3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
28 * This means, that your constructor is used only for newly allocated
29 * slabs and you must pass objects with the same intializations to
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
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
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>
91 #include <linux/poison.h>
92 #include <linux/swap.h>
93 #include <linux/cache.h>
94 #include <linux/interrupt.h>
95 #include <linux/init.h>
96 #include <linux/compiler.h>
97 #include <linux/cpuset.h>
98 #include <linux/seq_file.h>
99 #include <linux/notifier.h>
100 #include <linux/kallsyms.h>
101 #include <linux/cpu.h>
102 #include <linux/sysctl.h>
103 #include <linux/module.h>
104 #include <linux/rcupdate.h>
105 #include <linux/string.h>
106 #include <linux/uaccess.h>
107 #include <linux/nodemask.h>
108 #include <linux/mempolicy.h>
109 #include <linux/mutex.h>
110 #include <linux/fault-inject.h>
111 #include <linux/rtmutex.h>
112 #include <linux/reciprocal_div.h>
114 #include <asm/cacheflush.h>
115 #include <asm/tlbflush.h>
116 #include <asm/page.h>
119 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL,
120 * SLAB_RED_ZONE & SLAB_POISON.
121 * 0 for faster, smaller code (especially in the critical paths).
123 * STATS - 1 to collect stats for /proc/slabinfo.
124 * 0 for faster, smaller code (especially in the critical paths).
126 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
129 #ifdef CONFIG_DEBUG_SLAB
132 #define FORCED_DEBUG 1
136 #define FORCED_DEBUG 0
139 /* Shouldn't this be in a header file somewhere? */
140 #define BYTES_PER_WORD sizeof(void *)
142 #ifndef cache_line_size
143 #define cache_line_size() L1_CACHE_BYTES
146 #ifndef ARCH_KMALLOC_MINALIGN
148 * Enforce a minimum alignment for the kmalloc caches.
149 * Usually, the kmalloc caches are cache_line_size() aligned, except when
150 * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned.
151 * Some archs want to perform DMA into kmalloc caches and need a guaranteed
152 * alignment larger than BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that.
153 * Note that this flag disables some debug features.
155 #define ARCH_KMALLOC_MINALIGN 0
158 #ifndef ARCH_SLAB_MINALIGN
160 * Enforce a minimum alignment for all caches.
161 * Intended for archs that get misalignment faults even for BYTES_PER_WORD
162 * aligned buffers. Includes ARCH_KMALLOC_MINALIGN.
163 * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables
164 * some debug features.
166 #define ARCH_SLAB_MINALIGN 0
169 #ifndef ARCH_KMALLOC_FLAGS
170 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
173 /* Legal flag mask for kmem_cache_create(). */
175 # define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \
176 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
178 SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \
179 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
180 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
182 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
183 SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \
184 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
185 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD)
191 * Bufctl's are used for linking objs within a slab
194 * This implementation relies on "struct page" for locating the cache &
195 * slab an object belongs to.
196 * This allows the bufctl structure to be small (one int), but limits
197 * the number of objects a slab (not a cache) can contain when off-slab
198 * bufctls are used. The limit is the size of the largest general cache
199 * that does not use off-slab slabs.
200 * For 32bit archs with 4 kB pages, is this 56.
201 * This is not serious, as it is only for large objects, when it is unwise
202 * to have too many per slab.
203 * Note: This limit can be raised by introducing a general cache whose size
204 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
207 typedef unsigned int kmem_bufctl_t;
208 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
209 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
210 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
211 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
216 * Manages the objs in a slab. Placed either at the beginning of mem allocated
217 * for a slab, or allocated from an general cache.
218 * Slabs are chained into three list: fully used, partial, fully free slabs.
221 struct list_head list;
222 unsigned long colouroff;
223 void *s_mem; /* including colour offset */
224 unsigned int inuse; /* num of objs active in slab */
226 unsigned short nodeid;
232 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
233 * arrange for kmem_freepages to be called via RCU. This is useful if
234 * we need to approach a kernel structure obliquely, from its address
235 * obtained without the usual locking. We can lock the structure to
236 * stabilize it and check it's still at the given address, only if we
237 * can be sure that the memory has not been meanwhile reused for some
238 * other kind of object (which our subsystem's lock might corrupt).
240 * rcu_read_lock before reading the address, then rcu_read_unlock after
241 * taking the spinlock within the structure expected at that address.
243 * We assume struct slab_rcu can overlay struct slab when destroying.
246 struct rcu_head head;
247 struct kmem_cache *cachep;
255 * - LIFO ordering, to hand out cache-warm objects from _alloc
256 * - reduce the number of linked list operations
257 * - reduce spinlock operations
259 * The limit is stored in the per-cpu structure to reduce the data cache
266 unsigned int batchcount;
267 unsigned int touched;
270 * Must have this definition in here for the proper
271 * alignment of array_cache. Also simplifies accessing
273 * [0] is for gcc 2.95. It should really be [].
278 * bootstrap: The caches do not work without cpuarrays anymore, but the
279 * cpuarrays are allocated from the generic caches...
281 #define BOOT_CPUCACHE_ENTRIES 1
282 struct arraycache_init {
283 struct array_cache cache;
284 void *entries[BOOT_CPUCACHE_ENTRIES];
288 * The slab lists for all objects.
291 struct list_head slabs_partial; /* partial list first, better asm code */
292 struct list_head slabs_full;
293 struct list_head slabs_free;
294 unsigned long free_objects;
295 unsigned int free_limit;
296 unsigned int colour_next; /* Per-node cache coloring */
297 spinlock_t list_lock;
298 struct array_cache *shared; /* shared per node */
299 struct array_cache **alien; /* on other nodes */
300 unsigned long next_reap; /* updated without locking */
301 int free_touched; /* updated without locking */
305 * Need this for bootstrapping a per node allocator.
307 #define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1)
308 struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
309 #define CACHE_CACHE 0
311 #define SIZE_L3 (1 + MAX_NUMNODES)
313 static int drain_freelist(struct kmem_cache *cache,
314 struct kmem_list3 *l3, int tofree);
315 static void free_block(struct kmem_cache *cachep, void **objpp, int len,
317 static int enable_cpucache(struct kmem_cache *cachep);
318 static void cache_reap(struct work_struct *unused);
321 * This function must be completely optimized away if a constant is passed to
322 * it. Mostly the same as what is in linux/slab.h except it returns an index.
324 static __always_inline int index_of(const size_t size)
326 extern void __bad_size(void);
328 if (__builtin_constant_p(size)) {
336 #include "linux/kmalloc_sizes.h"
344 static int slab_early_init = 1;
346 #define INDEX_AC index_of(sizeof(struct arraycache_init))
347 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
349 static void kmem_list3_init(struct kmem_list3 *parent)
351 INIT_LIST_HEAD(&parent->slabs_full);
352 INIT_LIST_HEAD(&parent->slabs_partial);
353 INIT_LIST_HEAD(&parent->slabs_free);
354 parent->shared = NULL;
355 parent->alien = NULL;
356 parent->colour_next = 0;
357 spin_lock_init(&parent->list_lock);
358 parent->free_objects = 0;
359 parent->free_touched = 0;
362 #define MAKE_LIST(cachep, listp, slab, nodeid) \
364 INIT_LIST_HEAD(listp); \
365 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
368 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
370 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
371 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
372 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
382 /* 1) per-cpu data, touched during every alloc/free */
383 struct array_cache *array[NR_CPUS];
384 /* 2) Cache tunables. Protected by cache_chain_mutex */
385 unsigned int batchcount;
389 unsigned int buffer_size;
390 u32 reciprocal_buffer_size;
391 /* 3) touched by every alloc & free from the backend */
393 unsigned int flags; /* constant flags */
394 unsigned int num; /* # of objs per slab */
396 /* 4) cache_grow/shrink */
397 /* order of pgs per slab (2^n) */
398 unsigned int gfporder;
400 /* force GFP flags, e.g. GFP_DMA */
403 size_t colour; /* cache colouring range */
404 unsigned int colour_off; /* colour offset */
405 struct kmem_cache *slabp_cache;
406 unsigned int slab_size;
407 unsigned int dflags; /* dynamic flags */
409 /* constructor func */
410 void (*ctor) (void *, struct kmem_cache *, unsigned long);
412 /* de-constructor func */
413 void (*dtor) (void *, struct kmem_cache *, unsigned long);
415 /* 5) cache creation/removal */
417 struct list_head next;
421 unsigned long num_active;
422 unsigned long num_allocations;
423 unsigned long high_mark;
425 unsigned long reaped;
426 unsigned long errors;
427 unsigned long max_freeable;
428 unsigned long node_allocs;
429 unsigned long node_frees;
430 unsigned long node_overflow;
438 * If debugging is enabled, then the allocator can add additional
439 * fields and/or padding to every object. buffer_size contains the total
440 * object size including these internal fields, the following two
441 * variables contain the offset to the user object and its size.
447 * We put nodelists[] at the end of kmem_cache, because we want to size
448 * this array to nr_node_ids slots instead of MAX_NUMNODES
449 * (see kmem_cache_init())
450 * We still use [MAX_NUMNODES] and not [1] or [0] because cache_cache
451 * is statically defined, so we reserve the max number of nodes.
453 struct kmem_list3 *nodelists[MAX_NUMNODES];
455 * Do not add fields after nodelists[]
459 #define CFLGS_OFF_SLAB (0x80000000UL)
460 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
462 #define BATCHREFILL_LIMIT 16
464 * Optimization question: fewer reaps means less probability for unnessary
465 * cpucache drain/refill cycles.
467 * OTOH the cpuarrays can contain lots of objects,
468 * which could lock up otherwise freeable slabs.
