (g_cpucache_up == PARTIAL_AC) {
2075 set_up_list3s(cachep, SIZE_L3);
2076 g_cpucache_up = PARTIAL_L3;
2077 } else {
2078 int node;
2079 for_each_node_state(node, N_NORMAL_MEMORY) {
2080 cachep->nodelists[node] =
2081 kmalloc_node(sizeof(struct kmem_list3),
2082 GFP_KERNEL, node);
2083 BUG_ON(!cachep->nodelists[node]);
2084 kmem_list3_init(cachep->nodelists[node]);
2088 cachep->nodelists[numa_node_id()]->next_reap =
2089 jiffies + REAPTIMEOUT_LIST3 +
2090 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2092 cpu_cache_get(cachep)->avail = 0;
2093 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2094 cpu_cache_get(cachep)->batchcount = 1;
2095 cpu_cache_get(cachep)->touched = 0;
2096 cachep->batchcount = 1;
2097 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2098 return 0;
2102 * kmem_cache_create - Create a cache.
2103 * @name: A string which is used in /proc/slabinfo to identify this cache.
2104 * @size: The size of objects to be created in this cache.
2105 * @align: The required alignment for the objects.
2106 * @flags: SLAB flags
2107 * @ctor: A constructor for the objects.
2109 * Returns a ptr to the cache on success, NULL on failure.
2110 * Cannot be called within a int, but can be interrupted.
2111 * The @ctor is run when new pages are allocated by the cache.
2113 * @name must be valid until the cache is destroyed. This implies that
2114 * the module calling this has to destroy the cache before getting unloaded.
2116 * The flags are
2118 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2119 * to catch references to uninitialised memory.
2121 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2122 * for buffer overruns.
2124 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2125 * cacheline. This can be beneficial if you're counting cycles as closely
2126 * as davem.
2128 struct kmem_cache *
2129 kmem_cache_create (const char *name, size_t size, size_t align,
2130 unsigned long flags,
2131 void (*ctor)(struct kmem_cache *, void *))
2133 size_t left_over, slab_size, ralign;
2134 struct kmem_cache *cachep = NULL, *pc;
2137 * Sanity checks... these are all serious usage bugs.
2139 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
2140 size > KMALLOC_MAX_SIZE) {
2141 printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__,
2142 name);
2143 BUG();
2147 * We use cache_chain_mutex to ensure a consistent view of
2148 * cpu_online_map as well. Please see cpuup_callback
2150 mutex_lock(&cache_chain_mutex);
2152 list_for_each_entry(pc, &cache_chain, next) {
2153 char tmp;
2154 int res;
2157 * This happens when the module gets unloaded and doesn't
2158 * destroy its slab cache and no-one else reuses the vmalloc
2159 * area of the module. Print a warning.
2161 res = probe_kernel_address(pc->name, tmp);
2162 if (res) {
2163 printk(KERN_ERR
2164 "SLAB: cache with size %d has lost its name\n",
2165 pc->buffer_size);
2166 continue;
2169 if (!strcmp(pc->name, name)) {
2170 printk(KERN_ERR
2171 "kmem_cache_create: duplicate cache %s\n", name);
2172 dump_stack();
2173 goto oops;
2177 #if DEBUG
2178 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
2179 #if FORCED_DEBUG
2181 * Enable redzoning and last user accounting, except for caches with
2182 * large objects, if the increased size would increase the object size
2183 * above the next power of two: caches with object sizes just above a
2184 * power of two have a significant amount of internal fragmentation.
2186 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2187 2 * sizeof(unsigned long long)))
2188 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2189 if (!(flags & SLAB_DESTROY_BY_RCU))
2190 flags |= SLAB_POISON;
2191 #endif
2192 if (flags & SLAB_DESTROY_BY_RCU)
2193 BUG_ON(flags & SLAB_POISON);
2194 #endif
2196 * Always checks flags, a caller might be expecting debug support which
2197 * isn't available.
2199 BUG_ON(flags & ~CREATE_MASK);
2202 * Check that size is in terms of words. This is needed to avoid
2203 * unaligned accesses for some archs when redzoning is used, and makes
2204 * sure any on-slab bufctl's are also correctly aligned.
2206 if (size & (BYTES_PER_WORD - 1)) {
2207 size += (BYTES_PER_WORD - 1);
2208 size &= ~(BYTES_PER_WORD - 1);
2211 /* calculate the final buffer alignment: */
2213 /* 1) arch recommendation: can be overridden for debug */
2214 if (flags & SLAB_HWCACHE_ALIGN) {
2216 * Default alignment: as specified by the arch code. Except if
2217 * an object is really small, then squeeze multiple objects into
2218 * one cacheline.
2220 ralign = cache_line_size();
2221 while (size <= ralign / 2)
2222 ralign /= 2;
2223 } else {
2224 ralign = BYTES_PER_WORD;
2228 * Redzoning and user store require word alignment or possibly larger.
2229 * Note this will be overridden by architecture or caller mandated
2230 * alignment if either is greater than BYTES_PER_WORD.
2232 if (flags & SLAB_STORE_USER)
2233 ralign = BYTES_PER_WORD;
2235 if (flags & SLAB_RED_ZONE) {
2236 ralign = REDZONE_ALIGN;
2237 /* If redzoning, ensure that the second redzone is suitably
2238 * aligned, by adjusting the object size accordingly. */
2239 size += REDZONE_ALIGN - 1;
2240 size &= ~(REDZONE_ALIGN - 1);
2243 /* 2) arch mandated alignment */
2244 if (ralign < ARCH_SLAB_MINALIGN) {
2245 ralign = ARCH_SLAB_MINALIGN;
2247 /* 3) caller mandated alignment */
2248 if (ralign < align) {
2249 ralign = align;
2251 /* disable debug if necessary */
2252 if (ralign > __alignof__(unsigned long long))
2253 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2255 * 4) Store it.
2257 align = ralign;
2259 /* Get cache's description obj. */
2260 cachep = kmem_cache_zalloc(&cache_cache, GFP_KERNEL);
2261 if (!cachep)
2262 goto oops;
2264 #if DEBUG
2265 cachep->obj_size = size;
2268 * Both debugging options require word-alignment which is calculated
2269 * into align above.
2271 if (flags & SLAB_RED_ZONE) {
2272 /* add space for red zone words */
2273 cachep->obj_offset += sizeof(unsigned long long);
2274 size += 2 * sizeof(unsigned long long);
2276 if (flags & SLAB_STORE_USER) {
2277 /* user store requires one word storage behind the end of
2278 * the real object. But if the second red zone needs to be
2279 * aligned to 64 bits, we must allow that much space.
2281 if (flags & SLAB_RED_ZONE)
2282 size += REDZONE_ALIGN;
2283 else
2284 size += BYTES_PER_WORD;
2286 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2287 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
2288 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2289 cachep->obj_offset += PAGE_SIZE - size;
2290 size = PAGE_SIZE;
2292 #endif
2293 #endif
2296 * Determine if the slab management is 'on' or 'off' slab.
2297 * (bootstrapping cannot cope with offslab caches so don't do
2298 * it too early on.)
2300 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init)
2302 * Size is large, assume best to place the slab management obj
2303 * off-slab (should allow better packing of objs).
2305 flags |= CFLGS_OFF_SLAB;
2307 size = ALIGN(size, align);
2309 left_over = calculate_slab_order(cachep, size, align, flags);
2311 if (!cachep->num) {
2312 printk(KERN_ERR
2313 "kmem_cache_create: couldn't create cache %s.\n", name);
2314 kmem_cache_free(&cache_cache, cachep);
2315 cachep = NULL;
2316 goto oops;
2318 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2319 + sizeof(struct slab), align);
2322 * If the slab has been placed off-slab, and we have enough space then
2323 * move it on-slab. This is at the expense of any extra colouring.
2325 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2326 flags &= ~CFLGS_OFF_SLAB;
2327 left_over -= slab_size;
2330 if (flags & CFLGS_OFF_SLAB) {
2331 /* really off slab. No need for manual alignment */
2332 slab_size =
2333 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2336 cachep->colour_off = cache_line_size();
2337 /* Offset must be a multiple of the alignment. */
2338 if (cachep->colour_off < align)
2339 cachep->colour_off = align;
2340 cachep->colour = left_over / cachep->colour_off;
2341 cachep->slab_size = slab_size;
2342 cachep->flags = flags;
2343 cachep->gfpflags = 0;
2344 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2345 cachep->gfpflags |= GFP_DMA;
2346 cachep->buffer_size = size;
2347 cachep->reciprocal_buffer_size = reciprocal_value(size);
2349 if (flags & CFLGS_OFF_SLAB) {
2350 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2352 * This is a possibility for one of the malloc_sizes caches.
