hvcs: remove driver_data direct access of struct device
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / mm / slab.c
blob18e3164de09a9ce3b5c69b9fd53b6533b39e9486
1 /*
2 * linux/mm/slab.c
3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
28 * This means, that your constructor is used only for newly allocated
29 * slabs and you must pass objects with the same initializations to
30 * kmem_cache_free.
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
38 * partial slabs
39 * empty slabs with no allocated objects
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
53 * The c_cpuarray may not be read with enabled local interrupts -
54 * it's changed with a smp_call_function().
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
58 * Several members in struct kmem_cache and struct slab never change, they
59 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
66 * his patch.
68 * Further notes from the original documentation:
70 * 11 April '97. Started multi-threading - markhe
71 * The global cache-chain is protected by the mutex 'cache_chain_mutex'.
72 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
76 * At present, each engine can be growing a cache. This should be blocked.
78 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
89 #include <linux/slab.h>
90 #include <linux/mm.h>
91 #include <linux/poison.h>
92 #include <linux/swap.h>
93 #include <linux/cache.h>
94 #include <linux/interrupt.h>
95 #include <linux/init.h>
96 #include <linux/compiler.h>
97 #include <linux/cpuset.h>
98 #include <linux/proc_fs.h>
99 #include <linux/seq_file.h>
100 #include <linux/notifier.h>
101 #include <linux/kallsyms.h>
102 #include <linux/cpu.h>
103 #include <linux/sysctl.h>
104 #include <linux/module.h>
105 #include <linux/kmemtrace.h>
106 #include <linux/rcupdate.h>
107 #include <linux/string.h>
108 #include <linux/uaccess.h>
109 #include <linux/nodemask.h>
110 #include <linux/kmemleak.h>
111 #include <linux/mempolicy.h>
112 #include <linux/mutex.h>
113 #include <linux/fault-inject.h>
114 #include <linux/rtmutex.h>
115 #include <linux/reciprocal_div.h>
116 #include <linux/debugobjects.h>
118 #include <asm/cacheflush.h>
119 #include <asm/tlbflush.h>
120 #include <asm/page.h>
123 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
124 * 0 for faster, smaller code (especially in the critical paths).
126 * STATS - 1 to collect stats for /proc/slabinfo.
127 * 0 for faster, smaller code (especially in the critical paths).
129 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
132 #ifdef CONFIG_DEBUG_SLAB
133 #define DEBUG 1
134 #define STATS 1
135 #define FORCED_DEBUG 1
136 #else
137 #define DEBUG 0
138 #define STATS 0
139 #define FORCED_DEBUG 0
140 #endif
142 /* Shouldn't this be in a header file somewhere? */
143 #define BYTES_PER_WORD sizeof(void *)
144 #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
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 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE)
183 #else
184 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
185 SLAB_CACHE_DMA | \
186 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
187 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
188 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE)
189 #endif
192 * kmem_bufctl_t:
194 * Bufctl's are used for linking objs within a slab
195 * linked offsets.
197 * This implementation relies on "struct page" for locating the cache &
198 * slab an object belongs to.
199 * This allows the bufctl structure to be small (one int), but limits
200 * the number of objects a slab (not a cache) can contain when off-slab
201 * bufctls are used. The limit is the size of the largest general cache
202 * that does not use off-slab slabs.
203 * For 32bit archs with 4 kB pages, is this 56.
204 * This is not serious, as it is only for large objects, when it is unwise
205 * to have too many per slab.
206 * Note: This limit can be raised by introducing a general cache whose size
207 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
210 typedef unsigned int kmem_bufctl_t;
211 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
212 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
213 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
214 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
217 * struct slab
219 * Manages the objs in a slab. Placed either at the beginning of mem allocated
220 * for a slab, or allocated from an general cache.
221 * Slabs are chained into three list: fully used, partial, fully free slabs.
223 struct slab {
224 struct list_head list;
225 unsigned long colouroff;
226 void *s_mem; /* including colour offset */
227 unsigned int inuse; /* num of objs active in slab */
228 kmem_bufctl_t free;
229 unsigned short nodeid;
233 * struct slab_rcu
235 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
236 * arrange for kmem_freepages to be called via RCU. This is useful if
237 * we need to approach a kernel structure obliquely, from its address
238 * obtained without the usual locking. We can lock the structure to
239 * stabilize it and check it's still at the given address, only if we
240 * can be sure that the memory has not been meanwhile reused for some
241 * other kind of object (which our subsystem's lock might corrupt).
243 * rcu_read_lock before reading the address, then rcu_read_unlock after
244 * taking the spinlock within the structure expected at that address.
246 * We assume struct slab_rcu can overlay struct slab when destroying.
248 struct slab_rcu {
249 struct rcu_head head;
250 struct kmem_cache *cachep;
251 void *addr;
255 * struct array_cache
257 * Purpose:
258 * - LIFO ordering, to hand out cache-warm objects from _alloc
259 * - reduce the number of linked list operations
260 * - reduce spinlock operations
262 * The limit is stored in the per-cpu structure to reduce the data cache
263 * footprint.
266 struct array_cache {
267 unsigned int avail;
268 unsigned int limit;
269 unsigned int batchcount;
270 unsigned int touched;
271 spinlock_t lock;
272 void *entry[]; /*
273 * Must have this definition in here for the proper
274 * alignment of array_cache. Also simplifies accessing
275 * the entries.
280 * bootstrap: The caches do not work without cpuarrays anymore, but the
281 * cpuarrays are allocated from the generic caches...
283 #define BOOT_CPUCACHE_ENTRIES 1
284 struct arraycache_init {
285 struct array_cache cache;
286 void *entries[BOOT_CPUCACHE_ENTRIES];
290 * The slab lists for all objects.
292 struct kmem_list3 {
293 struct list_head slabs_partial; /* partial list first, better asm code */
294 struct list_head slabs_full;
295 struct list_head slabs_free;
296 unsigned long free_objects;
297 unsigned int free_limit;
298 unsigned int colour_next; /* Per-node cache coloring */
299 spinlock_t list_lock;
300 struct array_cache *shared; /* shared per node */
301 struct array_cache **alien; /* on other nodes */
302 unsigned long next_reap; /* updated without locking */
303 int free_touched; /* updated without locking */
307 * The slab allocator is initialized with interrupts disabled. Therefore, make
308 * sure early boot allocations don't accidentally enable interrupts.
310 static gfp_t slab_gfp_mask __read_mostly = SLAB_GFP_BOOT_MASK;
313 * Need this for bootstrapping a per node allocator.
315 #define NUM_INIT_LISTS (3 * MAX_NUMNODES)
316 struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
317 #define CACHE_CACHE 0
318 #define SIZE_AC MAX_NUMNODES
319 #define SIZE_L3 (2 * MAX_NUMNODES)
321 static int drain_freelist(struct kmem_cache *cache,
322 struct kmem_list3 *l3, int tofree);
323 static void free_block(struct kmem_cache *cachep, void **objpp, int len,
324 int node);
325 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
326 static void cache_reap(struct work_struct *unused);
329 * This function must be completely optimized away if a constant is passed to
330 * it. Mostly the same as what is in linux/slab.h except it returns an index.
332 static __always_inline int index_of(const size_t size)
334 extern void __bad_size(void);
336 if (__builtin_constant_p(size)) {
337 int i = 0;
339 #define CACHE(x) \
340 if (size <=x) \
341 return i; \
342 else \
343 i++;
344 #include <linux/kmalloc_sizes.h>
345 #undef CACHE
346 __bad_size();
347 } else
348 __bad_size();
349 return 0;
352 static int slab_early_init = 1;
354 #define INDEX_AC index_of(sizeof(struct arraycache_init))
355 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
357 static void kmem_list3_init(struct kmem_list3 *parent)
359 INIT_LIST_HEAD(&parent->slabs_full);
360 INIT_LIST_HEAD(&parent->slabs_partial);
361 INIT_LIST_HEAD(&parent->slabs_free);
362 parent->shared = NULL;
363 parent->alien = NULL;
364 parent->colour_next = 0;
365 spin_lock_init(&parent->list_lock);
366 parent->free_objects = 0;
367 parent->free_touched = 0;
370 #define MAKE_LIST(cachep, listp, slab, nodeid) \
371 do { \
372 INIT_LIST_HEAD(listp); \
373 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
374 } while (0)
376 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
377 do { \
378 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
379 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
380 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
381 } while (0)
384 * struct kmem_cache
386 * manages a cache.
389 struct kmem_cache {
390 /* 1) per-cpu data, touched during every alloc/free */
391 struct array_cache *array[NR_CPUS];
392 /* 2) Cache tunables. Protected by cache_chain_mutex */
393 unsigned int batchcount;
394 unsigned int limit;
395 unsigned int shared;
397 unsigned int buffer_size;
398 u32 reciprocal_buffer_size;
399 /* 3) touched by every alloc & free from the backend */
401 unsigned int flags; /* constant flags */
402 unsigned int num; /* # of objs per slab */
404 /* 4) cache_grow/shrink */
405 /* order of pgs per slab (2^n) */
406 unsigned int gfporder;
408 /* force GFP flags, e.g. GFP_DMA */
409 gfp_t gfpflags;
411 size_t colour; /* cache colouring range */
412 unsigned int colour_off; /* colour offset */
413 struct kmem_cache *slabp_cache;
414 unsigned int slab_size;
415 unsigned int dflags; /* dynamic flags */
417 /* constructor func */
418 void (*ctor)(void *obj);
420 /* 5) cache creation/removal */
421 const char *name;
422 struct list_head next;
424 /* 6) statistics */
425 #if STATS
426 unsigned long num_active;
427 unsigned long num_allocations;
428 unsigned long high_mark;
429 unsigned long grown;
430 unsigned long reaped;
431 unsigned long errors;
432 unsigned long max_freeable;
433 unsigned long node_allocs;
434 unsigned long node_frees;
435 unsigned long node_overflow;
436 atomic_t allochit;
437 atomic_t allocmiss;
438 atomic_t freehit;
439 atomic_t freemiss;
440 #endif
441 #if DEBUG
443 * If debugging is enabled, then the allocator can add additional
444 * fields and/or padding to every object. buffer_size contains the total
445 * object size including these internal fields, the following two
446 * variables contain the offset to the user object and its size.
448 int obj_offset;
449 int obj_size;
450 #endif
452 * We put nodelists[] at the end of kmem_cache, because we want to size
453 * this array to nr_node_ids slots instead of MAX_NUMNODES
454 * (see kmem_cache_init())
455 * We still use [MAX_NUMNODES] and not [1] or [0] because cache_cache
456 * is statically defined, so we reserve the max number of nodes.
458 struct kmem_list3 *nodelists[MAX_NUMNODES];
460 * Do not add fields after nodelists[]
464 #define CFLGS_OFF_SLAB (0x80000000UL)
465 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
467 #define BATCHREFILL_LIMIT 16
469 * Optimization question: fewer reaps means less probability for unnessary
470 * cpucache drain/refill cycles.
472 * OTOH the cpuarrays can contain lots of objects,
473 * which could lock up otherwise freeable slabs.
475 #define REAPTIMEOUT_CPUC (2*HZ)
476 #define REAPTIMEOUT_LIST3 (4*HZ)
478 #if STATS
479 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
480 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
481 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
482 #define STATS_INC_GROWN(x) ((x)->grown++)
483 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
484 #define STATS_SET_HIGH(x) \
485 do { \
486 if ((x)->num_active > (x)->high_mark) \
487 (x)->high_mark = (x)->num_active; \
488 } while (0)
489 #define STATS_INC_ERR(x) ((x)->errors++)
490 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
491 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
492 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
493 #define STATS_SET_FREEABLE(x, i) \
494 do { \
495 if ((x)->max_freeable < i) \
496 (x)->max_freeable = i; \
497 } while (0)
498 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
499 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
500 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
501 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
502 #else
503 #define STATS_INC_ACTIVE(x) do { } while (0)
504 #define STATS_DEC_ACTIVE(x) do { } while (0)
505 #define STATS_INC_ALLOCED(x) do { } while (0)
506 #define STATS_INC_GROWN(x) do { } while (0)
507 #define STATS_ADD_REAPED(x,y) do { } while (0)
508 #define STATS_SET_HIGH(x) do { } while (0)
509 #define STATS_INC_ERR(x) do { } while (0)
510 #define STATS_INC_NODEALLOCS(x) do { } while (0)
511 #define STATS_INC_NODEFREES(x) do { } while (0)
512 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
513 #define STATS_SET_FREEABLE(x, i) do { } while (0)
514 #define STATS_INC_ALLOCHIT(x) do { } while (0)
515 #define STATS_INC_ALLOCMISS(x) do { } while (0)
516 #define STATS_INC_FREEHIT(x) do { } while (0)
517 #define STATS_INC_FREEMISS(x) do { } while (0)
518 #endif
520 #if DEBUG
523 * memory layout of objects:
524 * 0 : objp
525 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
526 * the end of an object is aligned with the end of the real
527 * allocation. Catches writes behind the end of the allocation.
528 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
529 * redzone word.
530 * cachep->obj_offset: The real object.
531 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
532 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
533 * [BYTES_PER_WORD long]
535 static int obj_offset(struct kmem_cache *cachep)
537 return cachep->obj_offset;
540 static int obj_size(struct kmem_cache *cachep)
542 return cachep->obj_size;
545 static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
547 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
548 return (unsigned long long*) (objp + obj_offset(cachep) -
549 sizeof(unsigned long long));
552 static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
554 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
555 if (cachep->flags & SLAB_STORE_USER)
556 return (unsigned long long *)(objp + cachep->buffer_size -
557 sizeof(unsigned long long) -
558 REDZONE_ALIGN);
559 return (unsigned long long *) (objp + cachep->buffer_size -
560 sizeof(unsigned long long));
563 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
565 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
566 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
569 #else
571 #define obj_offset(x) 0
572 #define obj_size(cachep) (cachep->buffer_size)
573 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
574 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
575 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
577 #endif
579 #ifdef CONFIG_KMEMTRACE
580 size_t slab_buffer_size(struct kmem_cache *cachep)
582 return cachep->buffer_size;
584 EXPORT_SYMBOL(slab_buffer_size);
585 #endif
588 * Do not go above this order unless 0 objects fit into the slab.
590 #define BREAK_GFP_ORDER_HI 1
591 #define BREAK_GFP_ORDER_LO 0
592 static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
595 * Functions for storing/retrieving the cachep and or slab from the page
596 * allocator. These are used to find the slab an obj belongs to. With kfree(),
597 * these are used to find the cache which an obj belongs to.
599 static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
601 page->lru.next = (struct list_head *)cache;
604 static inline struct kmem_cache *page_get_cache(struct page *page)
606 page = compound_head(page);
607 BUG_ON(!PageSlab(page));
608 return (struct kmem_cache *)page->lru.next;
611 static inline void page_set_slab(struct page *page, struct slab *slab)
613 page->lru.prev = (struct list_head *)slab;
616 static inline struct slab *page_get_slab(struct page *page)
618 BUG_ON(!PageSlab(page));
619 return (struct slab *)page->lru.prev;
622 static inline struct kmem_cache *virt_to_cache(const void *obj)
624 struct page *page = virt_to_head_page(obj);
625 return page_get_cache(page);
628 static inline struct slab *virt_to_slab(const void *obj)
630 struct page *page = virt_to_head_page(obj);
631 return page_get_slab(page);
634 static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
635 unsigned int idx)
637 return slab->s_mem + cache->buffer_size * idx;
641 * We want to avoid an expensive divide : (offset / cache->buffer_size)
642 * Using the fact that buffer_size is a constant for a particular cache,
643 * we can replace (offset / cache->buffer_size) by
644 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
646 static inline unsigned int obj_to_index(const struct kmem_cache *cache,
647 const struct slab *slab, void *obj)
649 u32 offset = (obj - slab->s_mem);
650 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
654 * These are the default caches for kmalloc. Custom caches can have other sizes.