470 #define REAPTIMEOUT_CPUC (2*HZ)
471 #define REAPTIMEOUT_LIST3 (4*HZ)
474 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
475 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
476 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
477 #define STATS_INC_GROWN(x) ((x)->grown++)
478 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
479 #define STATS_SET_HIGH(x) \
481 if ((x)->num_active > (x)->high_mark) \
482 (x)->high_mark = (x)->num_active; \
484 #define STATS_INC_ERR(x) ((x)->errors++)
485 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
486 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
487 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
488 #define STATS_SET_FREEABLE(x, i) \
490 if ((x)->max_freeable < i) \
491 (x)->max_freeable = i; \
493 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
494 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
495 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
496 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
498 #define STATS_INC_ACTIVE(x) do { } while (0)
499 #define STATS_DEC_ACTIVE(x) do { } while (0)
500 #define STATS_INC_ALLOCED(x) do { } while (0)
501 #define STATS_INC_GROWN(x) do { } while (0)
502 #define STATS_ADD_REAPED(x,y) do { } while (0)
503 #define STATS_SET_HIGH(x) do { } while (0)
504 #define STATS_INC_ERR(x) do { } while (0)
505 #define STATS_INC_NODEALLOCS(x) do { } while (0)
506 #define STATS_INC_NODEFREES(x) do { } while (0)
507 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
508 #define STATS_SET_FREEABLE(x, i) do { } while (0)
509 #define STATS_INC_ALLOCHIT(x) do { } while (0)
510 #define STATS_INC_ALLOCMISS(x) do { } while (0)
511 #define STATS_INC_FREEHIT(x) do { } while (0)
512 #define STATS_INC_FREEMISS(x) do { } while (0)
518 * memory layout of objects:
520 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
521 * the end of an object is aligned with the end of the real
522 * allocation. Catches writes behind the end of the allocation.
523 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
525 * cachep->obj_offset: The real object.
526 * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
527 * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address
528 * [BYTES_PER_WORD long]
530 static int obj_offset(struct kmem_cache *cachep)
532 return cachep->obj_offset;
535 static int obj_size(struct kmem_cache *cachep)
537 return cachep->obj_size;
540 static unsigned long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
542 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
543 return (unsigned long*) (objp+obj_offset(cachep)-BYTES_PER_WORD);
546 static unsigned long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
548 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
549 if (cachep->flags & SLAB_STORE_USER)
550 return (unsigned long *)(objp + cachep->buffer_size -
552 return (unsigned long *)(objp + cachep->buffer_size - BYTES_PER_WORD);
555 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
557 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
558 return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD);
563 #define obj_offset(x) 0
564 #define obj_size(cachep) (cachep->buffer_size)
565 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;})
566 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;})
567 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
572 * Maximum size of an obj (in 2^order pages) and absolute limit for the gfp
575 #if defined(CONFIG_LARGE_ALLOCS)
576 #define MAX_OBJ_ORDER 13 /* up to 32Mb */
577 #define MAX_GFP_ORDER 13 /* up to 32Mb */
578 #elif defined(CONFIG_MMU)
579 #define MAX_OBJ_ORDER 5 /* 32 pages */
580 #define MAX_GFP_ORDER 5 /* 32 pages */
582 #define MAX_OBJ_ORDER 8 /* up to 1Mb */
583 #define MAX_GFP_ORDER 8 /* up to 1Mb */
587 * Do not go above this order unless 0 objects fit into the slab.
589 #define BREAK_GFP_ORDER_HI 1
590 #define BREAK_GFP_ORDER_LO 0
591 static int slab_break_gfp_order = BREAK_GFP_ORDER_LO;
594 * Functions for storing/retrieving the cachep and or slab from the page
595 * allocator. These are used to find the slab an obj belongs to. With kfree(),
596 * these are used to find the cache which an obj belongs to.
598 static inline void page_set_cache(struct page *page, struct kmem_cache *cache)
600 page->lru.next = (struct list_head *)cache;
603 static inline struct kmem_cache *page_get_cache(struct page *page)
605 page = compound_head(page);
606 BUG_ON(!PageSlab(page));
607 return (struct kmem_cache *)page->lru.next;
610 static inline void page_set_slab(struct page *page, struct slab *slab)
612 page->lru.prev = (struct list_head *)slab;
615 static inline struct slab *page_get_slab(struct page *page)
617 page = compound_head(page);
618 BUG_ON(!PageSlab(page));
619 return (struct slab *)page->lru.prev;
622 static inline struct kmem_cache *virt_to_cache(const void *obj)
624 struct page *page = virt_to_page(obj);
625 return page_get_cache(page);
628 static inline struct slab *virt_to_slab(const void *obj)
630 struct page *page = virt_to_page(obj);
631 return page_get_slab(page);
634 static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
637 return slab->s_mem + cache->buffer_size * idx;
641 * We want to avoid an expensive divide : (offset / cache->buffer_size)
642 * Using the fact that buffer_size is a constant for a particular cache,
643 * we can replace (offset / cache->buffer_size) by
644 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
646 static inline unsigned int obj_to_index(const struct kmem_cache *cache,
647 const struct slab *slab, void *obj)
649 u32 offset = (obj - slab->s_mem);
650 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
654 * These are the default caches for kmalloc. Custom caches can have other sizes.
656 struct cache_sizes malloc_sizes[] = {
657 #define CACHE(x) { .cs_size = (x) },
658 #include <linux/kmalloc_sizes.h>
662 EXPORT_SYMBOL(malloc_sizes);
664 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
670 static struct cache_names __initdata cache_names[] = {
671 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
672 #include <linux/kmalloc_sizes.h>
677 static struct arraycache_init initarray_cache __initdata =
678 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
679 static struct arraycache_init initarray_generic =
680 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
682 /* internal cache of cache description objs */
683 static struct kmem_cache cache_cache = {
685 .limit = BOOT_CPUCACHE_ENTRIES,
687 .buffer_size = sizeof(struct kmem_cache),
688 .name = "kmem_cache",
691 #define BAD_ALIEN_MAGIC 0x01020304ul
693 #ifdef CONFIG_LOCKDEP
696 * Slab sometimes uses the kmalloc slabs to store the slab headers
697 * for other slabs "off slab".
698 * The locking for this is tricky in that it nests within the locks
699 * of all other slabs in a few places; to deal with this special
700 * locking we put on-slab caches into a separate lock-class.
702 * We set lock class for alien array caches which are up during init.
703 * The lock annotation will be lost if all cpus of a node goes down and
704 * then comes back up during hotplug
706 static struct lock_class_key on_slab_l3_key;
707 static struct lock_class_key on_slab_alc_key;
709 static inline void init_lock_keys(void)
713 struct cache_sizes *s = malloc_sizes;
715 while (s->cs_size != ULONG_MAX) {
717 struct array_cache **alc;
719 struct kmem_list3 *l3 = s->cs_cachep->nodelists[q];
720 if (!l3 || OFF_SLAB(s->cs_cachep))
722 lockdep_set_class(&l3->list_lock, &on_slab_l3_key);
725 * FIXME: This check for BAD_ALIEN_MAGIC
726 * should go away when common slab code is taught to
727 * work even without alien caches.
728 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
729 * for alloc_alien_cache,
731 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
735 lockdep_set_class(&alc[r]->lock,
743 static inline void init_lock_keys(void)
749 * 1. Guard access to the cache-chain.
750 * 2. Protect sanity of cpu_online_map against cpu hotplug events
752 static DEFINE_MUTEX(cache_chain_mutex);
753 static struct list_head cache_chain;
756 * chicken and egg problem: delay the per-cpu array allocation
757 * until the general caches are up.
767 * used by boot code to determine if it can use slab based allocator
769 int slab_is_available(void)
771 return g_cpucache_up == FULL;
774 static DEFINE_PER_CPU(struct delayed_work, reap_work);
776 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
778 return cachep->array[smp_processor_id()];
781 static inline struct kmem_cache *__find_general_cachep(size_t size,
784 struct cache_sizes *csizep = malloc_sizes;
787 /* This happens if someone tries to call
788 * kmem_cache_create(), or __kmalloc(), before
789 * the generic caches are initialized.
791 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
793 while (size > csizep->cs_size)
797 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
798 * has cs_{dma,}cachep==NULL. Thus no special case
799 * for large kmalloc calls required.
801 #ifdef CONFIG_ZONE_DMA
802 if (unlikely(gfpflags & GFP_DMA))
803 return csizep->cs_dmacachep;
805 return csizep->cs_cachep;
808 static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
810 return __find_general_cachep(size, gfpflags);
813 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
815 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
819 * Calculate the number of objects and left-over bytes for a given buffer size.
821 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
822 size_t align, int flags, size_t *left_over,
827 size_t slab_size = PAGE_SIZE << gfporder;
830 * The slab management structure can be either off the slab or
831 * on it. For the latter case, the memory allocated for a
835 * - One kmem_bufctl_t for each object
836 * - Padding to respect alignment of @align
837 * - @buffer_size bytes for each object
839 * If the slab management structure is off the slab, then the
840 * alignment will already be calculated into the size. Because
841 * the slabs are all pages aligned, the objects will be at the
842 * correct alignment when allocated.
844 if (flags & CFLGS_OFF_SLAB) {
846 nr_objs = slab_size / buffer_size;
848 if (nr_objs > SLAB_LIMIT)
849 nr_objs = SLAB_LIMIT;
852 * Ignore padding for the initial guess. The padding
853 * is at most @align-1 bytes, and @buffer_size is at
854 * least @align. In the worst case, this result will
855 * be one greater than the number of objects that fit
856 * into the memory allocation when taking the padding
859 nr_objs = (slab_size - sizeof(struct slab)) /
860 (buffer_size + sizeof(kmem_bufctl_t));
863 * This calculated number will be either the right
864 * amount, or one greater than what we want.
866 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
870 if (nr_objs > SLAB_LIMIT)
871 nr_objs = SLAB_LIMIT;
873 mgmt_size = slab_mgmt_size(nr_objs, align);
876 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
879 #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg)
881 static void __slab_error(const char *function, struct kmem_cache *cachep,
884 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
885 function, cachep->name, msg);
890 * By default on NUMA we use alien caches to stage the freeing of
891 * objects allocated from other nodes. This causes massive memory
892 * inefficiencies when using fake NUMA setup to split memory into a
893 * large number of small nodes, so it can be disabled on the command
897 static int use_alien_caches __read_mostly = 1;
898 static int __init noaliencache_setup(char *s)
900 use_alien_caches = 0;
903 __setup("noaliencache", noaliencache_setup);
907 * Special reaping functions for NUMA systems called from cache_reap().
908 * These take care of doing round robin flushing of alien caches (containing
909 * objects freed on different nodes from which they were allocated) and the
910 * flushing of remote pcps by calling drain_node_pages.