2353 * But since we go off slab only for object size greater than
2354 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2355 * this should not happen at all.
2356 * But leave a BUG_ON for some lucky dude.
2358 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
2360 cachep->ctor = ctor;
2361 cachep->name = name;
2363 if (setup_cpu_cache(cachep)) {
2364 __kmem_cache_destroy(cachep);
2365 cachep = NULL;
2366 goto oops;
2369 /* cache setup completed, link it into the list */
2370 list_add(&cachep->next, &cache_chain);
2371 oops:
2372 if (!cachep && (flags & SLAB_PANIC))
2373 panic("kmem_cache_create(): failed to create slab `%s'\n",
2374 name);
2375 mutex_unlock(&cache_chain_mutex);
2376 return cachep;
2378 EXPORT_SYMBOL(kmem_cache_create);
2380 #if DEBUG
2381 static void check_irq_off(void)
2383 BUG_ON(!irqs_disabled());
2386 static void check_irq_on(void)
2388 BUG_ON(irqs_disabled());
2391 static void check_spinlock_acquired(struct kmem_cache *cachep)
2393 #ifdef CONFIG_SMP
2394 check_irq_off();
2395 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
2396 #endif
2399 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2401 #ifdef CONFIG_SMP
2402 check_irq_off();
2403 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2404 #endif
2407 #else
2408 #define check_irq_off() do { } while(0)
2409 #define check_irq_on() do { } while(0)
2410 #define check_spinlock_acquired(x) do { } while(0)
2411 #define check_spinlock_acquired_node(x, y) do { } while(0)
2412 #endif
2414 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2415 struct array_cache *ac,
2416 int force, int node);
2418 static void do_drain(void *arg)
2420 struct kmem_cache *cachep = arg;
2421 struct array_cache *ac;
2422 int node = numa_node_id();
2424 check_irq_off();
2425 ac = cpu_cache_get(cachep);
2426 spin_lock(&cachep->nodelists[node]->list_lock);
2427 free_block(cachep, ac->entry, ac->avail, node);
2428 spin_unlock(&cachep->nodelists[node]->list_lock);
2429 ac->avail = 0;
2432 static void drain_cpu_caches(struct kmem_cache *cachep)
2434 struct kmem_list3 *l3;
2435 int node;
2437 on_each_cpu(do_drain, cachep, 1, 1);
2438 check_irq_on();
2439 for_each_online_node(node) {
2440 l3 = cachep->nodelists[node];
2441 if (l3 && l3->alien)
2442 drain_alien_cache(cachep, l3->alien);
2445 for_each_online_node(node) {
2446 l3 = cachep->nodelists[node];
2447 if (l3)
2448 drain_array(cachep, l3, l3->shared, 1, node);
2453 * Remove slabs from the list of free slabs.
2454 * Specify the number of slabs to drain in tofree.
2456 * Returns the actual number of slabs released.
2458 static int drain_freelist(struct kmem_cache *cache,
2459 struct kmem_list3 *l3, int tofree)
2461 struct list_head *p;
2462 int nr_freed;
2463 struct slab *slabp;
2465 nr_freed = 0;
2466 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
2468 spin_lock_irq(&l3->list_lock);
2469 p = l3->slabs_free.prev;
2470 if (p == &l3->slabs_free) {
2471 spin_unlock_irq(&l3->list_lock);
2472 goto out;
2475 slabp = list_entry(p, struct slab, list);
2476 #if DEBUG
2477 BUG_ON(slabp->inuse);
2478 #endif
2479 list_del(&slabp->list);
2481 * Safe to drop the lock. The slab is no longer linked
2482 * to the cache.
2484 l3->free_objects -= cache->num;
2485 spin_unlock_irq(&l3->list_lock);
2486 slab_destroy(cache, slabp);
2487 nr_freed++;
2489 out:
2490 return nr_freed;
2493 /* Called with cache_chain_mutex held to protect against cpu hotplug */
2494 static int __cache_shrink(struct kmem_cache *cachep)
2496 int ret = 0, i = 0;
2497 struct kmem_list3 *l3;
2499 drain_cpu_caches(cachep);
2501 check_irq_on();
2502 for_each_online_node(i) {
2503 l3 = cachep->nodelists[i];
2504 if (!l3)
2505 continue;
2507 drain_freelist(cachep, l3, l3->free_objects);
2509 ret += !list_empty(&l3->slabs_full) ||
2510 !list_empty(&l3->slabs_partial);
2512 return (ret ? 1 : 0);
2516 * kmem_cache_shrink - Shrink a cache.
2517 * @cachep: The cache to shrink.
2519 * Releases as many slabs as possible for a cache.
2520 * To help debugging, a zero exit status indicates all slabs were released.
2522 int kmem_cache_shrink(struct kmem_cache *cachep)
2524 int ret;
2525 BUG_ON(!cachep || in_interrupt());
2527 mutex_lock(&cache_chain_mutex);
2528 ret = __cache_shrink(cachep);
2529 mutex_unlock(&cache_chain_mutex);
2530 return ret;
2532 EXPORT_SYMBOL(kmem_cache_shrink);
2535 * kmem_cache_destroy - delete a cache
2536 * @cachep: the cache to destroy
2538 * Remove a &struct kmem_cache object from the slab cache.
2540 * It is expected this function will be called by a module when it is
2541 * unloaded. This will remove the cache completely, and avoid a duplicate
2542 * cache being allocated each time a module is loaded and unloaded, if the
2543 * module doesn't have persistent in-kernel storage across loads and unloads.
2545 * The cache must be empty before calling this function.
2547 * The caller must guarantee that noone will allocate memory from the cache
2548 * during the kmem_cache_destroy().
2550 void kmem_cache_destroy(struct kmem_cache *cachep)
2552 BUG_ON(!cachep || in_interrupt());
2554 /* Find the cache in the chain of caches. */
2555 mutex_lock(&cache_chain_mutex);
2557 * the chain is never empty, cache_cache is never destroyed
2559 list_del(&cachep->next);
2560 if (__cache_shrink(cachep)) {
2561 slab_error(cachep, "Can't free all objects");
2562 list_add(&cachep->next, &cache_chain);
2563 mutex_unlock(&cache_chain_mutex);
2564 return;
2567 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
2568 synchronize_rcu();
2570 __kmem_cache_destroy(cachep);
2571 mutex_unlock(&cache_chain_mutex);
2573 EXPORT_SYMBOL(kmem_cache_destroy);
2576 * Get the memory for a slab management obj.
2577 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2578 * always come from malloc_sizes caches. The slab descriptor cannot
2579 * come from the same cache which is getting created because,
2580 * when we are searching for an appropriate cache for these
2581 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2582 * If we are creating a malloc_sizes cache here it would not be visible to
2583 * kmem_find_general_cachep till the initialization is complete.
2584 * Hence we cannot have slabp_cache same as the original cache.
2586 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2587 int colour_off, gfp_t local_flags,
2588 int nodeid)
2590 struct slab *slabp;
2592 if (OFF_SLAB(cachep)) {
2593 /* Slab management obj is off-slab. */
2594 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2595 local_flags & ~GFP_THISNODE, nodeid);
2596 if (!slabp)
2597 return NULL;
2598 } else {
2599 slabp = objp + colour_off;
2600 colour_off += cachep->slab_size;
2602 slabp->inuse = 0;
2603 slabp->colouroff = colour_off;
2604 slabp->s_mem = objp + colour_off;
2605 slabp->nodeid = nodeid;
2606 return slabp;
2609 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2611 return (kmem_bufctl_t *) (slabp + 1);
2614 static void cache_init_objs(struct kmem_cache *cachep,
2615 struct slab *slabp)
2617 int i;
2619 for (i = 0; i < cachep->num; i++) {
2620 void *objp = index_to_obj(cachep, slabp, i);
2621 #if DEBUG
2622 /* need to poison the objs? */
2623 if (cachep->flags & SLAB_POISON)
2624 poison_obj(cachep, objp, POISON_FREE);
2625 if (cachep->flags & SLAB_STORE_USER)
2626 *dbg_userword(cachep, objp) = NULL;
2628 if (cachep->flags & SLAB_RED_ZONE) {
2629 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2630 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2633 * Constructors are not allowed to allocate memory from the same
2634 * cache which they are a constructor for. Otherwise, deadlock.