656 struct cache_sizes malloc_sizes[] = {
657 #define CACHE(x) { .cs_size = (x) },
658 #include <linux/kmalloc_sizes.h>
659 CACHE(ULONG_MAX)
660 #undef CACHE
662 EXPORT_SYMBOL(malloc_sizes);
664 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
665 struct cache_names {
666 char *name;
667 char *name_dma;
670 static struct cache_names __initdata cache_names[] = {
671 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
672 #include <linux/kmalloc_sizes.h>
673 {NULL,}
674 #undef CACHE
677 static struct arraycache_init initarray_cache __initdata =
678 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
679 static struct arraycache_init initarray_generic =
680 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
682 /* internal cache of cache description objs */
683 static struct kmem_cache cache_cache = {
684 .batchcount = 1,
685 .limit = BOOT_CPUCACHE_ENTRIES,
686 .shared = 1,
687 .buffer_size = sizeof(struct kmem_cache),
688 .name = "kmem_cache",
691 #define BAD_ALIEN_MAGIC 0x01020304ul
693 #ifdef CONFIG_LOCKDEP
696 * Slab sometimes uses the kmalloc slabs to store the slab headers
697 * for other slabs "off slab".
698 * The locking for this is tricky in that it nests within the locks
699 * of all other slabs in a few places; to deal with this special
700 * locking we put on-slab caches into a separate lock-class.
702 * We set lock class for alien array caches which are up during init.
703 * The lock annotation will be lost if all cpus of a node goes down and
704 * then comes back up during hotplug
706 static struct lock_class_key on_slab_l3_key;
707 static struct lock_class_key on_slab_alc_key;
709 static inline void init_lock_keys(void)
712 int q;
713 struct cache_sizes *s = malloc_sizes;
715 while (s->cs_size != ULONG_MAX) {
716 for_each_node(q) {
717 struct array_cache **alc;
718 int r;
719 struct kmem_list3 *l3 = s->cs_cachep->nodelists[q];
720 if (!l3 || OFF_SLAB(s->cs_cachep))
721 continue;
722 lockdep_set_class(&l3->list_lock, &on_slab_l3_key);
723 alc = l3->alien;
725 * FIXME: This check for BAD_ALIEN_MAGIC
726 * should go away when common slab code is taught to
727 * work even without alien caches.
728 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
729 * for alloc_alien_cache,
731 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
732 continue;
733 for_each_node(r) {
734 if (alc[r])
735 lockdep_set_class(&alc[r]->lock,
736 &on_slab_alc_key);
739 s++;
742 #else
743 static inline void init_lock_keys(void)
746 #endif
749 * Guard access to the cache-chain.
751 static DEFINE_MUTEX(cache_chain_mutex);
752 static struct list_head cache_chain;
755 * chicken and egg problem: delay the per-cpu array allocation
756 * until the general caches are up.
758 static enum {
759 NONE,
760 PARTIAL_AC,
761 PARTIAL_L3,
762 EARLY,
763 FULL
764 } g_cpucache_up;
767 * used by boot code to determine if it can use slab based allocator
769 int slab_is_available(void)
771 return g_cpucache_up >= EARLY;
774 static DEFINE_PER_CPU(struct delayed_work, reap_work);
776 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
778 return cachep->array[smp_processor_id()];
781 static inline struct kmem_cache *__find_general_cachep(size_t size,
782 gfp_t gfpflags)
784 struct cache_sizes *csizep = malloc_sizes;
786 #if DEBUG
787 /* This happens if someone tries to call
788 * kmem_cache_create(), or __kmalloc(), before
789 * the generic caches are initialized.
791 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
792 #endif
793 if (!size)
794 return ZERO_SIZE_PTR;
796 while (size > csizep->cs_size)
797 csizep++;
800 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
801 * has cs_{dma,}cachep==NULL. Thus no special case
802 * for large kmalloc calls required.
804 #ifdef CONFIG_ZONE_DMA
805 if (unlikely(gfpflags & GFP_DMA))
806 return csizep->cs_dmacachep;
807 #endif
808 return csizep->cs_cachep;
811 static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
813 return __find_general_cachep(size, gfpflags);
816 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
818 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
822 * Calculate the number of objects and left-over bytes for a given buffer size.
824 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
825 size_t align, int flags, size_t *left_over,
826 unsigned int *num)
828 int nr_objs;
829 size_t mgmt_size;
830 size_t slab_size = PAGE_SIZE << gfporder;
833 * The slab management structure can be either off the slab or
834 * on it. For the latter case, the memory allocated for a
835 * slab is used for:
837 * - The struct slab
838 * - One kmem_bufctl_t for each object
839 * - Padding to respect alignment of @align
840 * - @buffer_size bytes for each object
842 * If the slab management structure is off the slab, then the
843 * alignment will already be calculated into the size. Because
844 * the slabs are all pages aligned, the objects will be at the
845 * correct alignment when allocated.
847 if (flags & CFLGS_OFF_SLAB) {
848 mgmt_size = 0;
849 nr_objs = slab_size / buffer_size;
851 if (nr_objs > SLAB_LIMIT)
852 nr_objs = SLAB_LIMIT;
853 } else {
855 * Ignore padding for the initial guess. The padding
856 * is at most @align-1 bytes, and @buffer_size is at
857 * least @align. In the worst case, this result will
858 * be one greater than the number of objects that fit
859 * into the memory allocation when taking the padding
860 * into account.
862 nr_objs = (slab_size - sizeof(struct slab)) /
863 (buffer_size + sizeof(kmem_bufctl_t));
866 * This calculated number will be either the right
867 * amount, or one greater than what we want.
869 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
870 > slab_size)
871 nr_objs--;
873 if (nr_objs > SLAB_LIMIT)
874 nr_objs = SLAB_LIMIT;
876 mgmt_size = slab_mgmt_size(nr_objs, align);
878 *num = nr_objs;
879 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
882 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
884 static void __slab_error(const char *function, struct kmem_cache *cachep,
885 char *msg)
887 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
888 function, cachep->name, msg);
889 dump_stack();
893 * By default on NUMA we use alien caches to stage the freeing of
894 * objects allocated from other nodes. This causes massive memory
895 * inefficiencies when using fake NUMA setup to split memory into a
896 * large number of small nodes, so it can be disabled on the command
897 * line
900 static int use_alien_caches __read_mostly = 1;
901 static int numa_platform __read_mostly = 1;
902 static int __init noaliencache_setup(char *s)
904 use_alien_caches = 0;
905 return 1;
907 __setup("noaliencache", noaliencache_setup);
909 #ifdef CONFIG_NUMA
911 * Special reaping functions for NUMA systems called from cache_reap().
912 * These take care of doing round robin flushing of alien caches (containing
913 * objects freed on different nodes from which they were allocated) and the
914 * flushing of remote pcps by calling drain_node_pages.
916 static DEFINE_PER_CPU(unsigned long, reap_node);
918 static void init_reap_node(int cpu)
920 int node;
922 node = next_node(cpu_to_node(cpu), node_online_map);
923 if (node == MAX_NUMNODES)
924 node = first_node(node_online_map);
926 per_cpu(reap_node, cpu) = node;
929 static void next_reap_node(void)
931 int node = __get_cpu_var(reap_node);
933 node = next_node(node, node_online_map);
934 if (unlikely(node >= MAX_NUMNODES))
935 node = first_node(node_online_map);
936 __get_cpu_var(reap_node) = node;
939 #else
940 #define init_reap_node(cpu) do { } while (0)
941 #define next_reap_node(void) do { } while (0)
942 #endif
945 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
946 * via the workqueue/eventd.
947 * Add the CPU number into the expiration time to minimize the possibility of
948 * the CPUs getting into lockstep and contending for the global cache chain
949 * lock.
951 static void __cpuinit start_cpu_timer(int cpu)
953 struct delayed_work *reap_work = &per_cpu(reap_work, cpu);
956 * When this gets called from do_initcalls via cpucache_init(),
957 * init_workqueues() has already run, so keventd will be setup
958 * at that time.
960 if (keventd_up() && reap_work->work.func == NULL) {
961 init_reap_node(cpu);
962 INIT_DELAYED_WORK(reap_work, cache_reap);
963 schedule_delayed_work_on(cpu, reap_work,
964 __round_jiffies_relative(HZ, cpu));
968 static struct array_cache *alloc_arraycache(int node, int entries,
969 int batchcount, gfp_t gfp)
971 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
972 struct array_cache *nc = NULL;
974 nc = kmalloc_node(memsize, gfp, node);
976 * The array_cache structures contain pointers to free object.
977 * However, when such objects are allocated or transfered to another
978 * cache the pointers are not cleared and they could be counted as
979 * valid references during a kmemleak scan. Therefore, kmemleak must
980 * not scan such objects.
982 kmemleak_no_scan(nc);
983 if (nc) {
984 nc->avail = 0;
985 nc->limit = entries;
986 nc->batchcount = batchcount;
987 nc->touched = 0;
988 spin_lock_init(&nc->lock);
990 return nc;
994 * Transfer objects in one arraycache to another.
995 * Locking must be handled by the caller.
997 * Return the number of entries transferred.
999 static int transfer_objects(struct array_cache *to,
1000 struct array_cache *from, unsigned int max)
1002 /* Figure out how many entries to transfer */
1003 int nr = min(min(from->avail, max), to->limit - to->avail);
1005 if (!nr)
1006 return 0;
1008 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
1009 sizeof(void *) *nr);
1011 from->avail -= nr;
1012 to->avail += nr;
1013 to->touched = 1;
1014 return nr;
1017 #ifndef CONFIG_NUMA
1019 #define drain_alien_cache(cachep, alien) do { } while (0)
1020 #define reap_alien(cachep, l3) do { } while (0)
1022 static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
1024 return (struct array_cache **)BAD_ALIEN_MAGIC;
1027 static inline void free_alien_cache(struct array_cache **ac_ptr)
1031 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1033 return 0;
1036 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
1037 gfp_t flags)
1039 return NULL;
1042 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
1043 gfp_t flags, int nodeid)
1045 return NULL;
1048 #else /* CONFIG_NUMA */
1050 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
1051 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
1053 static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
1055 struct array_cache **ac_ptr;
1056 int memsize = sizeof(void *) * nr_node_ids;
1057 int i;
1059 if (limit > 1)
1060 limit = 12;
1061 ac_ptr = kmalloc_node(memsize, gfp, node);
1062 if (ac_ptr) {
1063 for_each_node(i) {
1064 if (i == node || !node_online(i)) {
1065 ac_ptr[i] = NULL;
1066 continue;
1068 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
1069 if (!ac_ptr[i]) {
1070 for (i--; i >= 0; i--)
1071 kfree(ac_ptr[i]);
1072 kfree(ac_ptr);
1073 return NULL;
1077 return ac_ptr;
1080 static void free_alien_cache(struct array_cache **ac_ptr)
1082 int i;
1084 if (!ac_ptr)
1085 return;
1086 for_each_node(i)
1087 kfree(ac_ptr[i]);
1088 kfree(ac_ptr);
1091 static void __drain_alien_cache(struct kmem_cache *cachep,
1092 struct array_cache *ac, int node)
1094 struct kmem_list3 *rl3 = cachep->nodelists[node];
1096 if (ac->avail) {
1097 spin_lock(&rl3->list_lock);
1099 * Stuff objects into the remote nodes shared array first.
1100 * That way we could avoid the overhead of putting the objects
1101 * into the free lists and getting them back later.
1103 if (rl3->shared)
1104 transfer_objects(rl3->shared, ac, ac->limit);
1106 free_block(cachep, ac->entry, ac->avail, node);
1107 ac->avail = 0;
1108 spin_unlock(&rl3->list_lock);
1113 * Called from cache_reap() to regularly drain alien caches round robin.
1115 static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1117 int node = __get_cpu_var(reap_node);
1119 if (l3->alien) {
1120 struct array_cache *ac = l3->alien[node];
1122 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1123 __drain_alien_cache(cachep, ac, node);
1124 spin_unlock_irq(&ac->lock);
1129 static void drain_alien_cache(struct kmem_cache *cachep,
1130 struct array_cache **alien)
1132 int i = 0;
1133 struct array_cache *ac;
1134 unsigned long flags;
1136 for_each_online_node(i) {
1137 ac = alien[i];
1138 if (ac) {
1139 spin_lock_irqsave(&ac->lock, flags);
1140 __drain_alien_cache(cachep, ac, i);
1141 spin_unlock_irqrestore(&ac->lock, flags);
1146 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1148 struct slab *slabp = virt_to_slab(objp);
1149 int nodeid = slabp->nodeid;
1150 struct kmem_list3 *l3;
1151 struct array_cache *alien = NULL;
1152 int node;
1154 node = numa_node_id();
1157 * Make sure we are not freeing a object from another node to the array
1158 * cache on this cpu.
1160 if (likely(slabp->nodeid == node))
1161 return 0;
1163 l3 = cachep->nodelists[node];
1164 STATS_INC_NODEFREES(cachep);
1165 if (l3->alien && l3->alien[nodeid]) {
1166 alien = l3->alien[nodeid];
1167 spin_lock(&alien->lock);
1168 if (unlikely(alien->avail == alien->limit)) {
1169 STATS_INC_ACOVERFLOW(cachep);
1170 __drain_alien_cache(cachep, alien, nodeid);
1172 alien->entry[alien->avail++] = objp;
1173 spin_unlock(&alien->lock);
1174 } else {
1175 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1176 free_block(cachep, &objp, 1, nodeid);
1177 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1179 return 1;
1181 #endif
1183 static void __cpuinit cpuup_canceled(long cpu)
1185 struct kmem_cache *cachep;
1186 struct kmem_list3 *l3 = NULL;
1187 int node = cpu_to_node(cpu);
1188 const struct cpumask *mask = cpumask_of_node(node);
1190 list_for_each_entry(cachep, &cache_chain, next) {
1191 struct array_cache *nc;
1192 struct array_cache *shared;
1193 struct array_cache **alien;
1195 /* cpu is dead; no one can alloc from it. */
1196 nc = cachep->array[cpu];
1197 cachep->array[cpu] = NULL;
1198 l3 = cachep->nodelists[node];
1200 if (!l3)
1201 goto free_array_cache;
1203 spin_lock_irq(&l3->list_lock);
1205 /* Free limit for this kmem_list3 */
1206 l3->free_limit -= cachep->batchcount;
1207 if (nc)
1208 free_block(cachep, nc->entry, nc->avail, node);
1210 if (!cpus_empty(*mask)) {
1211 spin_unlock_irq(&l3->list_lock);
1212 goto free_array_cache;
1215 shared = l3->shared;
1216 if (shared) {
1217 free_block(cachep, shared->entry,
1218 shared->avail, node);
1219 l3->shared = NULL;
1222 alien = l3->alien;
1223 l3->alien = NULL;
1225 spin_unlock_irq(&l3->list_lock);
1227 kfree(shared);
1228 if (alien) {
1229 drain_alien_cache(cachep, alien);
1230 free_alien_cache(alien);
1232 free_array_cache:
1233 kfree(nc);
1236 * In the previous loop, all the objects were freed to
1237 * the respective cache's slabs, now we can go ahead and
1238 * shrink each nodelist to its limit.
