arm: mxs: include asm/processor.h for cpu_relax()
[linux-2.6/linux-acpi-2.6/ibm-acpi-2.6.git] / mm / slab.c
blobbcfa4987c8ae2c9fc273f1baf75d4461ed743052
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/rcupdate.h>
106 #include <linux/string.h>
107 #include <linux/uaccess.h>
108 #include <linux/nodemask.h>
109 #include <linux/kmemleak.h>
110 #include <linux/mempolicy.h>
111 #include <linux/mutex.h>
112 #include <linux/fault-inject.h>
113 #include <linux/rtmutex.h>
114 #include <linux/reciprocal_div.h>
115 #include <linux/debugobjects.h>
116 #include <linux/kmemcheck.h>
117 #include <linux/memory.h>
118 #include <linux/prefetch.h>
120 #include <asm/cacheflush.h>
121 #include <asm/tlbflush.h>
122 #include <asm/page.h>
125 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
126 * 0 for faster, smaller code (especially in the critical paths).
128 * STATS - 1 to collect stats for /proc/slabinfo.
129 * 0 for faster, smaller code (especially in the critical paths).
131 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
134 #ifdef CONFIG_DEBUG_SLAB
135 #define DEBUG 1
136 #define STATS 1
137 #define FORCED_DEBUG 1
138 #else
139 #define DEBUG 0
140 #define STATS 0
141 #define FORCED_DEBUG 0
142 #endif
144 /* Shouldn't this be in a header file somewhere? */
145 #define BYTES_PER_WORD sizeof(void *)
146 #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
148 #ifndef ARCH_KMALLOC_FLAGS
149 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
150 #endif
152 /* Legal flag mask for kmem_cache_create(). */
153 #if DEBUG
154 # define CREATE_MASK (SLAB_RED_ZONE | \
155 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
156 SLAB_CACHE_DMA | \
157 SLAB_STORE_USER | \
158 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
159 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
160 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
161 #else
162 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
163 SLAB_CACHE_DMA | \
164 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
165 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
166 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
167 #endif
170 * kmem_bufctl_t:
172 * Bufctl's are used for linking objs within a slab
173 * linked offsets.
175 * This implementation relies on "struct page" for locating the cache &
176 * slab an object belongs to.
177 * This allows the bufctl structure to be small (one int), but limits
178 * the number of objects a slab (not a cache) can contain when off-slab
179 * bufctls are used. The limit is the size of the largest general cache
180 * that does not use off-slab slabs.
181 * For 32bit archs with 4 kB pages, is this 56.
182 * This is not serious, as it is only for large objects, when it is unwise
183 * to have too many per slab.
184 * Note: This limit can be raised by introducing a general cache whose size
185 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
188 typedef unsigned int kmem_bufctl_t;
189 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
190 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
191 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
192 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
195 * struct slab_rcu
197 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
198 * arrange for kmem_freepages to be called via RCU. This is useful if
199 * we need to approach a kernel structure obliquely, from its address
200 * obtained without the usual locking. We can lock the structure to
201 * stabilize it and check it's still at the given address, only if we
202 * can be sure that the memory has not been meanwhile reused for some
203 * other kind of object (which our subsystem's lock might corrupt).
205 * rcu_read_lock before reading the address, then rcu_read_unlock after
206 * taking the spinlock within the structure expected at that address.
208 struct slab_rcu {
209 struct rcu_head head;
210 struct kmem_cache *cachep;
211 void *addr;
215 * struct slab
217 * Manages the objs in a slab. Placed either at the beginning of mem allocated
218 * for a slab, or allocated from an general cache.
219 * Slabs are chained into three list: fully used, partial, fully free slabs.
221 struct slab {
222 union {
223 struct {
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;
231 struct slab_rcu __slab_cover_slab_rcu;
236 * struct array_cache
238 * Purpose:
239 * - LIFO ordering, to hand out cache-warm objects from _alloc
240 * - reduce the number of linked list operations
241 * - reduce spinlock operations
243 * The limit is stored in the per-cpu structure to reduce the data cache
244 * footprint.
247 struct array_cache {
248 unsigned int avail;
249 unsigned int limit;
250 unsigned int batchcount;
251 unsigned int touched;
252 spinlock_t lock;
253 void *entry[]; /*
254 * Must have this definition in here for the proper
255 * alignment of array_cache. Also simplifies accessing
256 * the entries.
261 * bootstrap: The caches do not work without cpuarrays anymore, but the
262 * cpuarrays are allocated from the generic caches...
264 #define BOOT_CPUCACHE_ENTRIES 1
265 struct arraycache_init {
266 struct array_cache cache;
267 void *entries[BOOT_CPUCACHE_ENTRIES];
271 * The slab lists for all objects.
273 struct kmem_list3 {
274 struct list_head slabs_partial; /* partial list first, better asm code */
275 struct list_head slabs_full;
276 struct list_head slabs_free;
277 unsigned long free_objects;
278 unsigned int free_limit;
279 unsigned int colour_next; /* Per-node cache coloring */
280 spinlock_t list_lock;
281 struct array_cache *shared; /* shared per node */
282 struct array_cache **alien; /* on other nodes */
283 unsigned long next_reap; /* updated without locking */
284 int free_touched; /* updated without locking */
288 * Need this for bootstrapping a per node allocator.
290 #define NUM_INIT_LISTS (3 * MAX_NUMNODES)
291 static struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
292 #define CACHE_CACHE 0
293 #define SIZE_AC MAX_NUMNODES
294 #define SIZE_L3 (2 * MAX_NUMNODES)
296 static int drain_freelist(struct kmem_cache *cache,
297 struct kmem_list3 *l3, int tofree);
298 static void free_block(struct kmem_cache *cachep, void **objpp, int len,
299 int node);
300 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
301 static void cache_reap(struct work_struct *unused);
304 * This function must be completely optimized away if a constant is passed to
305 * it. Mostly the same as what is in linux/slab.h except it returns an index.
307 static __always_inline int index_of(const size_t size)
309 extern void __bad_size(void);
311 if (__builtin_constant_p(size)) {
312 int i = 0;
314 #define CACHE(x) \
315 if (size <=x) \
316 return i; \
317 else \
318 i++;
319 #include <linux/kmalloc_sizes.h>
320 #undef CACHE
321 __bad_size();
322 } else
323 __bad_size();
324 return 0;
327 static int slab_early_init = 1;
329 #define INDEX_AC index_of(sizeof(struct arraycache_init))
330 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
332 static void kmem_list3_init(struct kmem_list3 *parent)
334 INIT_LIST_HEAD(&parent->slabs_full);
335 INIT_LIST_HEAD(&parent->slabs_partial);
336 INIT_LIST_HEAD(&parent->slabs_free);
337 parent->shared = NULL;
338 parent->alien = NULL;
339 parent->colour_next = 0;
340 spin_lock_init(&parent->list_lock);
341 parent->free_objects = 0;
342 parent->free_touched = 0;
345 #define MAKE_LIST(cachep, listp, slab, nodeid) \
346 do { \
347 INIT_LIST_HEAD(listp); \
348 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
349 } while (0)
351 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
352 do { \
353 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
354 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
355 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
356 } while (0)
358 #define CFLGS_OFF_SLAB (0x80000000UL)
359 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
361 #define BATCHREFILL_LIMIT 16
363 * Optimization question: fewer reaps means less probability for unnessary
364 * cpucache drain/refill cycles.
366 * OTOH the cpuarrays can contain lots of objects,
367 * which could lock up otherwise freeable slabs.
369 #define REAPTIMEOUT_CPUC (2*HZ)
370 #define REAPTIMEOUT_LIST3 (4*HZ)
372 #if STATS
373 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
374 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
375 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
376 #define STATS_INC_GROWN(x) ((x)->grown++)
377 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
378 #define STATS_SET_HIGH(x) \
379 do { \
380 if ((x)->num_active > (x)->high_mark) \
381 (x)->high_mark = (x)->num_active; \
382 } while (0)
383 #define STATS_INC_ERR(x) ((x)->errors++)
384 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
385 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
386 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
387 #define STATS_SET_FREEABLE(x, i) \
388 do { \
389 if ((x)->max_freeable < i) \
390 (x)->max_freeable = i; \
391 } while (0)
392 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
393 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
394 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
395 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
396 #else
397 #define STATS_INC_ACTIVE(x) do { } while (0)
398 #define STATS_DEC_ACTIVE(x) do { } while (0)
399 #define STATS_INC_ALLOCED(x) do { } while (0)
400 #define STATS_INC_GROWN(x) do { } while (0)
401 #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
402 #define STATS_SET_HIGH(x) do { } while (0)
403 #define STATS_INC_ERR(x) do { } while (0)
404 #define STATS_INC_NODEALLOCS(x) do { } while (0)
405 #define STATS_INC_NODEFREES(x) do { } while (0)
406 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
407 #define STATS_SET_FREEABLE(x, i) do { } while (0)
408 #define STATS_INC_ALLOCHIT(x) do { } while (0)
409 #define STATS_INC_ALLOCMISS(x) do { } while (0)
410 #define STATS_INC_FREEHIT(x) do { } while (0)
411 #define STATS_INC_FREEMISS(x) do { } while (0)
412 #endif
414 #if DEBUG
417 * memory layout of objects:
418 * 0 : objp
419 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
420 * the end of an object is aligned with the end of the real
421 * allocation. Catches writes behind the end of the allocation.
422 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
423 * redzone word.
424 * cachep->obj_offset: The real object.
425 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
426 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
427 * [BYTES_PER_WORD long]
429 static int obj_offset(struct kmem_cache *cachep)
431 return cachep->obj_offset;
434 static int obj_size(struct kmem_cache *cachep)
436 return cachep->obj_size;
439 static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
441 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
442 return (unsigned long long*) (objp + obj_offset(cachep) -
443 sizeof(unsigned long long));
446 static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
448 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
449 if (cachep->flags & SLAB_STORE_USER)
450 return (unsigned long long *)(objp + cachep->buffer_size -
451 sizeof(unsigned long long) -
452 REDZONE_ALIGN);
453 return (unsigned long long *) (objp + cachep->buffer_size -
454 sizeof(unsigned long long));
457 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
459 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
460 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
463 #else
465 #define obj_offset(x) 0
466 #define obj_size(cachep) (cachep->buffer_size)
467 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
468 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
469 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
471 #endif
473 #ifdef CONFIG_TRACING
474 size_t slab_buffer_size(struct kmem_cache *cachep)
476 return cachep->buffer_size;
478 EXPORT_SYMBOL(slab_buffer_size);
479 #endif
482 * Do not go above this order unless 0 objects fit into the slab.
484 #define BREAK_GFP_ORDER_HI 1
485 #define BREAK_GFP_ORDER_LO 0
486 static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
489 * Functions for storing/retrieving the cachep and or slab from the page
490 * allocator. These are used to find the slab an obj belongs to. With kfree(),
491 * these are used to find the cache which an obj belongs to.
493 static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
495 page->lru.next = (struct list_head *)cache;
498 static inline struct kmem_cache *page_get_cache(struct page *page)
500 page = compound_head(page);
501 BUG_ON(!PageSlab(page));
502 return (struct kmem_cache *)page->lru.next;
505 static inline void page_set_slab(struct page *page, struct slab *slab)
507 page->lru.prev = (struct list_head *)slab;
510 static inline struct slab *page_get_slab(struct page *page)
512 BUG_ON(!PageSlab(page));
513 return (struct slab *)page->lru.prev;
516 static inline struct kmem_cache *virt_to_cache(const void *obj)
518 struct page *page = virt_to_head_page(obj);
519 return page_get_cache(page);
522 static inline struct slab *virt_to_slab(const void *obj)
524 struct page *page = virt_to_head_page(obj);
525 return page_get_slab(page);
528 static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
529 unsigned int idx)
531 return slab->s_mem + cache->buffer_size * idx;
535 * We want to avoid an expensive divide : (offset / cache->buffer_size)
536 * Using the fact that buffer_size is a constant for a particular cache,
537 * we can replace (offset / cache->buffer_size) by
538 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
540 static inline unsigned int obj_to_index(const struct kmem_cache *cache,
541 const struct slab *slab, void *obj)
543 u32 offset = (obj - slab->s_mem);
544 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
548 * These are the default caches for kmalloc. Custom caches can have other sizes.
550 struct cache_sizes malloc_sizes[] = {
551 #define CACHE(x) { .cs_size = (x) },
552 #include <linux/kmalloc_sizes.h>
553 CACHE(ULONG_MAX)
554 #undef CACHE
556 EXPORT_SYMBOL(malloc_sizes);
558 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
559 struct cache_names {
560 char *name;
561 char *name_dma;
564 static struct cache_names __initdata cache_names[] = {
565 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
566 #include <linux/kmalloc_sizes.h>
567 {NULL,}
568 #undef CACHE
571 static struct arraycache_init initarray_cache __initdata =
572 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
573 static struct arraycache_init initarray_generic =
574 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
576 /* internal cache of cache description objs */
577 static struct kmem_cache cache_cache = {
578 .batchcount = 1,
579 .limit = BOOT_CPUCACHE_ENTRIES,
580 .shared = 1,
581 .buffer_size = sizeof(struct kmem_cache),
582 .name = "kmem_cache",
585 #define BAD_ALIEN_MAGIC 0x01020304ul
588 * chicken and egg problem: delay the per-cpu array allocation
589 * until the general caches are up.
591 static enum {
592 NONE,
593 PARTIAL_AC,
594 PARTIAL_L3,
595 EARLY,
596 FULL
597 } g_cpucache_up;
600 * used by boot code to determine if it can use slab based allocator
602 int slab_is_available(void)
604 return g_cpucache_up >= EARLY;
607 #ifdef CONFIG_LOCKDEP
610 * Slab sometimes uses the kmalloc slabs to store the slab headers
611 * for other slabs "off slab".
612 * The locking for this is tricky in that it nests within the locks
613 * of all other slabs in a few places; to deal with this special
614 * locking we put on-slab caches into a separate lock-class.
616 * We set lock class for alien array caches which are up during init.
617 * The lock annotation will be lost if all cpus of a node goes down and
618 * then comes back up during hotplug
620 static struct lock_class_key on_slab_l3_key;
621 static struct lock_class_key on_slab_alc_key;
623 static void init_node_lock_keys(int q)
625 struct cache_sizes *s = malloc_sizes;
627 if (g_cpucache_up != FULL)
628 return;
630 for (s = malloc_sizes; s->cs_size != ULONG_MAX; s++) {
631 struct array_cache **alc;
632 struct kmem_list3 *l3;
633 int r;
635 l3 = s->cs_cachep->nodelists[q];
636 if (!l3 || OFF_SLAB(s->cs_cachep))
637 continue;
638 lockdep_set_class(&l3->list_lock, &on_slab_l3_key);
639 alc = l3->alien;
641 * FIXME: This check for BAD_ALIEN_MAGIC
642 * should go away when common slab code is taught to
643 * work even without alien caches.
644 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
645 * for alloc_alien_cache,
647 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
648 continue;
649 for_each_node(r) {
650 if (alc[r])
651 lockdep_set_class(&alc[r]->lock,
652 &on_slab_alc_key);
657 static inline void init_lock_keys(void)
659 int node;
661 for_each_node(node)
662 init_node_lock_keys(node);
664 #else
665 static void init_node_lock_keys(int q)
669 static inline void init_lock_keys(void)
672 #endif
675 * Guard access to the cache-chain.