2635 * They must also be threaded.
2637 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2638 cachep->ctor(cachep, objp + obj_offset(cachep));
2640 if (cachep->flags & SLAB_RED_ZONE) {
2641 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2642 slab_error(cachep, "constructor overwrote the"
2643 " end of an object");
2644 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2645 slab_error(cachep, "constructor overwrote the"
2646 " start of an object");
2648 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2649 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2650 kernel_map_pages(virt_to_page(objp),
2651 cachep->buffer_size / PAGE_SIZE, 0);
2652 #else
2653 if (cachep->ctor)
2654 cachep->ctor(cachep, objp);
2655 #endif
2656 slab_bufctl(slabp)[i] = i + 1;
2658 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2659 slabp->free = 0;
2662 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2664 if (CONFIG_ZONE_DMA_FLAG) {
2665 if (flags & GFP_DMA)
2666 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2667 else
2668 BUG_ON(cachep->gfpflags & GFP_DMA);
2672 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2673 int nodeid)
2675 void *objp = index_to_obj(cachep, slabp, slabp->free);
2676 kmem_bufctl_t next;
2678 slabp->inuse++;
2679 next = slab_bufctl(slabp)[slabp->free];
2680 #if DEBUG
2681 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2682 WARN_ON(slabp->nodeid != nodeid);
2683 #endif
2684 slabp->free = next;
2686 return objp;
2689 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2690 void *objp, int nodeid)
2692 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2694 #if DEBUG
2695 /* Verify that the slab belongs to the intended node */
2696 WARN_ON(slabp->nodeid != nodeid);
2698 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2699 printk(KERN_ERR "slab: double free detected in cache "
2700 "'%s', objp %p\n", cachep->name, objp);
2701 BUG();
2703 #endif
2704 slab_bufctl(slabp)[objnr] = slabp->free;
2705 slabp->free = objnr;
2706 slabp->inuse--;
2710 * Map pages beginning at addr to the given cache and slab. This is required
2711 * for the slab allocator to be able to lookup the cache and slab of a
2712 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2714 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2715 void *addr)
2717 int nr_pages;
2718 struct page *page;
2720 page = virt_to_page(addr);
2722 nr_pages = 1;
2723 if (likely(!PageCompound(page)))
2724 nr_pages <<= cache->gfporder;
2726 do {
2727 page_set_cache(page, cache);
2728 page_set_slab(page, slab);
2729 page++;
2730 } while (--nr_pages);
2734 * Grow (by 1) the number of slabs within a cache. This is called by
2735 * kmem_cache_alloc() when there are no active objs left in a cache.
2737 static int cache_grow(struct kmem_cache *cachep,
2738 gfp_t flags, int nodeid, void *objp)
2740 struct slab *slabp;
2741 size_t offset;
2742 gfp_t local_flags;
2743 struct kmem_list3 *l3;
2746 * Be lazy and only check for valid flags here, keeping it out of the
2747 * critical path in kmem_cache_alloc().
2749 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2750 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2752 /* Take the l3 list lock to change the colour_next on this node */
2753 check_irq_off();
2754 l3 = cachep->nodelists[nodeid];
2755 spin_lock(&l3->list_lock);
2757 /* Get colour for the slab, and cal the next value. */
2758 offset = l3->colour_next;
2759 l3->colour_next++;
2760 if (l3->colour_next >= cachep->colour)
2761 l3->colour_next = 0;
2762 spin_unlock(&l3->list_lock);
2764 offset *= cachep->colour_off;
2766 if (local_flags & __GFP_WAIT)
2767 local_irq_enable();
2770 * The test for missing atomic flag is performed here, rather than
2771 * the more obvious place, simply to reduce the critical path length
2772 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2773 * will eventually be caught here (where it matters).
2775 kmem_flagcheck(cachep, flags);
2778 * Get mem for the objs. Attempt to allocate a physical page from
2779 * 'nodeid'.
2781 if (!objp)
2782 objp = kmem_getpages(cachep, local_flags, nodeid);
2783 if (!objp)
2784 goto failed;
2786 /* Get slab management. */
2787 slabp = alloc_slabmgmt(cachep, objp, offset,
2788 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2789 if (!slabp)
2790 goto opps1;
2792 slabp->nodeid = nodeid;
2793 slab_map_pages(cachep, slabp, objp);
2795 cache_init_objs(cachep, slabp);
2797 if (local_flags & __GFP_WAIT)
2798 local_irq_disable();
2799 check_irq_off();
2800 spin_lock(&l3->list_lock);
2802 /* Make slab active. */
2803 list_add_tail(&slabp->list, &(l3->slabs_free));
2804 STATS_INC_GROWN(cachep);
2805 l3->free_objects += cachep->num;
2806 spin_unlock(&l3->list_lock);
2807 return 1;
2808 opps1:
2809 kmem_freepages(cachep, objp);
2810 failed:
2811 if (local_flags & __GFP_WAIT)
2812 local_irq_disable();
2813 return 0;
2816 #if DEBUG
2819 * Perform extra freeing checks:
2820 * - detect bad pointers.
2821 * - POISON/RED_ZONE checking
2823 static void kfree_debugcheck(const void *objp)
2825 if (!virt_addr_valid(objp)) {
2826 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2827 (unsigned long)objp);
2828 BUG();
2832 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2834 unsigned long long redzone1, redzone2;
2836 redzone1 = *dbg_redzone1(cache, obj);
2837 redzone2 = *dbg_redzone2(cache, obj);
2840 * Redzone is ok.
2842 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2843 return;
2845 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2846 slab_error(cache, "double free detected");
2847 else
2848 slab_error(cache, "memory outside object was overwritten");
2850 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2851 obj, redzone1, redzone2);
2854 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2855 void *caller)
2857 struct page *page;
2858 unsigned int objnr;
2859 struct slab *slabp;
2861 objp -= obj_offset(cachep);
2862 kfree_debugcheck(objp);
2863 page = virt_to_head_page(objp);
2865 slabp = page_get_slab(page);
2867 if (cachep->flags & SLAB_RED_ZONE) {
2868 verify_redzone_free(cachep, objp);
2869 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2870 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2872 if (cachep->flags & SLAB_STORE_USER)
2873 *dbg_userword(cachep, objp) = caller;
2875 objnr = obj_to_index(cachep, slabp, objp);
2877 BUG_ON(objnr >= cachep->num);
2878 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
2880 #ifdef CONFIG_DEBUG_SLAB_LEAK
2881 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2882 #endif
2883 if (cachep->flags & SLAB_POISON) {
2884 #ifdef CONFIG_DEBUG_PAGEALLOC
2885 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
2886 store_stackinfo(cachep, objp, (unsigned long)caller);
2887 kernel_map_pages(virt_to_page(objp),
2888 cachep->buffer_size / PAGE_SIZE, 0);
2889 } else {
2890 poison_obj(cachep, objp, POISON_FREE);
2892 #else
2893 poison_obj(cachep, objp, POISON_FREE);
2894 #endif
2896 return objp;
2899 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
2901 kmem_bufctl_t i;
2902 int entries = 0;
2904 /* Check slab's freelist to see if this obj is there. */
2905 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2906 entries++;
2907 if (entries > cachep->num || i >= cachep->num)
2908 goto bad;
2910 if (entries != cachep->num - slabp->inuse) {
2911 bad:
2912 printk(KERN_ERR "slab: Internal list corruption detected in "
2913 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2914 cachep->name, cachep->num, slabp, slabp->inuse);
2915 for (i = 0;
2916 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
2917 i++) {
2918 if (i % 16 == 0)
2919 printk("\n%03x:", i);
2920 printk(" %02x", ((unsigned char *)slabp)[i]);
2922 printk("\n");
2923 BUG();
2926 #else
2927 #define kfree_debugcheck(x) do { } while(0)
2928 #define cache_free_debugcheck(x,objp,z) (objp)
2929 #define check_slabp(x,y) do { } while(0)
2930 #endif
2932 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
2934 int batchcount;
2935 struct kmem_list3 *l3;
2936 struct array_cache *ac;
2937 int node;
2939 node = numa_node_id();
2941 check_irq_off();
2942 ac = cpu_cache_get(cachep);
2943 retry:
2944 batchcount = ac->batchcount;
2945 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2947 * If there was little recent activity on this cache, then
2948 * perform only a partial refill. Otherwise we could generate
2949 * refill bouncing.