1240 list_for_each_entry(cachep, &cache_chain, next) {
1241 l3 = cachep->nodelists[node];
1242 if (!l3)
1243 continue;
1244 drain_freelist(cachep, l3, l3->free_objects);
1248 static int __cpuinit cpuup_prepare(long cpu)
1250 struct kmem_cache *cachep;
1251 struct kmem_list3 *l3 = NULL;
1252 int node = cpu_to_node(cpu);
1253 const int memsize = sizeof(struct kmem_list3);
1256 * We need to do this right in the beginning since
1257 * alloc_arraycache's are going to use this list.
1258 * kmalloc_node allows us to add the slab to the right
1259 * kmem_list3 and not this cpu's kmem_list3
1262 list_for_each_entry(cachep, &cache_chain, next) {
1264 * Set up the size64 kmemlist for cpu before we can
1265 * begin anything. Make sure some other cpu on this
1266 * node has not already allocated this
1268 if (!cachep->nodelists[node]) {
1269 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1270 if (!l3)
1271 goto bad;
1272 kmem_list3_init(l3);
1273 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1274 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1277 * The l3s don't come and go as CPUs come and
1278 * go. cache_chain_mutex is sufficient
1279 * protection here.
1281 cachep->nodelists[node] = l3;
1284 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1285 cachep->nodelists[node]->free_limit =
1286 (1 + nr_cpus_node(node)) *
1287 cachep->batchcount + cachep->num;
1288 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1292 * Now we can go ahead with allocating the shared arrays and
1293 * array caches
1295 list_for_each_entry(cachep, &cache_chain, next) {
1296 struct array_cache *nc;
1297 struct array_cache *shared = NULL;
1298 struct array_cache **alien = NULL;
1300 nc = alloc_arraycache(node, cachep->limit,
1301 cachep->batchcount, GFP_KERNEL);
1302 if (!nc)
1303 goto bad;
1304 if (cachep->shared) {
1305 shared = alloc_arraycache(node,
1306 cachep->shared * cachep->batchcount,
1307 0xbaadf00d, GFP_KERNEL);
1308 if (!shared) {
1309 kfree(nc);
1310 goto bad;
1313 if (use_alien_caches) {
1314 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
1315 if (!alien) {
1316 kfree(shared);
1317 kfree(nc);
1318 goto bad;
1321 cachep->array[cpu] = nc;
1322 l3 = cachep->nodelists[node];
1323 BUG_ON(!l3);
1325 spin_lock_irq(&l3->list_lock);
1326 if (!l3->shared) {
1328 * We are serialised from CPU_DEAD or
1329 * CPU_UP_CANCELLED by the cpucontrol lock
1331 l3->shared = shared;
1332 shared = NULL;
1334 #ifdef CONFIG_NUMA
1335 if (!l3->alien) {
1336 l3->alien = alien;
1337 alien = NULL;
1339 #endif
1340 spin_unlock_irq(&l3->list_lock);
1341 kfree(shared);
1342 free_alien_cache(alien);
1344 return 0;
1345 bad:
1346 cpuup_canceled(cpu);
1347 return -ENOMEM;
1350 static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1351 unsigned long action, void *hcpu)
1353 long cpu = (long)hcpu;
1354 int err = 0;
1356 switch (action) {
1357 case CPU_UP_PREPARE:
1358 case CPU_UP_PREPARE_FROZEN:
1359 mutex_lock(&cache_chain_mutex);
1360 err = cpuup_prepare(cpu);
1361 mutex_unlock(&cache_chain_mutex);
1362 break;
1363 case CPU_ONLINE:
1364 case CPU_ONLINE_FROZEN:
1365 start_cpu_timer(cpu);
1366 break;
1367 #ifdef CONFIG_HOTPLUG_CPU
1368 case CPU_DOWN_PREPARE:
1369 case CPU_DOWN_PREPARE_FROZEN:
1371 * Shutdown cache reaper. Note that the cache_chain_mutex is
1372 * held so that if cache_reap() is invoked it cannot do
1373 * anything expensive but will only modify reap_work
1374 * and reschedule the timer.
1376 cancel_rearming_delayed_work(&per_cpu(reap_work, cpu));
1377 /* Now the cache_reaper is guaranteed to be not running. */
1378 per_cpu(reap_work, cpu).work.func = NULL;
1379 break;
1380 case CPU_DOWN_FAILED:
1381 case CPU_DOWN_FAILED_FROZEN:
1382 start_cpu_timer(cpu);
1383 break;
1384 case CPU_DEAD:
1385 case CPU_DEAD_FROZEN:
1387 * Even if all the cpus of a node are down, we don't free the
1388 * kmem_list3 of any cache. This to avoid a race between
1389 * cpu_down, and a kmalloc allocation from another cpu for
1390 * memory from the node of the cpu going down. The list3
1391 * structure is usually allocated from kmem_cache_create() and
1392 * gets destroyed at kmem_cache_destroy().
1394 /* fall through */
1395 #endif
1396 case CPU_UP_CANCELED:
1397 case CPU_UP_CANCELED_FROZEN:
1398 mutex_lock(&cache_chain_mutex);
1399 cpuup_canceled(cpu);
1400 mutex_unlock(&cache_chain_mutex);
1401 break;
1403 return err ? NOTIFY_BAD : NOTIFY_OK;
1406 static struct notifier_block __cpuinitdata cpucache_notifier = {
1407 &cpuup_callback, NULL, 0
1411 * swap the static kmem_list3 with kmalloced memory
1413 static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1414 int nodeid)
1416 struct kmem_list3 *ptr;
1418 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid);
1419 BUG_ON(!ptr);
1421 memcpy(ptr, list, sizeof(struct kmem_list3));
1423 * Do not assume that spinlocks can be initialized via memcpy:
1425 spin_lock_init(&ptr->list_lock);
1427 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1428 cachep->nodelists[nodeid] = ptr;
1432 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1433 * size of kmem_list3.
1435 static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1437 int node;
1439 for_each_online_node(node) {
1440 cachep->nodelists[node] = &initkmem_list3[index + node];
1441 cachep->nodelists[node]->next_reap = jiffies +
1442 REAPTIMEOUT_LIST3 +
1443 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1448 * Initialisation. Called after the page allocator have been initialised and
1449 * before smp_init().
1451 void __init kmem_cache_init(void)
1453 size_t left_over;
1454 struct cache_sizes *sizes;
1455 struct cache_names *names;
1456 int i;
1457 int order;
1458 int node;
1460 if (num_possible_nodes() == 1) {
1461 use_alien_caches = 0;
1462 numa_platform = 0;
1465 for (i = 0; i < NUM_INIT_LISTS; i++) {
1466 kmem_list3_init(&initkmem_list3[i]);
1467 if (i < MAX_NUMNODES)
1468 cache_cache.nodelists[i] = NULL;
1470 set_up_list3s(&cache_cache, CACHE_CACHE);
1473 * Fragmentation resistance on low memory - only use bigger
1474 * page orders on machines with more than 32MB of memory.
1476 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1477 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1479 /* Bootstrap is tricky, because several objects are allocated
1480 * from caches that do not exist yet:
1481 * 1) initialize the cache_cache cache: it contains the struct
1482 * kmem_cache structures of all caches, except cache_cache itself:
1483 * cache_cache is statically allocated.
1484 * Initially an __init data area is used for the head array and the
1485 * kmem_list3 structures, it's replaced with a kmalloc allocated
1486 * array at the end of the bootstrap.
1487 * 2) Create the first kmalloc cache.
1488 * The struct kmem_cache for the new cache is allocated normally.
1489 * An __init data area is used for the head array.
1490 * 3) Create the remaining kmalloc caches, with minimally sized
1491 * head arrays.
1492 * 4) Replace the __init data head arrays for cache_cache and the first
1493 * kmalloc cache with kmalloc allocated arrays.
1494 * 5) Replace the __init data for kmem_list3 for cache_cache and
1495 * the other cache's with kmalloc allocated memory.
1496 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1499 node = numa_node_id();
1501 /* 1) create the cache_cache */
1502 INIT_LIST_HEAD(&cache_chain);
1503 list_add(&cache_cache.next, &cache_chain);
1504 cache_cache.colour_off = cache_line_size();
1505 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
1506 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node];
1509 * struct kmem_cache size depends on nr_node_ids, which
1510 * can be less than MAX_NUMNODES.
1512 cache_cache.buffer_size = offsetof(struct kmem_cache, nodelists) +
1513 nr_node_ids * sizeof(struct kmem_list3 *);
1514 #if DEBUG
1515 cache_cache.obj_size = cache_cache.buffer_size;
1516 #endif
1517 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1518 cache_line_size());
1519 cache_cache.reciprocal_buffer_size =
1520 reciprocal_value(cache_cache.buffer_size);
1522 for (order = 0; order < MAX_ORDER; order++) {
1523 cache_estimate(order, cache_cache.buffer_size,
1524 cache_line_size(), 0, &left_over, &cache_cache.num);
1525 if (cache_cache.num)
1526 break;
1528 BUG_ON(!cache_cache.num);
1529 cache_cache.gfporder = order;
1530 cache_cache.colour = left_over / cache_cache.colour_off;
1531 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1532 sizeof(struct slab), cache_line_size());
1534 /* 2+3) create the kmalloc caches */
1535 sizes = malloc_sizes;
1536 names = cache_names;
1539 * Initialize the caches that provide memory for the array cache and the
1540 * kmem_list3 structures first. Without this, further allocations will
1541 * bug.
1544 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
1545 sizes[INDEX_AC].cs_size,
1546 ARCH_KMALLOC_MINALIGN,
1547 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1548 NULL);
1550 if (INDEX_AC != INDEX_L3) {
1551 sizes[INDEX_L3].cs_cachep =
1552 kmem_cache_create(names[INDEX_L3].name,
1553 sizes[INDEX_L3].cs_size,
1554 ARCH_KMALLOC_MINALIGN,
1555 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1556 NULL);
1559 slab_early_init = 0;
1561 while (sizes->cs_size != ULONG_MAX) {
1563 * For performance, all the general caches are L1 aligned.
1564 * This should be particularly beneficial on SMP boxes, as it
1565 * eliminates "false sharing".
1566 * Note for systems short on memory removing the alignment will
1567 * allow tighter packing of the smaller caches.
1569 if (!sizes->cs_cachep) {
1570 sizes->cs_cachep = kmem_cache_create(names->name,
1571 sizes->cs_size,
1572 ARCH_KMALLOC_MINALIGN,
1573 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1574 NULL);
1576 #ifdef CONFIG_ZONE_DMA
1577 sizes->cs_dmacachep = kmem_cache_create(
1578 names->name_dma,
1579 sizes->cs_size,
1580 ARCH_KMALLOC_MINALIGN,
1581 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1582 SLAB_PANIC,
1583 NULL);
1584 #endif
1585 sizes++;
1586 names++;
1588 /* 4) Replace the bootstrap head arrays */
1590 struct array_cache *ptr;
1592 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1594 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1595 memcpy(ptr, cpu_cache_get(&cache_cache),
1596 sizeof(struct arraycache_init));
1598 * Do not assume that spinlocks can be initialized via memcpy:
1600 spin_lock_init(&ptr->lock);
1602 cache_cache.array[smp_processor_id()] = ptr;
1604 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1606 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1607 != &initarray_generic.cache);
1608 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1609 sizeof(struct arraycache_init));
1611 * Do not assume that spinlocks can be initialized via memcpy:
1613 spin_lock_init(&ptr->lock);
1615 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1616 ptr;
1618 /* 5) Replace the bootstrap kmem_list3's */
1620 int nid;
1622 for_each_online_node(nid) {
1623 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid);
1625 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1626 &initkmem_list3[SIZE_AC + nid], nid);
1628 if (INDEX_AC != INDEX_L3) {
1629 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1630 &initkmem_list3[SIZE_L3 + nid], nid);
1635 g_cpucache_up = EARLY;
1637 /* Annotate slab for lockdep -- annotate the malloc caches */
1638 init_lock_keys();
1641 void __init kmem_cache_init_late(void)
1643 struct kmem_cache *cachep;
1646 * Interrupts are enabled now so all GFP allocations are safe.
1648 slab_gfp_mask = __GFP_BITS_MASK;
1650 /* 6) resize the head arrays to their final sizes */
1651 mutex_lock(&cache_chain_mutex);
1652 list_for_each_entry(cachep, &cache_chain, next)
1653 if (enable_cpucache(cachep, GFP_NOWAIT))
1654 BUG();
1655 mutex_unlock(&cache_chain_mutex);
1657 /* Done! */
1658 g_cpucache_up = FULL;
1661 * Register a cpu startup notifier callback that initializes
1662 * cpu_cache_get for all new cpus
1664 register_cpu_notifier(&cpucache_notifier);
1667 * The reap timers are started later, with a module init call: That part
1668 * of the kernel is not yet operational.
1672 static int __init cpucache_init(void)
1674 int cpu;
1677 * Register the timers that return unneeded pages to the page allocator
1679 for_each_online_cpu(cpu)
1680 start_cpu_timer(cpu);
1681 return 0;
1683 __initcall(cpucache_init);
1686 * Interface to system's page allocator. No need to hold the cache-lock.
1688 * If we requested dmaable memory, we will get it. Even if we
1689 * did not request dmaable memory, we might get it, but that
1690 * would be relatively rare and ignorable.
1692 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1694 struct page *page;
1695 int nr_pages;
1696 int i;
1698 #ifndef CONFIG_MMU
1700 * Nommu uses slab's for process anonymous memory allocations, and thus
1701 * requires __GFP_COMP to properly refcount higher order allocations
1703 flags |= __GFP_COMP;
1704 #endif
1706 flags |= cachep->gfpflags;
1707 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1708 flags |= __GFP_RECLAIMABLE;
1710 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
1711 if (!page)
1712 return NULL;
1714 nr_pages = (1 << cachep->gfporder);
1715 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1716 add_zone_page_state(page_zone(page),
1717 NR_SLAB_RECLAIMABLE, nr_pages);
1718 else
1719 add_zone_page_state(page_zone(page),
1720 NR_SLAB_UNRECLAIMABLE, nr_pages);
1721 for (i = 0; i < nr_pages; i++)
1722 __SetPageSlab(page + i);
1723 return page_address(page);
1727 * Interface to system's page release.