912 static DEFINE_PER_CPU(unsigned long, reap_node);
914 static void init_reap_node(int cpu)
918 node = next_node(cpu_to_node(cpu), node_online_map);
919 if (node == MAX_NUMNODES)
920 node = first_node(node_online_map);
922 per_cpu(reap_node, cpu) = node;
925 static void next_reap_node(void)
927 int node = __get_cpu_var(reap_node);
930 * Also drain per cpu pages on remote zones
932 if (node != numa_node_id())
933 drain_node_pages(node);
935 node = next_node(node, node_online_map);
936 if (unlikely(node >= MAX_NUMNODES))
937 node = first_node(node_online_map);
938 __get_cpu_var(reap_node) = node;
942 #define init_reap_node(cpu) do { } while (0)
943 #define next_reap_node(void) do { } while (0)
947 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
948 * via the workqueue/eventd.
949 * Add the CPU number into the expiration time to minimize the possibility of
950 * the CPUs getting into lockstep and contending for the global cache chain
953 static void __devinit start_cpu_timer(int cpu)
955 struct delayed_work *reap_work = &per_cpu(reap_work, cpu);
958 * When this gets called from do_initcalls via cpucache_init(),
959 * init_workqueues() has already run, so keventd will be setup
962 if (keventd_up() && reap_work->work.func == NULL) {
964 INIT_DELAYED_WORK(reap_work, cache_reap);
965 schedule_delayed_work_on(cpu, reap_work,
966 __round_jiffies_relative(HZ, cpu));
970 static struct array_cache *alloc_arraycache(int node, int entries,
973 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
974 struct array_cache *nc = NULL;
976 nc = kmalloc_node(memsize, GFP_KERNEL, node);
980 nc->batchcount = batchcount;
982 spin_lock_init(&nc->lock);
988 * Transfer objects in one arraycache to another.
989 * Locking must be handled by the caller.
991 * Return the number of entries transferred.
993 static int transfer_objects(struct array_cache *to,
994 struct array_cache *from, unsigned int max)
996 /* Figure out how many entries to transfer */
997 int nr = min(min(from->avail, max), to->limit - to->avail);
1002 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
1003 sizeof(void *) *nr);
1013 #define drain_alien_cache(cachep, alien) do { } while (0)
1014 #define reap_alien(cachep, l3) do { } while (0)
1016 static inline struct array_cache **alloc_alien_cache(int node, int limit)
1018 return (struct array_cache **)BAD_ALIEN_MAGIC;
1021 static inline void free_alien_cache(struct array_cache **ac_ptr)
1025 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1030 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
1036 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
1037 gfp_t flags, int nodeid)
1042 #else /* CONFIG_NUMA */
1044 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
1045 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
1047 static struct array_cache **alloc_alien_cache(int node, int limit)
1049 struct array_cache **ac_ptr;
1050 int memsize = sizeof(void *) * nr_node_ids;
1055 ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node);
1058 if (i == node || !node_online(i)) {
1062 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d);
1064 for (i--; i <= 0; i--)
1074 static void free_alien_cache(struct array_cache **ac_ptr)
1085 static void __drain_alien_cache(struct kmem_cache *cachep,
1086 struct array_cache *ac, int node)
1088 struct kmem_list3 *rl3 = cachep->nodelists[node];
1091 spin_lock(&rl3->list_lock);
1093 * Stuff objects into the remote nodes shared array first.
1094 * That way we could avoid the overhead of putting the objects
1095 * into the free lists and getting them back later.
1098 transfer_objects(rl3->shared, ac, ac->limit);
1100 free_block(cachep, ac->entry, ac->avail, node);
1102 spin_unlock(&rl3->list_lock);
1107 * Called from cache_reap() to regularly drain alien caches round robin.
1109 static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1111 int node = __get_cpu_var(reap_node);
1114 struct array_cache *ac = l3->alien[node];
1116 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1117 __drain_alien_cache(cachep, ac, node);
1118 spin_unlock_irq(&ac->lock);
1123 static void drain_alien_cache(struct kmem_cache *cachep,
1124 struct array_cache **alien)
1127 struct array_cache *ac;
1128 unsigned long flags;
1130 for_each_online_node(i) {
1133 spin_lock_irqsave(&ac->lock, flags);
1134 __drain_alien_cache(cachep, ac, i);
1135 spin_unlock_irqrestore(&ac->lock, flags);
1140 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1142 struct slab *slabp = virt_to_slab(objp);
1143 int nodeid = slabp->nodeid;
1144 struct kmem_list3 *l3;
1145 struct array_cache *alien = NULL;
1148 node = numa_node_id();
1151 * Make sure we are not freeing a object from another node to the array
1152 * cache on this cpu.
1154 if (likely(slabp->nodeid == node))
1157 l3 = cachep->nodelists[node];
1158 STATS_INC_NODEFREES(cachep);
1159 if (l3->alien && l3->alien[nodeid]) {
1160 alien = l3->alien[nodeid];
1161 spin_lock(&alien->lock);
1162 if (unlikely(alien->avail == alien->limit)) {
1163 STATS_INC_ACOVERFLOW(cachep);
1164 __drain_alien_cache(cachep, alien, nodeid);
1166 alien->entry[alien->avail++] = objp;
1167 spin_unlock(&alien->lock);
1169 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1170 free_block(cachep, &objp, 1, nodeid);
1171 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1177 static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1178 unsigned long action, void *hcpu)
1180 long cpu = (long)hcpu;
1181 struct kmem_cache *cachep;
1182 struct kmem_list3 *l3 = NULL;
1183 int node = cpu_to_node(cpu);
1184 int memsize = sizeof(struct kmem_list3);
1187 case CPU_UP_PREPARE:
1188 mutex_lock(&cache_chain_mutex);
1190 * We need to do this right in the beginning since
1191 * alloc_arraycache's are going to use this list.
1192 * kmalloc_node allows us to add the slab to the right
1193 * kmem_list3 and not this cpu's kmem_list3
1196 list_for_each_entry(cachep, &cache_chain, next) {
1198 * Set up the size64 kmemlist for cpu before we can
1199 * begin anything. Make sure some other cpu on this
1200 * node has not already allocated this
1202 if (!cachep->nodelists[node]) {
1203 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1206 kmem_list3_init(l3);
1207 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1208 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1211 * The l3s don't come and go as CPUs come and
1212 * go. cache_chain_mutex is sufficient
1215 cachep->nodelists[node] = l3;
1218 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1219 cachep->nodelists[node]->free_limit =
1220 (1 + nr_cpus_node(node)) *
1221 cachep->batchcount + cachep->num;
1222 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1226 * Now we can go ahead with allocating the shared arrays and
1229 list_for_each_entry(cachep, &cache_chain, next) {
1230 struct array_cache *nc;
1231 struct array_cache *shared = NULL;
1232 struct array_cache **alien = NULL;
1234 nc = alloc_arraycache(node, cachep->limit,
1235 cachep->batchcount);
1238 if (cachep->shared) {
1239 shared = alloc_arraycache(node,
1240 cachep->shared * cachep->batchcount,
1245 if (use_alien_caches) {
1246 alien = alloc_alien_cache(node, cachep->limit);
1250 cachep->array[cpu] = nc;
1251 l3 = cachep->nodelists[node];
1254 spin_lock_irq(&l3->list_lock);
1257 * We are serialised from CPU_DEAD or
1258 * CPU_UP_CANCELLED by the cpucontrol lock
1260 l3->shared = shared;
1269 spin_unlock_irq(&l3->list_lock);
1271 free_alien_cache(alien);
1275 mutex_unlock(&cache_chain_mutex);
1276 start_cpu_timer(cpu);
1278 #ifdef CONFIG_HOTPLUG_CPU
1279 case CPU_DOWN_PREPARE:
1280 mutex_lock(&cache_chain_mutex);
1282 case CPU_DOWN_FAILED:
1283 mutex_unlock(&cache_chain_mutex);
1287 * Even if all the cpus of a node are down, we don't free the
1288 * kmem_list3 of any cache. This to avoid a race between
1289 * cpu_down, and a kmalloc allocation from another cpu for
1290 * memory from the node of the cpu going down. The list3
1291 * structure is usually allocated from kmem_cache_create() and
1292 * gets destroyed at kmem_cache_destroy().
1296 case CPU_UP_CANCELED:
1297 list_for_each_entry(cachep, &cache_chain, next) {
1298 struct array_cache *nc;
1299 struct array_cache *shared;
1300 struct array_cache **alien;
1303 mask = node_to_cpumask(node);
1304 /* cpu is dead; no one can alloc from it. */
1305 nc = cachep->array[cpu];
1306 cachep->array[cpu] = NULL;
1307 l3 = cachep->nodelists[node];
1310 goto free_array_cache;
1312 spin_lock_irq(&l3->list_lock);
1314 /* Free limit for this kmem_list3 */
1315 l3->free_limit -= cachep->batchcount;
1317 free_block(cachep, nc->entry, nc->avail, node);
1319 if (!cpus_empty(mask)) {
1320 spin_unlock_irq(&l3->list_lock);
1321 goto free_array_cache;
1324 shared = l3->shared;
1326 free_block(cachep, shared->entry,
1327 shared->avail, node);
1334 spin_unlock_irq(&l3->list_lock);
1338 drain_alien_cache(cachep, alien);
1339 free_alien_cache(alien);
1345 * In the previous loop, all the objects were freed to
1346 * the respective cache's slabs, now we can go ahead and
1347 * shrink each nodelist to its limit.
1349 list_for_each_entry(cachep, &cache_chain, next) {
1350 l3 = cachep->nodelists[node];
1353 drain_freelist(cachep, l3, l3->free_objects);
1355 mutex_unlock(&cache_chain_mutex);
1363 static struct notifier_block __cpuinitdata cpucache_notifier = {
1364 &cpuup_callback, NULL, 0
1368 * swap the static kmem_list3 with kmalloced memory
1370 static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1373 struct kmem_list3 *ptr;
1375 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid);
1378 local_irq_disable();
1379 memcpy(ptr, list, sizeof(struct kmem_list3));
1381 * Do not assume that spinlocks can be initialized via memcpy:
1383 spin_lock_init(&ptr->list_lock);
1385 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1386 cachep->nodelists[nodeid] = ptr;
1391 * Initialisation. Called after the page allocator have been initialised and
1392 * before smp_init().