677 static DEFINE_MUTEX(cache_chain_mutex);
678 static struct list_head cache_chain;
680 static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
682 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
684 return cachep->array[smp_processor_id()];
687 static inline struct kmem_cache *__find_general_cachep(size_t size,
688 gfp_t gfpflags)
690 struct cache_sizes *csizep = malloc_sizes;
692 #if DEBUG
693 /* This happens if someone tries to call
694 * kmem_cache_create(), or __kmalloc(), before
695 * the generic caches are initialized.
697 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
698 #endif
699 if (!size)
700 return ZERO_SIZE_PTR;
702 while (size > csizep->cs_size)
703 csizep++;
706 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
707 * has cs_{dma,}cachep==NULL. Thus no special case
708 * for large kmalloc calls required.
710 #ifdef CONFIG_ZONE_DMA
711 if (unlikely(gfpflags & GFP_DMA))
712 return csizep->cs_dmacachep;
713 #endif
714 return csizep->cs_cachep;
717 static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
719 return __find_general_cachep(size, gfpflags);
722 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
724 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
728 * Calculate the number of objects and left-over bytes for a given buffer size.
730 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
731 size_t align, int flags, size_t *left_over,
732 unsigned int *num)
734 int nr_objs;
735 size_t mgmt_size;
736 size_t slab_size = PAGE_SIZE << gfporder;
739 * The slab management structure can be either off the slab or
740 * on it. For the latter case, the memory allocated for a
741 * slab is used for:
743 * - The struct slab
744 * - One kmem_bufctl_t for each object
745 * - Padding to respect alignment of @align
746 * - @buffer_size bytes for each object
748 * If the slab management structure is off the slab, then the
749 * alignment will already be calculated into the size. Because
750 * the slabs are all pages aligned, the objects will be at the
751 * correct alignment when allocated.
753 if (flags & CFLGS_OFF_SLAB) {
754 mgmt_size = 0;
755 nr_objs = slab_size / buffer_size;
757 if (nr_objs > SLAB_LIMIT)
758 nr_objs = SLAB_LIMIT;
759 } else {
761 * Ignore padding for the initial guess. The padding
762 * is at most @align-1 bytes, and @buffer_size is at
763 * least @align. In the worst case, this result will
764 * be one greater than the number of objects that fit
765 * into the memory allocation when taking the padding
766 * into account.
768 nr_objs = (slab_size - sizeof(struct slab)) /
769 (buffer_size + sizeof(kmem_bufctl_t));
772 * This calculated number will be either the right
773 * amount, or one greater than what we want.
775 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
776 > slab_size)
777 nr_objs--;
779 if (nr_objs > SLAB_LIMIT)
780 nr_objs = SLAB_LIMIT;
782 mgmt_size = slab_mgmt_size(nr_objs, align);
784 *num = nr_objs;
785 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
788 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
790 static void __slab_error(const char *function, struct kmem_cache *cachep,
791 char *msg)
793 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
794 function, cachep->name, msg);
795 dump_stack();
799 * By default on NUMA we use alien caches to stage the freeing of
800 * objects allocated from other nodes. This causes massive memory
801 * inefficiencies when using fake NUMA setup to split memory into a
802 * large number of small nodes, so it can be disabled on the command
803 * line
806 static int use_alien_caches __read_mostly = 1;
807 static int __init noaliencache_setup(char *s)
809 use_alien_caches = 0;
810 return 1;
812 __setup("noaliencache", noaliencache_setup);
814 #ifdef CONFIG_NUMA
816 * Special reaping functions for NUMA systems called from cache_reap().
817 * These take care of doing round robin flushing of alien caches (containing
818 * objects freed on different nodes from which they were allocated) and the
819 * flushing of remote pcps by calling drain_node_pages.
821 static DEFINE_PER_CPU(unsigned long, slab_reap_node);
823 static void init_reap_node(int cpu)
825 int node;
827 node = next_node(cpu_to_mem(cpu), node_online_map);
828 if (node == MAX_NUMNODES)
829 node = first_node(node_online_map);
831 per_cpu(slab_reap_node, cpu) = node;
834 static void next_reap_node(void)
836 int node = __this_cpu_read(slab_reap_node);
838 node = next_node(node, node_online_map);
839 if (unlikely(node >= MAX_NUMNODES))
840 node = first_node(node_online_map);
841 __this_cpu_write(slab_reap_node, node);
844 #else
845 #define init_reap_node(cpu) do { } while (0)
846 #define next_reap_node(void) do { } while (0)
847 #endif
850 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
851 * via the workqueue/eventd.
852 * Add the CPU number into the expiration time to minimize the possibility of
853 * the CPUs getting into lockstep and contending for the global cache chain
854 * lock.
856 static void __cpuinit start_cpu_timer(int cpu)
858 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
861 * When this gets called from do_initcalls via cpucache_init(),
862 * init_workqueues() has already run, so keventd will be setup
863 * at that time.
865 if (keventd_up() && reap_work->work.func == NULL) {
866 init_reap_node(cpu);
867 INIT_DELAYED_WORK_DEFERRABLE(reap_work, cache_reap);
868 schedule_delayed_work_on(cpu, reap_work,
869 __round_jiffies_relative(HZ, cpu));
873 static struct array_cache *alloc_arraycache(int node, int entries,
874 int batchcount, gfp_t gfp)
876 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
877 struct array_cache *nc = NULL;
879 nc = kmalloc_node(memsize, gfp, node);
881 * The array_cache structures contain pointers to free object.
882 * However, when such objects are allocated or transferred to another
883 * cache the pointers are not cleared and they could be counted as
884 * valid references during a kmemleak scan. Therefore, kmemleak must
885 * not scan such objects.
887 kmemleak_no_scan(nc);
888 if (nc) {
889 nc->avail = 0;
890 nc->limit = entries;
891 nc->batchcount = batchcount;
892 nc->touched = 0;
893 spin_lock_init(&nc->lock);
895 return nc;
899 * Transfer objects in one arraycache to another.
900 * Locking must be handled by the caller.
902 * Return the number of entries transferred.
904 static int transfer_objects(struct array_cache *to,
905 struct array_cache *from, unsigned int max)
907 /* Figure out how many entries to transfer */
908 int nr = min3(from->avail, max, to->limit - to->avail);
910 if (!nr)
911 return 0;
913 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
914 sizeof(void *) *nr);
916 from->avail -= nr;
917 to->avail += nr;
918 return nr;
921 #ifndef CONFIG_NUMA
923 #define drain_alien_cache(cachep, alien) do { } while (0)
924 #define reap_alien(cachep, l3) do { } while (0)
926 static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
928 return (struct array_cache **)BAD_ALIEN_MAGIC;
931 static inline void free_alien_cache(struct array_cache **ac_ptr)
935 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
937 return 0;
940 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
941 gfp_t flags)
943 return NULL;
946 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
947 gfp_t flags, int nodeid)
949 return NULL;
952 #else /* CONFIG_NUMA */
954 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
955 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
957 static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
959 struct array_cache **ac_ptr;
960 int memsize = sizeof(void *) * nr_node_ids;
961 int i;
963 if (limit > 1)
964 limit = 12;
965 ac_ptr = kzalloc_node(memsize, gfp, node);
966 if (ac_ptr) {
967 for_each_node(i) {
968 if (i == node || !node_online(i))
969 continue;
970 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
971 if (!ac_ptr[i]) {
972 for (i--; i >= 0; i--)
973 kfree(ac_ptr[i]);
974 kfree(ac_ptr);
975 return NULL;
979 return ac_ptr;
982 static void free_alien_cache(struct array_cache **ac_ptr)
984 int i;
986 if (!ac_ptr)
987 return;
988 for_each_node(i)
989 kfree(ac_ptr[i]);
990 kfree(ac_ptr);
993 static void __drain_alien_cache(struct kmem_cache *cachep,
994 struct array_cache *ac, int node)
996 struct kmem_list3 *rl3 = cachep->nodelists[node];
998 if (ac->avail) {
999 spin_lock(&rl3->list_lock);
1001 * Stuff objects into the remote nodes shared array first.
1002 * That way we could avoid the overhead of putting the objects
1003 * into the free lists and getting them back later.
1005 if (rl3->shared)
1006 transfer_objects(rl3->shared, ac, ac->limit);
1008 free_block(cachep, ac->entry, ac->avail, node);
1009 ac->avail = 0;
1010 spin_unlock(&rl3->list_lock);
1015 * Called from cache_reap() to regularly drain alien caches round robin.
1017 static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1019 int node = __this_cpu_read(slab_reap_node);
1021 if (l3->alien) {
1022 struct array_cache *ac = l3->alien[node];
1024 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1025 __drain_alien_cache(cachep, ac, node);
1026 spin_unlock_irq(&ac->lock);
1031 static void drain_alien_cache(struct kmem_cache *cachep,
1032 struct array_cache **alien)
1034 int i = 0;
1035 struct array_cache *ac;
1036 unsigned long flags;
1038 for_each_online_node(i) {
1039 ac = alien[i];
1040 if (ac) {
1041 spin_lock_irqsave(&ac->lock, flags);
1042 __drain_alien_cache(cachep, ac, i);
1043 spin_unlock_irqrestore(&ac->lock, flags);
1048 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1050 struct slab *slabp = virt_to_slab(objp);
1051 int nodeid = slabp->nodeid;
1052 struct kmem_list3 *l3;
1053 struct array_cache *alien = NULL;
1054 int node;
1056 node = numa_mem_id();
1059 * Make sure we are not freeing a object from another node to the array
1060 * cache on this cpu.
1062 if (likely(slabp->nodeid == node))
1063 return 0;
1065 l3 = cachep->nodelists[node];
1066 STATS_INC_NODEFREES(cachep);
1067 if (l3->alien && l3->alien[nodeid]) {
1068 alien = l3->alien[nodeid];
1069 spin_lock(&alien->lock);
1070 if (unlikely(alien->avail == alien->limit)) {
1071 STATS_INC_ACOVERFLOW(cachep);
1072 __drain_alien_cache(cachep, alien, nodeid);
1074 alien->entry[alien->avail++] = objp;
1075 spin_unlock(&alien->lock);
1076 } else {
1077 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1078 free_block(cachep, &objp, 1, nodeid);
1079 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1081 return 1;
1083 #endif
1086 * Allocates and initializes nodelists for a node on each slab cache, used for
1087 * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3
1088 * will be allocated off-node since memory is not yet online for the new node.
1089 * When hotplugging memory or a cpu, existing nodelists are not replaced if
1090 * already in use.
1092 * Must hold cache_chain_mutex.
1094 static int init_cache_nodelists_node(int node)
1096 struct kmem_cache *cachep;
1097 struct kmem_list3 *l3;
1098 const int memsize = sizeof(struct kmem_list3);
1100 list_for_each_entry(cachep, &cache_chain, next) {
1102 * Set up the size64 kmemlist for cpu before we can
1103 * begin anything. Make sure some other cpu on this
1104 * node has not already allocated this
1106 if (!cachep->nodelists[node]) {
1107 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1108 if (!l3)
1109 return -ENOMEM;
1110 kmem_list3_init(l3);
1111 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1112 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1115 * The l3s don't come and go as CPUs come and
1116 * go. cache_chain_mutex is sufficient
1117 * protection here.
1119 cachep->nodelists[node] = l3;
1122 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1123 cachep->nodelists[node]->free_limit =
1124 (1 + nr_cpus_node(node)) *
1125 cachep->batchcount + cachep->num;
1126 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1128 return 0;
1131 static void __cpuinit cpuup_canceled(long cpu)
1133 struct kmem_cache *cachep;
1134 struct kmem_list3 *l3 = NULL;
1135 int node = cpu_to_mem(cpu);
1136 const struct cpumask *mask = cpumask_of_node(node);
1138 list_for_each_entry(cachep, &cache_chain, next) {
1139 struct array_cache *nc;
1140 struct array_cache *shared;
1141 struct array_cache **alien;
1143 /* cpu is dead; no one can alloc from it. */
1144 nc = cachep->array[cpu];
1145 cachep->array[cpu] = NULL;
1146 l3 = cachep->nodelists[node];
1148 if (!l3)
1149 goto free_array_cache;
1151 spin_lock_irq(&l3->list_lock);
1153 /* Free limit for this kmem_list3 */
1154 l3->free_limit -= cachep->batchcount;
1155 if (nc)
1156 free_block(cachep, nc->entry, nc->avail, node);
1158 if (!cpumask_empty(mask)) {
1159 spin_unlock_irq(&l3->list_lock);
1160 goto free_array_cache;
1163 shared = l3->shared;
1164 if (shared) {
1165 free_block(cachep, shared->entry,
1166 shared->avail, node);
1167 l3->shared = NULL;
1170 alien = l3->alien;
1171 l3->alien = NULL;
1173 spin_unlock_irq(&l3->list_lock);
1175 kfree(shared);
1176 if (alien) {
1177 drain_alien_cache(cachep, alien);
1178 free_alien_cache(alien);
1180 free_array_cache:
1181 kfree(nc);
1184 * In the previous loop, all the objects were freed to
1185 * the respective cache's slabs, now we can go ahead and
1186 * shrink each nodelist to its limit.
1188 list_for_each_entry(cachep, &cache_chain, next) {
1189 l3 = cachep->nodelists[node];
1190 if (!l3)
1191 continue;
1192 drain_freelist(cachep, l3, l3->free_objects);
1196 static int __cpuinit cpuup_prepare(long cpu)
1198 struct kmem_cache *cachep;
1199 struct kmem_list3 *l3 = NULL;
1200 int node = cpu_to_mem(cpu);
1201 int err;
1204 * We need to do this right in the beginning since
1205 * alloc_arraycache's are going to use this list.
1206 * kmalloc_node allows us to add the slab to the right
1207 * kmem_list3 and not this cpu's kmem_list3
1209 err = init_cache_nodelists_node(node);
1210 if (err < 0)
1211 goto bad;
1214 * Now we can go ahead with allocating the shared arrays and
1215 * array caches
1217 list_for_each_entry(cachep, &cache_chain, next) {
1218 struct array_cache *nc;
1219 struct array_cache *shared = NULL;
1220 struct array_cache **alien = NULL;
1222 nc = alloc_arraycache(node, cachep->limit,
1223 cachep->batchcount, GFP_KERNEL);
1224 if (!nc)
1225 goto bad;
1226 if (cachep->shared) {
1227 shared = alloc_arraycache(node,
1228 cachep->shared * cachep->batchcount,
1229 0xbaadf00d, GFP_KERNEL);
1230 if (!shared) {
1231 kfree(nc);
1232 goto bad;
1235 if (use_alien_caches) {
1236 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
1237 if (!alien) {
1238 kfree(shared);
1239 kfree(nc);
1240 goto bad;
1243 cachep->array[cpu] = nc;
1244 l3 = cachep->nodelists[node];
1245 BUG_ON(!l3);
1247 spin_lock_irq(&l3->list_lock);
1248 if (!l3->shared) {
1250 * We are serialised from CPU_DEAD or
1251 * CPU_UP_CANCELLED by the cpucontrol lock
1253 l3->shared = shared;
1254 shared = NULL;
1256 #ifdef CONFIG_NUMA
1257 if (!l3->alien) {
1258 l3->alien = alien;
1259 alien = NULL;
1261 #endif
1262 spin_unlock_irq(&l3->list_lock);
1263 kfree(shared);
1264 free_alien_cache(alien);
1266 init_node_lock_keys(node);
1268 return 0;
1269 bad:
1270 cpuup_canceled(cpu);
1271 return -ENOMEM;
1274 static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1275 unsigned long action, void *hcpu)
1277 long cpu = (long)hcpu;
1278 int err = 0;
1280 switch (action) {
1281 case CPU_UP_PREPARE:
1282 case CPU_UP_PREPARE_FROZEN:
1283 mutex_lock(&cache_chain_mutex);
1284 err = cpuup_prepare(cpu);
1285 mutex_unlock(&cache_chain_mutex);
1286 break;
1287 case CPU_ONLINE:
1288 case CPU_ONLINE_FROZEN:
1289 start_cpu_timer(cpu);
1290 break;
1291 #ifdef CONFIG_HOTPLUG_CPU
1292 case CPU_DOWN_PREPARE:
1293 case CPU_DOWN_PREPARE_FROZEN:
1295 * Shutdown cache reaper. Note that the cache_chain_mutex is
1296 * held so that if cache_reap() is invoked it cannot do
1297 * anything expensive but will only modify reap_work
1298 * and reschedule the timer.