1729 static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1731 unsigned long i = (1 << cachep->gfporder);
1732 struct page *page = virt_to_page(addr);
1733 const unsigned long nr_freed = i;
1735 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1736 sub_zone_page_state(page_zone(page),
1737 NR_SLAB_RECLAIMABLE, nr_freed);
1738 else
1739 sub_zone_page_state(page_zone(page),
1740 NR_SLAB_UNRECLAIMABLE, nr_freed);
1741 while (i--) {
1742 BUG_ON(!PageSlab(page));
1743 __ClearPageSlab(page);
1744 page++;
1746 if (current->reclaim_state)
1747 current->reclaim_state->reclaimed_slab += nr_freed;
1748 free_pages((unsigned long)addr, cachep->gfporder);
1751 static void kmem_rcu_free(struct rcu_head *head)
1753 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1754 struct kmem_cache *cachep = slab_rcu->cachep;
1756 kmem_freepages(cachep, slab_rcu->addr);
1757 if (OFF_SLAB(cachep))
1758 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1761 #if DEBUG
1763 #ifdef CONFIG_DEBUG_PAGEALLOC
1764 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1765 unsigned long caller)
1767 int size = obj_size(cachep);
1769 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1771 if (size < 5 * sizeof(unsigned long))
1772 return;
1774 *addr++ = 0x12345678;
1775 *addr++ = caller;
1776 *addr++ = smp_processor_id();
1777 size -= 3 * sizeof(unsigned long);
1779 unsigned long *sptr = &caller;
1780 unsigned long svalue;
1782 while (!kstack_end(sptr)) {
1783 svalue = *sptr++;
1784 if (kernel_text_address(svalue)) {
1785 *addr++ = svalue;
1786 size -= sizeof(unsigned long);
1787 if (size <= sizeof(unsigned long))
1788 break;
1793 *addr++ = 0x87654321;
1795 #endif
1797 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1799 int size = obj_size(cachep);
1800 addr = &((char *)addr)[obj_offset(cachep)];
1802 memset(addr, val, size);
1803 *(unsigned char *)(addr + size - 1) = POISON_END;
1806 static void dump_line(char *data, int offset, int limit)
1808 int i;
1809 unsigned char error = 0;
1810 int bad_count = 0;
1812 printk(KERN_ERR "%03x:", offset);
1813 for (i = 0; i < limit; i++) {
1814 if (data[offset + i] != POISON_FREE) {
1815 error = data[offset + i];
1816 bad_count++;
1818 printk(" %02x", (unsigned char)data[offset + i]);
1820 printk("\n");
1822 if (bad_count == 1) {
1823 error ^= POISON_FREE;
1824 if (!(error & (error - 1))) {
1825 printk(KERN_ERR "Single bit error detected. Probably "
1826 "bad RAM.\n");
1827 #ifdef CONFIG_X86
1828 printk(KERN_ERR "Run memtest86+ or a similar memory "
1829 "test tool.\n");
1830 #else
1831 printk(KERN_ERR "Run a memory test tool.\n");
1832 #endif
1836 #endif
1838 #if DEBUG
1840 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1842 int i, size;
1843 char *realobj;
1845 if (cachep->flags & SLAB_RED_ZONE) {
1846 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
1847 *dbg_redzone1(cachep, objp),
1848 *dbg_redzone2(cachep, objp));
1851 if (cachep->flags & SLAB_STORE_USER) {
1852 printk(KERN_ERR "Last user: [<%p>]",
1853 *dbg_userword(cachep, objp));
1854 print_symbol("(%s)",
1855 (unsigned long)*dbg_userword(cachep, objp));
1856 printk("\n");
1858 realobj = (char *)objp + obj_offset(cachep);
1859 size = obj_size(cachep);
1860 for (i = 0; i < size && lines; i += 16, lines--) {
1861 int limit;
1862 limit = 16;
1863 if (i + limit > size)
1864 limit = size - i;
1865 dump_line(realobj, i, limit);
1869 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1871 char *realobj;
1872 int size, i;
1873 int lines = 0;
1875 realobj = (char *)objp + obj_offset(cachep);
1876 size = obj_size(cachep);
1878 for (i = 0; i < size; i++) {
1879 char exp = POISON_FREE;
1880 if (i == size - 1)
1881 exp = POISON_END;
1882 if (realobj[i] != exp) {
1883 int limit;
1884 /* Mismatch ! */
1885 /* Print header */
1886 if (lines == 0) {
1887 printk(KERN_ERR
1888 "Slab corruption: %s start=%p, len=%d\n",
1889 cachep->name, realobj, size);
1890 print_objinfo(cachep, objp, 0);
1892 /* Hexdump the affected line */
1893 i = (i / 16) * 16;
1894 limit = 16;
1895 if (i + limit > size)
1896 limit = size - i;
1897 dump_line(realobj, i, limit);
1898 i += 16;
1899 lines++;
1900 /* Limit to 5 lines */
1901 if (lines > 5)
1902 break;
1905 if (lines != 0) {
1906 /* Print some data about the neighboring objects, if they
1907 * exist:
1909 struct slab *slabp = virt_to_slab(objp);
1910 unsigned int objnr;
1912 objnr = obj_to_index(cachep, slabp, objp);
1913 if (objnr) {
1914 objp = index_to_obj(cachep, slabp, objnr - 1);
1915 realobj = (char *)objp + obj_offset(cachep);
1916 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1917 realobj, size);
1918 print_objinfo(cachep, objp, 2);
1920 if (objnr + 1 < cachep->num) {
1921 objp = index_to_obj(cachep, slabp, objnr + 1);
1922 realobj = (char *)objp + obj_offset(cachep);
1923 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1924 realobj, size);
1925 print_objinfo(cachep, objp, 2);
1929 #endif
1931 #if DEBUG
1932 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
1934 int i;
1935 for (i = 0; i < cachep->num; i++) {
1936 void *objp = index_to_obj(cachep, slabp, i);
1938 if (cachep->flags & SLAB_POISON) {
1939 #ifdef CONFIG_DEBUG_PAGEALLOC
1940 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1941 OFF_SLAB(cachep))
1942 kernel_map_pages(virt_to_page(objp),
1943 cachep->buffer_size / PAGE_SIZE, 1);
1944 else
1945 check_poison_obj(cachep, objp);
1946 #else
1947 check_poison_obj(cachep, objp);
1948 #endif
1950 if (cachep->flags & SLAB_RED_ZONE) {
1951 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1952 slab_error(cachep, "start of a freed object "
1953 "was overwritten");
1954 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1955 slab_error(cachep, "end of a freed object "
1956 "was overwritten");
1960 #else
1961 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
1964 #endif
1967 * slab_destroy - destroy and release all objects in a slab
1968 * @cachep: cache pointer being destroyed
1969 * @slabp: slab pointer being destroyed
1971 * Destroy all the objs in a slab, and release the mem back to the system.
1972 * Before calling the slab must have been unlinked from the cache. The
1973 * cache-lock is not held/needed.
1975 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
1977 void *addr = slabp->s_mem - slabp->colouroff;
1979 slab_destroy_debugcheck(cachep, slabp);
1980 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1981 struct slab_rcu *slab_rcu;
1983 slab_rcu = (struct slab_rcu *)slabp;
1984 slab_rcu->cachep = cachep;
1985 slab_rcu->addr = addr;
1986 call_rcu(&slab_rcu->head, kmem_rcu_free);
1987 } else {
1988 kmem_freepages(cachep, addr);
1989 if (OFF_SLAB(cachep))
1990 kmem_cache_free(cachep->slabp_cache, slabp);
1994 static void __kmem_cache_destroy(struct kmem_cache *cachep)
1996 int i;
1997 struct kmem_list3 *l3;
1999 for_each_online_cpu(i)
2000 kfree(cachep->array[i]);
2002 /* NUMA: free the list3 structures */
2003 for_each_online_node(i) {
2004 l3 = cachep->nodelists[i];
2005 if (l3) {
2006 kfree(l3->shared);
2007 free_alien_cache(l3->alien);
2008 kfree(l3);
2011 kmem_cache_free(&cache_cache, cachep);
2016 * calculate_slab_order - calculate size (page order) of slabs
2017 * @cachep: pointer to the cache that is being created
2018 * @size: size of objects to be created in this cache.
2019 * @align: required alignment for the objects.
2020 * @flags: slab allocation flags
2022 * Also calculates the number of objects per slab.
2024 * This could be made much more intelligent. For now, try to avoid using
2025 * high order pages for slabs. When the gfp() functions are more friendly
2026 * towards high-order requests, this should be changed.
2028 static size_t calculate_slab_order(struct kmem_cache *cachep,
2029 size_t size, size_t align, unsigned long flags)
2031 unsigned long offslab_limit;
2032 size_t left_over = 0;
2033 int gfporder;
2035 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
2036 unsigned int num;
2037 size_t remainder;
2039 cache_estimate(gfporder, size, align, flags, &remainder, &num);
2040 if (!num)
2041 continue;
2043 if (flags & CFLGS_OFF_SLAB) {
2045 * Max number of objs-per-slab for caches which
2046 * use off-slab slabs. Needed to avoid a possible
2047 * looping condition in cache_grow().
2049 offslab_limit = size - sizeof(struct slab);
2050 offslab_limit /= sizeof(kmem_bufctl_t);
2052 if (num > offslab_limit)
2053 break;
2056 /* Found something acceptable - save it away */
2057 cachep->num = num;
2058 cachep->gfporder = gfporder;
2059 left_over = remainder;
2062 * A VFS-reclaimable slab tends to have most allocations
2063 * as GFP_NOFS and we really don't want to have to be allocating
2064 * higher-order pages when we are unable to shrink dcache.
2066 if (flags & SLAB_RECLAIM_ACCOUNT)
2067 break;
2070 * Large number of objects is good, but very large slabs are
2071 * currently bad for the gfp()s.
2073 if (gfporder >= slab_break_gfp_order)
2074 break;
2077 * Acceptable internal fragmentation?
2079 if (left_over * 8 <= (PAGE_SIZE << gfporder))
2080 break;
2082 return left_over;
2085 static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
2087 if (g_cpucache_up == FULL)
2088 return enable_cpucache(cachep, gfp);
2090 if (g_cpucache_up == NONE) {
2092 * Note: the first kmem_cache_create must create the cache
2093 * that's used by kmalloc(24), otherwise the creation of
2094 * further caches will BUG().
2096 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2099 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2100 * the first cache, then we need to set up all its list3s,
2101 * otherwise the creation of further caches will BUG().
2103 set_up_list3s(cachep, SIZE_AC);
2104 if (INDEX_AC == INDEX_L3)
2105 g_cpucache_up = PARTIAL_L3;
2106 else
2107 g_cpucache_up = PARTIAL_AC;
2108 } else {
2109 cachep->array[smp_processor_id()] =
2110 kmalloc(sizeof(struct arraycache_init), gfp);
2112 if (g_cpucache_up == PARTIAL_AC) {
2113 set_up_list3s(cachep, SIZE_L3);
2114 g_cpucache_up = PARTIAL_L3;
2115 } else {
2116 int node;
2117 for_each_online_node(node) {
2118 cachep->nodelists[node] =
2119 kmalloc_node(sizeof(struct kmem_list3),
2120 gfp, node);
2121 BUG_ON(!cachep->nodelists[node]);
2122 kmem_list3_init(cachep->nodelists[node]);
2126 cachep->nodelists[numa_node_id()]->next_reap =
2127 jiffies + REAPTIMEOUT_LIST3 +
2128 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2130 cpu_cache_get(cachep)->avail = 0;
2131 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2132 cpu_cache_get(cachep)->batchcount = 1;
2133 cpu_cache_get(cachep)->touched = 0;
2134 cachep->batchcount = 1;
2135 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2136 return 0;
2140 * kmem_cache_create - Create a cache.
2141 * @name: A string which is used in /proc/slabinfo to identify this cache.
2142 * @size: The size of objects to be created in this cache.
2143 * @align: The required alignment for the objects.
2144 * @flags: SLAB flags
2145 * @ctor: A constructor for the objects.
2147 * Returns a ptr to the cache on success, NULL on failure.
2148 * Cannot be called within a int, but can be interrupted.
2149 * The @ctor is run when new pages are allocated by the cache.
2151 * @name must be valid until the cache is destroyed. This implies that
2152 * the module calling this has to destroy the cache before getting unloaded.
2153 * Note that kmem_cache_name() is not guaranteed to return the same pointer,
2154 * therefore applications must manage it themselves.
2156 * The flags are
2158 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2159 * to catch references to uninitialised memory.
2161 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2162 * for buffer overruns.
2164 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2165 * cacheline. This can be beneficial if you're counting cycles as closely
2166 * as davem.
2168 struct kmem_cache *
2169 kmem_cache_create (const char *name, size_t size, size_t align,
2170 unsigned long flags, void (*ctor)(void *))
2172 size_t left_over, slab_size, ralign;
2173 struct kmem_cache *cachep = NULL, *pc;
2174 gfp_t gfp;
2177 * Sanity checks... these are all serious usage bugs.
2179 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
2180 size > KMALLOC_MAX_SIZE) {
2181 printk(KERN_ERR "%s: Early error in slab %s\n", __func__,
2182 name);
2183 BUG();
2187 * We use cache_chain_mutex to ensure a consistent view of
2188 * cpu_online_mask as well. Please see cpuup_callback
2190 if (slab_is_available()) {
2191 get_online_cpus();
2192 mutex_lock(&cache_chain_mutex);
2195 list_for_each_entry(pc, &cache_chain, next) {
2196 char tmp;
2197 int res;
2200 * This happens when the module gets unloaded and doesn't
2201 * destroy its slab cache and no-one else reuses the vmalloc
2202 * area of the module. Print a warning.
2204 res = probe_kernel_address(pc->name, tmp);
2205 if (res) {
2206 printk(KERN_ERR
2207 "SLAB: cache with size %d has lost its name\n",
2208 pc->buffer_size);
2209 continue;
2212 if (!strcmp(pc->name, name)) {
2213 printk(KERN_ERR
2214 "kmem_cache_create: duplicate cache %s\n", name);
2215 dump_stack();
2216 goto oops;
2220 #if DEBUG
2221 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
2222 #if FORCED_DEBUG
2224 * Enable redzoning and last user accounting, except for caches with
2225 * large objects, if the increased size would increase the object size
2226 * above the next power of two: caches with object sizes just above a
2227 * power of two have a significant amount of internal fragmentation.
2229 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2230 2 * sizeof(unsigned long long)))
2231 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2232 if (!(flags & SLAB_DESTROY_BY_RCU))
2233 flags |= SLAB_POISON;
2234 #endif
2235 if (flags & SLAB_DESTROY_BY_RCU)
2236 BUG_ON(flags & SLAB_POISON);
2237 #endif
2239 * Always checks flags, a caller might be expecting debug support which
2240 * isn't available.
2242 BUG_ON(flags & ~CREATE_MASK);
2245 * Check that size is in terms of words. This is needed to avoid
2246 * unaligned accesses for some archs when redzoning is used, and makes
2247 * sure any on-slab bufctl's are also correctly aligned.
2249 if (size & (BYTES_PER_WORD - 1)) {
2250 size += (BYTES_PER_WORD - 1);
2251 size &= ~(BYTES_PER_WORD - 1);
2254 /* calculate the final buffer alignment: */
2256 /* 1) arch recommendation: can be overridden for debug */
2257 if (flags & SLAB_HWCACHE_ALIGN) {
2259 * Default alignment: as specified by the arch code. Except if
2260 * an object is really small, then squeeze multiple objects into
2261 * one cacheline.
2263 ralign = cache_line_size();
2264 while (size <= ralign / 2)
2265 ralign /= 2;
2266 } else {
2267 ralign = BYTES_PER_WORD;
2271 * Redzoning and user store require word alignment or possibly larger.
2272 * Note this will be overridden by architecture or caller mandated
2273 * alignment if either is greater than BYTES_PER_WORD.
2275 if (flags & SLAB_STORE_USER)
2276 ralign = BYTES_PER_WORD;
2278 if (flags & SLAB_RED_ZONE) {
2279 ralign = REDZONE_ALIGN;
2280 /* If redzoning, ensure that the second redzone is suitably
2281 * aligned, by adjusting the object size accordingly. */
2282 size += REDZONE_ALIGN - 1;
2283 size &= ~(REDZONE_ALIGN - 1);
2286 /* 2) arch mandated alignment */
2287 if (ralign < ARCH_SLAB_MINALIGN) {
2288 ralign = ARCH_SLAB_MINALIGN;
2290 /* 3) caller mandated alignment */
2291 if (ralign < align) {
2292 ralign = align;
2294 /* disable debug if necessary */
2295 if (ralign > __alignof__(unsigned long long))
2296 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2298 * 4) Store it.
2300 align = ralign;
2302 if (slab_is_available())
2303 gfp = GFP_KERNEL;
2304 else
2305 gfp = GFP_NOWAIT;
2307 /* Get cache's description obj. */
2308 cachep = kmem_cache_zalloc(&cache_cache, gfp);
2309 if (!cachep)
2310 goto oops;
2312 #if DEBUG
2313 cachep->obj_size = size;
2316 * Both debugging options require word-alignment which is calculated
2317 * into align above.