1394 void __init kmem_cache_init(void)
1397 struct cache_sizes *sizes;
1398 struct cache_names *names;
1403 if (num_possible_nodes() == 1)
1404 use_alien_caches = 0;
1406 for (i = 0; i < NUM_INIT_LISTS; i++) {
1407 kmem_list3_init(&initkmem_list3[i]);
1408 if (i < MAX_NUMNODES)
1409 cache_cache.nodelists[i] = NULL;
1413 * Fragmentation resistance on low memory - only use bigger
1414 * page orders on machines with more than 32MB of memory.
1416 if (num_physpages > (32 << 20) >> PAGE_SHIFT)
1417 slab_break_gfp_order = BREAK_GFP_ORDER_HI;
1419 /* Bootstrap is tricky, because several objects are allocated
1420 * from caches that do not exist yet:
1421 * 1) initialize the cache_cache cache: it contains the struct
1422 * kmem_cache structures of all caches, except cache_cache itself:
1423 * cache_cache is statically allocated.
1424 * Initially an __init data area is used for the head array and the
1425 * kmem_list3 structures, it's replaced with a kmalloc allocated
1426 * array at the end of the bootstrap.
1427 * 2) Create the first kmalloc cache.
1428 * The struct kmem_cache for the new cache is allocated normally.
1429 * An __init data area is used for the head array.
1430 * 3) Create the remaining kmalloc caches, with minimally sized
1432 * 4) Replace the __init data head arrays for cache_cache and the first
1433 * kmalloc cache with kmalloc allocated arrays.
1434 * 5) Replace the __init data for kmem_list3 for cache_cache and
1435 * the other cache's with kmalloc allocated memory.
1436 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1439 node = numa_node_id();
1441 /* 1) create the cache_cache */
1442 INIT_LIST_HEAD(&cache_chain);
1443 list_add(&cache_cache.next, &cache_chain);
1444 cache_cache.colour_off = cache_line_size();
1445 cache_cache.array[smp_processor_id()] = &initarray_cache.cache;
1446 cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE];
1449 * struct kmem_cache size depends on nr_node_ids, which
1450 * can be less than MAX_NUMNODES.
1452 cache_cache.buffer_size = offsetof(struct kmem_cache, nodelists) +
1453 nr_node_ids * sizeof(struct kmem_list3 *);
1455 cache_cache.obj_size = cache_cache.buffer_size;
1457 cache_cache.buffer_size = ALIGN(cache_cache.buffer_size,
1459 cache_cache.reciprocal_buffer_size =
1460 reciprocal_value(cache_cache.buffer_size);
1462 for (order = 0; order < MAX_ORDER; order++) {
1463 cache_estimate(order, cache_cache.buffer_size,
1464 cache_line_size(), 0, &left_over, &cache_cache.num);
1465 if (cache_cache.num)
1468 BUG_ON(!cache_cache.num);
1469 cache_cache.gfporder = order;
1470 cache_cache.colour = left_over / cache_cache.colour_off;
1471 cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) +
1472 sizeof(struct slab), cache_line_size());
1474 /* 2+3) create the kmalloc caches */
1475 sizes = malloc_sizes;
1476 names = cache_names;
1479 * Initialize the caches that provide memory for the array cache and the
1480 * kmem_list3 structures first. Without this, further allocations will
1484 sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name,
1485 sizes[INDEX_AC].cs_size,
1486 ARCH_KMALLOC_MINALIGN,
1487 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1490 if (INDEX_AC != INDEX_L3) {
1491 sizes[INDEX_L3].cs_cachep =
1492 kmem_cache_create(names[INDEX_L3].name,
1493 sizes[INDEX_L3].cs_size,
1494 ARCH_KMALLOC_MINALIGN,
1495 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1499 slab_early_init = 0;
1501 while (sizes->cs_size != ULONG_MAX) {
1503 * For performance, all the general caches are L1 aligned.
1504 * This should be particularly beneficial on SMP boxes, as it
1505 * eliminates "false sharing".
1506 * Note for systems short on memory removing the alignment will
1507 * allow tighter packing of the smaller caches.
1509 if (!sizes->cs_cachep) {
1510 sizes->cs_cachep = kmem_cache_create(names->name,
1512 ARCH_KMALLOC_MINALIGN,
1513 ARCH_KMALLOC_FLAGS|SLAB_PANIC,
1516 #ifdef CONFIG_ZONE_DMA
1517 sizes->cs_dmacachep = kmem_cache_create(
1520 ARCH_KMALLOC_MINALIGN,
1521 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA|
1528 /* 4) Replace the bootstrap head arrays */
1530 struct array_cache *ptr;
1532 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1534 local_irq_disable();
1535 BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache);
1536 memcpy(ptr, cpu_cache_get(&cache_cache),
1537 sizeof(struct arraycache_init));
1539 * Do not assume that spinlocks can be initialized via memcpy:
1541 spin_lock_init(&ptr->lock);
1543 cache_cache.array[smp_processor_id()] = ptr;
1546 ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
1548 local_irq_disable();
1549 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1550 != &initarray_generic.cache);
1551 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1552 sizeof(struct arraycache_init));
1554 * Do not assume that spinlocks can be initialized via memcpy:
1556 spin_lock_init(&ptr->lock);
1558 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1562 /* 5) Replace the bootstrap kmem_list3's */
1566 /* Replace the static kmem_list3 structures for the boot cpu */
1567 init_list(&cache_cache, &initkmem_list3[CACHE_CACHE], node);
1569 for_each_online_node(nid) {
1570 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1571 &initkmem_list3[SIZE_AC + nid], nid);
1573 if (INDEX_AC != INDEX_L3) {
1574 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1575 &initkmem_list3[SIZE_L3 + nid], nid);
1580 /* 6) resize the head arrays to their final sizes */
1582 struct kmem_cache *cachep;
1583 mutex_lock(&cache_chain_mutex);
1584 list_for_each_entry(cachep, &cache_chain, next)
1585 if (enable_cpucache(cachep))
1587 mutex_unlock(&cache_chain_mutex);
1590 /* Annotate slab for lockdep -- annotate the malloc caches */
1595 g_cpucache_up = FULL;
1598 * Register a cpu startup notifier callback that initializes
1599 * cpu_cache_get for all new cpus
1601 register_cpu_notifier(&cpucache_notifier);
1604 * The reap timers are started later, with a module init call: That part
1605 * of the kernel is not yet operational.
1609 static int __init cpucache_init(void)
1614 * Register the timers that return unneeded pages to the page allocator
1616 for_each_online_cpu(cpu)
1617 start_cpu_timer(cpu);
1620 __initcall(cpucache_init);
1623 * Interface to system's page allocator. No need to hold the cache-lock.
1625 * If we requested dmaable memory, we will get it. Even if we
1626 * did not request dmaable memory, we might get it, but that
1627 * would be relatively rare and ignorable.
1629 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1637 * Nommu uses slab's for process anonymous memory allocations, and thus
1638 * requires __GFP_COMP to properly refcount higher order allocations
1640 flags |= __GFP_COMP;
1643 flags |= cachep->gfpflags;
1645 page = alloc_pages_node(nodeid, flags, cachep->gfporder);
1649 nr_pages = (1 << cachep->gfporder);
1650 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1651 add_zone_page_state(page_zone(page),
1652 NR_SLAB_RECLAIMABLE, nr_pages);
1654 add_zone_page_state(page_zone(page),
1655 NR_SLAB_UNRECLAIMABLE, nr_pages);
1656 for (i = 0; i < nr_pages; i++)
1657 __SetPageSlab(page + i);
1658 return page_address(page);
1662 * Interface to system's page release.