2951 batchcount = BATCHREFILL_LIMIT;
2953 l3 = cachep->nodelists[node];
2955 BUG_ON(ac->avail > 0 || !l3);
2956 spin_lock(&l3->list_lock);
2958 /* See if we can refill from the shared array */
2959 if (l3->shared && transfer_objects(ac, l3->shared, batchcount))
2960 goto alloc_done;
2962 while (batchcount > 0) {
2963 struct list_head *entry;
2964 struct slab *slabp;
2965 /* Get slab alloc is to come from. */
2966 entry = l3->slabs_partial.next;
2967 if (entry == &l3->slabs_partial) {
2968 l3->free_touched = 1;
2969 entry = l3->slabs_free.next;
2970 if (entry == &l3->slabs_free)
2971 goto must_grow;
2974 slabp = list_entry(entry, struct slab, list);
2975 check_slabp(cachep, slabp);
2976 check_spinlock_acquired(cachep);
2979 * The slab was either on partial or free list so
2980 * there must be at least one object available for
2981 * allocation.
2983 BUG_ON(slabp->inuse < 0 || slabp->inuse >= cachep->num);
2985 while (slabp->inuse < cachep->num && batchcount--) {
2986 STATS_INC_ALLOCED(cachep);
2987 STATS_INC_ACTIVE(cachep);
2988 STATS_SET_HIGH(cachep);
2990 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
2991 node);
2993 check_slabp(cachep, slabp);
2995 /* move slabp to correct slabp list: */
2996 list_del(&slabp->list);
2997 if (slabp->free == BUFCTL_END)
2998 list_add(&slabp->list, &l3->slabs_full);
2999 else
3000 list_add(&slabp->list, &l3->slabs_partial);
3003 must_grow:
3004 l3->free_objects -= ac->avail;
3005 alloc_done:
3006 spin_unlock(&l3->list_lock);
3008 if (unlikely(!ac->avail)) {
3009 int x;
3010 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
3012 /* cache_grow can reenable interrupts, then ac could change. */
3013 ac = cpu_cache_get(cachep);
3014 if (!x && ac->avail == 0) /* no objects in sight? abort */
3015 return NULL;
3017 if (!ac->avail) /* objects refilled by interrupt? */
3018 goto retry;
3020 ac->touched = 1;
3021 return ac->entry[--ac->avail];
3024 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3025 gfp_t flags)
3027 might_sleep_if(flags & __GFP_WAIT);
3028 #if DEBUG
3029 kmem_flagcheck(cachep, flags);
3030 #endif
3033 #if DEBUG
3034 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3035 gfp_t flags, void *objp, void *caller)
3037 if (!objp)
3038 return objp;
3039 if (cachep->flags & SLAB_POISON) {
3040 #ifdef CONFIG_DEBUG_PAGEALLOC
3041 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
3042 kernel_map_pages(virt_to_page(objp),
3043 cachep->buffer_size / PAGE_SIZE, 1);
3044 else
3045 check_poison_obj(cachep, objp);
3046 #else
3047 check_poison_obj(cachep, objp);
3048 #endif
3049 poison_obj(cachep, objp, POISON_INUSE);
3051 if (cachep->flags & SLAB_STORE_USER)
3052 *dbg_userword(cachep, objp) = caller;
3054 if (cachep->flags & SLAB_RED_ZONE) {
3055 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3056 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3057 slab_error(cachep, "double free, or memory outside"
3058 " object was overwritten");
3059 printk(KERN_ERR
3060 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3061 objp, *dbg_redzone1(cachep, objp),
3062 *dbg_redzone2(cachep, objp));
3064 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3065 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3067 #ifdef CONFIG_DEBUG_SLAB_LEAK
3069 struct slab *slabp;
3070 unsigned objnr;
3072 slabp = page_get_slab(virt_to_head_page(objp));
3073 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3074 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3076 #endif
3077 objp += obj_offset(cachep);
3078 if (cachep->ctor && cachep->flags & SLAB_POISON)
3079 cachep->ctor(cachep, objp);
3080 #if ARCH_SLAB_MINALIGN
3081 if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) {
3082 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3083 objp, ARCH_SLAB_MINALIGN);
3085 #endif
3086 return objp;
3088 #else
3089 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3090 #endif
3092 #ifdef CONFIG_FAILSLAB
3094 static struct failslab_attr {
3096 struct fault_attr attr;
3098 u32 ignore_gfp_wait;
3099 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3100 struct dentry *ignore_gfp_wait_file;
3101 #endif
3103 } failslab = {
3104 .attr = FAULT_ATTR_INITIALIZER,
3105 .ignore_gfp_wait = 1,
3108 static int __init setup_failslab(char *str)
3110 return setup_fault_attr(&failslab.attr, str);
3112 __setup("failslab=", setup_failslab);
3114 static int should_failslab(struct kmem_cache *cachep, gfp_t flags)
3116 if (cachep == &cache_cache)
3117 return 0;
3118 if (flags & __GFP_NOFAIL)
3119 return 0;
3120 if (failslab.ignore_gfp_wait && (flags & __GFP_WAIT))
3121 return 0;
3123 return should_fail(&failslab.attr, obj_size(cachep));
3126 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3128 static int __init failslab_debugfs(void)
3130 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
3131 struct dentry *dir;
3132 int err;
3134 err = init_fault_attr_dentries(&failslab.attr, "failslab");
3135 if (err)
3136 return err;
3137 dir = failslab.attr.dentries.dir;
3139 failslab.ignore_gfp_wait_file =
3140 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3141 &failslab.ignore_gfp_wait);
3143 if (!failslab.ignore_gfp_wait_file) {
3144 err = -ENOMEM;
3145 debugfs_remove(failslab.ignore_gfp_wait_file);
3146 cleanup_fault_attr_dentries(&failslab.attr);
3149 return err;
3152 late_initcall(failslab_debugfs);
3154 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3156 #else /* CONFIG_FAILSLAB */
3158 static inline int should_failslab(struct kmem_cache *cachep, gfp_t flags)
3160 return 0;
3163 #endif /* CONFIG_FAILSLAB */
3165 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3167 void *objp;
3168 struct array_cache *ac;
3170 check_irq_off();
3172 ac = cpu_cache_get(cachep);
3173 if (likely(ac->avail)) {
3174 STATS_INC_ALLOCHIT(cachep);
3175 ac->touched = 1;
3176 objp = ac->entry[--ac->avail];
3177 } else {
3178 STATS_INC_ALLOCMISS(cachep);
3179 objp = cache_alloc_refill(cachep, flags);
3181 return objp;
3184 #ifdef CONFIG_NUMA
3186 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3188 * If we are in_interrupt, then process context, including cpusets and
3189 * mempolicy, may not apply and should not be used for allocation policy.
3191 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3193 int nid_alloc, nid_here;
3195 if (in_interrupt() || (flags & __GFP_THISNODE))
3196 return NULL;
3197 nid_alloc = nid_here = numa_node_id();
3198 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3199 nid_alloc = cpuset_mem_spread_node();
3200 else if (current->mempolicy)
3201 nid_alloc = slab_node(current->mempolicy);
3202 if (nid_alloc != nid_here)
3203 return ____cache_alloc_node(cachep, flags, nid_alloc);
3204 return NULL;
3208 * Fallback function if there was no memory available and no objects on a
3209 * certain node and fall back is permitted. First we scan all the
3210 * available nodelists for available objects. If that fails then we
3211 * perform an allocation without specifying a node. This allows the page
3212 * allocator to do its reclaim / fallback magic. We then insert the
3213 * slab into the proper nodelist and then allocate from it.
3215 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3217 struct zonelist *zonelist;
3218 gfp_t local_flags;
3219 struct zone **z;
3220 void *obj = NULL;
3221 int nid;
3223 if (flags & __GFP_THISNODE)
3224 return NULL;
3226 zonelist = &NODE_DATA(slab_node(current->mempolicy))
3227 ->node_zonelists[gfp_zone(flags)];
3228 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3230 retry:
3232 * Look through allowed nodes for objects available
3233 * from existing per node queues.
3235 for (z = zonelist->zones; *z && !obj; z++) {
3236 nid = zone_to_nid(*z);
3238 if (cpuset_zone_allowed_hardwall(*z, flags) &&
3239 cache->nodelists[nid] &&
3240 cache->nodelists[nid]->free_objects)
3241 obj = ____cache_alloc_node(cache,
3242 flags | GFP_THISNODE, nid);
3245 if (!obj) {
3247 * This allocation will be performed within the constraints
3248 * of the current cpuset / memory policy requirements.