2319 if (flags & SLAB_RED_ZONE) {
2320 /* add space for red zone words */
2321 cachep->obj_offset += sizeof(unsigned long long);
2322 size += 2 * sizeof(unsigned long long);
2324 if (flags & SLAB_STORE_USER) {
2325 /* user store requires one word storage behind the end of
2326 * the real object. But if the second red zone needs to be
2327 * aligned to 64 bits, we must allow that much space.
2329 if (flags & SLAB_RED_ZONE)
2330 size += REDZONE_ALIGN;
2331 else
2332 size += BYTES_PER_WORD;
2334 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2335 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
2336 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2337 cachep->obj_offset += PAGE_SIZE - size;
2338 size = PAGE_SIZE;
2340 #endif
2341 #endif
2344 * Determine if the slab management is 'on' or 'off' slab.
2345 * (bootstrapping cannot cope with offslab caches so don't do
2346 * it too early on.)
2348 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init)
2350 * Size is large, assume best to place the slab management obj
2351 * off-slab (should allow better packing of objs).
2353 flags |= CFLGS_OFF_SLAB;
2355 size = ALIGN(size, align);
2357 left_over = calculate_slab_order(cachep, size, align, flags);
2359 if (!cachep->num) {
2360 printk(KERN_ERR
2361 "kmem_cache_create: couldn't create cache %s.\n", name);
2362 kmem_cache_free(&cache_cache, cachep);
2363 cachep = NULL;
2364 goto oops;
2366 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2367 + sizeof(struct slab), align);
2370 * If the slab has been placed off-slab, and we have enough space then
2371 * move it on-slab. This is at the expense of any extra colouring.
2373 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2374 flags &= ~CFLGS_OFF_SLAB;
2375 left_over -= slab_size;
2378 if (flags & CFLGS_OFF_SLAB) {
2379 /* really off slab. No need for manual alignment */
2380 slab_size =
2381 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2384 cachep->colour_off = cache_line_size();
2385 /* Offset must be a multiple of the alignment. */
2386 if (cachep->colour_off < align)
2387 cachep->colour_off = align;
2388 cachep->colour = left_over / cachep->colour_off;
2389 cachep->slab_size = slab_size;
2390 cachep->flags = flags;
2391 cachep->gfpflags = 0;
2392 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2393 cachep->gfpflags |= GFP_DMA;
2394 cachep->buffer_size = size;
2395 cachep->reciprocal_buffer_size = reciprocal_value(size);
2397 if (flags & CFLGS_OFF_SLAB) {
2398 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2400 * This is a possibility for one of the malloc_sizes caches.
2401 * But since we go off slab only for object size greater than
2402 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2403 * this should not happen at all.
2404 * But leave a BUG_ON for some lucky dude.
2406 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
2408 cachep->ctor = ctor;
2409 cachep->name = name;
2411 if (setup_cpu_cache(cachep, gfp)) {
2412 __kmem_cache_destroy(cachep);
2413 cachep = NULL;
2414 goto oops;
2417 /* cache setup completed, link it into the list */
2418 list_add(&cachep->next, &cache_chain);
2419 oops:
2420 if (!cachep && (flags & SLAB_PANIC))
2421 panic("kmem_cache_create(): failed to create slab `%s'\n",
2422 name);
2423 if (slab_is_available()) {
2424 mutex_unlock(&cache_chain_mutex);
2425 put_online_cpus();
2427 return cachep;
2429 EXPORT_SYMBOL(kmem_cache_create);
2431 #if DEBUG
2432 static void check_irq_off(void)
2434 BUG_ON(!irqs_disabled());
2437 static void check_irq_on(void)
2439 BUG_ON(irqs_disabled());
2442 static void check_spinlock_acquired(struct kmem_cache *cachep)
2444 #ifdef CONFIG_SMP
2445 check_irq_off();
2446 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
2447 #endif
2450 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2452 #ifdef CONFIG_SMP
2453 check_irq_off();
2454 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2455 #endif
2458 #else
2459 #define check_irq_off() do { } while(0)
2460 #define check_irq_on() do { } while(0)
2461 #define check_spinlock_acquired(x) do { } while(0)
2462 #define check_spinlock_acquired_node(x, y) do { } while(0)
2463 #endif
2465 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2466 struct array_cache *ac,
2467 int force, int node);
2469 static void do_drain(void *arg)
2471 struct kmem_cache *cachep = arg;
2472 struct array_cache *ac;
2473 int node = numa_node_id();
2475 check_irq_off();
2476 ac = cpu_cache_get(cachep);
2477 spin_lock(&cachep->nodelists[node]->list_lock);
2478 free_block(cachep, ac->entry, ac->avail, node);
2479 spin_unlock(&cachep->nodelists[node]->list_lock);
2480 ac->avail = 0;
2483 static void drain_cpu_caches(struct kmem_cache *cachep)
2485 struct kmem_list3 *l3;
2486 int node;
2488 on_each_cpu(do_drain, cachep, 1);
2489 check_irq_on();
2490 for_each_online_node(node) {
2491 l3 = cachep->nodelists[node];
2492 if (l3 && l3->alien)
2493 drain_alien_cache(cachep, l3->alien);
2496 for_each_online_node(node) {
2497 l3 = cachep->nodelists[node];
2498 if (l3)
2499 drain_array(cachep, l3, l3->shared, 1, node);
2504 * Remove slabs from the list of free slabs.
2505 * Specify the number of slabs to drain in tofree.
2507 * Returns the actual number of slabs released.
2509 static int drain_freelist(struct kmem_cache *cache,
2510 struct kmem_list3 *l3, int tofree)
2512 struct list_head *p;
2513 int nr_freed;
2514 struct slab *slabp;
2516 nr_freed = 0;
2517 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
2519 spin_lock_irq(&l3->list_lock);
2520 p = l3->slabs_free.prev;
2521 if (p == &l3->slabs_free) {
2522 spin_unlock_irq(&l3->list_lock);
2523 goto out;
2526 slabp = list_entry(p, struct slab, list);
2527 #if DEBUG
2528 BUG_ON(slabp->inuse);
2529 #endif
2530 list_del(&slabp->list);
2532 * Safe to drop the lock. The slab is no longer linked
2533 * to the cache.
2535 l3->free_objects -= cache->num;
2536 spin_unlock_irq(&l3->list_lock);
2537 slab_destroy(cache, slabp);
2538 nr_freed++;
2540 out:
2541 return nr_freed;
2544 /* Called with cache_chain_mutex held to protect against cpu hotplug */
2545 static int __cache_shrink(struct kmem_cache *cachep)
2547 int ret = 0, i = 0;
2548 struct kmem_list3 *l3;
2550 drain_cpu_caches(cachep);
2552 check_irq_on();
2553 for_each_online_node(i) {
2554 l3 = cachep->nodelists[i];
2555 if (!l3)
2556 continue;
2558 drain_freelist(cachep, l3, l3->free_objects);
2560 ret += !list_empty(&l3->slabs_full) ||
2561 !list_empty(&l3->slabs_partial);
2563 return (ret ? 1 : 0);
2567 * kmem_cache_shrink - Shrink a cache.
2568 * @cachep: The cache to shrink.
2570 * Releases as many slabs as possible for a cache.
2571 * To help debugging, a zero exit status indicates all slabs were released.
2573 int kmem_cache_shrink(struct kmem_cache *cachep)
2575 int ret;
2576 BUG_ON(!cachep || in_interrupt());
2578 get_online_cpus();
2579 mutex_lock(&cache_chain_mutex);
2580 ret = __cache_shrink(cachep);
2581 mutex_unlock(&cache_chain_mutex);
2582 put_online_cpus();
2583 return ret;
2585 EXPORT_SYMBOL(kmem_cache_shrink);
2588 * kmem_cache_destroy - delete a cache
2589 * @cachep: the cache to destroy
2591 * Remove a &struct kmem_cache object from the slab cache.
2593 * It is expected this function will be called by a module when it is
2594 * unloaded. This will remove the cache completely, and avoid a duplicate
2595 * cache being allocated each time a module is loaded and unloaded, if the
2596 * module doesn't have persistent in-kernel storage across loads and unloads.
2598 * The cache must be empty before calling this function.
2600 * The caller must guarantee that noone will allocate memory from the cache
2601 * during the kmem_cache_destroy().
2603 void kmem_cache_destroy(struct kmem_cache *cachep)
2605 BUG_ON(!cachep || in_interrupt());
2607 /* Find the cache in the chain of caches. */
2608 get_online_cpus();
2609 mutex_lock(&cache_chain_mutex);
2611 * the chain is never empty, cache_cache is never destroyed
2613 list_del(&cachep->next);
2614 if (__cache_shrink(cachep)) {
2615 slab_error(cachep, "Can't free all objects");
2616 list_add(&cachep->next, &cache_chain);
2617 mutex_unlock(&cache_chain_mutex);
2618 put_online_cpus();
2619 return;
2622 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
2623 synchronize_rcu();
2625 __kmem_cache_destroy(cachep);
2626 mutex_unlock(&cache_chain_mutex);
2627 put_online_cpus();
2629 EXPORT_SYMBOL(kmem_cache_destroy);
2632 * Get the memory for a slab management obj.
2633 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2634 * always come from malloc_sizes caches. The slab descriptor cannot
2635 * come from the same cache which is getting created because,
2636 * when we are searching for an appropriate cache for these
2637 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2638 * If we are creating a malloc_sizes cache here it would not be visible to
2639 * kmem_find_general_cachep till the initialization is complete.
2640 * Hence we cannot have slabp_cache same as the original cache.
2642 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2643 int colour_off, gfp_t local_flags,
2644 int nodeid)
2646 struct slab *slabp;
2648 if (OFF_SLAB(cachep)) {
2649 /* Slab management obj is off-slab. */
2650 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2651 local_flags, nodeid);
2653 * If the first object in the slab is leaked (it's allocated
2654 * but no one has a reference to it), we want to make sure
2655 * kmemleak does not treat the ->s_mem pointer as a reference
2656 * to the object. Otherwise we will not report the leak.
2658 kmemleak_scan_area(slabp, offsetof(struct slab, list),
2659 sizeof(struct list_head), local_flags);
2660 if (!slabp)
2661 return NULL;
2662 } else {
2663 slabp = objp + colour_off;
2664 colour_off += cachep->slab_size;
2666 slabp->inuse = 0;
2667 slabp->colouroff = colour_off;
2668 slabp->s_mem = objp + colour_off;
2669 slabp->nodeid = nodeid;
2670 slabp->free = 0;
2671 return slabp;
2674 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2676 return (kmem_bufctl_t *) (slabp + 1);
2679 static void cache_init_objs(struct kmem_cache *cachep,
2680 struct slab *slabp)
2682 int i;
2684 for (i = 0; i < cachep->num; i++) {
2685 void *objp = index_to_obj(cachep, slabp, i);
2686 #if DEBUG
2687 /* need to poison the objs? */
2688 if (cachep->flags & SLAB_POISON)
2689 poison_obj(cachep, objp, POISON_FREE);
2690 if (cachep->flags & SLAB_STORE_USER)
2691 *dbg_userword(cachep, objp) = NULL;
2693 if (cachep->flags & SLAB_RED_ZONE) {
2694 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2695 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2698 * Constructors are not allowed to allocate memory from the same
2699 * cache which they are a constructor for. Otherwise, deadlock.
2700 * They must also be threaded.
2702 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2703 cachep->ctor(objp + obj_offset(cachep));
2705 if (cachep->flags & SLAB_RED_ZONE) {
2706 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2707 slab_error(cachep, "constructor overwrote the"
2708 " end of an object");
2709 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2710 slab_error(cachep, "constructor overwrote the"
2711 " start of an object");
2713 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2714 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2715 kernel_map_pages(virt_to_page(objp),
2716 cachep->buffer_size / PAGE_SIZE, 0);
2717 #else
2718 if (cachep->ctor)
2719 cachep->ctor(objp);
2720 #endif
2721 slab_bufctl(slabp)[i] = i + 1;
2723 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2726 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2728 if (CONFIG_ZONE_DMA_FLAG) {
2729 if (flags & GFP_DMA)
2730 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2731 else
2732 BUG_ON(cachep->gfpflags & GFP_DMA);
2736 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2737 int nodeid)
2739 void *objp = index_to_obj(cachep, slabp, slabp->free);
2740 kmem_bufctl_t next;
2742 slabp->inuse++;
2743 next = slab_bufctl(slabp)[slabp->free];
2744 #if DEBUG
2745 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2746 WARN_ON(slabp->nodeid != nodeid);
2747 #endif
2748 slabp->free = next;
2750 return objp;
2753 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2754 void *objp, int nodeid)
2756 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2758 #if DEBUG
2759 /* Verify that the slab belongs to the intended node */
2760 WARN_ON(slabp->nodeid != nodeid);
2762 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2763 printk(KERN_ERR "slab: double free detected in cache "
2764 "'%s', objp %p\n", cachep->name, objp);
2765 BUG();
2767 #endif
2768 slab_bufctl(slabp)[objnr] = slabp->free;
2769 slabp->free = objnr;
2770 slabp->inuse--;
2774 * Map pages beginning at addr to the given cache and slab. This is required
2775 * for the slab allocator to be able to lookup the cache and slab of a
2776 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2778 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2779 void *addr)
2781 int nr_pages;
2782 struct page *page;
2784 page = virt_to_page(addr);
2786 nr_pages = 1;
2787 if (likely(!PageCompound(page)))
2788 nr_pages <<= cache->gfporder;
2790 do {
2791 page_set_cache(page, cache);
2792 page_set_slab(page, slab);
2793 page++;
2794 } while (--nr_pages);
2798 * Grow (by 1) the number of slabs within a cache. This is called by
2799 * kmem_cache_alloc() when there are no active objs left in a cache.
2801 static int cache_grow(struct kmem_cache *cachep,
2802 gfp_t flags, int nodeid, void *objp)
2804 struct slab *slabp;
2805 size_t offset;
2806 gfp_t local_flags;
2807 struct kmem_list3 *l3;
2810 * Be lazy and only check for valid flags here, keeping it out of the
2811 * critical path in kmem_cache_alloc().
2813 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2814 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2816 /* Take the l3 list lock to change the colour_next on this node */
2817 check_irq_off();
2818 l3 = cachep->nodelists[nodeid];
2819 spin_lock(&l3->list_lock);
2821 /* Get colour for the slab, and cal the next value. */
2822 offset = l3->colour_next;
2823 l3->colour_next++;
2824 if (l3->colour_next >= cachep->colour)
2825 l3->colour_next = 0;
2826 spin_unlock(&l3->list_lock);
2828 offset *= cachep->colour_off;
2830 if (local_flags & __GFP_WAIT)
2831 local_irq_enable();
2834 * The test for missing atomic flag is performed here, rather than
2835 * the more obvious place, simply to reduce the critical path length
2836 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2837 * will eventually be caught here (where it matters).
2839 kmem_flagcheck(cachep, flags);
2842 * Get mem for the objs. Attempt to allocate a physical page from
2843 * 'nodeid'.
2845 if (!objp)
2846 objp = kmem_getpages(cachep, local_flags, nodeid);
2847 if (!objp)
2848 goto failed;
2850 /* Get slab management. */
2851 slabp = alloc_slabmgmt(cachep, objp, offset,
2852 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2853 if (!slabp)
2854 goto opps1;
2856 slab_map_pages(cachep, slabp, objp);
2858 cache_init_objs(cachep, slabp);
2860 if (local_flags & __GFP_WAIT)
2861 local_irq_disable();
2862 check_irq_off();
2863 spin_lock(&l3->list_lock);
2865 /* Make slab active. */
2866 list_add_tail(&slabp->list, &(l3->slabs_free));
2867 STATS_INC_GROWN(cachep);
2868 l3->free_objects += cachep->num;
2869 spin_unlock(&l3->list_lock);
2870 return 1;
2871 opps1:
2872 kmem_freepages(cachep, objp);
2873 failed:
2874 if (local_flags & __GFP_WAIT)
2875 local_irq_disable();
2876 return 0;
2879 #if DEBUG
2882 * Perform extra freeing checks:
2883 * - detect bad pointers.