1664 static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1666 unsigned long i = (1 << cachep->gfporder);
1667 struct page *page = virt_to_page(addr);
1668 const unsigned long nr_freed = i;
1670 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1671 sub_zone_page_state(page_zone(page),
1672 NR_SLAB_RECLAIMABLE, nr_freed);
1674 sub_zone_page_state(page_zone(page),
1675 NR_SLAB_UNRECLAIMABLE, nr_freed);
1677 BUG_ON(!PageSlab(page));
1678 __ClearPageSlab(page);
1681 if (current->reclaim_state)
1682 current->reclaim_state->reclaimed_slab += nr_freed;
1683 free_pages((unsigned long)addr, cachep->gfporder);
1686 static void kmem_rcu_free(struct rcu_head *head)
1688 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1689 struct kmem_cache *cachep = slab_rcu->cachep;
1691 kmem_freepages(cachep, slab_rcu->addr);
1692 if (OFF_SLAB(cachep))
1693 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1698 #ifdef CONFIG_DEBUG_PAGEALLOC
1699 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1700 unsigned long caller)
1702 int size = obj_size(cachep);
1704 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1706 if (size < 5 * sizeof(unsigned long))
1709 *addr++ = 0x12345678;
1711 *addr++ = smp_processor_id();
1712 size -= 3 * sizeof(unsigned long);
1714 unsigned long *sptr = &caller;
1715 unsigned long svalue;
1717 while (!kstack_end(sptr)) {
1719 if (kernel_text_address(svalue)) {
1721 size -= sizeof(unsigned long);
1722 if (size <= sizeof(unsigned long))
1728 *addr++ = 0x87654321;
1732 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1734 int size = obj_size(cachep);
1735 addr = &((char *)addr)[obj_offset(cachep)];
1737 memset(addr, val, size);
1738 *(unsigned char *)(addr + size - 1) = POISON_END;
1741 static void dump_line(char *data, int offset, int limit)
1744 unsigned char error = 0;
1747 printk(KERN_ERR "%03x:", offset);
1748 for (i = 0; i < limit; i++) {
1749 if (data[offset + i] != POISON_FREE) {
1750 error = data[offset + i];
1753 printk(" %02x", (unsigned char)data[offset + i]);
1757 if (bad_count == 1) {
1758 error ^= POISON_FREE;
1759 if (!(error & (error - 1))) {
1760 printk(KERN_ERR "Single bit error detected. Probably "
1763 printk(KERN_ERR "Run memtest86+ or a similar memory "
1766 printk(KERN_ERR "Run a memory test tool.\n");
1775 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1780 if (cachep->flags & SLAB_RED_ZONE) {
1781 printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n",
1782 *dbg_redzone1(cachep, objp),
1783 *dbg_redzone2(cachep, objp));
1786 if (cachep->flags & SLAB_STORE_USER) {
1787 printk(KERN_ERR "Last user: [<%p>]",
1788 *dbg_userword(cachep, objp));
1789 print_symbol("(%s)",
1790 (unsigned long)*dbg_userword(cachep, objp));
1793 realobj = (char *)objp + obj_offset(cachep);
1794 size = obj_size(cachep);
1795 for (i = 0; i < size && lines; i += 16, lines--) {
1798 if (i + limit > size)
1800 dump_line(realobj, i, limit);
1804 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1810 realobj = (char *)objp + obj_offset(cachep);
1811 size = obj_size(cachep);
1813 for (i = 0; i < size; i++) {
1814 char exp = POISON_FREE;
1817 if (realobj[i] != exp) {
1823 "Slab corruption: %s start=%p, len=%d\n",
1824 cachep->name, realobj, size);
1825 print_objinfo(cachep, objp, 0);
1827 /* Hexdump the affected line */
1830 if (i + limit > size)
1832 dump_line(realobj, i, limit);
1835 /* Limit to 5 lines */
1841 /* Print some data about the neighboring objects, if they
1844 struct slab *slabp = virt_to_slab(objp);
1847 objnr = obj_to_index(cachep, slabp, objp);
1849 objp = index_to_obj(cachep, slabp, objnr - 1);
1850 realobj = (char *)objp + obj_offset(cachep);
1851 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1853 print_objinfo(cachep, objp, 2);
1855 if (objnr + 1 < cachep->num) {
1856 objp = index_to_obj(cachep, slabp, objnr + 1);
1857 realobj = (char *)objp + obj_offset(cachep);
1858 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1860 print_objinfo(cachep, objp, 2);
1868 * slab_destroy_objs - destroy a slab and its objects
1869 * @cachep: cache pointer being destroyed
1870 * @slabp: slab pointer being destroyed
1872 * Call the registered destructor for each object in a slab that is being
1875 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
1878 for (i = 0; i < cachep->num; i++) {
1879 void *objp = index_to_obj(cachep, slabp, i);
1881 if (cachep->flags & SLAB_POISON) {
1882 #ifdef CONFIG_DEBUG_PAGEALLOC
1883 if (cachep->buffer_size % PAGE_SIZE == 0 &&
1885 kernel_map_pages(virt_to_page(objp),
1886 cachep->buffer_size / PAGE_SIZE, 1);
1888 check_poison_obj(cachep, objp);
1890 check_poison_obj(cachep, objp);
1893 if (cachep->flags & SLAB_RED_ZONE) {
1894 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1895 slab_error(cachep, "start of a freed object "
1897 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1898 slab_error(cachep, "end of a freed object "
1901 if (cachep->dtor && !(cachep->flags & SLAB_POISON))
1902 (cachep->dtor) (objp + obj_offset(cachep), cachep, 0);
1906 static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp)
1910 for (i = 0; i < cachep->num; i++) {
1911 void *objp = index_to_obj(cachep, slabp, i);
1912 (cachep->dtor) (objp, cachep, 0);
1919 * slab_destroy - destroy and release all objects in a slab
1920 * @cachep: cache pointer being destroyed
1921 * @slabp: slab pointer being destroyed
1923 * Destroy all the objs in a slab, and release the mem back to the system.
1924 * Before calling the slab must have been unlinked from the cache. The
1925 * cache-lock is not held/needed.
1927 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
1929 void *addr = slabp->s_mem - slabp->colouroff;
1931 slab_destroy_objs(cachep, slabp);
1932 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1933 struct slab_rcu *slab_rcu;
1935 slab_rcu = (struct slab_rcu *)slabp;
1936 slab_rcu->cachep = cachep;
1937 slab_rcu->addr = addr;
1938 call_rcu(&slab_rcu->head, kmem_rcu_free);
1940 kmem_freepages(cachep, addr);
1941 if (OFF_SLAB(cachep))
1942 kmem_cache_free(cachep->slabp_cache, slabp);
1947 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1948 * size of kmem_list3.
1950 static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1954 for_each_online_node(node) {
1955 cachep->nodelists[node] = &initkmem_list3[index + node];
1956 cachep->nodelists[node]->next_reap = jiffies +
1958 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1962 static void __kmem_cache_destroy(struct kmem_cache *cachep)
1965 struct kmem_list3 *l3;
1967 for_each_online_cpu(i)
1968 kfree(cachep->array[i]);
1970 /* NUMA: free the list3 structures */
1971 for_each_online_node(i) {
1972 l3 = cachep->nodelists[i];
1975 free_alien_cache(l3->alien);
1979 kmem_cache_free(&cache_cache, cachep);
1984 * calculate_slab_order - calculate size (page order) of slabs
1985 * @cachep: pointer to the cache that is being created
1986 * @size: size of objects to be created in this cache.
1987 * @align: required alignment for the objects.
1988 * @flags: slab allocation flags
1990 * Also calculates the number of objects per slab.
1992 * This could be made much more intelligent. For now, try to avoid using
1993 * high order pages for slabs. When the gfp() functions are more friendly
1994 * towards high-order requests, this should be changed.
1996 static size_t calculate_slab_order(struct kmem_cache *cachep,
1997 size_t size, size_t align, unsigned long flags)
1999 unsigned long offslab_limit;
2000 size_t left_over = 0;
2003 for (gfporder = 0; gfporder <= MAX_GFP_ORDER; gfporder++) {
2007 cache_estimate(gfporder, size, align, flags, &remainder, &num);
2011 if (flags & CFLGS_OFF_SLAB) {
2013 * Max number of objs-per-slab for caches which
2014 * use off-slab slabs. Needed to avoid a possible
2015 * looping condition in cache_grow().
2017 offslab_limit = size - sizeof(struct slab);
2018 offslab_limit /= sizeof(kmem_bufctl_t);
2020 if (num > offslab_limit)
2024 /* Found something acceptable - save it away */
2026 cachep->gfporder = gfporder;
2027 left_over = remainder;
2030 * A VFS-reclaimable slab tends to have most allocations
2031 * as GFP_NOFS and we really don't want to have to be allocating
2032 * higher-order pages when we are unable to shrink dcache.
2034 if (flags & SLAB_RECLAIM_ACCOUNT)
2038 * Large number of objects is good, but very large slabs are
2039 * currently bad for the gfp()s.
2041 if (gfporder >= slab_break_gfp_order)
2045 * Acceptable internal fragmentation?
2047 if (left_over * 8 <= (PAGE_SIZE << gfporder))
2053 static int setup_cpu_cache(struct kmem_cache *cachep)
2055 if (g_cpucache_up == FULL)
2056 return enable_cpucache(cachep);
2058 if (g_cpucache_up == NONE) {
2060 * Note: the first kmem_cache_create must create the cache
2061 * that's used by kmalloc(24), otherwise the creation of
2062 * further caches will BUG().
2064 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2067 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2068 * the first cache, then we need to set up all its list3s,
2069 * otherwise the creation of further caches will BUG().
2071 set_up_list3s(cachep, SIZE_AC);
2072 if (INDEX_AC == INDEX_L3)
2073 g_cpucache_up = PARTIAL_L3;
2075 g_cpucache_up = PARTIAL_AC;
2077 cachep->array[smp_processor_id()] =
2078 kmalloc(sizeof(struct arraycache_init), GFP_KERNEL);
2080 if (g_cpucache_up == PARTIAL_AC) {
2081 set_up_list3s(cachep, SIZE_L3);
2082 g_cpucache_up = PARTIAL_L3;
2085 for_each_online_node(node) {
2086 cachep->nodelists[node] =
2087 kmalloc_node(sizeof(struct kmem_list3),
2089 BUG_ON(!cachep->nodelists[node]);
2090 kmem_list3_init(cachep->nodelists[node]);
2094 cachep->nodelists[numa_node_id()]->next_reap =
2095 jiffies + REAPTIMEOUT_LIST3 +
2096 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2098 cpu_cache_get(cachep)->avail = 0;
2099 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2100 cpu_cache_get(cachep)->batchcount = 1;
2101 cpu_cache_get(cachep)->touched = 0;
2102 cachep->batchcount = 1;
2103 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2108 * kmem_cache_create - Create a cache.
2109 * @name: A string which is used in /proc/slabinfo to identify this cache.
2110 * @size: The size of objects to be created in this cache.
2111 * @align: The required alignment for the objects.
2112 * @flags: SLAB flags
2113 * @ctor: A constructor for the objects.
2114 * @dtor: A destructor for the objects.
2116 * Returns a ptr to the cache on success, NULL on failure.
2117 * Cannot be called within a int, but can be interrupted.
2118 * The @ctor is run when new pages are allocated by the cache
2119 * and the @dtor is run before the pages are handed back.
2121 * @name must be valid until the cache is destroyed. This implies that
2122 * the module calling this has to destroy the cache before getting unloaded.
2126 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2127 * to catch references to uninitialised memory.
2129 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2130 * for buffer overruns.
2132 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2133 * cacheline. This can be beneficial if you're counting cycles as closely
2137 kmem_cache_create (const char *name, size_t size, size_t align,
2138 unsigned long flags,
2139 void (*ctor)(void*, struct kmem_cache *, unsigned long),
2140 void (*dtor)(void*, struct kmem_cache *, unsigned long))
2142 size_t left_over, slab_size, ralign;
2143 struct kmem_cache *cachep = NULL, *pc;
2146 * Sanity checks... these are all serious usage bugs.
2148 if (!name || in_interrupt() || (size < BYTES_PER_WORD) ||
2149 (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) {
2150 printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__,
2156 * We use cache_chain_mutex to ensure a consistent view of
2157 * cpu_online_map as well. Please see cpuup_callback
2159 mutex_lock(&cache_chain_mutex);
2161 list_for_each_entry(pc, &cache_chain, next) {
2166 * This happens when the module gets unloaded and doesn't
2167 * destroy its slab cache and no-one else reuses the vmalloc
2168 * area of the module. Print a warning.