1300 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1301 /* Now the cache_reaper is guaranteed to be not running. */
1302 per_cpu(slab_reap_work, cpu).work.func = NULL;
1303 break;
1304 case CPU_DOWN_FAILED:
1305 case CPU_DOWN_FAILED_FROZEN:
1306 start_cpu_timer(cpu);
1307 break;
1308 case CPU_DEAD:
1309 case CPU_DEAD_FROZEN:
1311 * Even if all the cpus of a node are down, we don't free the
1312 * kmem_list3 of any cache. This to avoid a race between
1313 * cpu_down, and a kmalloc allocation from another cpu for
1314 * memory from the node of the cpu going down. The list3
1315 * structure is usually allocated from kmem_cache_create() and
1316 * gets destroyed at kmem_cache_destroy().
1318 /* fall through */
1319 #endif
1320 case CPU_UP_CANCELED:
1321 case CPU_UP_CANCELED_FROZEN:
1322 mutex_lock(&cache_chain_mutex);
1323 cpuup_canceled(cpu);
1324 mutex_unlock(&cache_chain_mutex);
1325 break;
1327 return notifier_from_errno(err);
1330 static struct notifier_block __cpuinitdata cpucache_notifier = {
1331 &cpuup_callback, NULL, 0
1334 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1336 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1337 * Returns -EBUSY if all objects cannot be drained so that the node is not
1338 * removed.
1340 * Must hold cache_chain_mutex.
1342 static int __meminit drain_cache_nodelists_node(int node)
1344 struct kmem_cache *cachep;
1345 int ret = 0;
1347 list_for_each_entry(cachep, &cache_chain, next) {
1348 struct kmem_list3 *l3;
1350 l3 = cachep->nodelists[node];
1351 if (!l3)
1352 continue;
1354 drain_freelist(cachep, l3, l3->free_objects);
1356 if (!list_empty(&l3->slabs_full) ||
1357 !list_empty(&l3->slabs_partial)) {
1358 ret = -EBUSY;
1359 break;
1362 return ret;
1365 static int __meminit slab_memory_callback(struct notifier_block *self,
1366 unsigned long action, void *arg)
1368 struct memory_notify *mnb = arg;
1369 int ret = 0;
1370 int nid;
1372 nid = mnb->status_change_nid;
1373 if (nid < 0)
1374 goto out;
1376 switch (action) {
1377 case MEM_GOING_ONLINE:
1378 mutex_lock(&cache_chain_mutex);
1379 ret = init_cache_nodelists_node(nid);
1380 mutex_unlock(&cache_chain_mutex);
1381 break;
1382 case MEM_GOING_OFFLINE:
1383 mutex_lock(&cache_chain_mutex);
1384 ret = drain_cache_nodelists_node(nid);
1385 mutex_unlock(&cache_chain_mutex);
1386 break;
1387 case MEM_ONLINE:
1388 case MEM_OFFLINE:
1389 case MEM_CANCEL_ONLINE:
1390 case MEM_CANCEL_OFFLINE:
1391 break;
1393 out:
1394 return notifier_from_errno(ret);
1396 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1399 * swap the static kmem_list3 with kmalloced memory
1401 static void __init init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1402 int nodeid)
1404 struct kmem_list3 *ptr;
1406 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid);
1407 BUG_ON(!ptr);
1409 memcpy(ptr, list, sizeof(struct kmem_list3));
1411 * Do not assume that spinlocks can be initialized via memcpy:
1413 spin_lock_init(&ptr->list_lock);
1415 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1416 cachep->nodelists[nodeid] = ptr;
1420 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1421 * size of kmem_list3.
1423 static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1425 int node;
1427 for_each_online_node(node) {
1428 cachep->nodelists[node] = &initkmem_list3[index + node];
1429 cachep->nodelists[node]->next_reap = jiffies +
1430 REAPTIMEOUT_LIST3 +
1431 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1436 * Initialisation. Called after the page allocator have been initialised and
1437 * before smp_init().
1439 void __init kmem_cache_init(void)
1441 size_t left_over;
1442 struct cache_sizes *sizes;
1443 struct cache_names *names;
1444 int i;
1445 int order;
1446 int node;
1448 if (num_possible_nodes() == 1)
1449 use_alien_caches = 0;
1451 for (i = 0; i < NUM_INIT_LISTS; i++) {
1452 kmem_list3_init(&initkmem_list3[i]);
1453 if (i < MAX_NUMNODES)
1454 cache_cache.nodelists[i] = NULL;
1456 set_up_list3s(&cache_cache, CACHE_CACHE);
1459 * Fragmentation resistance on low memory - only use bigger
1460 * page orders on machines with more than 32MB of memory.
1462 if (totalram_pages > (32 << 20) >> PAGE_SHIFT)
1463 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1465 /* Bootstrap is tricky, because several objects are allocated
1466 * from caches that do not exist yet:
1467 * 1) initialize the cache_cache cache: it contains the struct
1468 * kmem_cache structures of all caches, except cache_cache itself:
1469 * cache_cache is statically allocated.
1470 * Initially an __init data area is used for the head array and the
1471 * kmem_list3 structures, it's replaced with a kmalloc allocated
1472 * array at the end of the bootstrap.
1473 * 2) Create the first kmalloc cache.
1474 * The struct kmem_cache for the new cache is allocated normally.
1475 * An __init data area is used for the head array.
1476 * 3) Create the remaining kmalloc caches, with minimally sized
1477 * head arrays.
1478 * 4) Replace the __init data head arrays for cache_cache and the first
1479 * kmalloc cache with kmalloc allocated arrays.
1480 * 5) Replace the __init data for kmem_list3 for cache_cache and
1481 * the other cache's with kmalloc allocated memory.
1482 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1485 node = numa_mem_id();
1487 /* 1) create the cache_cache */
1488 INIT_LIST_HEAD(&cache_chain);
1489 list_add(&cache_cache.next, &cache_chain);
1490 cache_cache.colour_off = cache_line_size();
1491 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
1492 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node];
1495 * struct kmem_cache size depends on nr_node_ids, which
1496 * can be less than MAX_NUMNODES.
1498 cache_cache.buffer_size = offsetof(struct kmem_cache, nodelists) +
1499 nr_node_ids * sizeof(struct kmem_list3 *);
1500 #if DEBUG
1501 cache_cache.obj_size = cache_cache.buffer_size;
1502 #endif
1503 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1504 cache_line_size());
1505 cache_cache.reciprocal_buffer_size =
1506 reciprocal_value(cache_cache.buffer_size);
1508 for (order = 0; order < MAX_ORDER; order++) {
1509 cache_estimate(order, cache_cache.buffer_size,
1510 cache_line_size(), 0, &left_over, &cache_cache.num);
1511 if (cache_cache.num)
1512 break;
1514 BUG_ON(!cache_cache.num);
1515 cache_cache.gfporder = order;
1516 cache_cache.colour = left_over / cache_cache.colour_off;
1517 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1518 sizeof(struct slab), cache_line_size());
1520 /* 2+3) create the kmalloc caches */
1521 sizes = malloc_sizes;
1522 names = cache_names;
1525 * Initialize the caches that provide memory for the array cache and the
1526 * kmem_list3 structures first. Without this, further allocations will
1527 * bug.
1530 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
1531 sizes[INDEX_AC].cs_size,
1532 ARCH_KMALLOC_MINALIGN,
1533 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1534 NULL);
1536 if (INDEX_AC != INDEX_L3) {
1537 sizes[INDEX_L3].cs_cachep =
1538 kmem_cache_create(names[INDEX_L3].name,
1539 sizes[INDEX_L3].cs_size,
1540 ARCH_KMALLOC_MINALIGN,
1541 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1542 NULL);
1545 slab_early_init = 0;
1547 while (sizes->cs_size != ULONG_MAX) {
1549 * For performance, all the general caches are L1 aligned.
1550 * This should be particularly beneficial on SMP boxes, as it
1551 * eliminates "false sharing".
1552 * Note for systems short on memory removing the alignment will
1553 * allow tighter packing of the smaller caches.
1555 if (!sizes->cs_cachep) {
1556 sizes->cs_cachep = kmem_cache_create(names->name,
1557 sizes->cs_size,
1558 ARCH_KMALLOC_MINALIGN,
1559 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1560 NULL);
1562 #ifdef CONFIG_ZONE_DMA
1563 sizes->cs_dmacachep = kmem_cache_create(
1564 names->name_dma,
1565 sizes->cs_size,
1566 ARCH_KMALLOC_MINALIGN,
1567 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1568 SLAB_PANIC,
1569 NULL);
1570 #endif
1571 sizes++;
1572 names++;
1574 /* 4) Replace the bootstrap head arrays */
1576 struct array_cache *ptr;
1578 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1580 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1581 memcpy(ptr, cpu_cache_get(&cache_cache),
1582 sizeof(struct arraycache_init));
1584 * Do not assume that spinlocks can be initialized via memcpy:
1586 spin_lock_init(&ptr->lock);
1588 cache_cache.array[smp_processor_id()] = ptr;
1590 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1592 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1593 != &initarray_generic.cache);
1594 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1595 sizeof(struct arraycache_init));
1597 * Do not assume that spinlocks can be initialized via memcpy:
1599 spin_lock_init(&ptr->lock);
1601 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1602 ptr;
1604 /* 5) Replace the bootstrap kmem_list3's */
1606 int nid;
1608 for_each_online_node(nid) {
1609 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid);
1611 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1612 &initkmem_list3[SIZE_AC + nid], nid);
1614 if (INDEX_AC != INDEX_L3) {
1615 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1616 &initkmem_list3[SIZE_L3 + nid], nid);
1621 g_cpucache_up = EARLY;
1624 void __init kmem_cache_init_late(void)
1626 struct kmem_cache *cachep;
1628 /* 6) resize the head arrays to their final sizes */
1629 mutex_lock(&cache_chain_mutex);
1630 list_for_each_entry(cachep, &cache_chain, next)
1631 if (enable_cpucache(cachep, GFP_NOWAIT))
1632 BUG();
1633 mutex_unlock(&cache_chain_mutex);
1635 /* Done! */
1636 g_cpucache_up = FULL;
1638 /* Annotate slab for lockdep -- annotate the malloc caches */
1639 init_lock_keys();
1642 * Register a cpu startup notifier callback that initializes
1643 * cpu_cache_get for all new cpus
1645 register_cpu_notifier(&cpucache_notifier);
1647 #ifdef CONFIG_NUMA
1649 * Register a memory hotplug callback that initializes and frees
1650 * nodelists.
1652 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1653 #endif
1656 * The reap timers are started later, with a module init call: That part
1657 * of the kernel is not yet operational.
1661 static int __init cpucache_init(void)
1663 int cpu;
1666 * Register the timers that return unneeded pages to the page allocator
1668 for_each_online_cpu(cpu)
1669 start_cpu_timer(cpu);
1670 return 0;
1672 __initcall(cpucache_init);
1675 * Interface to system's page allocator. No need to hold the cache-lock.
1677 * If we requested dmaable memory, we will get it. Even if we
1678 * did not request dmaable memory, we might get it, but that
1679 * would be relatively rare and ignorable.
1681 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1683 struct page *page;
1684 int nr_pages;
1685 int i;
1687 #ifndef CONFIG_MMU
1689 * Nommu uses slab's for process anonymous memory allocations, and thus
1690 * requires __GFP_COMP to properly refcount higher order allocations
1692 flags |= __GFP_COMP;
1693 #endif
1695 flags |= cachep->gfpflags;
1696 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1697 flags |= __GFP_RECLAIMABLE;
1699 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1700 if (!page)
1701 return NULL;
1703 nr_pages = (1 << cachep->gfporder);
1704 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1705 add_zone_page_state(page_zone(page),
1706 NR_SLAB_RECLAIMABLE, nr_pages);
1707 else
1708 add_zone_page_state(page_zone(page),
1709 NR_SLAB_UNRECLAIMABLE, nr_pages);
1710 for (i = 0; i < nr_pages; i++)
1711 __SetPageSlab(page + i);
1713 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1714 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1716 if (cachep->ctor)
1717 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1718 else
1719 kmemcheck_mark_unallocated_pages(page, nr_pages);
1722 return page_address(page);
1726 * Interface to system's page release.