2884 * - POISON/RED_ZONE checking
2886 static void kfree_debugcheck(const void *objp)
2888 if (!virt_addr_valid(objp)) {
2889 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2890 (unsigned long)objp);
2891 BUG();
2895 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2897 unsigned long long redzone1, redzone2;
2899 redzone1 = *dbg_redzone1(cache, obj);
2900 redzone2 = *dbg_redzone2(cache, obj);
2903 * Redzone is ok.
2905 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2906 return;
2908 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2909 slab_error(cache, "double free detected");
2910 else
2911 slab_error(cache, "memory outside object was overwritten");
2913 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2914 obj, redzone1, redzone2);
2917 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2918 void *caller)
2920 struct page *page;
2921 unsigned int objnr;
2922 struct slab *slabp;
2924 BUG_ON(virt_to_cache(objp) != cachep);
2926 objp -= obj_offset(cachep);
2927 kfree_debugcheck(objp);
2928 page = virt_to_head_page(objp);
2930 slabp = page_get_slab(page);
2932 if (cachep->flags & SLAB_RED_ZONE) {
2933 verify_redzone_free(cachep, objp);
2934 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2935 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2937 if (cachep->flags & SLAB_STORE_USER)
2938 *dbg_userword(cachep, objp) = caller;
2940 objnr = obj_to_index(cachep, slabp, objp);
2942 BUG_ON(objnr >= cachep->num);
2943 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
2945 #ifdef CONFIG_DEBUG_SLAB_LEAK
2946 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2947 #endif
2948 if (cachep->flags & SLAB_POISON) {
2949 #ifdef CONFIG_DEBUG_PAGEALLOC
2950 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
2951 store_stackinfo(cachep, objp, (unsigned long)caller);
2952 kernel_map_pages(virt_to_page(objp),
2953 cachep->buffer_size / PAGE_SIZE, 0);
2954 } else {
2955 poison_obj(cachep, objp, POISON_FREE);
2957 #else
2958 poison_obj(cachep, objp, POISON_FREE);
2959 #endif
2961 return objp;
2964 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
2966 kmem_bufctl_t i;
2967 int entries = 0;
2969 /* Check slab's freelist to see if this obj is there. */
2970 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2971 entries++;
2972 if (entries > cachep->num || i >= cachep->num)
2973 goto bad;
2975 if (entries != cachep->num - slabp->inuse) {
2976 bad:
2977 printk(KERN_ERR "slab: Internal list corruption detected in "
2978 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2979 cachep->name, cachep->num, slabp, slabp->inuse);
2980 for (i = 0;
2981 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
2982 i++) {
2983 if (i % 16 == 0)
2984 printk("\n%03x:", i);
2985 printk(" %02x", ((unsigned char *)slabp)[i]);
2987 printk("\n");
2988 BUG();
2991 #else
2992 #define kfree_debugcheck(x) do { } while(0)
2993 #define cache_free_debugcheck(x,objp,z) (objp)
2994 #define check_slabp(x,y) do { } while(0)
2995 #endif
2997 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
2999 int batchcount;
3000 struct kmem_list3 *l3;
3001 struct array_cache *ac;
3002 int node;
3004 retry:
3005 check_irq_off();
3006 node = numa_node_id();
3007 ac = cpu_cache_get(cachep);
3008 batchcount = ac->batchcount;
3009 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
3011 * If there was little recent activity on this cache, then
3012 * perform only a partial refill. Otherwise we could generate
3013 * refill bouncing.
3015 batchcount = BATCHREFILL_LIMIT;
3017 l3 = cachep->nodelists[node];
3019 BUG_ON(ac->avail > 0 || !l3);
3020 spin_lock(&l3->list_lock);
3022 /* See if we can refill from the shared array */
3023 if (l3->shared && transfer_objects(ac, l3->shared, batchcount))
3024 goto alloc_done;
3026 while (batchcount > 0) {
3027 struct list_head *entry;
3028 struct slab *slabp;
3029 /* Get slab alloc is to come from. */
3030 entry = l3->slabs_partial.next;
3031 if (entry == &l3->slabs_partial) {
3032 l3->free_touched = 1;
3033 entry = l3->slabs_free.next;
3034 if (entry == &l3->slabs_free)
3035 goto must_grow;
3038 slabp = list_entry(entry, struct slab, list);
3039 check_slabp(cachep, slabp);
3040 check_spinlock_acquired(cachep);
3043 * The slab was either on partial or free list so
3044 * there must be at least one object available for
3045 * allocation.
3047 BUG_ON(slabp->inuse >= cachep->num);
3049 while (slabp->inuse < cachep->num && batchcount--) {
3050 STATS_INC_ALLOCED(cachep);
3051 STATS_INC_ACTIVE(cachep);
3052 STATS_SET_HIGH(cachep);
3054 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
3055 node);
3057 check_slabp(cachep, slabp);
3059 /* move slabp to correct slabp list: */
3060 list_del(&slabp->list);
3061 if (slabp->free == BUFCTL_END)
3062 list_add(&slabp->list, &l3->slabs_full);
3063 else
3064 list_add(&slabp->list, &l3->slabs_partial);
3067 must_grow:
3068 l3->free_objects -= ac->avail;
3069 alloc_done:
3070 spin_unlock(&l3->list_lock);
3072 if (unlikely(!ac->avail)) {
3073 int x;
3074 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
3076 /* cache_grow can reenable interrupts, then ac could change. */
3077 ac = cpu_cache_get(cachep);
3078 if (!x && ac->avail == 0) /* no objects in sight? abort */
3079 return NULL;
3081 if (!ac->avail) /* objects refilled by interrupt? */
3082 goto retry;
3084 ac->touched = 1;
3085 return ac->entry[--ac->avail];
3088 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3089 gfp_t flags)
3091 might_sleep_if(flags & __GFP_WAIT);
3092 #if DEBUG
3093 kmem_flagcheck(cachep, flags);
3094 #endif
3097 #if DEBUG
3098 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3099 gfp_t flags, void *objp, void *caller)
3101 if (!objp)
3102 return objp;
3103 if (cachep->flags & SLAB_POISON) {
3104 #ifdef CONFIG_DEBUG_PAGEALLOC
3105 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
3106 kernel_map_pages(virt_to_page(objp),
3107 cachep->buffer_size / PAGE_SIZE, 1);
3108 else
3109 check_poison_obj(cachep, objp);
3110 #else
3111 check_poison_obj(cachep, objp);
3112 #endif
3113 poison_obj(cachep, objp, POISON_INUSE);
3115 if (cachep->flags & SLAB_STORE_USER)
3116 *dbg_userword(cachep, objp) = caller;
3118 if (cachep->flags & SLAB_RED_ZONE) {
3119 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3120 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3121 slab_error(cachep, "double free, or memory outside"
3122 " object was overwritten");
3123 printk(KERN_ERR
3124 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3125 objp, *dbg_redzone1(cachep, objp),
3126 *dbg_redzone2(cachep, objp));
3128 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3129 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3131 #ifdef CONFIG_DEBUG_SLAB_LEAK
3133 struct slab *slabp;
3134 unsigned objnr;
3136 slabp = page_get_slab(virt_to_head_page(objp));
3137 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3138 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3140 #endif
3141 objp += obj_offset(cachep);
3142 if (cachep->ctor && cachep->flags & SLAB_POISON)
3143 cachep->ctor(objp);
3144 #if ARCH_SLAB_MINALIGN
3145 if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) {
3146 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3147 objp, ARCH_SLAB_MINALIGN);
3149 #endif
3150 return objp;
3152 #else
3153 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3154 #endif
3156 static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
3158 if (cachep == &cache_cache)
3159 return false;
3161 return should_failslab(obj_size(cachep), flags);
3164 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3166 void *objp;
3167 struct array_cache *ac;
3169 check_irq_off();
3171 ac = cpu_cache_get(cachep);
3172 if (likely(ac->avail)) {
3173 STATS_INC_ALLOCHIT(cachep);
3174 ac->touched = 1;
3175 objp = ac->entry[--ac->avail];
3176 } else {
3177 STATS_INC_ALLOCMISS(cachep);
3178 objp = cache_alloc_refill(cachep, flags);
3181 * To avoid a false negative, if an object that is in one of the
3182 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3183 * treat the array pointers as a reference to the object.
3185 kmemleak_erase(&ac->entry[ac->avail]);
3186 return objp;
3189 #ifdef CONFIG_NUMA
3191 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3193 * If we are in_interrupt, then process context, including cpusets and
3194 * mempolicy, may not apply and should not be used for allocation policy.
3196 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3198 int nid_alloc, nid_here;
3200 if (in_interrupt() || (flags & __GFP_THISNODE))
3201 return NULL;
3202 nid_alloc = nid_here = numa_node_id();
3203 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3204 nid_alloc = cpuset_mem_spread_node();
3205 else if (current->mempolicy)
3206 nid_alloc = slab_node(current->mempolicy);
3207 if (nid_alloc != nid_here)
3208 return ____cache_alloc_node(cachep, flags, nid_alloc);
3209 return NULL;
3213 * Fallback function if there was no memory available and no objects on a
3214 * certain node and fall back is permitted. First we scan all the
3215 * available nodelists for available objects. If that fails then we
3216 * perform an allocation without specifying a node. This allows the page
3217 * allocator to do its reclaim / fallback magic. We then insert the
3218 * slab into the proper nodelist and then allocate from it.
3220 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3222 struct zonelist *zonelist;
3223 gfp_t local_flags;
3224 struct zoneref *z;
3225 struct zone *zone;
3226 enum zone_type high_zoneidx = gfp_zone(flags);
3227 void *obj = NULL;
3228 int nid;
3230 if (flags & __GFP_THISNODE)
3231 return NULL;
3233 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
3234 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3236 retry:
3238 * Look through allowed nodes for objects available
3239 * from existing per node queues.
3241 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3242 nid = zone_to_nid(zone);
3244 if (cpuset_zone_allowed_hardwall(zone, flags) &&
3245 cache->nodelists[nid] &&
3246 cache->nodelists[nid]->free_objects) {
3247 obj = ____cache_alloc_node(cache,
3248 flags | GFP_THISNODE, nid);
3249 if (obj)
3250 break;
3254 if (!obj) {
3256 * This allocation will be performed within the constraints
3257 * of the current cpuset / memory policy requirements.
3258 * We may trigger various forms of reclaim on the allowed
3259 * set and go into memory reserves if necessary.
3261 if (local_flags & __GFP_WAIT)
3262 local_irq_enable();
3263 kmem_flagcheck(cache, flags);
3264 obj = kmem_getpages(cache, local_flags, -1);
3265 if (local_flags & __GFP_WAIT)
3266 local_irq_disable();
3267 if (obj) {
3269 * Insert into the appropriate per node queues
3271 nid = page_to_nid(virt_to_page(obj));
3272 if (cache_grow(cache, flags, nid, obj)) {
3273 obj = ____cache_alloc_node(cache,
3274 flags | GFP_THISNODE, nid);
3275 if (!obj)
3277 * Another processor may allocate the
3278 * objects in the slab since we are
3279 * not holding any locks.
3281 goto retry;
3282 } else {
3283 /* cache_grow already freed obj */
3284 obj = NULL;
3288 return obj;
3292 * A interface to enable slab creation on nodeid
3294 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3295 int nodeid)
3297 struct list_head *entry;
3298 struct slab *slabp;
3299 struct kmem_list3 *l3;
3300 void *obj;
3301 int x;
3303 l3 = cachep->nodelists[nodeid];
3304 BUG_ON(!l3);
3306 retry:
3307 check_irq_off();
3308 spin_lock(&l3->list_lock);
3309 entry = l3->slabs_partial.next;
3310 if (entry == &l3->slabs_partial) {
3311 l3->free_touched = 1;
3312 entry = l3->slabs_free.next;
3313 if (entry == &l3->slabs_free)
3314 goto must_grow;
3317 slabp = list_entry(entry, struct slab, list);
3318 check_spinlock_acquired_node(cachep, nodeid);
3319 check_slabp(cachep, slabp);
3321 STATS_INC_NODEALLOCS(cachep);
3322 STATS_INC_ACTIVE(cachep);
3323 STATS_SET_HIGH(cachep);
3325 BUG_ON(slabp->inuse == cachep->num);
3327 obj = slab_get_obj(cachep, slabp, nodeid);
3328 check_slabp(cachep, slabp);
3329 l3->free_objects--;
3330 /* move slabp to correct slabp list: */
3331 list_del(&slabp->list);
3333 if (slabp->free == BUFCTL_END)
3334 list_add(&slabp->list, &l3->slabs_full);
3335 else
3336 list_add(&slabp->list, &l3->slabs_partial);
3338 spin_unlock(&l3->list_lock);
3339 goto done;
3341 must_grow:
3342 spin_unlock(&l3->list_lock);
3343 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
3344 if (x)
3345 goto retry;
3347 return fallback_alloc(cachep, flags);
3349 done:
3350 return obj;
3354 * kmem_cache_alloc_node - Allocate an object on the specified node
3355 * @cachep: The cache to allocate from.
3356 * @flags: See kmalloc().
3357 * @nodeid: node number of the target node.
3358 * @caller: return address of caller, used for debug information
3360 * Identical to kmem_cache_alloc but it will allocate memory on the given
3361 * node, which can improve the performance for cpu bound structures.
3363 * Fallback to other node is possible if __GFP_THISNODE is not set.
3365 static __always_inline void *
3366 __cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3367 void *caller)
3369 unsigned long save_flags;
3370 void *ptr;
3372 flags &= slab_gfp_mask;
3374 lockdep_trace_alloc(flags);
3376 if (slab_should_failslab(cachep, flags))
3377 return NULL;
3379 cache_alloc_debugcheck_before(cachep, flags);
3380 local_irq_save(save_flags);
3382 if (unlikely(nodeid == -1))
3383 nodeid = numa_node_id();
3385 if (unlikely(!cachep->nodelists[nodeid])) {
3386 /* Node not bootstrapped yet */
3387 ptr = fallback_alloc(cachep, flags);
3388 goto out;
3391 if (nodeid == numa_node_id()) {
3393 * Use the locally cached objects if possible.
3394 * However ____cache_alloc does not allow fallback
3395 * to other nodes. It may fail while we still have
3396 * objects on other nodes available.
3398 ptr = ____cache_alloc(cachep, flags);
3399 if (ptr)
3400 goto out;
3402 /* ___cache_alloc_node can fall back to other nodes */
3403 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3404 out:
3405 local_irq_restore(save_flags);
3406 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3407 kmemleak_alloc_recursive(ptr, obj_size(cachep), 1, cachep->flags,
3408 flags);
3410 if (unlikely((flags & __GFP_ZERO) && ptr))
3411 memset(ptr, 0, obj_size(cachep));
3413 return ptr;
3416 static __always_inline void *
3417 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3419 void *objp;
3421 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3422 objp = alternate_node_alloc(cache, flags);
3423 if (objp)
3424 goto out;
3426 objp = ____cache_alloc(cache, flags);
3429 * We may just have run out of memory on the local node.