3249 * We may trigger various forms of reclaim on the allowed
3250 * set and go into memory reserves if necessary.
3252 if (local_flags & __GFP_WAIT)
3253 local_irq_enable();
3254 kmem_flagcheck(cache, flags);
3255 obj = kmem_getpages(cache, flags, -1);
3256 if (local_flags & __GFP_WAIT)
3257 local_irq_disable();
3258 if (obj) {
3260 * Insert into the appropriate per node queues
3262 nid = page_to_nid(virt_to_page(obj));
3263 if (cache_grow(cache, flags, nid, obj)) {
3264 obj = ____cache_alloc_node(cache,
3265 flags | GFP_THISNODE, nid);
3266 if (!obj)
3268 * Another processor may allocate the
3269 * objects in the slab since we are
3270 * not holding any locks.
3272 goto retry;
3273 } else {
3274 /* cache_grow already freed obj */
3275 obj = NULL;
3279 return obj;
3283 * A interface to enable slab creation on nodeid
3285 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3286 int nodeid)
3288 struct list_head *entry;
3289 struct slab *slabp;
3290 struct kmem_list3 *l3;
3291 void *obj;
3292 int x;
3294 l3 = cachep->nodelists[nodeid];
3295 BUG_ON(!l3);
3297 retry:
3298 check_irq_off();
3299 spin_lock(&l3->list_lock);
3300 entry = l3->slabs_partial.next;
3301 if (entry == &l3->slabs_partial) {
3302 l3->free_touched = 1;
3303 entry = l3->slabs_free.next;
3304 if (entry == &l3->slabs_free)
3305 goto must_grow;
3308 slabp = list_entry(entry, struct slab, list);
3309 check_spinlock_acquired_node(cachep, nodeid);
3310 check_slabp(cachep, slabp);
3312 STATS_INC_NODEALLOCS(cachep);
3313 STATS_INC_ACTIVE(cachep);
3314 STATS_SET_HIGH(cachep);
3316 BUG_ON(slabp->inuse == cachep->num);
3318 obj = slab_get_obj(cachep, slabp, nodeid);
3319 check_slabp(cachep, slabp);
3320 l3->free_objects--;
3321 /* move slabp to correct slabp list: */
3322 list_del(&slabp->list);
3324 if (slabp->free == BUFCTL_END)
3325 list_add(&slabp->list, &l3->slabs_full);
3326 else
3327 list_add(&slabp->list, &l3->slabs_partial);
3329 spin_unlock(&l3->list_lock);
3330 goto done;
3332 must_grow:
3333 spin_unlock(&l3->list_lock);
3334 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
3335 if (x)
3336 goto retry;
3338 return fallback_alloc(cachep, flags);
3340 done:
3341 return obj;
3345 * kmem_cache_alloc_node - Allocate an object on the specified node
3346 * @cachep: The cache to allocate from.
3347 * @flags: See kmalloc().
3348 * @nodeid: node number of the target node.
3349 * @caller: return address of caller, used for debug information
3351 * Identical to kmem_cache_alloc but it will allocate memory on the given
3352 * node, which can improve the performance for cpu bound structures.
3354 * Fallback to other node is possible if __GFP_THISNODE is not set.
3356 static __always_inline void *
3357 __cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3358 void *caller)
3360 unsigned long save_flags;
3361 void *ptr;
3363 if (should_failslab(cachep, flags))
3364 return NULL;
3366 cache_alloc_debugcheck_before(cachep, flags);
3367 local_irq_save(save_flags);
3369 if (unlikely(nodeid == -1))
3370 nodeid = numa_node_id();
3372 if (unlikely(!cachep->nodelists[nodeid])) {
3373 /* Node not bootstrapped yet */
3374 ptr = fallback_alloc(cachep, flags);
3375 goto out;
3378 if (nodeid == numa_node_id()) {
3380 * Use the locally cached objects if possible.
3381 * However ____cache_alloc does not allow fallback
3382 * to other nodes. It may fail while we still have
3383 * objects on other nodes available.
3385 ptr = ____cache_alloc(cachep, flags);
3386 if (ptr)
3387 goto out;
3389 /* ___cache_alloc_node can fall back to other nodes */
3390 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3391 out:
3392 local_irq_restore(save_flags);
3393 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3395 if (unlikely((flags & __GFP_ZERO) && ptr))
3396 memset(ptr, 0, obj_size(cachep));
3398 return ptr;
3401 static __always_inline void *
3402 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3404 void *objp;
3406 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3407 objp = alternate_node_alloc(cache, flags);
3408 if (objp)
3409 goto out;
3411 objp = ____cache_alloc(cache, flags);
3414 * We may just have run out of memory on the local node.
3415 * ____cache_alloc_node() knows how to locate memory on other nodes
3417 if (!objp)
3418 objp = ____cache_alloc_node(cache, flags, numa_node_id());
3420 out:
3421 return objp;
3423 #else
3425 static __always_inline void *
3426 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3428 return ____cache_alloc(cachep, flags);
3431 #endif /* CONFIG_NUMA */
3433 static __always_inline void *
3434 __cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3436 unsigned long save_flags;
3437 void *objp;
3439 if (should_failslab(cachep, flags))
3440 return NULL;
3442 cache_alloc_debugcheck_before(cachep, flags);
3443 local_irq_save(save_flags);
3444 objp = __do_cache_alloc(cachep, flags);
3445 local_irq_restore(save_flags);
3446 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3447 prefetchw(objp);
3449 if (unlikely((flags & __GFP_ZERO) && objp))
3450 memset(objp, 0, obj_size(cachep));
3452 return objp;
3456 * Caller needs to acquire correct kmem_list's list_lock
3458 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3459 int node)
3461 int i;
3462 struct kmem_list3 *l3;
3464 for (i = 0; i < nr_objects; i++) {
3465 void *objp = objpp[i];
3466 struct slab *slabp;
3468 slabp = virt_to_slab(objp);
3469 l3 = cachep->nodelists[node];
3470 list_del(&slabp->list);
3471 check_spinlock_acquired_node(cachep, node);
3472 check_slabp(cachep, slabp);
3473 slab_put_obj(cachep, slabp, objp, node);
3474 STATS_DEC_ACTIVE(cachep);
3475 l3->free_objects++;
3476 check_slabp(cachep, slabp);
3478 /* fixup slab chains */
3479 if (slabp->inuse == 0) {
3480 if (l3->free_objects > l3->free_limit) {
3481 l3->free_objects -= cachep->num;
3482 /* No need to drop any previously held
3483 * lock here, even if we have a off-slab slab
3484 * descriptor it is guaranteed to come from
3485 * a different cache, refer to comments before
3486 * alloc_slabmgmt.
3488 slab_destroy(cachep, slabp);
3489 } else {
3490 list_add(&slabp->list, &l3->slabs_free);
3492 } else {
3493 /* Unconditionally move a slab to the end of the
3494 * partial list on free - maximum time for the
3495 * other objects to be freed, too.
3497 list_add_tail(&slabp->list, &l3->slabs_partial);
3502 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3504 int batchcount;
3505 struct kmem_list3 *l3;
3506 int node = numa_node_id();
3508 batchcount = ac->batchcount;
3509 #if DEBUG
3510 BUG_ON(!batchcount || batchcount > ac->avail);
3511 #endif
3512 check_irq_off();
3513 l3 = cachep->nodelists[node];
3514 spin_lock(&l3->list_lock);
3515 if (l3->shared) {
3516 struct array_cache *shared_array = l3->shared;
3517 int max = shared_array->limit - shared_array->avail;
3518 if (max) {
3519 if (batchcount > max)
3520 batchcount = max;
3521 memcpy(&(shared_array->entry[shared_array->avail]),
3522 ac->entry, sizeof(void *) * batchcount);
3523 shared_array->avail += batchcount;
3524 goto free_done;
3528 free_block(cachep, ac->entry, batchcount, node);
3529 free_done:
3530 #if STATS
3532 int i = 0;
3533 struct list_head *p;
3535 p = l3->slabs_free.next;
3536 while (p != &(l3->slabs_free)) {
3537 struct slab *slabp;
3539 slabp = list_entry(p, struct slab, list);
3540 BUG_ON(slabp->inuse);
3542 i++;
3543 p = p->next;
3545 STATS_SET_FREEABLE(cachep, i);
3547 #endif
3548 spin_unlock(&l3->list_lock);
3549 ac->avail -= batchcount;
3550 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3554 * Release an obj back to its cache. If the obj has a constructed state, it must
3555 * be in this state _before_ it is released. Called with disabled ints.