1728 static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1730 unsigned long i = (1 << cachep->gfporder);
1731 struct page *page = virt_to_page(addr);
1732 const unsigned long nr_freed = i;
1734 kmemcheck_free_shadow(page, cachep->gfporder);
1736 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1737 sub_zone_page_state(page_zone(page),
1738 NR_SLAB_RECLAIMABLE, nr_freed);
1739 else
1740 sub_zone_page_state(page_zone(page),
1741 NR_SLAB_UNRECLAIMABLE, nr_freed);
1742 while (i--) {
1743 BUG_ON(!PageSlab(page));
1744 __ClearPageSlab(page);
1745 page++;
1747 if (current->reclaim_state)
1748 current->reclaim_state->reclaimed_slab += nr_freed;
1749 free_pages((unsigned long)addr, cachep->gfporder);
1752 static void kmem_rcu_free(struct rcu_head *head)
1754 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1755 struct kmem_cache *cachep = slab_rcu->cachep;
1757 kmem_freepages(cachep, slab_rcu->addr);
1758 if (OFF_SLAB(cachep))
1759 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1762 #if DEBUG
1764 #ifdef CONFIG_DEBUG_PAGEALLOC
1765 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1766 unsigned long caller)
1768 int size = obj_size(cachep);
1770 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1772 if (size < 5 * sizeof(unsigned long))
1773 return;
1775 *addr++ = 0x12345678;
1776 *addr++ = caller;
1777 *addr++ = smp_processor_id();
1778 size -= 3 * sizeof(unsigned long);
1780 unsigned long *sptr = &caller;
1781 unsigned long svalue;
1783 while (!kstack_end(sptr)) {
1784 svalue = *sptr++;
1785 if (kernel_text_address(svalue)) {
1786 *addr++ = svalue;
1787 size -= sizeof(unsigned long);
1788 if (size <= sizeof(unsigned long))
1789 break;
1794 *addr++ = 0x87654321;
1796 #endif
1798 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1800 int size = obj_size(cachep);
1801 addr = &((char *)addr)[obj_offset(cachep)];
1803 memset(addr, val, size);
1804 *(unsigned char *)(addr + size - 1) = POISON_END;
1807 static void dump_line(char *data, int offset, int limit)
1809 int i;
1810 unsigned char error = 0;
1811 int bad_count = 0;
1813 printk(KERN_ERR "%03x:", offset);
1814 for (i = 0; i < limit; i++) {
1815 if (data[offset + i] != POISON_FREE) {
1816 error = data[offset + i];
1817 bad_count++;
1819 printk(" %02x", (unsigned char)data[offset + i]);
1821 printk("\n");
1823 if (bad_count == 1) {
1824 error ^= POISON_FREE;
1825 if (!(error & (error - 1))) {
1826 printk(KERN_ERR "Single bit error detected. Probably "
1827 "bad RAM.\n");
1828 #ifdef CONFIG_X86
1829 printk(KERN_ERR "Run memtest86+ or a similar memory "
1830 "test tool.\n");
1831 #else
1832 printk(KERN_ERR "Run a memory test tool.\n");
1833 #endif
1837 #endif
1839 #if DEBUG
1841 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1843 int i, size;
1844 char *realobj;
1846 if (cachep->flags & SLAB_RED_ZONE) {
1847 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
1848 *dbg_redzone1(cachep, objp),
1849 *dbg_redzone2(cachep, objp));
1852 if (cachep->flags & SLAB_STORE_USER) {
1853 printk(KERN_ERR "Last user: [<%p>]",
1854 *dbg_userword(cachep, objp));
1855 print_symbol("(%s)",
1856 (unsigned long)*dbg_userword(cachep, objp));
1857 printk("\n");
1859 realobj = (char *)objp + obj_offset(cachep);
1860 size = obj_size(cachep);
1861 for (i = 0; i < size && lines; i += 16, lines--) {
1862 int limit;
1863 limit = 16;
1864 if (i + limit > size)
1865 limit = size - i;
1866 dump_line(realobj, i, limit);
1870 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1872 char *realobj;
1873 int size, i;
1874 int lines = 0;
1876 realobj = (char *)objp + obj_offset(cachep);
1877 size = obj_size(cachep);
1879 for (i = 0; i < size; i++) {
1880 char exp = POISON_FREE;
1881 if (i == size - 1)
1882 exp = POISON_END;
1883 if (realobj[i] != exp) {
1884 int limit;
1885 /* Mismatch ! */
1886 /* Print header */
1887 if (lines == 0) {
1888 printk(KERN_ERR
1889 "Slab corruption: %s start=%p, len=%d\n",
1890 cachep->name, realobj, size);
1891 print_objinfo(cachep, objp, 0);
1893 /* Hexdump the affected line */
1894 i = (i / 16) * 16;
1895 limit = 16;
1896 if (i + limit > size)
1897 limit = size - i;
1898 dump_line(realobj, i, limit);
1899 i += 16;
1900 lines++;
1901 /* Limit to 5 lines */
1902 if (lines > 5)
1903 break;
1906 if (lines != 0) {
1907 /* Print some data about the neighboring objects, if they
1908 * exist:
1910 struct slab *slabp = virt_to_slab(objp);
1911 unsigned int objnr;
1913 objnr = obj_to_index(cachep, slabp, objp);
1914 if (objnr) {
1915 objp = index_to_obj(cachep, slabp, objnr - 1);
1916 realobj = (char *)objp + obj_offset(cachep);
1917 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1918 realobj, size);
1919 print_objinfo(cachep, objp, 2);
1921 if (objnr + 1 < cachep->num) {
1922 objp = index_to_obj(cachep, slabp, objnr + 1);
1923 realobj = (char *)objp + obj_offset(cachep);
1924 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1925 realobj, size);
1926 print_objinfo(cachep, objp, 2);
1930 #endif
1932 #if DEBUG
1933 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
1935 int i;
1936 for (i = 0; i < cachep->num; i++) {
1937 void *objp = index_to_obj(cachep, slabp, i);
1939 if (cachep->flags & SLAB_POISON) {
1940 #ifdef CONFIG_DEBUG_PAGEALLOC
1941 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1942 OFF_SLAB(cachep))
1943 kernel_map_pages(virt_to_page(objp),
1944 cachep->buffer_size / PAGE_SIZE, 1);
1945 else
1946 check_poison_obj(cachep, objp);
1947 #else
1948 check_poison_obj(cachep, objp);
1949 #endif
1951 if (cachep->flags & SLAB_RED_ZONE) {
1952 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1953 slab_error(cachep, "start of a freed object "
1954 "was overwritten");
1955 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1956 slab_error(cachep, "end of a freed object "
1957 "was overwritten");
1961 #else
1962 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
1965 #endif
1968 * slab_destroy - destroy and release all objects in a slab
1969 * @cachep: cache pointer being destroyed
1970 * @slabp: slab pointer being destroyed
1972 * Destroy all the objs in a slab, and release the mem back to the system.
1973 * Before calling the slab must have been unlinked from the cache. The
1974 * cache-lock is not held/needed.
1976 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
1978 void *addr = slabp->s_mem - slabp->colouroff;
1980 slab_destroy_debugcheck(cachep, slabp);
1981 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1982 struct slab_rcu *slab_rcu;
1984 slab_rcu = (struct slab_rcu *)slabp;
1985 slab_rcu->cachep = cachep;
1986 slab_rcu->addr = addr;
1987 call_rcu(&slab_rcu->head, kmem_rcu_free);
1988 } else {
1989 kmem_freepages(cachep, addr);
1990 if (OFF_SLAB(cachep))
1991 kmem_cache_free(cachep->slabp_cache, slabp);
1995 static void __kmem_cache_destroy(struct kmem_cache *cachep)
1997 int i;
1998 struct kmem_list3 *l3;
2000 for_each_online_cpu(i)
2001 kfree(cachep->array[i]);
2003 /* NUMA: free the list3 structures */
2004 for_each_online_node(i) {
2005 l3 = cachep->nodelists[i];
2006 if (l3) {
2007 kfree(l3->shared);
2008 free_alien_cache(l3->alien);
2009 kfree(l3);
2012 kmem_cache_free(&cache_cache, cachep);
2017 * calculate_slab_order - calculate size (page order) of slabs
2018 * @cachep: pointer to the cache that is being created
2019 * @size: size of objects to be created in this cache.
2020 * @align: required alignment for the objects.
2021 * @flags: slab allocation flags
2023 * Also calculates the number of objects per slab.
2025 * This could be made much more intelligent. For now, try to avoid using
2026 * high order pages for slabs. When the gfp() functions are more friendly
2027 * towards high-order requests, this should be changed.
2029 static size_t calculate_slab_order(struct kmem_cache *cachep,
2030 size_t size, size_t align, unsigned long flags)
2032 unsigned long offslab_limit;
2033 size_t left_over = 0;
2034 int gfporder;
2036 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
2037 unsigned int num;
2038 size_t remainder;
2040 cache_estimate(gfporder, size, align, flags, &remainder, &num);
2041 if (!num)
2042 continue;
2044 if (flags & CFLGS_OFF_SLAB) {
2046 * Max number of objs-per-slab for caches which
2047 * use off-slab slabs. Needed to avoid a possible
2048 * looping condition in cache_grow().
2050 offslab_limit = size - sizeof(struct slab);
2051 offslab_limit /= sizeof(kmem_bufctl_t);
2053 if (num > offslab_limit)
2054 break;
2057 /* Found something acceptable - save it away */
2058 cachep->num = num;
2059 cachep->gfporder = gfporder;
2060 left_over = remainder;
2063 * A VFS-reclaimable slab tends to have most allocations
2064 * as GFP_NOFS and we really don't want to have to be allocating
2065 * higher-order pages when we are unable to shrink dcache.
2067 if (flags & SLAB_RECLAIM_ACCOUNT)
2068 break;
2071 * Large number of objects is good, but very large slabs are
2072 * currently bad for the gfp()s.
2074 if (gfporder >= slab_break_gfp_order)
2075 break;
2078 * Acceptable internal fragmentation?
2080 if (left_over * 8 <= (PAGE_SIZE << gfporder))
2081 break;
2083 return left_over;
2086 static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
2088 if (g_cpucache_up == FULL)
2089 return enable_cpucache(cachep, gfp);
2091 if (g_cpucache_up == NONE) {
2093 * Note: the first kmem_cache_create must create the cache
2094 * that's used by kmalloc(24), otherwise the creation of
2095 * further caches will BUG().
2097 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2100 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2101 * the first cache, then we need to set up all its list3s,
2102 * otherwise the creation of further caches will BUG().
2104 set_up_list3s(cachep, SIZE_AC);
2105 if (INDEX_AC == INDEX_L3)
2106 g_cpucache_up = PARTIAL_L3;
2107 else
2108 g_cpucache_up = PARTIAL_AC;
2109 } else {
2110 cachep->array[smp_processor_id()] =
2111 kmalloc(sizeof(struct arraycache_init), gfp);
2113 if (g_cpucache_up == PARTIAL_AC) {
2114 set_up_list3s(cachep, SIZE_L3);
2115 g_cpucache_up = PARTIAL_L3;
2116 } else {
2117 int node;
2118 for_each_online_node(node) {
2119 cachep->nodelists[node] =
2120 kmalloc_node(sizeof(struct kmem_list3),
2121 gfp, node);
2122 BUG_ON(!cachep->nodelists[node]);
2123 kmem_list3_init(cachep->nodelists[node]);
2127 cachep->nodelists[numa_mem_id()]->next_reap =
2128 jiffies + REAPTIMEOUT_LIST3 +
2129 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2131 cpu_cache_get(cachep)->avail = 0;
2132 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2133 cpu_cache_get(cachep)->batchcount = 1;
2134 cpu_cache_get(cachep)->touched = 0;
2135 cachep->batchcount = 1;
2136 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2137 return 0;
2141 * kmem_cache_create - Create a cache.
2142 * @name: A string which is used in /proc/slabinfo to identify this cache.
2143 * @size: The size of objects to be created in this cache.
2144 * @align: The required alignment for the objects.
2145 * @flags: SLAB flags
2146 * @ctor: A constructor for the objects.
2148 * Returns a ptr to the cache on success, NULL on failure.
2149 * Cannot be called within a int, but can be interrupted.
2150 * The @ctor is run when new pages are allocated by the cache.
2152 * @name must be valid until the cache is destroyed. This implies that
2153 * the module calling this has to destroy the cache before getting unloaded.
2155 * The flags are
2157 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2158 * to catch references to uninitialised memory.
2160 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2161 * for buffer overruns.
2163 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2164 * cacheline. This can be beneficial if you're counting cycles as closely
2165 * as davem.
2167 struct kmem_cache *
2168 kmem_cache_create (const char *name, size_t size, size_t align,
2169 unsigned long flags, void (*ctor)(void *))
2171 size_t left_over, slab_size, ralign;
2172 struct kmem_cache *cachep = NULL, *pc;
2173 gfp_t gfp;
2176 * Sanity checks... these are all serious usage bugs.
2178 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
2179 size > KMALLOC_MAX_SIZE) {
2180 printk(KERN_ERR "%s: Early error in slab %s\n", __func__,
2181 name);
2182 BUG();
2186 * We use cache_chain_mutex to ensure a consistent view of
2187 * cpu_online_mask as well. Please see cpuup_callback
2189 if (slab_is_available()) {
2190 get_online_cpus();
2191 mutex_lock(&cache_chain_mutex);
2194 list_for_each_entry(pc, &cache_chain, next) {
2195 char tmp;
2196 int res;
2199 * This happens when the module gets unloaded and doesn't
2200 * destroy its slab cache and no-one else reuses the vmalloc
2201 * area of the module. Print a warning.
2203 res = probe_kernel_address(pc->name, tmp);
2204 if (res) {
2205 printk(KERN_ERR
2206 "SLAB: cache with size %d has lost its name\n",
2207 pc->buffer_size);
2208 continue;
2211 if (!strcmp(pc->name, name)) {
2212 printk(KERN_ERR
2213 "kmem_cache_create: duplicate cache %s\n", name);
2214 dump_stack();
2215 goto oops;
2219 #if DEBUG
2220 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
2221 #if FORCED_DEBUG
2223 * Enable redzoning and last user accounting, except for caches with
2224 * large objects, if the increased size would increase the object size
2225 * above the next power of two: caches with object sizes just above a
2226 * power of two have a significant amount of internal fragmentation.
2228 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2229 2 * sizeof(unsigned long long)))
2230 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2231 if (!(flags & SLAB_DESTROY_BY_RCU))
2232 flags |= SLAB_POISON;
2233 #endif
2234 if (flags & SLAB_DESTROY_BY_RCU)
2235 BUG_ON(flags & SLAB_POISON);
2236 #endif
2238 * Always checks flags, a caller might be expecting debug support which
2239 * isn't available.
2241 BUG_ON(flags & ~CREATE_MASK);
2244 * Check that size is in terms of words. This is needed to avoid
2245 * unaligned accesses for some archs when redzoning is used, and makes
2246 * sure any on-slab bufctl's are also correctly aligned.
2248 if (size & (BYTES_PER_WORD - 1)) {
2249 size += (BYTES_PER_WORD - 1);
2250 size &= ~(BYTES_PER_WORD - 1);
2253 /* calculate the final buffer alignment: */
2255 /* 1) arch recommendation: can be overridden for debug */
2256 if (flags & SLAB_HWCACHE_ALIGN) {
2258 * Default alignment: as specified by the arch code. Except if
2259 * an object is really small, then squeeze multiple objects into
2260 * one cacheline.
2262 ralign = cache_line_size();
2263 while (size <= ralign / 2)
2264 ralign /= 2;
2265 } else {
2266 ralign = BYTES_PER_WORD;
2270 * Redzoning and user store require word alignment or possibly larger.
2271 * Note this will be overridden by architecture or caller mandated
2272 * alignment if either is greater than BYTES_PER_WORD.
2274 if (flags & SLAB_STORE_USER)
2275 ralign = BYTES_PER_WORD;
2277 if (flags & SLAB_RED_ZONE) {
2278 ralign = REDZONE_ALIGN;
2279 /* If redzoning, ensure that the second redzone is suitably
2280 * aligned, by adjusting the object size accordingly. */
2281 size += REDZONE_ALIGN - 1;
2282 size &= ~(REDZONE_ALIGN - 1);
2285 /* 2) arch mandated alignment */
2286 if (ralign < ARCH_SLAB_MINALIGN) {
2287 ralign = ARCH_SLAB_MINALIGN;
2289 /* 3) caller mandated alignment */
2290 if (ralign < align) {
2291 ralign = align;
2293 /* disable debug if necessary */
2294 if (ralign > __alignof__(unsigned long long))
2295 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2297 * 4) Store it.
2299 align = ralign;
2301 if (slab_is_available())
2302 gfp = GFP_KERNEL;
2303 else
2304 gfp = GFP_NOWAIT;
2306 /* Get cache's description obj. */
2307 cachep = kmem_cache_zalloc(&cache_cache, gfp);
2308 if (!cachep)
2309 goto oops;
2311 #if DEBUG
2312 cachep->obj_size = size;
2315 * Both debugging options require word-alignment which is calculated
2316 * into align above.
2318 if (flags & SLAB_RED_ZONE) {
2319 /* add space for red zone words */
2320 cachep->obj_offset += sizeof(unsigned long long);
2321 size += 2 * sizeof(unsigned long long);
2323 if (flags & SLAB_STORE_USER) {
2324 /* user store requires one word storage behind the end of
2325 * the real object. But if the second red zone needs to be
2326 * aligned to 64 bits, we must allow that much space.