3430 * ____cache_alloc_node() knows how to locate memory on other nodes
3432 if (!objp)
3433 objp = ____cache_alloc_node(cache, flags, numa_node_id());
3435 out:
3436 return objp;
3438 #else
3440 static __always_inline void *
3441 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3443 return ____cache_alloc(cachep, flags);
3446 #endif /* CONFIG_NUMA */
3448 static __always_inline void *
3449 __cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3451 unsigned long save_flags;
3452 void *objp;
3454 flags &= slab_gfp_mask;
3456 lockdep_trace_alloc(flags);
3458 if (slab_should_failslab(cachep, flags))
3459 return NULL;
3461 cache_alloc_debugcheck_before(cachep, flags);
3462 local_irq_save(save_flags);
3463 objp = __do_cache_alloc(cachep, flags);
3464 local_irq_restore(save_flags);
3465 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3466 kmemleak_alloc_recursive(objp, obj_size(cachep), 1, cachep->flags,
3467 flags);
3468 prefetchw(objp);
3470 if (unlikely((flags & __GFP_ZERO) && objp))
3471 memset(objp, 0, obj_size(cachep));
3473 return objp;
3477 * Caller needs to acquire correct kmem_list's list_lock
3479 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3480 int node)
3482 int i;
3483 struct kmem_list3 *l3;
3485 for (i = 0; i < nr_objects; i++) {
3486 void *objp = objpp[i];
3487 struct slab *slabp;
3489 slabp = virt_to_slab(objp);
3490 l3 = cachep->nodelists[node];
3491 list_del(&slabp->list);
3492 check_spinlock_acquired_node(cachep, node);
3493 check_slabp(cachep, slabp);
3494 slab_put_obj(cachep, slabp, objp, node);
3495 STATS_DEC_ACTIVE(cachep);
3496 l3->free_objects++;
3497 check_slabp(cachep, slabp);
3499 /* fixup slab chains */
3500 if (slabp->inuse == 0) {
3501 if (l3->free_objects > l3->free_limit) {
3502 l3->free_objects -= cachep->num;
3503 /* No need to drop any previously held
3504 * lock here, even if we have a off-slab slab
3505 * descriptor it is guaranteed to come from
3506 * a different cache, refer to comments before
3507 * alloc_slabmgmt.
3509 slab_destroy(cachep, slabp);
3510 } else {
3511 list_add(&slabp->list, &l3->slabs_free);
3513 } else {
3514 /* Unconditionally move a slab to the end of the
3515 * partial list on free - maximum time for the
3516 * other objects to be freed, too.
3518 list_add_tail(&slabp->list, &l3->slabs_partial);
3523 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3525 int batchcount;
3526 struct kmem_list3 *l3;
3527 int node = numa_node_id();
3529 batchcount = ac->batchcount;
3530 #if DEBUG
3531 BUG_ON(!batchcount || batchcount > ac->avail);
3532 #endif
3533 check_irq_off();
3534 l3 = cachep->nodelists[node];
3535 spin_lock(&l3->list_lock);
3536 if (l3->shared) {
3537 struct array_cache *shared_array = l3->shared;
3538 int max = shared_array->limit - shared_array->avail;
3539 if (max) {
3540 if (batchcount > max)
3541 batchcount = max;
3542 memcpy(&(shared_array->entry[shared_array->avail]),
3543 ac->entry, sizeof(void *) * batchcount);
3544 shared_array->avail += batchcount;
3545 goto free_done;
3549 free_block(cachep, ac->entry, batchcount, node);
3550 free_done:
3551 #if STATS
3553 int i = 0;
3554 struct list_head *p;
3556 p = l3->slabs_free.next;
3557 while (p != &(l3->slabs_free)) {
3558 struct slab *slabp;
3560 slabp = list_entry(p, struct slab, list);
3561 BUG_ON(slabp->inuse);
3563 i++;
3564 p = p->next;
3566 STATS_SET_FREEABLE(cachep, i);
3568 #endif
3569 spin_unlock(&l3->list_lock);
3570 ac->avail -= batchcount;
3571 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3575 * Release an obj back to its cache. If the obj has a constructed state, it must
3576 * be in this state _before_ it is released. Called with disabled ints.
3578 static inline void __cache_free(struct kmem_cache *cachep, void *objp)
3580 struct array_cache *ac = cpu_cache_get(cachep);
3582 check_irq_off();
3583 kmemleak_free_recursive(objp, cachep->flags);
3584 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3587 * Skip calling cache_free_alien() when the platform is not numa.
3588 * This will avoid cache misses that happen while accessing slabp (which
3589 * is per page memory reference) to get nodeid. Instead use a global
3590 * variable to skip the call, which is mostly likely to be present in
3591 * the cache.
3593 if (numa_platform && cache_free_alien(cachep, objp))
3594 return;
3596 if (likely(ac->avail < ac->limit)) {
3597 STATS_INC_FREEHIT(cachep);
3598 ac->entry[ac->avail++] = objp;
3599 return;
3600 } else {
3601 STATS_INC_FREEMISS(cachep);
3602 cache_flusharray(cachep, ac);
3603 ac->entry[ac->avail++] = objp;
3608 * kmem_cache_alloc - Allocate an object
3609 * @cachep: The cache to allocate from.
3610 * @flags: See kmalloc().
3612 * Allocate an object from this cache. The flags are only relevant
3613 * if the cache has no available objects.
3615 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3617 void *ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3619 trace_kmem_cache_alloc(_RET_IP_, ret,
3620 obj_size(cachep), cachep->buffer_size, flags);
3622 return ret;
3624 EXPORT_SYMBOL(kmem_cache_alloc);
3626 #ifdef CONFIG_KMEMTRACE
3627 void *kmem_cache_alloc_notrace(struct kmem_cache *cachep, gfp_t flags)
3629 return __cache_alloc(cachep, flags, __builtin_return_address(0));
3631 EXPORT_SYMBOL(kmem_cache_alloc_notrace);
3632 #endif
3635 * kmem_ptr_validate - check if an untrusted pointer might be a slab entry.
3636 * @cachep: the cache we're checking against
3637 * @ptr: pointer to validate
3639 * This verifies that the untrusted pointer looks sane;
3640 * it is _not_ a guarantee that the pointer is actually
3641 * part of the slab cache in question, but it at least
3642 * validates that the pointer can be dereferenced and
3643 * looks half-way sane.
3645 * Currently only used for dentry validation.
3647 int kmem_ptr_validate(struct kmem_cache *cachep, const void *ptr)
3649 unsigned long addr = (unsigned long)ptr;
3650 unsigned long min_addr = PAGE_OFFSET;
3651 unsigned long align_mask = BYTES_PER_WORD - 1;
3652 unsigned long size = cachep->buffer_size;
3653 struct page *page;
3655 if (unlikely(addr < min_addr))
3656 goto out;
3657 if (unlikely(addr > (unsigned long)high_memory - size))
3658 goto out;
3659 if (unlikely(addr & align_mask))
3660 goto out;
3661 if (unlikely(!kern_addr_valid(addr)))
3662 goto out;
3663 if (unlikely(!kern_addr_valid(addr + size - 1)))
3664 goto out;
3665 page = virt_to_page(ptr);
3666 if (unlikely(!PageSlab(page)))
3667 goto out;
3668 if (unlikely(page_get_cache(page) != cachep))
3669 goto out;
3670 return 1;
3671 out:
3672 return 0;
3675 #ifdef CONFIG_NUMA
3676 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3678 void *ret = __cache_alloc_node(cachep, flags, nodeid,
3679 __builtin_return_address(0));
3681 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3682 obj_size(cachep), cachep->buffer_size,
3683 flags, nodeid);
3685 return ret;
3687 EXPORT_SYMBOL(kmem_cache_alloc_node);
3689 #ifdef CONFIG_KMEMTRACE
3690 void *kmem_cache_alloc_node_notrace(struct kmem_cache *cachep,
3691 gfp_t flags,
3692 int nodeid)
3694 return __cache_alloc_node(cachep, flags, nodeid,
3695 __builtin_return_address(0));
3697 EXPORT_SYMBOL(kmem_cache_alloc_node_notrace);
3698 #endif
3700 static __always_inline void *
3701 __do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
3703 struct kmem_cache *cachep;
3704 void *ret;
3706 cachep = kmem_find_general_cachep(size, flags);
3707 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3708 return cachep;
3709 ret = kmem_cache_alloc_node_notrace(cachep, flags, node);
3711 trace_kmalloc_node((unsigned long) caller, ret,
3712 size, cachep->buffer_size, flags, node);
3714 return ret;
3717 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_KMEMTRACE)
3718 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3720 return __do_kmalloc_node(size, flags, node,
3721 __builtin_return_address(0));
3723 EXPORT_SYMBOL(__kmalloc_node);
3725 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3726 int node, unsigned long caller)
3728 return __do_kmalloc_node(size, flags, node, (void *)caller);
3730 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3731 #else
3732 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3734 return __do_kmalloc_node(size, flags, node, NULL);
3736 EXPORT_SYMBOL(__kmalloc_node);
3737 #endif /* CONFIG_DEBUG_SLAB */
3738 #endif /* CONFIG_NUMA */
3741 * __do_kmalloc - allocate memory
3742 * @size: how many bytes of memory are required.
3743 * @flags: the type of memory to allocate (see kmalloc).
3744 * @caller: function caller for debug tracking of the caller
3746 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3747 void *caller)
3749 struct kmem_cache *cachep;
3750 void *ret;
3752 /* If you want to save a few bytes .text space: replace
3753 * __ with kmem_.
3754 * Then kmalloc uses the uninlined functions instead of the inline
3755 * functions.
3757 cachep = __find_general_cachep(size, flags);
3758 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3759 return cachep;
3760 ret = __cache_alloc(cachep, flags, caller);
3762 trace_kmalloc((unsigned long) caller, ret,
3763 size, cachep->buffer_size, flags);
3765 return ret;
3769 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_KMEMTRACE)
3770 void *__kmalloc(size_t size, gfp_t flags)
3772 return __do_kmalloc(size, flags, __builtin_return_address(0));
3774 EXPORT_SYMBOL(__kmalloc);
3776 void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3778 return __do_kmalloc(size, flags, (void *)caller);
3780 EXPORT_SYMBOL(__kmalloc_track_caller);
3782 #else
3783 void *__kmalloc(size_t size, gfp_t flags)
3785 return __do_kmalloc(size, flags, NULL);
3787 EXPORT_SYMBOL(__kmalloc);
3788 #endif
3791 * kmem_cache_free - Deallocate an object
3792 * @cachep: The cache the allocation was from.
3793 * @objp: The previously allocated object.
3795 * Free an object which was previously allocated from this
3796 * cache.
3798 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3800 unsigned long flags;
3802 local_irq_save(flags);
3803 debug_check_no_locks_freed(objp, obj_size(cachep));
3804 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3805 debug_check_no_obj_freed(objp, obj_size(cachep));
3806 __cache_free(cachep, objp);
3807 local_irq_restore(flags);
3809 trace_kmem_cache_free(_RET_IP_, objp);
3811 EXPORT_SYMBOL(kmem_cache_free);
3814 * kfree - free previously allocated memory
3815 * @objp: pointer returned by kmalloc.
3817 * If @objp is NULL, no operation is performed.
3819 * Don't free memory not originally allocated by kmalloc()
3820 * or you will run into trouble.
3822 void kfree(const void *objp)
3824 struct kmem_cache *c;
3825 unsigned long flags;
3827 trace_kfree(_RET_IP_, objp);
3829 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3830 return;
3831 local_irq_save(flags);
3832 kfree_debugcheck(objp);
3833 c = virt_to_cache(objp);
3834 debug_check_no_locks_freed(objp, obj_size(c));
3835 debug_check_no_obj_freed(objp, obj_size(c));
3836 __cache_free(c, (void *)objp);
3837 local_irq_restore(flags);
3839 EXPORT_SYMBOL(kfree);
3841 unsigned int kmem_cache_size(struct kmem_cache *cachep)
3843 return obj_size(cachep);
3845 EXPORT_SYMBOL(kmem_cache_size);
3847 const char *kmem_cache_name(struct kmem_cache *cachep)
3849 return cachep->name;
3851 EXPORT_SYMBOL_GPL(kmem_cache_name);
3854 * This initializes kmem_list3 or resizes various caches for all nodes.
3856 static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
3858 int node;
3859 struct kmem_list3 *l3;
3860 struct array_cache *new_shared;
3861 struct array_cache **new_alien = NULL;
3863 for_each_online_node(node) {
3865 if (use_alien_caches) {
3866 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
3867 if (!new_alien)
3868 goto fail;
3871 new_shared = NULL;
3872 if (cachep->shared) {
3873 new_shared = alloc_arraycache(node,
3874 cachep->shared*cachep->batchcount,
3875 0xbaadf00d, gfp);
3876 if (!new_shared) {
3877 free_alien_cache(new_alien);
3878 goto fail;
3882 l3 = cachep->nodelists[node];
3883 if (l3) {
3884 struct array_cache *shared = l3->shared;
3886 spin_lock_irq(&l3->list_lock);
3888 if (shared)
3889 free_block(cachep, shared->entry,
3890 shared->avail, node);
3892 l3->shared = new_shared;
3893 if (!l3->alien) {
3894 l3->alien = new_alien;
3895 new_alien = NULL;
3897 l3->free_limit = (1 + nr_cpus_node(node)) *
3898 cachep->batchcount + cachep->num;
3899 spin_unlock_irq(&l3->list_lock);
3900 kfree(shared);
3901 free_alien_cache(new_alien);
3902 continue;
3904 l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node);
3905 if (!l3) {
3906 free_alien_cache(new_alien);
3907 kfree(new_shared);
3908 goto fail;
3911 kmem_list3_init(l3);
3912 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3913 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3914 l3->shared = new_shared;
3915 l3->alien = new_alien;
3916 l3->free_limit = (1 + nr_cpus_node(node)) *
3917 cachep->batchcount + cachep->num;
3918 cachep->nodelists[node] = l3;
3920 return 0;
3922 fail:
3923 if (!cachep->next.next) {
3924 /* Cache is not active yet. Roll back what we did */
3925 node--;
3926 while (node >= 0) {
3927 if (cachep->nodelists[node]) {
3928 l3 = cachep->nodelists[node];
3930 kfree(l3->shared);
3931 free_alien_cache(l3->alien);
3932 kfree(l3);
3933 cachep->nodelists[node] = NULL;
3935 node--;
3938 return -ENOMEM;
3941 struct ccupdate_struct {
3942 struct kmem_cache *cachep;
3943 struct array_cache *new[NR_CPUS];
3946 static void do_ccupdate_local(void *info)
3948 struct ccupdate_struct *new = info;
3949 struct array_cache *old;
3951 check_irq_off();
3952 old = cpu_cache_get(new->cachep);
3954 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3955 new->new[smp_processor_id()] = old;
3958 /* Always called with the cache_chain_mutex held */
3959 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3960 int batchcount, int shared, gfp_t gfp)
3962 struct ccupdate_struct *new;
3963 int i;
3965 new = kzalloc(sizeof(*new), gfp);
3966 if (!new)
3967 return -ENOMEM;
3969 for_each_online_cpu(i) {
3970 new->new[i] = alloc_arraycache(cpu_to_node(i), limit,
3971 batchcount, gfp);
3972 if (!new->new[i]) {
3973 for (i--; i >= 0; i--)
3974 kfree(new->new[i]);
3975 kfree(new);
3976 return -ENOMEM;
3979 new->cachep = cachep;
3981 on_each_cpu(do_ccupdate_local, (void *)new, 1);
3983 check_irq_on();
3984 cachep->batchcount = batchcount;
3985 cachep->limit = limit;
3986 cachep->shared = shared;
3988 for_each_online_cpu(i) {
3989 struct array_cache *ccold = new->new[i];
3990 if (!ccold)
3991 continue;
3992 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3993 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
3994 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
3995 kfree(ccold);
3997 kfree(new);
3998 return alloc_kmemlist(cachep, gfp);
4001 /* Called with cache_chain_mutex held always */
4002 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
4004 int err;
4005 int limit, shared;
4008 * The head array serves three purposes:
4009 * - create a LIFO ordering, i.e. return objects that are cache-warm
4010 * - reduce the number of spinlock operations.