3557 static inline void __cache_free(struct kmem_cache *cachep, void *objp)
3559 struct array_cache *ac = cpu_cache_get(cachep);
3561 check_irq_off();
3562 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3565 * Skip calling cache_free_alien() when the platform is not numa.
3566 * This will avoid cache misses that happen while accessing slabp (which
3567 * is per page memory reference) to get nodeid. Instead use a global
3568 * variable to skip the call, which is mostly likely to be present in
3569 * the cache.
3571 if (numa_platform && cache_free_alien(cachep, objp))
3572 return;
3574 if (likely(ac->avail < ac->limit)) {
3575 STATS_INC_FREEHIT(cachep);
3576 ac->entry[ac->avail++] = objp;
3577 return;
3578 } else {
3579 STATS_INC_FREEMISS(cachep);
3580 cache_flusharray(cachep, ac);
3581 ac->entry[ac->avail++] = objp;
3586 * kmem_cache_alloc - Allocate an object
3587 * @cachep: The cache to allocate from.
3588 * @flags: See kmalloc().
3590 * Allocate an object from this cache. The flags are only relevant
3591 * if the cache has no available objects.
3593 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3595 return __cache_alloc(cachep, flags, __builtin_return_address(0));
3597 EXPORT_SYMBOL(kmem_cache_alloc);
3600 * kmem_ptr_validate - check if an untrusted pointer might
3601 * be a slab entry.
3602 * @cachep: the cache we're checking against
3603 * @ptr: pointer to validate
3605 * This verifies that the untrusted pointer looks sane:
3606 * it is _not_ a guarantee that the pointer is actually
3607 * part of the slab cache in question, but it at least
3608 * validates that the pointer can be dereferenced and
3609 * looks half-way sane.
3611 * Currently only used for dentry validation.
3613 int kmem_ptr_validate(struct kmem_cache *cachep, const void *ptr)
3615 unsigned long addr = (unsigned long)ptr;
3616 unsigned long min_addr = PAGE_OFFSET;
3617 unsigned long align_mask = BYTES_PER_WORD - 1;
3618 unsigned long size = cachep->buffer_size;
3619 struct page *page;
3621 if (unlikely(addr < min_addr))
3622 goto out;
3623 if (unlikely(addr > (unsigned long)high_memory - size))
3624 goto out;
3625 if (unlikely(addr & align_mask))
3626 goto out;
3627 if (unlikely(!kern_addr_valid(addr)))
3628 goto out;
3629 if (unlikely(!kern_addr_valid(addr + size - 1)))
3630 goto out;
3631 page = virt_to_page(ptr);
3632 if (unlikely(!PageSlab(page)))
3633 goto out;
3634 if (unlikely(page_get_cache(page) != cachep))
3635 goto out;
3636 return 1;
3637 out:
3638 return 0;
3641 #ifdef CONFIG_NUMA
3642 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3644 return __cache_alloc_node(cachep, flags, nodeid,
3645 __builtin_return_address(0));
3647 EXPORT_SYMBOL(kmem_cache_alloc_node);
3649 static __always_inline void *
3650 __do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
3652 struct kmem_cache *cachep;
3654 cachep = kmem_find_general_cachep(size, flags);
3655 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3656 return cachep;
3657 return kmem_cache_alloc_node(cachep, flags, node);
3660 #ifdef CONFIG_DEBUG_SLAB
3661 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3663 return __do_kmalloc_node(size, flags, node,
3664 __builtin_return_address(0));
3666 EXPORT_SYMBOL(__kmalloc_node);
3668 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3669 int node, void *caller)
3671 return __do_kmalloc_node(size, flags, node, caller);
3673 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3674 #else
3675 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3677 return __do_kmalloc_node(size, flags, node, NULL);
3679 EXPORT_SYMBOL(__kmalloc_node);
3680 #endif /* CONFIG_DEBUG_SLAB */
3681 #endif /* CONFIG_NUMA */
3684 * __do_kmalloc - allocate memory
3685 * @size: how many bytes of memory are required.
3686 * @flags: the type of memory to allocate (see kmalloc).
3687 * @caller: function caller for debug tracking of the caller
3689 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3690 void *caller)
3692 struct kmem_cache *cachep;
3694 /* If you want to save a few bytes .text space: replace
3695 * __ with kmem_.
3696 * Then kmalloc uses the uninlined functions instead of the inline
3697 * functions.
3699 cachep = __find_general_cachep(size, flags);
3700 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3701 return cachep;
3702 return __cache_alloc(cachep, flags, caller);
3706 #ifdef CONFIG_DEBUG_SLAB
3707 void *__kmalloc(size_t size, gfp_t flags)
3709 return __do_kmalloc(size, flags, __builtin_return_address(0));
3711 EXPORT_SYMBOL(__kmalloc);
3713 void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller)
3715 return __do_kmalloc(size, flags, caller);
3717 EXPORT_SYMBOL(__kmalloc_track_caller);
3719 #else
3720 void *__kmalloc(size_t size, gfp_t flags)
3722 return __do_kmalloc(size, flags, NULL);
3724 EXPORT_SYMBOL(__kmalloc);
3725 #endif
3728 * kmem_cache_free - Deallocate an object
3729 * @cachep: The cache the allocation was from.
3730 * @objp: The previously allocated object.
3732 * Free an object which was previously allocated from this
3733 * cache.
3735 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3737 unsigned long flags;
3739 BUG_ON(virt_to_cache(objp) != cachep);
3741 local_irq_save(flags);
3742 debug_check_no_locks_freed(objp, obj_size(cachep));
3743 __cache_free(cachep, objp);
3744 local_irq_restore(flags);
3746 EXPORT_SYMBOL(kmem_cache_free);
3749 * kfree - free previously allocated memory
3750 * @objp: pointer returned by kmalloc.
3752 * If @objp is NULL, no operation is performed.
3754 * Don't free memory not originally allocated by kmalloc()
3755 * or you will run into trouble.
3757 void kfree(const void *objp)
3759 struct kmem_cache *c;
3760 unsigned long flags;
3762 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3763 return;
3764 local_irq_save(flags);
3765 kfree_debugcheck(objp);
3766 c = virt_to_cache(objp);
3767 debug_check_no_locks_freed(objp, obj_size(c));
3768 __cache_free(c, (void *)objp);
3769 local_irq_restore(flags);
3771 EXPORT_SYMBOL(kfree);
3773 unsigned int kmem_cache_size(struct kmem_cache *cachep)
3775 return obj_size(cachep);
3777 EXPORT_SYMBOL(kmem_cache_size);
3779 const char *kmem_cache_name(struct kmem_cache *cachep)
3781 return cachep->name;
3783 EXPORT_SYMBOL_GPL(kmem_cache_name);
3786 * This initializes kmem_list3 or resizes varioius caches for all nodes.