2328 if (flags & SLAB_RED_ZONE)
2329 size += REDZONE_ALIGN;
2330 else
2331 size += BYTES_PER_WORD;
2333 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2334 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
2335 && cachep->obj_size > cache_line_size() && ALIGN(size, align) < PAGE_SIZE) {
2336 cachep->obj_offset += PAGE_SIZE - ALIGN(size, align);
2337 size = PAGE_SIZE;
2339 #endif
2340 #endif
2343 * Determine if the slab management is 'on' or 'off' slab.
2344 * (bootstrapping cannot cope with offslab caches so don't do
2345 * it too early on. Always use on-slab management when
2346 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2348 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
2349 !(flags & SLAB_NOLEAKTRACE))
2351 * Size is large, assume best to place the slab management obj
2352 * off-slab (should allow better packing of objs).
2354 flags |= CFLGS_OFF_SLAB;
2356 size = ALIGN(size, align);
2358 left_over = calculate_slab_order(cachep, size, align, flags);
2360 if (!cachep->num) {
2361 printk(KERN_ERR
2362 "kmem_cache_create: couldn't create cache %s.\n", name);
2363 kmem_cache_free(&cache_cache, cachep);
2364 cachep = NULL;
2365 goto oops;
2367 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2368 + sizeof(struct slab), align);
2371 * If the slab has been placed off-slab, and we have enough space then
2372 * move it on-slab. This is at the expense of any extra colouring.
2374 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2375 flags &= ~CFLGS_OFF_SLAB;
2376 left_over -= slab_size;
2379 if (flags & CFLGS_OFF_SLAB) {
2380 /* really off slab. No need for manual alignment */
2381 slab_size =
2382 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2384 #ifdef CONFIG_PAGE_POISONING
2385 /* If we're going to use the generic kernel_map_pages()
2386 * poisoning, then it's going to smash the contents of
2387 * the redzone and userword anyhow, so switch them off.
2389 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2390 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2391 #endif
2394 cachep->colour_off = cache_line_size();
2395 /* Offset must be a multiple of the alignment. */
2396 if (cachep->colour_off < align)
2397 cachep->colour_off = align;
2398 cachep->colour = left_over / cachep->colour_off;
2399 cachep->slab_size = slab_size;
2400 cachep->flags = flags;
2401 cachep->gfpflags = 0;
2402 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2403 cachep->gfpflags |= GFP_DMA;
2404 cachep->buffer_size = size;
2405 cachep->reciprocal_buffer_size = reciprocal_value(size);
2407 if (flags & CFLGS_OFF_SLAB) {
2408 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2410 * This is a possibility for one of the malloc_sizes caches.
2411 * But since we go off slab only for object size greater than
2412 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2413 * this should not happen at all.
2414 * But leave a BUG_ON for some lucky dude.
2416 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
2418 cachep->ctor = ctor;
2419 cachep->name = name;
2421 if (setup_cpu_cache(cachep, gfp)) {
2422 __kmem_cache_destroy(cachep);
2423 cachep = NULL;
2424 goto oops;
2427 /* cache setup completed, link it into the list */
2428 list_add(&cachep->next, &cache_chain);
2429 oops:
2430 if (!cachep && (flags & SLAB_PANIC))
2431 panic("kmem_cache_create(): failed to create slab `%s'\n",
2432 name);
2433 if (slab_is_available()) {
2434 mutex_unlock(&cache_chain_mutex);
2435 put_online_cpus();
2437 return cachep;
2439 EXPORT_SYMBOL(kmem_cache_create);
2441 #if DEBUG
2442 static void check_irq_off(void)
2444 BUG_ON(!irqs_disabled());
2447 static void check_irq_on(void)
2449 BUG_ON(irqs_disabled());
2452 static void check_spinlock_acquired(struct kmem_cache *cachep)
2454 #ifdef CONFIG_SMP
2455 check_irq_off();
2456 assert_spin_locked(&cachep->nodelists[numa_mem_id()]->list_lock);
2457 #endif
2460 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2462 #ifdef CONFIG_SMP
2463 check_irq_off();
2464 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2465 #endif
2468 #else
2469 #define check_irq_off() do { } while(0)
2470 #define check_irq_on() do { } while(0)
2471 #define check_spinlock_acquired(x) do { } while(0)
2472 #define check_spinlock_acquired_node(x, y) do { } while(0)
2473 #endif
2475 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2476 struct array_cache *ac,
2477 int force, int node);
2479 static void do_drain(void *arg)
2481 struct kmem_cache *cachep = arg;
2482 struct array_cache *ac;
2483 int node = numa_mem_id();
2485 check_irq_off();
2486 ac = cpu_cache_get(cachep);
2487 spin_lock(&cachep->nodelists[node]->list_lock);
2488 free_block(cachep, ac->entry, ac->avail, node);
2489 spin_unlock(&cachep->nodelists[node]->list_lock);
2490 ac->avail = 0;
2493 static void drain_cpu_caches(struct kmem_cache *cachep)
2495 struct kmem_list3 *l3;
2496 int node;
2498 on_each_cpu(do_drain, cachep, 1);
2499 check_irq_on();
2500 for_each_online_node(node) {
2501 l3 = cachep->nodelists[node];
2502 if (l3 && l3->alien)
2503 drain_alien_cache(cachep, l3->alien);
2506 for_each_online_node(node) {
2507 l3 = cachep->nodelists[node];
2508 if (l3)
2509 drain_array(cachep, l3, l3->shared, 1, node);
2514 * Remove slabs from the list of free slabs.
2515 * Specify the number of slabs to drain in tofree.
2517 * Returns the actual number of slabs released.
2519 static int drain_freelist(struct kmem_cache *cache,
2520 struct kmem_list3 *l3, int tofree)
2522 struct list_head *p;
2523 int nr_freed;
2524 struct slab *slabp;
2526 nr_freed = 0;
2527 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
2529 spin_lock_irq(&l3->list_lock);
2530 p = l3->slabs_free.prev;
2531 if (p == &l3->slabs_free) {
2532 spin_unlock_irq(&l3->list_lock);
2533 goto out;
2536 slabp = list_entry(p, struct slab, list);
2537 #if DEBUG
2538 BUG_ON(slabp->inuse);
2539 #endif
2540 list_del(&slabp->list);
2542 * Safe to drop the lock. The slab is no longer linked
2543 * to the cache.
2545 l3->free_objects -= cache->num;
2546 spin_unlock_irq(&l3->list_lock);
2547 slab_destroy(cache, slabp);
2548 nr_freed++;
2550 out:
2551 return nr_freed;
2554 /* Called with cache_chain_mutex held to protect against cpu hotplug */
2555 static int __cache_shrink(struct kmem_cache *cachep)
2557 int ret = 0, i = 0;
2558 struct kmem_list3 *l3;
2560 drain_cpu_caches(cachep);
2562 check_irq_on();
2563 for_each_online_node(i) {
2564 l3 = cachep->nodelists[i];
2565 if (!l3)
2566 continue;
2568 drain_freelist(cachep, l3, l3->free_objects);
2570 ret += !list_empty(&l3->slabs_full) ||
2571 !list_empty(&l3->slabs_partial);
2573 return (ret ? 1 : 0);
2577 * kmem_cache_shrink - Shrink a cache.
2578 * @cachep: The cache to shrink.
2580 * Releases as many slabs as possible for a cache.
2581 * To help debugging, a zero exit status indicates all slabs were released.
2583 int kmem_cache_shrink(struct kmem_cache *cachep)
2585 int ret;
2586 BUG_ON(!cachep || in_interrupt());
2588 get_online_cpus();
2589 mutex_lock(&cache_chain_mutex);
2590 ret = __cache_shrink(cachep);
2591 mutex_unlock(&cache_chain_mutex);
2592 put_online_cpus();
2593 return ret;
2595 EXPORT_SYMBOL(kmem_cache_shrink);
2598 * kmem_cache_destroy - delete a cache
2599 * @cachep: the cache to destroy
2601 * Remove a &struct kmem_cache object from the slab cache.
2603 * It is expected this function will be called by a module when it is
2604 * unloaded. This will remove the cache completely, and avoid a duplicate
2605 * cache being allocated each time a module is loaded and unloaded, if the
2606 * module doesn't have persistent in-kernel storage across loads and unloads.
2608 * The cache must be empty before calling this function.
2610 * The caller must guarantee that no one will allocate memory from the cache
2611 * during the kmem_cache_destroy().
2613 void kmem_cache_destroy(struct kmem_cache *cachep)
2615 BUG_ON(!cachep || in_interrupt());
2617 /* Find the cache in the chain of caches. */
2618 get_online_cpus();
2619 mutex_lock(&cache_chain_mutex);
2621 * the chain is never empty, cache_cache is never destroyed
2623 list_del(&cachep->next);
2624 if (__cache_shrink(cachep)) {
2625 slab_error(cachep, "Can't free all objects");
2626 list_add(&cachep->next, &cache_chain);
2627 mutex_unlock(&cache_chain_mutex);
2628 put_online_cpus();
2629 return;
2632 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
2633 rcu_barrier();
2635 __kmem_cache_destroy(cachep);
2636 mutex_unlock(&cache_chain_mutex);
2637 put_online_cpus();
2639 EXPORT_SYMBOL(kmem_cache_destroy);
2642 * Get the memory for a slab management obj.
2643 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2644 * always come from malloc_sizes caches. The slab descriptor cannot
2645 * come from the same cache which is getting created because,
2646 * when we are searching for an appropriate cache for these
2647 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2648 * If we are creating a malloc_sizes cache here it would not be visible to
2649 * kmem_find_general_cachep till the initialization is complete.
2650 * Hence we cannot have slabp_cache same as the original cache.
2652 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2653 int colour_off, gfp_t local_flags,
2654 int nodeid)
2656 struct slab *slabp;
2658 if (OFF_SLAB(cachep)) {
2659 /* Slab management obj is off-slab. */
2660 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2661 local_flags, nodeid);
2663 * If the first object in the slab is leaked (it's allocated
2664 * but no one has a reference to it), we want to make sure
2665 * kmemleak does not treat the ->s_mem pointer as a reference
2666 * to the object. Otherwise we will not report the leak.
2668 kmemleak_scan_area(&slabp->list, sizeof(struct list_head),
2669 local_flags);
2670 if (!slabp)
2671 return NULL;
2672 } else {
2673 slabp = objp + colour_off;
2674 colour_off += cachep->slab_size;
2676 slabp->inuse = 0;
2677 slabp->colouroff = colour_off;
2678 slabp->s_mem = objp + colour_off;
2679 slabp->nodeid = nodeid;
2680 slabp->free = 0;
2681 return slabp;
2684 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2686 return (kmem_bufctl_t *) (slabp + 1);
2689 static void cache_init_objs(struct kmem_cache *cachep,
2690 struct slab *slabp)
2692 int i;
2694 for (i = 0; i < cachep->num; i++) {
2695 void *objp = index_to_obj(cachep, slabp, i);
2696 #if DEBUG
2697 /* need to poison the objs? */
2698 if (cachep->flags & SLAB_POISON)
2699 poison_obj(cachep, objp, POISON_FREE);
2700 if (cachep->flags & SLAB_STORE_USER)
2701 *dbg_userword(cachep, objp) = NULL;
2703 if (cachep->flags & SLAB_RED_ZONE) {
2704 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2705 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2708 * Constructors are not allowed to allocate memory from the same
2709 * cache which they are a constructor for. Otherwise, deadlock.
2710 * They must also be threaded.
2712 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2713 cachep->ctor(objp + obj_offset(cachep));
2715 if (cachep->flags & SLAB_RED_ZONE) {
2716 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2717 slab_error(cachep, "constructor overwrote the"
2718 " end of an object");
2719 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2720 slab_error(cachep, "constructor overwrote the"
2721 " start of an object");
2723 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2724 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2725 kernel_map_pages(virt_to_page(objp),
2726 cachep->buffer_size / PAGE_SIZE, 0);
2727 #else
2728 if (cachep->ctor)
2729 cachep->ctor(objp);
2730 #endif
2731 slab_bufctl(slabp)[i] = i + 1;
2733 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2736 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2738 if (CONFIG_ZONE_DMA_FLAG) {
2739 if (flags & GFP_DMA)
2740 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2741 else
2742 BUG_ON(cachep->gfpflags & GFP_DMA);
2746 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2747 int nodeid)
2749 void *objp = index_to_obj(cachep, slabp, slabp->free);
2750 kmem_bufctl_t next;
2752 slabp->inuse++;
2753 next = slab_bufctl(slabp)[slabp->free];
2754 #if DEBUG
2755 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2756 WARN_ON(slabp->nodeid != nodeid);
2757 #endif
2758 slabp->free = next;
2760 return objp;
2763 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2764 void *objp, int nodeid)
2766 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2768 #if DEBUG
2769 /* Verify that the slab belongs to the intended node */
2770 WARN_ON(slabp->nodeid != nodeid);
2772 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2773 printk(KERN_ERR "slab: double free detected in cache "
2774 "'%s', objp %p\n", cachep->name, objp);
2775 BUG();
2777 #endif
2778 slab_bufctl(slabp)[objnr] = slabp->free;
2779 slabp->free = objnr;
2780 slabp->inuse--;
2784 * Map pages beginning at addr to the given cache and slab. This is required
2785 * for the slab allocator to be able to lookup the cache and slab of a
2786 * virtual address for kfree, ksize, and slab debugging.
2788 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2789 void *addr)
2791 int nr_pages;
2792 struct page *page;
2794 page = virt_to_page(addr);
2796 nr_pages = 1;
2797 if (likely(!PageCompound(page)))
2798 nr_pages <<= cache->gfporder;
2800 do {
2801 page_set_cache(page, cache);
2802 page_set_slab(page, slab);
2803 page++;
2804 } while (--nr_pages);
2808 * Grow (by 1) the number of slabs within a cache. This is called by
2809 * kmem_cache_alloc() when there are no active objs left in a cache.
2811 static int cache_grow(struct kmem_cache *cachep,
2812 gfp_t flags, int nodeid, void *objp)
2814 struct slab *slabp;
2815 size_t offset;
2816 gfp_t local_flags;
2817 struct kmem_list3 *l3;
2820 * Be lazy and only check for valid flags here, keeping it out of the
2821 * critical path in kmem_cache_alloc().
2823 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2824 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2826 /* Take the l3 list lock to change the colour_next on this node */
2827 check_irq_off();
2828 l3 = cachep->nodelists[nodeid];
2829 spin_lock(&l3->list_lock);
2831 /* Get colour for the slab, and cal the next value. */
2832 offset = l3->colour_next;
2833 l3->colour_next++;
2834 if (l3->colour_next >= cachep->colour)
2835 l3->colour_next = 0;
2836 spin_unlock(&l3->list_lock);
2838 offset *= cachep->colour_off;
2840 if (local_flags & __GFP_WAIT)
2841 local_irq_enable();
2844 * The test for missing atomic flag is performed here, rather than
2845 * the more obvious place, simply to reduce the critical path length
2846 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2847 * will eventually be caught here (where it matters).
2849 kmem_flagcheck(cachep, flags);
2852 * Get mem for the objs. Attempt to allocate a physical page from
2853 * 'nodeid'.