4011 * - reduce the number of linked list operations on the slab and
4012 * bufctl chains: array operations are cheaper.
4013 * The numbers are guessed, we should auto-tune as described by
4014 * Bonwick.
4016 if (cachep->buffer_size > 131072)
4017 limit = 1;
4018 else if (cachep->buffer_size > PAGE_SIZE)
4019 limit = 8;
4020 else if (cachep->buffer_size > 1024)
4021 limit = 24;
4022 else if (cachep->buffer_size > 256)
4023 limit = 54;
4024 else
4025 limit = 120;
4028 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
4029 * allocation behaviour: Most allocs on one cpu, most free operations
4030 * on another cpu. For these cases, an efficient object passing between
4031 * cpus is necessary. This is provided by a shared array. The array
4032 * replaces Bonwick's magazine layer.
4033 * On uniprocessor, it's functionally equivalent (but less efficient)
4034 * to a larger limit. Thus disabled by default.
4036 shared = 0;
4037 if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1)
4038 shared = 8;
4040 #if DEBUG
4042 * With debugging enabled, large batchcount lead to excessively long
4043 * periods with disabled local interrupts. Limit the batchcount
4045 if (limit > 32)
4046 limit = 32;
4047 #endif
4048 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp);
4049 if (err)
4050 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
4051 cachep->name, -err);
4052 return err;
4056 * Drain an array if it contains any elements taking the l3 lock only if
4057 * necessary. Note that the l3 listlock also protects the array_cache
4058 * if drain_array() is used on the shared array.
4060 void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
4061 struct array_cache *ac, int force, int node)
4063 int tofree;
4065 if (!ac || !ac->avail)
4066 return;
4067 if (ac->touched && !force) {
4068 ac->touched = 0;
4069 } else {
4070 spin_lock_irq(&l3->list_lock);
4071 if (ac->avail) {
4072 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4073 if (tofree > ac->avail)
4074 tofree = (ac->avail + 1) / 2;
4075 free_block(cachep, ac->entry, tofree, node);
4076 ac->avail -= tofree;
4077 memmove(ac->entry, &(ac->entry[tofree]),
4078 sizeof(void *) * ac->avail);
4080 spin_unlock_irq(&l3->list_lock);
4085 * cache_reap - Reclaim memory from caches.
4086 * @w: work descriptor
4088 * Called from workqueue/eventd every few seconds.
4089 * Purpose:
4090 * - clear the per-cpu caches for this CPU.
4091 * - return freeable pages to the main free memory pool.
4093 * If we cannot acquire the cache chain mutex then just give up - we'll try
4094 * again on the next iteration.
4096 static void cache_reap(struct work_struct *w)
4098 struct kmem_cache *searchp;
4099 struct kmem_list3 *l3;
4100 int node = numa_node_id();
4101 struct delayed_work *work = to_delayed_work(w);
4103 if (!mutex_trylock(&cache_chain_mutex))
4104 /* Give up. Setup the next iteration. */
4105 goto out;
4107 list_for_each_entry(searchp, &cache_chain, next) {
4108 check_irq_on();
4111 * We only take the l3 lock if absolutely necessary and we
4112 * have established with reasonable certainty that
4113 * we can do some work if the lock was obtained.
4115 l3 = searchp->nodelists[node];
4117 reap_alien(searchp, l3);
4119 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
4122 * These are racy checks but it does not matter
4123 * if we skip one check or scan twice.
4125 if (time_after(l3->next_reap, jiffies))
4126 goto next;
4128 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
4130 drain_array(searchp, l3, l3->shared, 0, node);
4132 if (l3->free_touched)
4133 l3->free_touched = 0;
4134 else {
4135 int freed;
4137 freed = drain_freelist(searchp, l3, (l3->free_limit +
4138 5 * searchp->num - 1) / (5 * searchp->num));
4139 STATS_ADD_REAPED(searchp, freed);
4141 next:
4142 cond_resched();
4144 check_irq_on();
4145 mutex_unlock(&cache_chain_mutex);
4146 next_reap_node();
4147 out:
4148 /* Set up the next iteration */
4149 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
4152 #ifdef CONFIG_SLABINFO
4154 static void print_slabinfo_header(struct seq_file *m)
4157 * Output format version, so at least we can change it
4158 * without _too_ many complaints.
4160 #if STATS
4161 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4162 #else
4163 seq_puts(m, "slabinfo - version: 2.1\n");
4164 #endif
4165 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4166 "<objperslab> <pagesperslab>");
4167 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4168 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4169 #if STATS
4170 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
4171 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
4172 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4173 #endif
4174 seq_putc(m, '\n');
4177 static void *s_start(struct seq_file *m, loff_t *pos)
4179 loff_t n = *pos;
4181 mutex_lock(&cache_chain_mutex);
4182 if (!n)
4183 print_slabinfo_header(m);
4185 return seq_list_start(&cache_chain, *pos);
4188 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4190 return seq_list_next(p, &cache_chain, pos);
4193 static void s_stop(struct seq_file *m, void *p)
4195 mutex_unlock(&cache_chain_mutex);
4198 static int s_show(struct seq_file *m, void *p)
4200 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
4201 struct slab *slabp;
4202 unsigned long active_objs;
4203 unsigned long num_objs;
4204 unsigned long active_slabs = 0;
4205 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4206 const char *name;
4207 char *error = NULL;
4208 int node;
4209 struct kmem_list3 *l3;
4211 active_objs = 0;
4212 num_slabs = 0;
4213 for_each_online_node(node) {
4214 l3 = cachep->nodelists[node];
4215 if (!l3)
4216 continue;
4218 check_irq_on();
4219 spin_lock_irq(&l3->list_lock);
4221 list_for_each_entry(slabp, &l3->slabs_full, list) {
4222 if (slabp->inuse != cachep->num && !error)
4223 error = "slabs_full accounting error";
4224 active_objs += cachep->num;
4225 active_slabs++;
4227 list_for_each_entry(slabp, &l3->slabs_partial, list) {
4228 if (slabp->inuse == cachep->num && !error)
4229 error = "slabs_partial inuse accounting error";
4230 if (!slabp->inuse && !error)
4231 error = "slabs_partial/inuse accounting error";
4232 active_objs += slabp->inuse;
4233 active_slabs++;
4235 list_for_each_entry(slabp, &l3->slabs_free, list) {
4236 if (slabp->inuse && !error)
4237 error = "slabs_free/inuse accounting error";
4238 num_slabs++;
4240 free_objects += l3->free_objects;
4241 if (l3->shared)
4242 shared_avail += l3->shared->avail;
4244 spin_unlock_irq(&l3->list_lock);
4246 num_slabs += active_slabs;
4247 num_objs = num_slabs * cachep->num;
4248 if (num_objs - active_objs != free_objects && !error)
4249 error = "free_objects accounting error";
4251 name = cachep->name;
4252 if (error)
4253 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4255 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
4256 name, active_objs, num_objs, cachep->buffer_size,
4257 cachep->num, (1 << cachep->gfporder));
4258 seq_printf(m, " : tunables %4u %4u %4u",
4259 cachep->limit, cachep->batchcount, cachep->shared);
4260 seq_printf(m, " : slabdata %6lu %6lu %6lu",
4261 active_slabs, num_slabs, shared_avail);
4262 #if STATS
4263 { /* list3 stats */
4264 unsigned long high = cachep->high_mark;
4265 unsigned long allocs = cachep->num_allocations;
4266 unsigned long grown = cachep->grown;
4267 unsigned long reaped = cachep->reaped;
4268 unsigned long errors = cachep->errors;
4269 unsigned long max_freeable = cachep->max_freeable;
4270 unsigned long node_allocs = cachep->node_allocs;
4271 unsigned long node_frees = cachep->node_frees;
4272 unsigned long overflows = cachep->node_overflow;
4274 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
4275 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown,
4276 reaped, errors, max_freeable, node_allocs,
4277 node_frees, overflows);
4279 /* cpu stats */
4281 unsigned long allochit = atomic_read(&cachep->allochit);
4282 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4283 unsigned long freehit = atomic_read(&cachep->freehit);
4284 unsigned long freemiss = atomic_read(&cachep->freemiss);
4286 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4287 allochit, allocmiss, freehit, freemiss);
4289 #endif
4290 seq_putc(m, '\n');
4291 return 0;
4295 * slabinfo_op - iterator that generates /proc/slabinfo
4297 * Output layout:
4298 * cache-name
4299 * num-active-objs
4300 * total-objs
4301 * object size
4302 * num-active-slabs
4303 * total-slabs
4304 * num-pages-per-slab
4305 * + further values on SMP and with statistics enabled
4308 static const struct seq_operations slabinfo_op = {
4309 .start = s_start,
4310 .next = s_next,
4311 .stop = s_stop,
4312 .show = s_show,
4315 #define MAX_SLABINFO_WRITE 128
4317 * slabinfo_write - Tuning for the slab allocator
4318 * @file: unused
4319 * @buffer: user buffer
4320 * @count: data length
4321 * @ppos: unused
4323 ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4324 size_t count, loff_t *ppos)
4326 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4327 int limit, batchcount, shared, res;
4328 struct kmem_cache *cachep;
4330 if (count > MAX_SLABINFO_WRITE)
4331 return -EINVAL;
4332 if (copy_from_user(&kbuf, buffer, count))
4333 return -EFAULT;
4334 kbuf[MAX_SLABINFO_WRITE] = '\0';
4336 tmp = strchr(kbuf, ' ');
4337 if (!tmp)
4338 return -EINVAL;
4339 *tmp = '\0';
4340 tmp++;
4341 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4342 return -EINVAL;
4344 /* Find the cache in the chain of caches. */
4345 mutex_lock(&cache_chain_mutex);
4346 res = -EINVAL;
4347 list_for_each_entry(cachep, &cache_chain, next) {
4348 if (!strcmp(cachep->name, kbuf)) {
4349 if (limit < 1 || batchcount < 1 ||
4350 batchcount > limit || shared < 0) {
4351 res = 0;
4352 } else {
4353 res = do_tune_cpucache(cachep, limit,
4354 batchcount, shared,
4355 GFP_KERNEL);
4357 break;
4360 mutex_unlock(&cache_chain_mutex);
4361 if (res >= 0)
4362 res = count;
4363 return res;
4366 static int slabinfo_open(struct inode *inode, struct file *file)
4368 return seq_open(file, &slabinfo_op);
4371 static const struct file_operations proc_slabinfo_operations = {
4372 .open = slabinfo_open,
4373 .read = seq_read,
4374 .write = slabinfo_write,
4375 .llseek = seq_lseek,
4376 .release = seq_release,
4379 #ifdef CONFIG_DEBUG_SLAB_LEAK
4381 static void *leaks_start(struct seq_file *m, loff_t *pos)
4383 mutex_lock(&cache_chain_mutex);
4384 return seq_list_start(&cache_chain, *pos);
4387 static inline int add_caller(unsigned long *n, unsigned long v)
4389 unsigned long *p;
4390 int l;
4391 if (!v)
4392 return 1;
4393 l = n[1];
4394 p = n + 2;
4395 while (l) {
4396 int i = l/2;
4397 unsigned long *q = p + 2 * i;
4398 if (*q == v) {
4399 q[1]++;
4400 return 1;
4402 if (*q > v) {
4403 l = i;
4404 } else {
4405 p = q + 2;
4406 l -= i + 1;
4409 if (++n[1] == n[0])
4410 return 0;
4411 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4412 p[0] = v;
4413 p[1] = 1;
4414 return 1;
4417 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4419 void *p;
4420 int i;
4421 if (n[0] == n[1])
4422 return;
4423 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4424 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4425 continue;
4426 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4427 return;
4431 static void show_symbol(struct seq_file *m, unsigned long address)
4433 #ifdef CONFIG_KALLSYMS
4434 unsigned long offset, size;
4435 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4437 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4438 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4439 if (modname[0])
4440 seq_printf(m, " [%s]", modname);
4441 return;
4443 #endif
4444 seq_printf(m, "%p", (void *)address);
4447 static int leaks_show(struct seq_file *m, void *p)
4449 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
4450 struct slab *slabp;
4451 struct kmem_list3 *l3;
4452 const char *name;
4453 unsigned long *n = m->private;
4454 int node;
4455 int i;
4457 if (!(cachep->flags & SLAB_STORE_USER))
4458 return 0;
4459 if (!(cachep->flags & SLAB_RED_ZONE))
4460 return 0;
4462 /* OK, we can do it */
4464 n[1] = 0;
4466 for_each_online_node(node) {
4467 l3 = cachep->nodelists[node];
4468 if (!l3)
4469 continue;
4471 check_irq_on();
4472 spin_lock_irq(&l3->list_lock);
4474 list_for_each_entry(slabp, &l3->slabs_full, list)
4475 handle_slab(n, cachep, slabp);
4476 list_for_each_entry(slabp, &l3->slabs_partial, list)
4477 handle_slab(n, cachep, slabp);
4478 spin_unlock_irq(&l3->list_lock);
4480 name = cachep->name;
4481 if (n[0] == n[1]) {
4482 /* Increase the buffer size */
4483 mutex_unlock(&cache_chain_mutex);
4484 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4485 if (!m->private) {
4486 /* Too bad, we are really out */
4487 m->private = n;
4488 mutex_lock(&cache_chain_mutex);
4489 return -ENOMEM;
4491 *(unsigned long *)m->private = n[0] * 2;
4492 kfree(n);
4493 mutex_lock(&cache_chain_mutex);
4494 /* Now make sure this entry will be retried */
4495 m->count = m->size;
4496 return 0;
4498 for (i = 0; i < n[1]; i++) {
4499 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4500 show_symbol(m, n[2*i+2]);
4501 seq_putc(m, '\n');
4504 return 0;
4507 static const struct seq_operations slabstats_op = {
4508 .start = leaks_start,
4509 .next = s_next,
4510 .stop = s_stop,
4511 .show = leaks_show,
4514 static int slabstats_open(struct inode *inode, struct file *file)
4516 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4517 int ret = -ENOMEM;
4518 if (n) {
4519 ret = seq_open(file, &slabstats_op);
4520 if (!ret) {
4521 struct seq_file *m = file->private_data;
4522 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4523 m->private = n;
4524 n = NULL;
4526 kfree(n);
4528 return ret;
4531 static const struct file_operations proc_slabstats_operations = {
4532 .open = slabstats_open,
4533 .read = seq_read,
4534 .llseek = seq_lseek,
4535 .release = seq_release_private,
4537 #endif
4539 static int __init slab_proc_init(void)
4541 proc_create("slabinfo",S_IWUSR|S_IRUGO,NULL,&proc_slabinfo_operations);
4542 #ifdef CONFIG_DEBUG_SLAB_LEAK
4543 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4544 #endif
4545 return 0;
4547 module_init(slab_proc_init);
4548 #endif
4551 * ksize - get the actual amount of memory allocated for a given object
4552 * @objp: Pointer to the object
4554 * kmalloc may internally round up allocations and return more memory
4555 * than requested. ksize() can be used to determine the actual amount of
4556 * memory allocated. The caller may use this additional memory, even though
4557 * a smaller amount of memory was initially specified with the kmalloc call.
4558 * The caller must guarantee that objp points to a valid object previously
4559 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4560 * must not be freed during the duration of the call.
4562 size_t ksize(const void *objp)
4564 BUG_ON(!objp);
4565 if (unlikely(objp == ZERO_SIZE_PTR))
4566 return 0;
4568 return obj_size(virt_to_cache(objp));
4570 EXPORT_SYMBOL(ksize);