3788 static int alloc_kmemlist(struct kmem_cache *cachep)
3790 int node;
3791 struct kmem_list3 *l3;
3792 struct array_cache *new_shared;
3793 struct array_cache **new_alien = NULL;
3795 for_each_node_state(node, N_NORMAL_MEMORY) {
3797 if (use_alien_caches) {
3798 new_alien = alloc_alien_cache(node, cachep->limit);
3799 if (!new_alien)
3800 goto fail;
3803 new_shared = NULL;
3804 if (cachep->shared) {
3805 new_shared = alloc_arraycache(node,
3806 cachep->shared*cachep->batchcount,
3807 0xbaadf00d);
3808 if (!new_shared) {
3809 free_alien_cache(new_alien);
3810 goto fail;
3814 l3 = cachep->nodelists[node];
3815 if (l3) {
3816 struct array_cache *shared = l3->shared;
3818 spin_lock_irq(&l3->list_lock);
3820 if (shared)
3821 free_block(cachep, shared->entry,
3822 shared->avail, node);
3824 l3->shared = new_shared;
3825 if (!l3->alien) {
3826 l3->alien = new_alien;
3827 new_alien = NULL;
3829 l3->free_limit = (1 + nr_cpus_node(node)) *
3830 cachep->batchcount + cachep->num;
3831 spin_unlock_irq(&l3->list_lock);
3832 kfree(shared);
3833 free_alien_cache(new_alien);
3834 continue;
3836 l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node);
3837 if (!l3) {
3838 free_alien_cache(new_alien);
3839 kfree(new_shared);
3840 goto fail;
3843 kmem_list3_init(l3);
3844 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3845 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3846 l3->shared = new_shared;
3847 l3->alien = new_alien;
3848 l3->free_limit = (1 + nr_cpus_node(node)) *
3849 cachep->batchcount + cachep->num;
3850 cachep->nodelists[node] = l3;
3852 return 0;
3854 fail:
3855 if (!cachep->next.next) {
3856 /* Cache is not active yet. Roll back what we did */
3857 node--;
3858 while (node >= 0) {
3859 if (cachep->nodelists[node]) {
3860 l3 = cachep->nodelists[node];
3862 kfree(l3->shared);
3863 free_alien_cache(l3->alien);
3864 kfree(l3);
3865 cachep->nodelists[node] = NULL;
3867 node--;
3870 return -ENOMEM;
3873 struct ccupdate_struct {
3874 struct kmem_cache *cachep;
3875 struct array_cache *new[NR_CPUS];
3878 static void do_ccupdate_local(void *info)
3880 struct ccupdate_struct *new = info;
3881 struct array_cache *old;
3883 check_irq_off();
3884 old = cpu_cache_get(new->cachep);
3886 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3887 new->new[smp_processor_id()] = old;
3890 /* Always called with the cache_chain_mutex held */
3891 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3892 int batchcount, int shared)
3894 struct ccupdate_struct *new;
3895 int i;
3897 new = kzalloc(sizeof(*new), GFP_KERNEL);
3898 if (!new)
3899 return -ENOMEM;
3901 for_each_online_cpu(i) {
3902 new->new[i] = alloc_arraycache(cpu_to_node(i), limit,
3903 batchcount);
3904 if (!new->new[i]) {
3905 for (i--; i >= 0; i--)
3906 kfree(new->new[i]);
3907 kfree(new);
3908 return -ENOMEM;
3911 new->cachep = cachep;
3913 on_each_cpu(do_ccupdate_local, (void *)new, 1, 1);
3915 check_irq_on();
3916 cachep->batchcount = batchcount;
3917 cachep->limit = limit;
3918 cachep->shared = shared;
3920 for_each_online_cpu(i) {
3921 struct array_cache *ccold = new->new[i];
3922 if (!ccold)
3923 continue;
3924 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3925 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
3926 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3927 kfree(ccold);
3929 kfree(new);
3930 return alloc_kmemlist(cachep);
3933 /* Called with cache_chain_mutex held always */
3934 static int enable_cpucache(struct kmem_cache *cachep)
3936 int err;
3937 int limit, shared;
3940 * The head array serves three purposes:
3941 * - create a LIFO ordering, i.e. return objects that are cache-warm
3942 * - reduce the number of spinlock operations.
3943 * - reduce the number of linked list operations on the slab and
3944 * bufctl chains: array operations are cheaper.
3945 * The numbers are guessed, we should auto-tune as described by
3946 * Bonwick.
3948 if (cachep->buffer_size > 131072)
3949 limit = 1;
3950 else if (cachep->buffer_size > PAGE_SIZE)
3951 limit = 8;
3952 else if (cachep->buffer_size > 1024)
3953 limit = 24;
3954 else if (cachep->buffer_size > 256)
3955 limit = 54;
3956 else
3957 limit = 120;
3960 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3961 * allocation behaviour: Most allocs on one cpu, most free operations
3962 * on another cpu. For these cases, an efficient object passing between
3963 * cpus is necessary. This is provided by a shared array. The array
3964 * replaces Bonwick's magazine layer.
3965 * On uniprocessor, it's functionally equivalent (but less efficient)
3966 * to a larger limit. Thus disabled by default.
3968 shared = 0;
3969 if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1)
3970 shared = 8;
3972 #if DEBUG
3974 * With debugging enabled, large batchcount lead to excessively long
3975 * periods with disabled local interrupts. Limit the batchcount
3977 if (limit > 32)
3978 limit = 32;
3979 #endif
3980 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
3981 if (err)
3982 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
3983 cachep->name, -err);
3984 return err;
3988 * Drain an array if it contains any elements taking the l3 lock only if
3989 * necessary. Note that the l3 listlock also protects the array_cache
3990 * if drain_array() is used on the shared array.
3992 void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
3993 struct array_cache *ac, int force, int node)
3995 int tofree;
3997 if (!ac || !ac->avail)
3998 return;
3999 if (ac->touched && !force) {
4000 ac->touched = 0;
4001 } else {
4002 spin_lock_irq(&l3->list_lock);
4003 if (ac->avail) {
4004 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4005 if (tofree > ac->avail)
4006 tofree = (ac->avail + 1) / 2;
4007 free_block(cachep, ac->entry, tofree, node);
4008 ac->avail -= tofree;
4009 memmove(ac->entry, &(ac->entry[tofree]),
4010 sizeof(void *) * ac->avail);
4012 spin_unlock_irq(&l3->list_lock);
4017 * cache_reap - Reclaim memory from caches.
4018 * @w: work descriptor
4020 * Called from workqueue/eventd every few seconds.
4021 * Purpose:
4022 * - clear the per-cpu caches for this CPU.
4023 * - return freeable pages to the main free memory pool.
4025 * If we cannot acquire the cache chain mutex then just give up - we'll try
4026 * again on the next iteration.
4028 static void cache_reap(struct work_struct *w)
4030 struct kmem_cache *searchp;
4031 struct kmem_list3 *l3;
4032 int node = numa_node_id();
4033 struct delayed_work *work =
4034 container_of(w, struct delayed_work, work);
4036 if (!mutex_trylock(&cache_chain_mutex))
4037 /* Give up. Setup the next iteration. */
4038 goto out;
4040 list_for_each_entry(searchp, &cache_chain, next) {
4041 check_irq_on();
4044 * We only take the l3 lock if absolutely necessary and we
4045 * have established with reasonable certainty that
4046 * we can do some work if the lock was obtained.
4048 l3 = searchp->nodelists[node];
4050 reap_alien(searchp, l3);
4052 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
4055 * These are racy checks but it does not matter
4056 * if we skip one check or scan twice.
4058 if (time_after(l3->next_reap, jiffies))
4059 goto next;
4061 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
4063 drain_array(searchp, l3, l3->shared, 0, node);
4065 if (l3->free_touched)
4066 l3->free_touched = 0;
4067 else {
4068 int freed;
4070 freed = drain_freelist(searchp, l3, (l3->free_limit +
4071 5 * searchp->num - 1) / (5 * searchp->num));
4072 STATS_ADD_REAPED(searchp, freed);
4074 next:
4075 cond_resched();
4077 check_irq_on();
4078 mutex_unlock(&cache_chain_mutex);
4079 next_reap_node();
4080 out:
4081 /* Set up the next iteration */
4082 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
4085 #ifdef CONFIG_PROC_FS
4087 static void print_slabinfo_header(struct seq_file *m)
4090 * Output format version, so at least we can change it
4091 * without _too_ many complaints.