2855 if (!objp)
2856 objp = kmem_getpages(cachep, local_flags, nodeid);
2857 if (!objp)
2858 goto failed;
2860 /* Get slab management. */
2861 slabp = alloc_slabmgmt(cachep, objp, offset,
2862 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2863 if (!slabp)
2864 goto opps1;
2866 slab_map_pages(cachep, slabp, objp);
2868 cache_init_objs(cachep, slabp);
2870 if (local_flags & __GFP_WAIT)
2871 local_irq_disable();
2872 check_irq_off();
2873 spin_lock(&l3->list_lock);
2875 /* Make slab active. */
2876 list_add_tail(&slabp->list, &(l3->slabs_free));
2877 STATS_INC_GROWN(cachep);
2878 l3->free_objects += cachep->num;
2879 spin_unlock(&l3->list_lock);
2880 return 1;
2881 opps1:
2882 kmem_freepages(cachep, objp);
2883 failed:
2884 if (local_flags & __GFP_WAIT)
2885 local_irq_disable();
2886 return 0;
2889 #if DEBUG
2892 * Perform extra freeing checks:
2893 * - detect bad pointers.
2894 * - POISON/RED_ZONE checking
2896 static void kfree_debugcheck(const void *objp)
2898 if (!virt_addr_valid(objp)) {
2899 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2900 (unsigned long)objp);
2901 BUG();
2905 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2907 unsigned long long redzone1, redzone2;
2909 redzone1 = *dbg_redzone1(cache, obj);
2910 redzone2 = *dbg_redzone2(cache, obj);
2913 * Redzone is ok.
2915 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2916 return;
2918 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2919 slab_error(cache, "double free detected");
2920 else
2921 slab_error(cache, "memory outside object was overwritten");
2923 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2924 obj, redzone1, redzone2);
2927 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2928 void *caller)
2930 struct page *page;
2931 unsigned int objnr;
2932 struct slab *slabp;
2934 BUG_ON(virt_to_cache(objp) != cachep);
2936 objp -= obj_offset(cachep);
2937 kfree_debugcheck(objp);
2938 page = virt_to_head_page(objp);
2940 slabp = page_get_slab(page);
2942 if (cachep->flags & SLAB_RED_ZONE) {
2943 verify_redzone_free(cachep, objp);
2944 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2945 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2947 if (cachep->flags & SLAB_STORE_USER)
2948 *dbg_userword(cachep, objp) = caller;
2950 objnr = obj_to_index(cachep, slabp, objp);
2952 BUG_ON(objnr >= cachep->num);
2953 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
2955 #ifdef CONFIG_DEBUG_SLAB_LEAK
2956 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2957 #endif
2958 if (cachep->flags & SLAB_POISON) {
2959 #ifdef CONFIG_DEBUG_PAGEALLOC
2960 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
2961 store_stackinfo(cachep, objp, (unsigned long)caller);
2962 kernel_map_pages(virt_to_page(objp),
2963 cachep->buffer_size / PAGE_SIZE, 0);
2964 } else {
2965 poison_obj(cachep, objp, POISON_FREE);
2967 #else
2968 poison_obj(cachep, objp, POISON_FREE);
2969 #endif
2971 return objp;
2974 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
2976 kmem_bufctl_t i;
2977 int entries = 0;
2979 /* Check slab's freelist to see if this obj is there. */
2980 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2981 entries++;
2982 if (entries > cachep->num || i >= cachep->num)
2983 goto bad;
2985 if (entries != cachep->num - slabp->inuse) {
2986 bad:
2987 printk(KERN_ERR "slab: Internal list corruption detected in "
2988 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2989 cachep->name, cachep->num, slabp, slabp->inuse);
2990 for (i = 0;
2991 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
2992 i++) {
2993 if (i % 16 == 0)
2994 printk("\n%03x:", i);
2995 printk(" %02x", ((unsigned char *)slabp)[i]);
2997 printk("\n");
2998 BUG();
3001 #else
3002 #define kfree_debugcheck(x) do { } while(0)
3003 #define cache_free_debugcheck(x,objp,z) (objp)
3004 #define check_slabp(x,y) do { } while(0)
3005 #endif
3007 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
3009 int batchcount;
3010 struct kmem_list3 *l3;
3011 struct array_cache *ac;
3012 int node;
3014 retry:
3015 check_irq_off();
3016 node = numa_mem_id();
3017 ac = cpu_cache_get(cachep);
3018 batchcount = ac->batchcount;
3019 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
3021 * If there was little recent activity on this cache, then
3022 * perform only a partial refill. Otherwise we could generate
3023 * refill bouncing.
3025 batchcount = BATCHREFILL_LIMIT;
3027 l3 = cachep->nodelists[node];
3029 BUG_ON(ac->avail > 0 || !l3);
3030 spin_lock(&l3->list_lock);
3032 /* See if we can refill from the shared array */
3033 if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) {
3034 l3->shared->touched = 1;
3035 goto alloc_done;
3038 while (batchcount > 0) {
3039 struct list_head *entry;
3040 struct slab *slabp;
3041 /* Get slab alloc is to come from. */
3042 entry = l3->slabs_partial.next;
3043 if (entry == &l3->slabs_partial) {
3044 l3->free_touched = 1;
3045 entry = l3->slabs_free.next;
3046 if (entry == &l3->slabs_free)
3047 goto must_grow;
3050 slabp = list_entry(entry, struct slab, list);
3051 check_slabp(cachep, slabp);
3052 check_spinlock_acquired(cachep);
3055 * The slab was either on partial or free list so
3056 * there must be at least one object available for
3057 * allocation.
3059 BUG_ON(slabp->inuse >= cachep->num);
3061 while (slabp->inuse < cachep->num && batchcount--) {
3062 STATS_INC_ALLOCED(cachep);
3063 STATS_INC_ACTIVE(cachep);
3064 STATS_SET_HIGH(cachep);
3066 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
3067 node);
3069 check_slabp(cachep, slabp);
3071 /* move slabp to correct slabp list: */
3072 list_del(&slabp->list);
3073 if (slabp->free == BUFCTL_END)
3074 list_add(&slabp->list, &l3->slabs_full);
3075 else
3076 list_add(&slabp->list, &l3->slabs_partial);
3079 must_grow:
3080 l3->free_objects -= ac->avail;
3081 alloc_done:
3082 spin_unlock(&l3->list_lock);
3084 if (unlikely(!ac->avail)) {
3085 int x;
3086 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
3088 /* cache_grow can reenable interrupts, then ac could change. */
3089 ac = cpu_cache_get(cachep);
3090 if (!x && ac->avail == 0) /* no objects in sight? abort */
3091 return NULL;
3093 if (!ac->avail) /* objects refilled by interrupt? */
3094 goto retry;
3096 ac->touched = 1;
3097 return ac->entry[--ac->avail];
3100 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3101 gfp_t flags)
3103 might_sleep_if(flags & __GFP_WAIT);
3104 #if DEBUG
3105 kmem_flagcheck(cachep, flags);
3106 #endif
3109 #if DEBUG
3110 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3111 gfp_t flags, void *objp, void *caller)
3113 if (!objp)
3114 return objp;
3115 if (cachep->flags & SLAB_POISON) {
3116 #ifdef CONFIG_DEBUG_PAGEALLOC
3117 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
3118 kernel_map_pages(virt_to_page(objp),
3119 cachep->buffer_size / PAGE_SIZE, 1);
3120 else
3121 check_poison_obj(cachep, objp);
3122 #else
3123 check_poison_obj(cachep, objp);
3124 #endif
3125 poison_obj(cachep, objp, POISON_INUSE);
3127 if (cachep->flags & SLAB_STORE_USER)
3128 *dbg_userword(cachep, objp) = caller;
3130 if (cachep->flags & SLAB_RED_ZONE) {
3131 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3132 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3133 slab_error(cachep, "double free, or memory outside"
3134 " object was overwritten");
3135 printk(KERN_ERR
3136 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3137 objp, *dbg_redzone1(cachep, objp),
3138 *dbg_redzone2(cachep, objp));
3140 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3141 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3143 #ifdef CONFIG_DEBUG_SLAB_LEAK
3145 struct slab *slabp;
3146 unsigned objnr;
3148 slabp = page_get_slab(virt_to_head_page(objp));
3149 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3150 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3152 #endif
3153 objp += obj_offset(cachep);
3154 if (cachep->ctor && cachep->flags & SLAB_POISON)
3155 cachep->ctor(objp);
3156 #if ARCH_SLAB_MINALIGN
3157 if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) {
3158 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3159 objp, ARCH_SLAB_MINALIGN);
3161 #endif
3162 return objp;
3164 #else
3165 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3166 #endif
3168 static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
3170 if (cachep == &cache_cache)
3171 return false;
3173 return should_failslab(obj_size(cachep), flags, cachep->flags);
3176 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3178 void *objp;
3179 struct array_cache *ac;
3181 check_irq_off();
3183 ac = cpu_cache_get(cachep);
3184 if (likely(ac->avail)) {
3185 STATS_INC_ALLOCHIT(cachep);
3186 ac->touched = 1;
3187 objp = ac->entry[--ac->avail];
3188 } else {
3189 STATS_INC_ALLOCMISS(cachep);
3190 objp = cache_alloc_refill(cachep, flags);
3192 * the 'ac' may be updated by cache_alloc_refill(),
3193 * and kmemleak_erase() requires its correct value.
3195 ac = cpu_cache_get(cachep);
3198 * To avoid a false negative, if an object that is in one of the
3199 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3200 * treat the array pointers as a reference to the object.
3202 if (objp)
3203 kmemleak_erase(&ac->entry[ac->avail]);
3204 return objp;
3207 #ifdef CONFIG_NUMA
3209 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3211 * If we are in_interrupt, then process context, including cpusets and
3212 * mempolicy, may not apply and should not be used for allocation policy.
3214 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3216 int nid_alloc, nid_here;
3218 if (in_interrupt() || (flags & __GFP_THISNODE))
3219 return NULL;
3220 nid_alloc = nid_here = numa_mem_id();
3221 get_mems_allowed();
3222 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3223 nid_alloc = cpuset_slab_spread_node();
3224 else if (current->mempolicy)
3225 nid_alloc = slab_node(current->mempolicy);
3226 put_mems_allowed();
3227 if (nid_alloc != nid_here)
3228 return ____cache_alloc_node(cachep, flags, nid_alloc);
3229 return NULL;
3233 * Fallback function if there was no memory available and no objects on a
3234 * certain node and fall back is permitted. First we scan all the
3235 * available nodelists for available objects. If that fails then we
3236 * perform an allocation without specifying a node. This allows the page
3237 * allocator to do its reclaim / fallback magic. We then insert the
3238 * slab into the proper nodelist and then allocate from it.
3240 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3242 struct zonelist *zonelist;
3243 gfp_t local_flags;
3244 struct zoneref *z;
3245 struct zone *zone;
3246 enum zone_type high_zoneidx = gfp_zone(flags);
3247 void *obj = NULL;
3248 int nid;
3250 if (flags & __GFP_THISNODE)
3251 return NULL;
3253 get_mems_allowed();
3254 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
3255 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3257 retry:
3259 * Look through allowed nodes for objects available
3260 * from existing per node queues.
3262 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3263 nid = zone_to_nid(zone);
3265 if (cpuset_zone_allowed_hardwall(zone, flags) &&
3266 cache->nodelists[nid] &&
3267 cache->nodelists[nid]->free_objects) {
3268 obj = ____cache_alloc_node(cache,
3269 flags | GFP_THISNODE, nid);
3270 if (obj)
3271 break;
3275 if (!obj) {
3277 * This allocation will be performed within the constraints
3278 * of the current cpuset / memory policy requirements.
3279 * We may trigger various forms of reclaim on the allowed
3280 * set and go into memory reserves if necessary.
3282 if (local_flags & __GFP_WAIT)
3283 local_irq_enable();
3284 kmem_flagcheck(cache, flags);
3285 obj = kmem_getpages(cache, local_flags, numa_mem_id());
3286 if (local_flags & __GFP_WAIT)
3287 local_irq_disable();
3288 if (obj) {
3290 * Insert into the appropriate per node queues
3292 nid = page_to_nid(virt_to_page(obj));
3293 if (cache_grow(cache, flags, nid, obj)) {
3294 obj = ____cache_alloc_node(cache,
3295 flags | GFP_THISNODE, nid);
3296 if (!obj)
3298 * Another processor may allocate the
3299 * objects in the slab since we are
3300 * not holding any locks.
3302 goto retry;
3303 } else {
3304 /* cache_grow already freed obj */
3305 obj = NULL;
3309 put_mems_allowed();
3310 return obj;
3314 * A interface to enable slab creation on nodeid
3316 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3317 int nodeid)
3319 struct list_head *entry;
3320 struct slab *slabp;
3321 struct kmem_list3 *l3;
3322 void *obj;
3323 int x;
3325 l3 = cachep->nodelists[nodeid];
3326 BUG_ON(!l3);
3328 retry:
3329 check_irq_off();
3330 spin_lock(&l3->list_lock);
3331 entry = l3->slabs_partial.next;
3332 if (entry == &l3->slabs_partial) {
3333 l3->free_touched = 1;
3334 entry = l3->slabs_free.next;
3335 if (entry == &l3->slabs_free)
3336 goto must_grow;
3339 slabp = list_entry(entry, struct slab, list);
3340 check_spinlock_acquired_node(cachep, nodeid);
3341 check_slabp(cachep, slabp);
3343 STATS_INC_NODEALLOCS(cachep);
3344 STATS_INC_ACTIVE(cachep);
3345 STATS_SET_HIGH(cachep);
3347 BUG_ON(slabp->inuse == cachep->num);
3349 obj = slab_get_obj(cachep, slabp, nodeid);
3350 check_slabp(cachep, slabp);
3351 l3->free_objects--;
3352 /* move slabp to correct slabp list: */
3353 list_del(&slabp->list);
3355 if (slabp->free == BUFCTL_END)
3356 list_add(&slabp->list, &l3->slabs_full);
3357 else
3358 list_add(&slabp->list, &l3->slabs_partial);
3360 spin_unlock(&l3->list_lock);
3361 goto done;
3363 must_grow:
3364 spin_unlock(&l3->list_lock);
3365 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
3366 if (x)
3367 goto retry;
3369 return fallback_alloc(cachep, flags);
3371 done:
3372 return obj;
3376 * kmem_cache_alloc_node - Allocate an object on the specified node
3377 * @cachep: The cache to allocate from.
3378 * @flags: See kmalloc().
3379 * @nodeid: node number of the target node.
3380 * @caller: return address of caller, used for debug information
3382 * Identical to kmem_cache_alloc but it will allocate memory on the given
3383 * node, which can improve the performance for cpu bound structures.
3385 * Fallback to other node is possible if __GFP_THISNODE is not set.
3387 static __always_inline void *
3388 __cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3389 void *caller)
3391 unsigned long save_flags;
3392 void *ptr;
3393 int slab_node = numa_mem_id();
3395 flags &= gfp_allowed_mask;
3397 lockdep_trace_alloc(flags);
3399 if (slab_should_failslab(cachep, flags))
3400 return NULL;
3402 cache_alloc_debugcheck_before(cachep, flags);
3403 local_irq_save(save_flags);
3405 if (nodeid == -1)
3406 nodeid = slab_node;
3408 if (unlikely(!cachep->nodelists[nodeid])) {
3409 /* Node not bootstrapped yet */
3410 ptr = fallback_alloc(cachep, flags);
3411 goto out;
3414 if (nodeid == slab_node) {
3416 * Use the locally cached objects if possible.
3417 * However ____cache_alloc does not allow fallback
3418 * to other nodes. It may fail while we still have
3419 * objects on other nodes available.