2170 res = probe_kernel_address(pc->name, tmp);
2172 printk("SLAB: cache with size %d has lost its name\n",
2177 if (!strcmp(pc->name, name)) {
2178 printk("kmem_cache_create: duplicate cache %s\n", name);
2185 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
2186 if ((flags & SLAB_DEBUG_INITIAL) && !ctor) {
2187 /* No constructor, but inital state check requested */
2188 printk(KERN_ERR "%s: No con, but init state check "
2189 "requested - %s\n", __FUNCTION__, name);
2190 flags &= ~SLAB_DEBUG_INITIAL;
2194 * Enable redzoning and last user accounting, except for caches with
2195 * large objects, if the increased size would increase the object size
2196 * above the next power of two: caches with object sizes just above a
2197 * power of two have a significant amount of internal fragmentation.
2199 if (size < 4096 || fls(size - 1) == fls(size-1 + 3 * BYTES_PER_WORD))
2200 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2201 if (!(flags & SLAB_DESTROY_BY_RCU))
2202 flags |= SLAB_POISON;
2204 if (flags & SLAB_DESTROY_BY_RCU)
2205 BUG_ON(flags & SLAB_POISON);
2207 if (flags & SLAB_DESTROY_BY_RCU)
2211 * Always checks flags, a caller might be expecting debug support which
2214 BUG_ON(flags & ~CREATE_MASK);
2217 * Check that size is in terms of words. This is needed to avoid
2218 * unaligned accesses for some archs when redzoning is used, and makes
2219 * sure any on-slab bufctl's are also correctly aligned.
2221 if (size & (BYTES_PER_WORD - 1)) {
2222 size += (BYTES_PER_WORD - 1);
2223 size &= ~(BYTES_PER_WORD - 1);
2226 /* calculate the final buffer alignment: */
2228 /* 1) arch recommendation: can be overridden for debug */
2229 if (flags & SLAB_HWCACHE_ALIGN) {
2231 * Default alignment: as specified by the arch code. Except if
2232 * an object is really small, then squeeze multiple objects into
2235 ralign = cache_line_size();
2236 while (size <= ralign / 2)
2239 ralign = BYTES_PER_WORD;
2243 * Redzoning and user store require word alignment. Note this will be
2244 * overridden by architecture or caller mandated alignment if either
2245 * is greater than BYTES_PER_WORD.
2247 if (flags & SLAB_RED_ZONE || flags & SLAB_STORE_USER)
2248 ralign = BYTES_PER_WORD;
2250 /* 2) arch mandated alignment */
2251 if (ralign < ARCH_SLAB_MINALIGN) {
2252 ralign = ARCH_SLAB_MINALIGN;
2254 /* 3) caller mandated alignment */
2255 if (ralign < align) {
2258 /* disable debug if necessary */
2259 if (ralign > BYTES_PER_WORD)
2260 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2266 /* Get cache's description obj. */
2267 cachep = kmem_cache_zalloc(&cache_cache, GFP_KERNEL);
2272 cachep->obj_size = size;
2275 * Both debugging options require word-alignment which is calculated
2278 if (flags & SLAB_RED_ZONE) {
2279 /* add space for red zone words */
2280 cachep->obj_offset += BYTES_PER_WORD;
2281 size += 2 * BYTES_PER_WORD;
2283 if (flags & SLAB_STORE_USER) {
2284 /* user store requires one word storage behind the end of
2287 size += BYTES_PER_WORD;
2289 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2290 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
2291 && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) {
2292 cachep->obj_offset += PAGE_SIZE - size;
2299 * Determine if the slab management is 'on' or 'off' slab.
2300 * (bootstrapping cannot cope with offslab caches so don't do
2303 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init)
2305 * Size is large, assume best to place the slab management obj
2306 * off-slab (should allow better packing of objs).
2308 flags |= CFLGS_OFF_SLAB;
2310 size = ALIGN(size, align);
2312 left_over = calculate_slab_order(cachep, size, align, flags);
2315 printk("kmem_cache_create: couldn't create cache %s.\n", name);
2316 kmem_cache_free(&cache_cache, cachep);
2320 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2321 + sizeof(struct slab), align);
2324 * If the slab has been placed off-slab, and we have enough space then
2325 * move it on-slab. This is at the expense of any extra colouring.
2327 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2328 flags &= ~CFLGS_OFF_SLAB;
2329 left_over -= slab_size;
2332 if (flags & CFLGS_OFF_SLAB) {
2333 /* really off slab. No need for manual alignment */
2335 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2338 cachep->colour_off = cache_line_size();
2339 /* Offset must be a multiple of the alignment. */
2340 if (cachep->colour_off < align)
2341 cachep->colour_off = align;
2342 cachep->colour = left_over / cachep->colour_off;
2343 cachep->slab_size = slab_size;
2344 cachep->flags = flags;
2345 cachep->gfpflags = 0;
2346 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2347 cachep->gfpflags |= GFP_DMA;
2348 cachep->buffer_size = size;
2349 cachep->reciprocal_buffer_size = reciprocal_value(size);
2351 if (flags & CFLGS_OFF_SLAB) {
2352 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2354 * This is a possibility for one of the malloc_sizes caches.
2355 * But since we go off slab only for object size greater than
2356 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2357 * this should not happen at all.
2358 * But leave a BUG_ON for some lucky dude.
2360 BUG_ON(!cachep->slabp_cache);
2362 cachep->ctor = ctor;
2363 cachep->dtor = dtor;
2364 cachep->name = name;
2366 if (setup_cpu_cache(cachep)) {
2367 __kmem_cache_destroy(cachep);
2372 /* cache setup completed, link it into the list */
2373 list_add(&cachep->next, &cache_chain);
2375 if (!cachep && (flags & SLAB_PANIC))
2376 panic("kmem_cache_create(): failed to create slab `%s'\n",
2378 mutex_unlock(&cache_chain_mutex);
2381 EXPORT_SYMBOL(kmem_cache_create);
2384 static void check_irq_off(void)
2386 BUG_ON(!irqs_disabled());
2389 static void check_irq_on(void)
2391 BUG_ON(irqs_disabled());
2394 static void check_spinlock_acquired(struct kmem_cache *cachep)
2398 assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock);
2402 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2406 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2411 #define check_irq_off() do { } while(0)
2412 #define check_irq_on() do { } while(0)
2413 #define check_spinlock_acquired(x) do { } while(0)
2414 #define check_spinlock_acquired_node(x, y) do { } while(0)
2417 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2418 struct array_cache *ac,
2419 int force, int node);
2421 static void do_drain(void *arg)
2423 struct kmem_cache *cachep = arg;
2424 struct array_cache *ac;
2425 int node = numa_node_id();
2428 ac = cpu_cache_get(cachep);
2429 spin_lock(&cachep->nodelists[node]->list_lock);
2430 free_block(cachep, ac->entry, ac->avail, node);
2431 spin_unlock(&cachep->nodelists[node]->list_lock);
2435 static void drain_cpu_caches(struct kmem_cache *cachep)
2437 struct kmem_list3 *l3;
2440 on_each_cpu(do_drain, cachep, 1, 1);
2442 for_each_online_node(node) {
2443 l3 = cachep->nodelists[node];
2444 if (l3 && l3->alien)
2445 drain_alien_cache(cachep, l3->alien);
2448 for_each_online_node(node) {
2449 l3 = cachep->nodelists[node];
2451 drain_array(cachep, l3, l3->shared, 1, node);
2456 * Remove slabs from the list of free slabs.
2457 * Specify the number of slabs to drain in tofree.
2459 * Returns the actual number of slabs released.
2461 static int drain_freelist(struct kmem_cache *cache,
2462 struct kmem_list3 *l3, int tofree)
2464 struct list_head *p;
2469 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
2471 spin_lock_irq(&l3->list_lock);
2472 p = l3->slabs_free.prev;
2473 if (p == &l3->slabs_free) {
2474 spin_unlock_irq(&l3->list_lock);
2478 slabp = list_entry(p, struct slab, list);
2480 BUG_ON(slabp->inuse);
2482 list_del(&slabp->list);
2484 * Safe to drop the lock. The slab is no longer linked
2487 l3->free_objects -= cache->num;
2488 spin_unlock_irq(&l3->list_lock);
2489 slab_destroy(cache, slabp);
2496 /* Called with cache_chain_mutex held to protect against cpu hotplug */
2497 static int __cache_shrink(struct kmem_cache *cachep)
2500 struct kmem_list3 *l3;
2502 drain_cpu_caches(cachep);
2505 for_each_online_node(i) {
2506 l3 = cachep->nodelists[i];
2510 drain_freelist(cachep, l3, l3->free_objects);
2512 ret += !list_empty(&l3->slabs_full) ||
2513 !list_empty(&l3->slabs_partial);
2515 return (ret ? 1 : 0);
2519 * kmem_cache_shrink - Shrink a cache.
2520 * @cachep: The cache to shrink.
2522 * Releases as many slabs as possible for a cache.
2523 * To help debugging, a zero exit status indicates all slabs were released.
2525 int kmem_cache_shrink(struct kmem_cache *cachep)
2528 BUG_ON(!cachep || in_interrupt());
2530 mutex_lock(&cache_chain_mutex);
2531 ret = __cache_shrink(cachep);
2532 mutex_unlock(&cache_chain_mutex);
2535 EXPORT_SYMBOL(kmem_cache_shrink);
2538 * kmem_cache_destroy - delete a cache
2539 * @cachep: the cache to destroy
2541 * Remove a &struct kmem_cache object from the slab cache.
2543 * It is expected this function will be called by a module when it is
2544 * unloaded. This will remove the cache completely, and avoid a duplicate
2545 * cache being allocated each time a module is loaded and unloaded, if the
2546 * module doesn't have persistent in-kernel storage across loads and unloads.
2548 * The cache must be empty before calling this function.
2550 * The caller must guarantee that noone will allocate memory from the cache
2551 * during the kmem_cache_destroy().
2553 void kmem_cache_destroy(struct kmem_cache *cachep)
2555 BUG_ON(!cachep || in_interrupt());
2557 /* Find the cache in the chain of caches. */
2558 mutex_lock(&cache_chain_mutex);
2560 * the chain is never empty, cache_cache is never destroyed
2562 list_del(&cachep->next);
2563 if (__cache_shrink(cachep)) {
2564 slab_error(cachep, "Can't free all objects");
2565 list_add(&cachep->next, &cache_chain);
2566 mutex_unlock(&cache_chain_mutex);
2570 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
2573 __kmem_cache_destroy(cachep);
2574 mutex_unlock(&cache_chain_mutex);
2576 EXPORT_SYMBOL(kmem_cache_destroy);
2579 * Get the memory for a slab management obj.