4093 #if STATS
4094 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4095 #else
4096 seq_puts(m, "slabinfo - version: 2.1\n");
4097 #endif
4098 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4099 "<objperslab> <pagesperslab>");
4100 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4101 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4102 #if STATS
4103 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
4104 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
4105 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4106 #endif
4107 seq_putc(m, '\n');
4110 static void *s_start(struct seq_file *m, loff_t *pos)
4112 loff_t n = *pos;
4114 mutex_lock(&cache_chain_mutex);
4115 if (!n)
4116 print_slabinfo_header(m);
4118 return seq_list_start(&cache_chain, *pos);
4121 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4123 return seq_list_next(p, &cache_chain, pos);
4126 static void s_stop(struct seq_file *m, void *p)
4128 mutex_unlock(&cache_chain_mutex);
4131 static int s_show(struct seq_file *m, void *p)
4133 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
4134 struct slab *slabp;
4135 unsigned long active_objs;
4136 unsigned long num_objs;
4137 unsigned long active_slabs = 0;
4138 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4139 const char *name;
4140 char *error = NULL;
4141 int node;
4142 struct kmem_list3 *l3;
4144 active_objs = 0;
4145 num_slabs = 0;
4146 for_each_online_node(node) {
4147 l3 = cachep->nodelists[node];
4148 if (!l3)
4149 continue;
4151 check_irq_on();
4152 spin_lock_irq(&l3->list_lock);
4154 list_for_each_entry(slabp, &l3->slabs_full, list) {
4155 if (slabp->inuse != cachep->num && !error)
4156 error = "slabs_full accounting error";
4157 active_objs += cachep->num;
4158 active_slabs++;
4160 list_for_each_entry(slabp, &l3->slabs_partial, list) {
4161 if (slabp->inuse == cachep->num && !error)
4162 error = "slabs_partial inuse accounting error";
4163 if (!slabp->inuse && !error)
4164 error = "slabs_partial/inuse accounting error";
4165 active_objs += slabp->inuse;
4166 active_slabs++;
4168 list_for_each_entry(slabp, &l3->slabs_free, list) {
4169 if (slabp->inuse && !error)
4170 error = "slabs_free/inuse accounting error";
4171 num_slabs++;
4173 free_objects += l3->free_objects;
4174 if (l3->shared)
4175 shared_avail += l3->shared->avail;
4177 spin_unlock_irq(&l3->list_lock);
4179 num_slabs += active_slabs;
4180 num_objs = num_slabs * cachep->num;
4181 if (num_objs - active_objs != free_objects && !error)
4182 error = "free_objects accounting error";
4184 name = cachep->name;
4185 if (error)
4186 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4188 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
4189 name, active_objs, num_objs, cachep->buffer_size,
4190 cachep->num, (1 << cachep->gfporder));
4191 seq_printf(m, " : tunables %4u %4u %4u",
4192 cachep->limit, cachep->batchcount, cachep->shared);
4193 seq_printf(m, " : slabdata %6lu %6lu %6lu",
4194 active_slabs, num_slabs, shared_avail);
4195 #if STATS
4196 { /* list3 stats */
4197 unsigned long high = cachep->high_mark;
4198 unsigned long allocs = cachep->num_allocations;
4199 unsigned long grown = cachep->grown;
4200 unsigned long reaped = cachep->reaped;
4201 unsigned long errors = cachep->errors;
4202 unsigned long max_freeable = cachep->max_freeable;
4203 unsigned long node_allocs = cachep->node_allocs;
4204 unsigned long node_frees = cachep->node_frees;
4205 unsigned long overflows = cachep->node_overflow;
4207 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
4208 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown,
4209 reaped, errors, max_freeable, node_allocs,
4210 node_frees, overflows);
4212 /* cpu stats */
4214 unsigned long allochit = atomic_read(&cachep->allochit);
4215 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4216 unsigned long freehit = atomic_read(&cachep->freehit);
4217 unsigned long freemiss = atomic_read(&cachep->freemiss);
4219 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4220 allochit, allocmiss, freehit, freemiss);
4222 #endif
4223 seq_putc(m, '\n');
4224 return 0;
4228 * slabinfo_op - iterator that generates /proc/slabinfo
4230 * Output layout:
4231 * cache-name
4232 * num-active-objs
4233 * total-objs
4234 * object size
4235 * num-active-slabs
4236 * total-slabs
4237 * num-pages-per-slab
4238 * + further values on SMP and with statistics enabled
4241 const struct seq_operations slabinfo_op = {
4242 .start = s_start,
4243 .next = s_next,
4244 .stop = s_stop,
4245 .show = s_show,
4248 #define MAX_SLABINFO_WRITE 128
4250 * slabinfo_write - Tuning for the slab allocator
4251 * @file: unused
4252 * @buffer: user buffer
4253 * @count: data length
4254 * @ppos: unused
4256 ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4257 size_t count, loff_t *ppos)
4259 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4260 int limit, batchcount, shared, res;
4261 struct kmem_cache *cachep;
4263 if (count > MAX_SLABINFO_WRITE)
4264 return -EINVAL;
4265 if (copy_from_user(&kbuf, buffer, count))
4266 return -EFAULT;
4267 kbuf[MAX_SLABINFO_WRITE] = '\0';
4269 tmp = strchr(kbuf, ' ');
4270 if (!tmp)
4271 return -EINVAL;
4272 *tmp = '\0';
4273 tmp++;
4274 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4275 return -EINVAL;
4277 /* Find the cache in the chain of caches. */
4278 mutex_lock(&cache_chain_mutex);
4279 res = -EINVAL;
4280 list_for_each_entry(cachep, &cache_chain, next) {
4281 if (!strcmp(cachep->name, kbuf)) {
4282 if (limit < 1 || batchcount < 1 ||
4283 batchcount > limit || shared < 0) {
4284 res = 0;
4285 } else {
4286 res = do_tune_cpucache(cachep, limit,
4287 batchcount, shared);
4289 break;
4292 mutex_unlock(&cache_chain_mutex);
4293 if (res >= 0)
4294 res = count;
4295 return res;
4298 #ifdef CONFIG_DEBUG_SLAB_LEAK
4300 static void *leaks_start(struct seq_file *m, loff_t *pos)
4302 mutex_lock(&cache_chain_mutex);
4303 return seq_list_start(&cache_chain, *pos);
4306 static inline int add_caller(unsigned long *n, unsigned long v)
4308 unsigned long *p;
4309 int l;
4310 if (!v)
4311 return 1;
4312 l = n[1];
4313 p = n + 2;
4314 while (l) {
4315 int i = l/2;
4316 unsigned long *q = p + 2 * i;
4317 if (*q == v) {
4318 q[1]++;
4319 return 1;
4321 if (*q > v) {
4322 l = i;
4323 } else {
4324 p = q + 2;
4325 l -= i + 1;
4328 if (++n[1] == n[0])
4329 return 0;
4330 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4331 p[0] = v;
4332 p[1] = 1;
4333 return 1;
4336 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4338 void *p;
4339 int i;
4340 if (n[0] == n[1])
4341 return;
4342 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4343 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4344 continue;
4345 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4346 return;
4350 static void show_symbol(struct seq_file *m, unsigned long address)
4352 #ifdef CONFIG_KALLSYMS
4353 unsigned long offset, size;
4354 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4356 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4357 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4358 if (modname[0])
4359 seq_printf(m, " [%s]", modname);
4360 return;
4362 #endif
4363 seq_printf(m, "%p", (void *)address);
4366 static int leaks_show(struct seq_file *m, void *p)
4368 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
4369 struct slab *slabp;
4370 struct kmem_list3 *l3;
4371 const char *name;
4372 unsigned long *n = m->private;
4373 int node;
4374 int i;
4376 if (!(cachep->flags & SLAB_STORE_USER))
4377 return 0;
4378 if (!(cachep->flags & SLAB_RED_ZONE))
4379 return 0;
4381 /* OK, we can do it */
4383 n[1] = 0;
4385 for_each_online_node(node) {
4386 l3 = cachep->nodelists[node];
4387 if (!l3)
4388 continue;
4390 check_irq_on();
4391 spin_lock_irq(&l3->list_lock);
4393 list_for_each_entry(slabp, &l3->slabs_full, list)
4394 handle_slab(n, cachep, slabp);
4395 list_for_each_entry(slabp, &l3->slabs_partial, list)
4396 handle_slab(n, cachep, slabp);
4397 spin_unlock_irq(&l3->list_lock);
4399 name = cachep->name;
4400 if (n[0] == n[1]) {
4401 /* Increase the buffer size */
4402 mutex_unlock(&cache_chain_mutex);
4403 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4404 if (!m->private) {
4405 /* Too bad, we are really out */
4406 m->private = n;
4407 mutex_lock(&cache_chain_mutex);
4408 return -ENOMEM;
4410 *(unsigned long *)m->private = n[0] * 2;
4411 kfree(n);
4412 mutex_lock(&cache_chain_mutex);
4413 /* Now make sure this entry will be retried */
4414 m->count = m->size;
4415 return 0;
4417 for (i = 0; i < n[1]; i++) {
4418 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4419 show_symbol(m, n[2*i+2]);
4420 seq_putc(m, '\n');
4423 return 0;
4426 const struct seq_operations slabstats_op = {
4427 .start = leaks_start,
4428 .next = s_next,
4429 .stop = s_stop,
4430 .show = leaks_show,
4432 #endif
4433 #endif
4436 * ksize - get the actual amount of memory allocated for a given object
4437 * @objp: Pointer to the object
4439 * kmalloc may internally round up allocations and return more memory
4440 * than requested. ksize() can be used to determine the actual amount of
4441 * memory allocated. The caller may use this additional memory, even though
4442 * a smaller amount of memory was initially specified with the kmalloc call.
4443 * The caller must guarantee that objp points to a valid object previously
4444 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4445 * must not be freed during the duration of the call.
4447 size_t ksize(const void *objp)
4449 BUG_ON(!objp);
4450 if (unlikely(objp == ZERO_SIZE_PTR))
4451 return 0;
4453 return obj_size(virt_to_cache(objp));