3421 ptr = ____cache_alloc(cachep, flags);
3422 if (ptr)
3423 goto out;
3425 /* ___cache_alloc_node can fall back to other nodes */
3426 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3427 out:
3428 local_irq_restore(save_flags);
3429 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3430 kmemleak_alloc_recursive(ptr, obj_size(cachep), 1, cachep->flags,
3431 flags);
3433 if (likely(ptr))
3434 kmemcheck_slab_alloc(cachep, flags, ptr, obj_size(cachep));
3436 if (unlikely((flags & __GFP_ZERO) && ptr))
3437 memset(ptr, 0, obj_size(cachep));
3439 return ptr;
3442 static __always_inline void *
3443 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3445 void *objp;
3447 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3448 objp = alternate_node_alloc(cache, flags);
3449 if (objp)
3450 goto out;
3452 objp = ____cache_alloc(cache, flags);
3455 * We may just have run out of memory on the local node.
3456 * ____cache_alloc_node() knows how to locate memory on other nodes
3458 if (!objp)
3459 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3461 out:
3462 return objp;
3464 #else
3466 static __always_inline void *
3467 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3469 return ____cache_alloc(cachep, flags);
3472 #endif /* CONFIG_NUMA */
3474 static __always_inline void *
3475 __cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3477 unsigned long save_flags;
3478 void *objp;
3480 flags &= gfp_allowed_mask;
3482 lockdep_trace_alloc(flags);
3484 if (slab_should_failslab(cachep, flags))
3485 return NULL;
3487 cache_alloc_debugcheck_before(cachep, flags);
3488 local_irq_save(save_flags);
3489 objp = __do_cache_alloc(cachep, flags);
3490 local_irq_restore(save_flags);
3491 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3492 kmemleak_alloc_recursive(objp, obj_size(cachep), 1, cachep->flags,
3493 flags);
3494 prefetchw(objp);
3496 if (likely(objp))
3497 kmemcheck_slab_alloc(cachep, flags, objp, obj_size(cachep));
3499 if (unlikely((flags & __GFP_ZERO) && objp))
3500 memset(objp, 0, obj_size(cachep));
3502 return objp;
3506 * Caller needs to acquire correct kmem_list's list_lock
3508 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3509 int node)
3511 int i;
3512 struct kmem_list3 *l3;
3514 for (i = 0; i < nr_objects; i++) {
3515 void *objp = objpp[i];
3516 struct slab *slabp;
3518 slabp = virt_to_slab(objp);
3519 l3 = cachep->nodelists[node];
3520 list_del(&slabp->list);
3521 check_spinlock_acquired_node(cachep, node);
3522 check_slabp(cachep, slabp);
3523 slab_put_obj(cachep, slabp, objp, node);
3524 STATS_DEC_ACTIVE(cachep);
3525 l3->free_objects++;
3526 check_slabp(cachep, slabp);
3528 /* fixup slab chains */
3529 if (slabp->inuse == 0) {
3530 if (l3->free_objects > l3->free_limit) {
3531 l3->free_objects -= cachep->num;
3532 /* No need to drop any previously held
3533 * lock here, even if we have a off-slab slab
3534 * descriptor it is guaranteed to come from
3535 * a different cache, refer to comments before
3536 * alloc_slabmgmt.
3538 slab_destroy(cachep, slabp);
3539 } else {
3540 list_add(&slabp->list, &l3->slabs_free);
3542 } else {
3543 /* Unconditionally move a slab to the end of the
3544 * partial list on free - maximum time for the
3545 * other objects to be freed, too.
3547 list_add_tail(&slabp->list, &l3->slabs_partial);
3552 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3554 int batchcount;
3555 struct kmem_list3 *l3;
3556 int node = numa_mem_id();
3558 batchcount = ac->batchcount;
3559 #if DEBUG
3560 BUG_ON(!batchcount || batchcount > ac->avail);
3561 #endif
3562 check_irq_off();
3563 l3 = cachep->nodelists[node];
3564 spin_lock(&l3->list_lock);
3565 if (l3->shared) {
3566 struct array_cache *shared_array = l3->shared;
3567 int max = shared_array->limit - shared_array->avail;
3568 if (max) {
3569 if (batchcount > max)
3570 batchcount = max;
3571 memcpy(&(shared_array->entry[shared_array->avail]),
3572 ac->entry, sizeof(void *) * batchcount);
3573 shared_array->avail += batchcount;
3574 goto free_done;
3578 free_block(cachep, ac->entry, batchcount, node);
3579 free_done:
3580 #if STATS
3582 int i = 0;
3583 struct list_head *p;
3585 p = l3->slabs_free.next;
3586 while (p != &(l3->slabs_free)) {
3587 struct slab *slabp;
3589 slabp = list_entry(p, struct slab, list);
3590 BUG_ON(slabp->inuse);
3592 i++;
3593 p = p->next;
3595 STATS_SET_FREEABLE(cachep, i);
3597 #endif
3598 spin_unlock(&l3->list_lock);
3599 ac->avail -= batchcount;
3600 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3604 * Release an obj back to its cache. If the obj has a constructed state, it must
3605 * be in this state _before_ it is released. Called with disabled ints.
3607 static inline void __cache_free(struct kmem_cache *cachep, void *objp)
3609 struct array_cache *ac = cpu_cache_get(cachep);
3611 check_irq_off();
3612 kmemleak_free_recursive(objp, cachep->flags);
3613 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3615 kmemcheck_slab_free(cachep, objp, obj_size(cachep));
3618 * Skip calling cache_free_alien() when the platform is not numa.
3619 * This will avoid cache misses that happen while accessing slabp (which
3620 * is per page memory reference) to get nodeid. Instead use a global
3621 * variable to skip the call, which is mostly likely to be present in
3622 * the cache.
3624 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3625 return;
3627 if (likely(ac->avail < ac->limit)) {
3628 STATS_INC_FREEHIT(cachep);
3629 ac->entry[ac->avail++] = objp;
3630 return;
3631 } else {
3632 STATS_INC_FREEMISS(cachep);
3633 cache_flusharray(cachep, ac);
3634 ac->entry[ac->avail++] = objp;
3639 * kmem_cache_alloc - Allocate an object
3640 * @cachep: The cache to allocate from.
3641 * @flags: See kmalloc().
3643 * Allocate an object from this cache. The flags are only relevant
3644 * if the cache has no available objects.
3646 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3648 void *ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3650 trace_kmem_cache_alloc(_RET_IP_, ret,
3651 obj_size(cachep), cachep->buffer_size, flags);
3653 return ret;
3655 EXPORT_SYMBOL(kmem_cache_alloc);
3657 #ifdef CONFIG_TRACING
3658 void *
3659 kmem_cache_alloc_trace(size_t size, struct kmem_cache *cachep, gfp_t flags)
3661 void *ret;
3663 ret = __cache_alloc(cachep, flags, __builtin_return_address(0));
3665 trace_kmalloc(_RET_IP_, ret,
3666 size, slab_buffer_size(cachep), flags);
3667 return ret;
3669 EXPORT_SYMBOL(kmem_cache_alloc_trace);
3670 #endif
3672 #ifdef CONFIG_NUMA
3673 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3675 void *ret = __cache_alloc_node(cachep, flags, nodeid,
3676 __builtin_return_address(0));
3678 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3679 obj_size(cachep), cachep->buffer_size,
3680 flags, nodeid);
3682 return ret;
3684 EXPORT_SYMBOL(kmem_cache_alloc_node);
3686 #ifdef CONFIG_TRACING
3687 void *kmem_cache_alloc_node_trace(size_t size,
3688 struct kmem_cache *cachep,
3689 gfp_t flags,
3690 int nodeid)
3692 void *ret;
3694 ret = __cache_alloc_node(cachep, flags, nodeid,
3695 __builtin_return_address(0));
3696 trace_kmalloc_node(_RET_IP_, ret,
3697 size, slab_buffer_size(cachep),
3698 flags, nodeid);
3699 return ret;
3701 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3702 #endif
3704 static __always_inline void *
3705 __do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
3707 struct kmem_cache *cachep;
3709 cachep = kmem_find_general_cachep(size, flags);
3710 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3711 return cachep;
3712 return kmem_cache_alloc_node_trace(size, cachep, flags, node);
3715 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3716 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3718 return __do_kmalloc_node(size, flags, node,
3719 __builtin_return_address(0));
3721 EXPORT_SYMBOL(__kmalloc_node);
3723 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3724 int node, unsigned long caller)
3726 return __do_kmalloc_node(size, flags, node, (void *)caller);
3728 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3729 #else
3730 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3732 return __do_kmalloc_node(size, flags, node, NULL);
3734 EXPORT_SYMBOL(__kmalloc_node);
3735 #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3736 #endif /* CONFIG_NUMA */
3739 * __do_kmalloc - allocate memory
3740 * @size: how many bytes of memory are required.
3741 * @flags: the type of memory to allocate (see kmalloc).
3742 * @caller: function caller for debug tracking of the caller
3744 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3745 void *caller)
3747 struct kmem_cache *cachep;
3748 void *ret;
3750 /* If you want to save a few bytes .text space: replace
3751 * __ with kmem_.
3752 * Then kmalloc uses the uninlined functions instead of the inline
3753 * functions.
3755 cachep = __find_general_cachep(size, flags);
3756 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3757 return cachep;
3758 ret = __cache_alloc(cachep, flags, caller);
3760 trace_kmalloc((unsigned long) caller, ret,
3761 size, cachep->buffer_size, flags);
3763 return ret;
3767 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3768 void *__kmalloc(size_t size, gfp_t flags)
3770 return __do_kmalloc(size, flags, __builtin_return_address(0));
3772 EXPORT_SYMBOL(__kmalloc);
3774 void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3776 return __do_kmalloc(size, flags, (void *)caller);
3778 EXPORT_SYMBOL(__kmalloc_track_caller);
3780 #else
3781 void *__kmalloc(size_t size, gfp_t flags)
3783 return __do_kmalloc(size, flags, NULL);
3785 EXPORT_SYMBOL(__kmalloc);
3786 #endif
3789 * kmem_cache_free - Deallocate an object
3790 * @cachep: The cache the allocation was from.
3791 * @objp: The previously allocated object.
3793 * Free an object which was previously allocated from this
3794 * cache.
3796 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3798 unsigned long flags;
3800 local_irq_save(flags);
3801 debug_check_no_locks_freed(objp, obj_size(cachep));
3802 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3803 debug_check_no_obj_freed(objp, obj_size(cachep));
3804 __cache_free(cachep, objp);
3805 local_irq_restore(flags);
3807 trace_kmem_cache_free(_RET_IP_, objp);
3809 EXPORT_SYMBOL(kmem_cache_free);
3812 * kfree - free previously allocated memory
3813 * @objp: pointer returned by kmalloc.
3815 * If @objp is NULL, no operation is performed.
3817 * Don't free memory not originally allocated by kmalloc()
3818 * or you will run into trouble.
3820 void kfree(const void *objp)
3822 struct kmem_cache *c;
3823 unsigned long flags;
3825 trace_kfree(_RET_IP_, objp);
3827 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3828 return;
3829 local_irq_save(flags);
3830 kfree_debugcheck(objp);
3831 c = virt_to_cache(objp);
3832 debug_check_no_locks_freed(objp, obj_size(c));
3833 debug_check_no_obj_freed(objp, obj_size(c));
3834 __cache_free(c, (void *)objp);
3835 local_irq_restore(flags);
3837 EXPORT_SYMBOL(kfree);
3839 unsigned int kmem_cache_size(struct kmem_cache *cachep)
3841 return obj_size(cachep);
3843 EXPORT_SYMBOL(kmem_cache_size);
3846 * This initializes kmem_list3 or resizes various caches for all nodes.
3848 static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
3850 int node;
3851 struct kmem_list3 *l3;
3852 struct array_cache *new_shared;
3853 struct array_cache **new_alien = NULL;
3855 for_each_online_node(node) {
3857 if (use_alien_caches) {
3858 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
3859 if (!new_alien)
3860 goto fail;
3863 new_shared = NULL;
3864 if (cachep->shared) {
3865 new_shared = alloc_arraycache(node,
3866 cachep->shared*cachep->batchcount,
3867 0xbaadf00d, gfp);
3868 if (!new_shared) {
3869 free_alien_cache(new_alien);
3870 goto fail;
3874 l3 = cachep->nodelists[node];
3875 if (l3) {
3876 struct array_cache *shared = l3->shared;
3878 spin_lock_irq(&l3->list_lock);
3880 if (shared)
3881 free_block(cachep, shared->entry,
3882 shared->avail, node);
3884 l3->shared = new_shared;
3885 if (!l3->alien) {
3886 l3->alien = new_alien;
3887 new_alien = NULL;
3889 l3->free_limit = (1 + nr_cpus_node(node)) *
3890 cachep->batchcount + cachep->num;
3891 spin_unlock_irq(&l3->list_lock);
3892 kfree(shared);
3893 free_alien_cache(new_alien);
3894 continue;
3896 l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node);
3897 if (!l3) {
3898 free_alien_cache(new_alien);
3899 kfree(new_shared);
3900 goto fail;
3903 kmem_list3_init(l3);
3904 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3905 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3906 l3->shared = new_shared;
3907 l3->alien = new_alien;
3908 l3->free_limit = (1 + nr_cpus_node(node)) *
3909 cachep->batchcount + cachep->num;
3910 cachep->nodelists[node] = l3;
3912 return 0;
3914 fail:
3915 if (!cachep->next.next) {
3916 /* Cache is not active yet. Roll back what we did */
3917 node--;
3918 while (node >= 0) {
3919 if (cachep->nodelists[node]) {
3920 l3 = cachep->nodelists[node];
3922 kfree(l3->shared);
3923 free_alien_cache(l3->alien);
3924 kfree(l3);
3925 cachep->nodelists[node] = NULL;
3927 node--;
3930 return -ENOMEM;
3933 struct ccupdate_struct {
3934 struct kmem_cache *cachep;
3935 struct array_cache *new[NR_CPUS];
3938 static void do_ccupdate_local(void *info)
3940 struct ccupdate_struct *new = info;
3941 struct array_cache *old;
3943 check_irq_off();
3944 old = cpu_cache_get(new->cachep);
3946 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3947 new->new[smp_processor_id()] = old;
3950 /* Always called with the cache_chain_mutex held */
3951 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3952 int batchcount, int shared, gfp_t gfp)
3954 struct ccupdate_struct *new;
3955 int i;
3957 new = kzalloc(sizeof(*new), gfp);
3958 if (!new)
3959 return -ENOMEM;
3961 for_each_online_cpu(i) {
3962 new->new[i] = alloc_arraycache(cpu_to_mem(i), limit,
3963 batchcount, gfp);
3964 if (!new->new[i]) {
3965 for (i--; i >= 0; i--)
3966 kfree(new->new[i]);
3967 kfree(new);
3968 return -ENOMEM;
3971 new->cachep = cachep;
3973 on_each_cpu(do_ccupdate_local, (void *)new, 1);
3975 check_irq_on();
3976 cachep->batchcount = batchcount;
3977 cachep->limit = limit;
3978 cachep->shared = shared;
3980 for_each_online_cpu(i) {
3981 struct array_cache *ccold = new->new[i];
3982 if (!ccold)
3983 continue;
3984 spin_lock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
3985 free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i));
3986 spin_unlock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
3987 kfree(ccold);
3989 kfree(new);
3990 return alloc_kmemlist(cachep, gfp);
3993 /* Called with cache_chain_mutex held always */
3994 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
3996 int err;
3997 int limit, shared;
4000 * The head array serves three purposes:
4001 * - create a LIFO ordering, i.e. return objects that are cache-warm
4002 * - reduce the number of spinlock operations.
4003 * - reduce the number of linked list operations on the slab and
4004 * bufctl chains: array operations are cheaper.