2580 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2581 * always come from malloc_sizes caches. The slab descriptor cannot
2582 * come from the same cache which is getting created because,
2583 * when we are searching for an appropriate cache for these
2584 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2585 * If we are creating a malloc_sizes cache here it would not be visible to
2586 * kmem_find_general_cachep till the initialization is complete.
2587 * Hence we cannot have slabp_cache same as the original cache.
2589 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2590 int colour_off, gfp_t local_flags,
2595 if (OFF_SLAB(cachep)) {
2596 /* Slab management obj is off-slab. */
2597 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2598 local_flags & ~GFP_THISNODE, nodeid);
2602 slabp = objp + colour_off;
2603 colour_off += cachep->slab_size;
2606 slabp->colouroff = colour_off;
2607 slabp->s_mem = objp + colour_off;
2608 slabp->nodeid = nodeid;
2612 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2614 return (kmem_bufctl_t *) (slabp + 1);
2617 static void cache_init_objs(struct kmem_cache *cachep,
2618 struct slab *slabp, unsigned long ctor_flags)
2622 for (i = 0; i < cachep->num; i++) {
2623 void *objp = index_to_obj(cachep, slabp, i);
2625 /* need to poison the objs? */
2626 if (cachep->flags & SLAB_POISON)
2627 poison_obj(cachep, objp, POISON_FREE);
2628 if (cachep->flags & SLAB_STORE_USER)
2629 *dbg_userword(cachep, objp) = NULL;
2631 if (cachep->flags & SLAB_RED_ZONE) {
2632 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2633 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2636 * Constructors are not allowed to allocate memory from the same
2637 * cache which they are a constructor for. Otherwise, deadlock.
2638 * They must also be threaded.
2640 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2641 cachep->ctor(objp + obj_offset(cachep), cachep,
2644 if (cachep->flags & SLAB_RED_ZONE) {
2645 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2646 slab_error(cachep, "constructor overwrote the"
2647 " end of an object");
2648 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2649 slab_error(cachep, "constructor overwrote the"
2650 " start of an object");
2652 if ((cachep->buffer_size % PAGE_SIZE) == 0 &&
2653 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2654 kernel_map_pages(virt_to_page(objp),
2655 cachep->buffer_size / PAGE_SIZE, 0);
2658 cachep->ctor(objp, cachep, ctor_flags);
2660 slab_bufctl(slabp)[i] = i + 1;
2662 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2666 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2668 if (CONFIG_ZONE_DMA_FLAG) {
2669 if (flags & GFP_DMA)
2670 BUG_ON(!(cachep->gfpflags & GFP_DMA));
2672 BUG_ON(cachep->gfpflags & GFP_DMA);
2676 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2679 void *objp = index_to_obj(cachep, slabp, slabp->free);
2683 next = slab_bufctl(slabp)[slabp->free];
2685 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2686 WARN_ON(slabp->nodeid != nodeid);
2693 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2694 void *objp, int nodeid)
2696 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2699 /* Verify that the slab belongs to the intended node */
2700 WARN_ON(slabp->nodeid != nodeid);
2702 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2703 printk(KERN_ERR "slab: double free detected in cache "
2704 "'%s', objp %p\n", cachep->name, objp);
2708 slab_bufctl(slabp)[objnr] = slabp->free;
2709 slabp->free = objnr;
2714 * Map pages beginning at addr to the given cache and slab. This is required
2715 * for the slab allocator to be able to lookup the cache and slab of a
2716 * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging.
2718 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2724 page = virt_to_page(addr);
2727 if (likely(!PageCompound(page)))
2728 nr_pages <<= cache->gfporder;
2731 page_set_cache(page, cache);
2732 page_set_slab(page, slab);
2734 } while (--nr_pages);
2738 * Grow (by 1) the number of slabs within a cache. This is called by
2739 * kmem_cache_alloc() when there are no active objs left in a cache.
2741 static int cache_grow(struct kmem_cache *cachep,
2742 gfp_t flags, int nodeid, void *objp)
2747 unsigned long ctor_flags;
2748 struct kmem_list3 *l3;
2751 * Be lazy and only check for valid flags here, keeping it out of the
2752 * critical path in kmem_cache_alloc().
2754 BUG_ON(flags & ~(GFP_DMA | GFP_LEVEL_MASK | __GFP_NO_GROW));
2755 if (flags & __GFP_NO_GROW)
2758 ctor_flags = SLAB_CTOR_CONSTRUCTOR;
2759 local_flags = (flags & GFP_LEVEL_MASK);
2760 if (!(local_flags & __GFP_WAIT))
2762 * Not allowed to sleep. Need to tell a constructor about
2763 * this - it might need to know...
2765 ctor_flags |= SLAB_CTOR_ATOMIC;
2767 /* Take the l3 list lock to change the colour_next on this node */
2769 l3 = cachep->nodelists[nodeid];
2770 spin_lock(&l3->list_lock);
2772 /* Get colour for the slab, and cal the next value. */
2773 offset = l3->colour_next;
2775 if (l3->colour_next >= cachep->colour)
2776 l3->colour_next = 0;
2777 spin_unlock(&l3->list_lock);
2779 offset *= cachep->colour_off;
2781 if (local_flags & __GFP_WAIT)
2785 * The test for missing atomic flag is performed here, rather than
2786 * the more obvious place, simply to reduce the critical path length
2787 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2788 * will eventually be caught here (where it matters).
2790 kmem_flagcheck(cachep, flags);
2793 * Get mem for the objs. Attempt to allocate a physical page from
2797 objp = kmem_getpages(cachep, flags, nodeid);
2801 /* Get slab management. */
2802 slabp = alloc_slabmgmt(cachep, objp, offset,
2803 local_flags & ~GFP_THISNODE, nodeid);
2807 slabp->nodeid = nodeid;
2808 slab_map_pages(cachep, slabp, objp);
2810 cache_init_objs(cachep, slabp, ctor_flags);
2812 if (local_flags & __GFP_WAIT)
2813 local_irq_disable();
2815 spin_lock(&l3->list_lock);
2817 /* Make slab active. */
2818 list_add_tail(&slabp->list, &(l3->slabs_free));
2819 STATS_INC_GROWN(cachep);
2820 l3->free_objects += cachep->num;
2821 spin_unlock(&l3->list_lock);
2824 kmem_freepages(cachep, objp);
2826 if (local_flags & __GFP_WAIT)
2827 local_irq_disable();
2834 * Perform extra freeing checks:
2835 * - detect bad pointers.
2836 * - POISON/RED_ZONE checking
2837 * - destructor calls, for caches with POISON+dtor
2839 static void kfree_debugcheck(const void *objp)
2841 if (!virt_addr_valid(objp)) {
2842 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2843 (unsigned long)objp);
2848 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2850 unsigned long redzone1, redzone2;
2852 redzone1 = *dbg_redzone1(cache, obj);
2853 redzone2 = *dbg_redzone2(cache, obj);
2858 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2861 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2862 slab_error(cache, "double free detected");
2864 slab_error(cache, "memory outside object was overwritten");
2866 printk(KERN_ERR "%p: redzone 1:0x%lx, redzone 2:0x%lx.\n",
2867 obj, redzone1, redzone2);
2870 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2877 objp -= obj_offset(cachep);
2878 kfree_debugcheck(objp);
2879 page = virt_to_page(objp);
2881 slabp = page_get_slab(page);
2883 if (cachep->flags & SLAB_RED_ZONE) {
2884 verify_redzone_free(cachep, objp);
2885 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2886 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2888 if (cachep->flags & SLAB_STORE_USER)
2889 *dbg_userword(cachep, objp) = caller;
2891 objnr = obj_to_index(cachep, slabp, objp);
2893 BUG_ON(objnr >= cachep->num);
2894 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
2896 if (cachep->flags & SLAB_DEBUG_INITIAL) {
2898 * Need to call the slab's constructor so the caller can
2899 * perform a verify of its state (debugging). Called without
2900 * the cache-lock held.