4005 * The numbers are guessed, we should auto-tune as described by
4006 * Bonwick.
4008 if (cachep->buffer_size > 131072)
4009 limit = 1;
4010 else if (cachep->buffer_size > PAGE_SIZE)
4011 limit = 8;
4012 else if (cachep->buffer_size > 1024)
4013 limit = 24;
4014 else if (cachep->buffer_size > 256)
4015 limit = 54;
4016 else
4017 limit = 120;
4020 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
4021 * allocation behaviour: Most allocs on one cpu, most free operations
4022 * on another cpu. For these cases, an efficient object passing between
4023 * cpus is necessary. This is provided by a shared array. The array
4024 * replaces Bonwick's magazine layer.
4025 * On uniprocessor, it's functionally equivalent (but less efficient)
4026 * to a larger limit. Thus disabled by default.
4028 shared = 0;
4029 if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1)
4030 shared = 8;
4032 #if DEBUG
4034 * With debugging enabled, large batchcount lead to excessively long
4035 * periods with disabled local interrupts. Limit the batchcount
4037 if (limit > 32)
4038 limit = 32;
4039 #endif
4040 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp);
4041 if (err)
4042 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
4043 cachep->name, -err);
4044 return err;
4048 * Drain an array if it contains any elements taking the l3 lock only if
4049 * necessary. Note that the l3 listlock also protects the array_cache
4050 * if drain_array() is used on the shared array.
4052 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
4053 struct array_cache *ac, int force, int node)
4055 int tofree;
4057 if (!ac || !ac->avail)
4058 return;
4059 if (ac->touched && !force) {
4060 ac->touched = 0;
4061 } else {
4062 spin_lock_irq(&l3->list_lock);
4063 if (ac->avail) {
4064 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4065 if (tofree > ac->avail)
4066 tofree = (ac->avail + 1) / 2;
4067 free_block(cachep, ac->entry, tofree, node);
4068 ac->avail -= tofree;
4069 memmove(ac->entry, &(ac->entry[tofree]),
4070 sizeof(void *) * ac->avail);
4072 spin_unlock_irq(&l3->list_lock);
4077 * cache_reap - Reclaim memory from caches.
4078 * @w: work descriptor
4080 * Called from workqueue/eventd every few seconds.
4081 * Purpose:
4082 * - clear the per-cpu caches for this CPU.
4083 * - return freeable pages to the main free memory pool.
4085 * If we cannot acquire the cache chain mutex then just give up - we'll try
4086 * again on the next iteration.
4088 static void cache_reap(struct work_struct *w)
4090 struct kmem_cache *searchp;
4091 struct kmem_list3 *l3;
4092 int node = numa_mem_id();
4093 struct delayed_work *work = to_delayed_work(w);
4095 if (!mutex_trylock(&cache_chain_mutex))
4096 /* Give up. Setup the next iteration. */
4097 goto out;
4099 list_for_each_entry(searchp, &cache_chain, next) {
4100 check_irq_on();
4103 * We only take the l3 lock if absolutely necessary and we
4104 * have established with reasonable certainty that
4105 * we can do some work if the lock was obtained.
4107 l3 = searchp->nodelists[node];
4109 reap_alien(searchp, l3);
4111 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
4114 * These are racy checks but it does not matter
4115 * if we skip one check or scan twice.
4117 if (time_after(l3->next_reap, jiffies))
4118 goto next;
4120 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
4122 drain_array(searchp, l3, l3->shared, 0, node);
4124 if (l3->free_touched)
4125 l3->free_touched = 0;
4126 else {
4127 int freed;
4129 freed = drain_freelist(searchp, l3, (l3->free_limit +
4130 5 * searchp->num - 1) / (5 * searchp->num));
4131 STATS_ADD_REAPED(searchp, freed);
4133 next:
4134 cond_resched();
4136 check_irq_on();
4137 mutex_unlock(&cache_chain_mutex);
4138 next_reap_node();
4139 out:
4140 /* Set up the next iteration */
4141 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
4144 #ifdef CONFIG_SLABINFO
4146 static void print_slabinfo_header(struct seq_file *m)
4149 * Output format version, so at least we can change it
4150 * without _too_ many complaints.
4152 #if STATS
4153 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4154 #else
4155 seq_puts(m, "slabinfo - version: 2.1\n");
4156 #endif
4157 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4158 "<objperslab> <pagesperslab>");
4159 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4160 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4161 #if STATS
4162 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
4163 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
4164 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4165 #endif
4166 seq_putc(m, '\n');
4169 static void *s_start(struct seq_file *m, loff_t *pos)
4171 loff_t n = *pos;
4173 mutex_lock(&cache_chain_mutex);
4174 if (!n)
4175 print_slabinfo_header(m);
4177 return seq_list_start(&cache_chain, *pos);
4180 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4182 return seq_list_next(p, &cache_chain, pos);
4185 static void s_stop(struct seq_file *m, void *p)
4187 mutex_unlock(&cache_chain_mutex);
4190 static int s_show(struct seq_file *m, void *p)
4192 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
4193 struct slab *slabp;
4194 unsigned long active_objs;
4195 unsigned long num_objs;
4196 unsigned long active_slabs = 0;
4197 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4198 const char *name;
4199 char *error = NULL;
4200 int node;
4201 struct kmem_list3 *l3;
4203 active_objs = 0;
4204 num_slabs = 0;
4205 for_each_online_node(node) {
4206 l3 = cachep->nodelists[node];
4207 if (!l3)
4208 continue;
4210 check_irq_on();
4211 spin_lock_irq(&l3->list_lock);
4213 list_for_each_entry(slabp, &l3->slabs_full, list) {
4214 if (slabp->inuse != cachep->num && !error)
4215 error = "slabs_full accounting error";
4216 active_objs += cachep->num;
4217 active_slabs++;
4219 list_for_each_entry(slabp, &l3->slabs_partial, list) {
4220 if (slabp->inuse == cachep->num && !error)
4221 error = "slabs_partial inuse accounting error";
4222 if (!slabp->inuse && !error)
4223 error = "slabs_partial/inuse accounting error";
4224 active_objs += slabp->inuse;
4225 active_slabs++;
4227 list_for_each_entry(slabp, &l3->slabs_free, list) {
4228 if (slabp->inuse && !error)
4229 error = "slabs_free/inuse accounting error";
4230 num_slabs++;
4232 free_objects += l3->free_objects;
4233 if (l3->shared)
4234 shared_avail += l3->shared->avail;
4236 spin_unlock_irq(&l3->list_lock);
4238 num_slabs += active_slabs;
4239 num_objs = num_slabs * cachep->num;
4240 if (num_objs - active_objs != free_objects && !error)
4241 error = "free_objects accounting error";
4243 name = cachep->name;
4244 if (error)
4245 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4247 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
4248 name, active_objs, num_objs, cachep->buffer_size,
4249 cachep->num, (1 << cachep->gfporder));
4250 seq_printf(m, " : tunables %4u %4u %4u",
4251 cachep->limit, cachep->batchcount, cachep->shared);
4252 seq_printf(m, " : slabdata %6lu %6lu %6lu",
4253 active_slabs, num_slabs, shared_avail);
4254 #if STATS
4255 { /* list3 stats */
4256 unsigned long high = cachep->high_mark;
4257 unsigned long allocs = cachep->num_allocations;
4258 unsigned long grown = cachep->grown;
4259 unsigned long reaped = cachep->reaped;
4260 unsigned long errors = cachep->errors;
4261 unsigned long max_freeable = cachep->max_freeable;
4262 unsigned long node_allocs = cachep->node_allocs;
4263 unsigned long node_frees = cachep->node_frees;
4264 unsigned long overflows = cachep->node_overflow;
4266 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4267 "%4lu %4lu %4lu %4lu %4lu",
4268 allocs, high, grown,
4269 reaped, errors, max_freeable, node_allocs,
4270 node_frees, overflows);
4272 /* cpu stats */
4274 unsigned long allochit = atomic_read(&cachep->allochit);
4275 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4276 unsigned long freehit = atomic_read(&cachep->freehit);
4277 unsigned long freemiss = atomic_read(&cachep->freemiss);
4279 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4280 allochit, allocmiss, freehit, freemiss);
4282 #endif
4283 seq_putc(m, '\n');
4284 return 0;
4288 * slabinfo_op - iterator that generates /proc/slabinfo
4290 * Output layout:
4291 * cache-name
4292 * num-active-objs
4293 * total-objs
4294 * object size
4295 * num-active-slabs
4296 * total-slabs
4297 * num-pages-per-slab
4298 * + further values on SMP and with statistics enabled
4301 static const struct seq_operations slabinfo_op = {
4302 .start = s_start,
4303 .next = s_next,
4304 .stop = s_stop,
4305 .show = s_show,
4308 #define MAX_SLABINFO_WRITE 128
4310 * slabinfo_write - Tuning for the slab allocator
4311 * @file: unused
4312 * @buffer: user buffer
4313 * @count: data length
4314 * @ppos: unused
4316 static ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4317 size_t count, loff_t *ppos)
4319 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4320 int limit, batchcount, shared, res;
4321 struct kmem_cache *cachep;
4323 if (count > MAX_SLABINFO_WRITE)
4324 return -EINVAL;
4325 if (copy_from_user(&kbuf, buffer, count))
4326 return -EFAULT;
4327 kbuf[MAX_SLABINFO_WRITE] = '\0';
4329 tmp = strchr(kbuf, ' ');
4330 if (!tmp)
4331 return -EINVAL;
4332 *tmp = '\0';
4333 tmp++;
4334 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4335 return -EINVAL;
4337 /* Find the cache in the chain of caches. */
4338 mutex_lock(&cache_chain_mutex);
4339 res = -EINVAL;
4340 list_for_each_entry(cachep, &cache_chain, next) {
4341 if (!strcmp(cachep->name, kbuf)) {
4342 if (limit < 1 || batchcount < 1 ||
4343 batchcount > limit || shared < 0) {
4344 res = 0;
4345 } else {
4346 res = do_tune_cpucache(cachep, limit,
4347 batchcount, shared,
4348 GFP_KERNEL);
4350 break;
4353 mutex_unlock(&cache_chain_mutex);
4354 if (res >= 0)
4355 res = count;
4356 return res;
4359 static int slabinfo_open(struct inode *inode, struct file *file)
4361 return seq_open(file, &slabinfo_op);
4364 static const struct file_operations proc_slabinfo_operations = {
4365 .open = slabinfo_open,
4366 .read = seq_read,
4367 .write = slabinfo_write,
4368 .llseek = seq_lseek,
4369 .release = seq_release,
4372 #ifdef CONFIG_DEBUG_SLAB_LEAK
4374 static void *leaks_start(struct seq_file *m, loff_t *pos)
4376 mutex_lock(&cache_chain_mutex);
4377 return seq_list_start(&cache_chain, *pos);
4380 static inline int add_caller(unsigned long *n, unsigned long v)
4382 unsigned long *p;
4383 int l;
4384 if (!v)
4385 return 1;
4386 l = n[1];
4387 p = n + 2;
4388 while (l) {
4389 int i = l/2;
4390 unsigned long *q = p + 2 * i;
4391 if (*q == v) {
4392 q[1]++;
4393 return 1;
4395 if (*q > v) {
4396 l = i;
4397 } else {
4398 p = q + 2;
4399 l -= i + 1;
4402 if (++n[1] == n[0])
4403 return 0;
4404 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4405 p[0] = v;
4406 p[1] = 1;
4407 return 1;
4410 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4412 void *p;
4413 int i;
4414 if (n[0] == n[1])
4415 return;
4416 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4417 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4418 continue;
4419 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4420 return;
4424 static void show_symbol(struct seq_file *m, unsigned long address)
4426 #ifdef CONFIG_KALLSYMS
4427 unsigned long offset, size;
4428 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4430 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4431 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4432 if (modname[0])
4433 seq_printf(m, " [%s]", modname);
4434 return;
4436 #endif
4437 seq_printf(m, "%p", (void *)address);
4440 static int leaks_show(struct seq_file *m, void *p)
4442 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next);
4443 struct slab *slabp;
4444 struct kmem_list3 *l3;
4445 const char *name;
4446 unsigned long *n = m->private;
4447 int node;
4448 int i;
4450 if (!(cachep->flags & SLAB_STORE_USER))
4451 return 0;
4452 if (!(cachep->flags & SLAB_RED_ZONE))
4453 return 0;
4455 /* OK, we can do it */
4457 n[1] = 0;
4459 for_each_online_node(node) {
4460 l3 = cachep->nodelists[node];
4461 if (!l3)
4462 continue;
4464 check_irq_on();
4465 spin_lock_irq(&l3->list_lock);
4467 list_for_each_entry(slabp, &l3->slabs_full, list)
4468 handle_slab(n, cachep, slabp);
4469 list_for_each_entry(slabp, &l3->slabs_partial, list)
4470 handle_slab(n, cachep, slabp);
4471 spin_unlock_irq(&l3->list_lock);
4473 name = cachep->name;
4474 if (n[0] == n[1]) {
4475 /* Increase the buffer size */
4476 mutex_unlock(&cache_chain_mutex);
4477 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4478 if (!m->private) {
4479 /* Too bad, we are really out */
4480 m->private = n;
4481 mutex_lock(&cache_chain_mutex);
4482 return -ENOMEM;
4484 *(unsigned long *)m->private = n[0] * 2;
4485 kfree(n);
4486 mutex_lock(&cache_chain_mutex);
4487 /* Now make sure this entry will be retried */
4488 m->count = m->size;
4489 return 0;
4491 for (i = 0; i < n[1]; i++) {
4492 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4493 show_symbol(m, n[2*i+2]);
4494 seq_putc(m, '\n');
4497 return 0;
4500 static const struct seq_operations slabstats_op = {
4501 .start = leaks_start,
4502 .next = s_next,
4503 .stop = s_stop,
4504 .show = leaks_show,
4507 static int slabstats_open(struct inode *inode, struct file *file)
4509 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4510 int ret = -ENOMEM;
4511 if (n) {
4512 ret = seq_open(file, &slabstats_op);
4513 if (!ret) {
4514 struct seq_file *m = file->private_data;
4515 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4516 m->private = n;
4517 n = NULL;
4519 kfree(n);
4521 return ret;
4524 static const struct file_operations proc_slabstats_operations = {
4525 .open = slabstats_open,
4526 .read = seq_read,
4527 .llseek = seq_lseek,
4528 .release = seq_release_private,
4530 #endif
4532 static int __init slab_proc_init(void)
4534 proc_create("slabinfo",S_IWUSR|S_IRUGO,NULL,&proc_slabinfo_operations);
4535 #ifdef CONFIG_DEBUG_SLAB_LEAK
4536 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4537 #endif
4538 return 0;
4540 module_init(slab_proc_init);
4541 #endif
4544 * ksize - get the actual amount of memory allocated for a given object
4545 * @objp: Pointer to the object
4547 * kmalloc may internally round up allocations and return more memory
4548 * than requested. ksize() can be used to determine the actual amount of
4549 * memory allocated. The caller may use this additional memory, even though
4550 * a smaller amount of memory was initially specified with the kmalloc call.
4551 * The caller must guarantee that objp points to a valid object previously
4552 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4553 * must not be freed during the duration of the call.
4555 size_t ksize(const void *objp)
4557 BUG_ON(!objp);
4558 if (unlikely(objp == ZERO_SIZE_PTR))
4559 return 0;
4561 return obj_size(virt_to_cache(objp));
4563 EXPORT_SYMBOL(ksize);