2902 cachep->ctor(objp + obj_offset(cachep),
2903 cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY);
2905 if (cachep->flags & SLAB_POISON && cachep->dtor) {
2906 /* we want to cache poison the object,
2907 * call the destruction callback
2909 cachep->dtor(objp + obj_offset(cachep), cachep, 0);
2911 #ifdef CONFIG_DEBUG_SLAB_LEAK
2912 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2914 if (cachep->flags & SLAB_POISON) {
2915 #ifdef CONFIG_DEBUG_PAGEALLOC
2916 if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
2917 store_stackinfo(cachep, objp, (unsigned long)caller);
2918 kernel_map_pages(virt_to_page(objp),
2919 cachep->buffer_size / PAGE_SIZE, 0);
2921 poison_obj(cachep, objp, POISON_FREE);
2924 poison_obj(cachep, objp, POISON_FREE);
2930 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
2935 /* Check slab's freelist to see if this obj is there. */
2936 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2938 if (entries > cachep->num || i >= cachep->num)
2941 if (entries != cachep->num - slabp->inuse) {
2943 printk(KERN_ERR "slab: Internal list corruption detected in "
2944 "cache '%s'(%d), slabp %p(%d). Hexdump:\n",
2945 cachep->name, cachep->num, slabp, slabp->inuse);
2947 i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t);
2950 printk("\n%03x:", i);
2951 printk(" %02x", ((unsigned char *)slabp)[i]);
2958 #define kfree_debugcheck(x) do { } while(0)
2959 #define cache_free_debugcheck(x,objp,z) (objp)
2960 #define check_slabp(x,y) do { } while(0)
2963 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
2966 struct kmem_list3 *l3;
2967 struct array_cache *ac;
2970 node = numa_node_id();
2973 ac = cpu_cache_get(cachep);
2975 batchcount = ac->batchcount;
2976 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2978 * If there was little recent activity on this cache, then
2979 * perform only a partial refill. Otherwise we could generate
2982 batchcount = BATCHREFILL_LIMIT;
2984 l3 = cachep->nodelists[node];
2986 BUG_ON(ac->avail > 0 || !l3);
2987 spin_lock(&l3->list_lock);
2989 /* See if we can refill from the shared array */
2990 if (l3->shared && transfer_objects(ac, l3->shared, batchcount))
2993 while (batchcount > 0) {
2994 struct list_head *entry;
2996 /* Get slab alloc is to come from. */
2997 entry = l3->slabs_partial.next;
2998 if (entry == &l3->slabs_partial) {
2999 l3->free_touched = 1;
3000 entry = l3->slabs_free.next;
3001 if (entry == &l3->slabs_free)
3005 slabp = list_entry(entry, struct slab, list);
3006 check_slabp(cachep, slabp);
3007 check_spinlock_acquired(cachep);
3010 * The slab was either on partial or free list so
3011 * there must be at least one object available for
3014 BUG_ON(slabp->inuse < 0 || slabp->inuse >= cachep->num);
3016 while (slabp->inuse < cachep->num && batchcount--) {
3017 STATS_INC_ALLOCED(cachep);
3018 STATS_INC_ACTIVE(cachep);
3019 STATS_SET_HIGH(cachep);
3021 ac->entry[ac->avail++] = slab_get_obj(cachep, slabp,
3024 check_slabp(cachep, slabp);
3026 /* move slabp to correct slabp list: */
3027 list_del(&slabp->list);
3028 if (slabp->free == BUFCTL_END)
3029 list_add(&slabp->list, &l3->slabs_full);
3031 list_add(&slabp->list, &l3->slabs_partial);
3035 l3->free_objects -= ac->avail;
3037 spin_unlock(&l3->list_lock);
3039 if (unlikely(!ac->avail)) {
3041 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
3043 /* cache_grow can reenable interrupts, then ac could change. */
3044 ac = cpu_cache_get(cachep);
3045 if (!x && ac->avail == 0) /* no objects in sight? abort */
3048 if (!ac->avail) /* objects refilled by interrupt? */
3052 return ac->entry[--ac->avail];
3055 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3058 might_sleep_if(flags & __GFP_WAIT);
3060 kmem_flagcheck(cachep, flags);
3065 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3066 gfp_t flags, void *objp, void *caller)
3070 if (cachep->flags & SLAB_POISON) {
3071 #ifdef CONFIG_DEBUG_PAGEALLOC
3072 if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
3073 kernel_map_pages(virt_to_page(objp),
3074 cachep->buffer_size / PAGE_SIZE, 1);
3076 check_poison_obj(cachep, objp);
3078 check_poison_obj(cachep, objp);
3080 poison_obj(cachep, objp, POISON_INUSE);
3082 if (cachep->flags & SLAB_STORE_USER)
3083 *dbg_userword(cachep, objp) = caller;
3085 if (cachep->flags & SLAB_RED_ZONE) {
3086 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3087 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3088 slab_error(cachep, "double free, or memory outside"
3089 " object was overwritten");
3091 "%p: redzone 1:0x%lx, redzone 2:0x%lx\n",
3092 objp, *dbg_redzone1(cachep, objp),
3093 *dbg_redzone2(cachep, objp));
3095 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3096 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3098 #ifdef CONFIG_DEBUG_SLAB_LEAK
3103 slabp = page_get_slab(virt_to_page(objp));
3104 objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size;
3105 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3108 objp += obj_offset(cachep);
3109 if (cachep->ctor && cachep->flags & SLAB_POISON) {
3110 unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR;
3112 if (!(flags & __GFP_WAIT))
3113 ctor_flags |= SLAB_CTOR_ATOMIC;
3115 cachep->ctor(objp, cachep, ctor_flags);
3117 #if ARCH_SLAB_MINALIGN
3118 if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) {
3119 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3120 objp, ARCH_SLAB_MINALIGN);
3126 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3129 #ifdef CONFIG_FAILSLAB
3131 static struct failslab_attr {
3133 struct fault_attr attr;
3135 u32 ignore_gfp_wait;
3136 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3137 struct dentry *ignore_gfp_wait_file;
3141 .attr = FAULT_ATTR_INITIALIZER,
3142 .ignore_gfp_wait = 1,
3145 static int __init setup_failslab(char *str)
3147 return setup_fault_attr(&failslab.attr, str);
3149 __setup("failslab=", setup_failslab);
3151 static int should_failslab(struct kmem_cache *cachep, gfp_t flags)
3153 if (cachep == &cache_cache)
3155 if (flags & __GFP_NOFAIL)
3157 if (failslab.ignore_gfp_wait && (flags & __GFP_WAIT))
3160 return should_fail(&failslab.attr, obj_size(cachep));
3163 #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS
3165 static int __init failslab_debugfs(void)
3167 mode_t mode = S_IFREG | S_IRUSR | S_IWUSR;
3171 err = init_fault_attr_dentries(&failslab.attr, "failslab");
3174 dir = failslab.attr.dentries.dir;
3176 failslab.ignore_gfp_wait_file =
3177 debugfs_create_bool("ignore-gfp-wait", mode, dir,
3178 &failslab.ignore_gfp_wait);
3180 if (!failslab.ignore_gfp_wait_file) {
3182 debugfs_remove(failslab.ignore_gfp_wait_file);
3183 cleanup_fault_attr_dentries(&failslab.attr);
3189 late_initcall(failslab_debugfs);
3191 #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */
3193 #else /* CONFIG_FAILSLAB */
3195 static inline int should_failslab(struct kmem_cache *cachep, gfp_t flags)
3200 #endif /* CONFIG_FAILSLAB */
3202 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3205 struct array_cache *ac;
3209 if (should_failslab(cachep, flags))
3212 ac = cpu_cache_get(cachep);
3213 if (likely(ac->avail)) {
3214 STATS_INC_ALLOCHIT(cachep);
3216 objp = ac->entry[--ac->avail];
3218 STATS_INC_ALLOCMISS(cachep);
3219 objp = cache_alloc_refill(cachep, flags);
3226 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3228 * If we are in_interrupt, then process context, including cpusets and
3229 * mempolicy, may not apply and should not be used for allocation policy.
3231 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3233 int nid_alloc, nid_here;
3235 if (in_interrupt() || (flags & __GFP_THISNODE))
3237 nid_alloc = nid_here = numa_node_id();
3238 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3239 nid_alloc = cpuset_mem_spread_node();
3240 else if (current->mempolicy)
3241 nid_alloc = slab_node(current->mempolicy);
3242 if (nid_alloc != nid_here)
3243 return ____cache_alloc_node(cachep, flags, nid_alloc);
3248 * Fallback function if there was no memory available and no objects on a
3249 * certain node and fall back is permitted. First we scan all the
3250 * available nodelists for available objects. If that fails then we
3251 * perform an allocation without specifying a node. This allows the page
3252 * allocator to do its reclaim / fallback magic. We then insert the
3253 * slab into the proper nodelist and then allocate from it.
3255 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3257 struct zonelist *zonelist;
3263 if (flags & __GFP_THISNODE)
3266 zonelist = &NODE_DATA(slab_node(current->mempolicy))
3267 ->node_zonelists[gfp_zone(flags)];
3268 local_flags = (flags & GFP_LEVEL_MASK);
3272 * Look through allowed nodes for objects available
3273 * from existing per node queues.
3275 for (z = zonelist->zones; *z && !obj; z++) {
3276 nid = zone_to_nid(*z);
3278 if (cpuset_zone_allowed_hardwall(*z, flags) &&
3279 cache->nodelists[nid] &&
3280 cache->nodelists[nid]->free_objects)
3281 obj = ____cache_alloc_node(cache,
3282 flags | GFP_THISNODE, nid);
3285 if (!obj && !(flags & __GFP_NO_GROW)) {
3287 * This allocation will be performed within the constraints
3288 * of the current cpuset / memory policy requirements.
3289 * We may trigger various forms of reclaim on the allowed
3290 * set and go into memory reserves if necessary.
3292 if (local_flags & __GFP_WAIT)
3294 kmem_flagcheck(cache, flags);
3295 obj = kmem_getpages(cache, flags, -1);
3296 if (local_flags & __GFP_WAIT)
3297 local_irq_disable();
3300 * Insert into the appropriate per node queues
3302 nid = page_to_nid(virt_to_page(obj));
3303 if (cache_grow(cache, flags, nid, obj)) {
3304 obj = ____cache_alloc_node(cache,
3305 flags | GFP_THISNODE, nid);
3308 * Another processor may allocate the
3309 * objects in the slab since we are
3310 * not holding any locks.
3314 /* cache_grow already freed obj */
3323 * A interface to enable slab creation on nodeid
3325 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3328 struct list_head *entry;
3330 struct kmem_list3 *l3;
3334 l3 = cachep->nodelists[nodeid];
3339 spin_lock(&l3->list_lock);
3340 entry = l3->slabs_partial.next;
3341 if (entry == &l3->slabs_partial) {
3342 l3->free_touched = 1;
3343 entry = l3->slabs_free.next;
3344 if (entry == &l3->slabs_free)
3348 slabp = list_entry(entry, struct slab, list);
3349 check_spinlock_acquired_node(cachep, nodeid);
3350 check_slabp(cachep, slabp);
3352 STATS_INC_NODEALLOCS(cachep);
3353 STATS_INC_ACTIVE(cachep);
3354 STATS_SET_HIGH(cachep);
3356 BUG_ON(slabp->inuse == cachep->num);
3358 obj = slab_get_obj(cachep, slabp, nodeid);
3359 check_slabp(cachep, slabp);
3361 /* move slabp to correct slabp list: */
3362 list_del(&slabp->list);
3364 if (slabp->free == BUFCTL_END)
3365 list_add(&slabp->list, &l3->slabs_full);
3367 list_add(&slabp->list, &l3->slabs_partial);
3369 spin_unlock(&l3->list_lock);
3373 spin_unlock(&l3->list_lock);
3374 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
3378 return fallback_alloc(cachep, flags);
3385 * kmem_cache_alloc_node - Allocate an object on the specified node
3386 * @cachep: The cache to allocate from.
3387 * @flags: See kmalloc().
3388 * @nodeid: node number of the target node.
3389 * @caller: return address of caller, used for debug information
3391 * Identical to kmem_cache_alloc but it will allocate memory on the given
3392 * node, which can improve the performance for cpu bound structures.
3394 * Fallback to other node is possible if __GFP_THISNODE is not set.
3396 static __always_inline void *
3397 __cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3400 unsigned long save_flags;
3403 cache_alloc_debugcheck_before(cachep, flags);
3404 local_irq_save(save_flags);
3406 if (unlikely(nodeid == -1))
3407 nodeid = numa_node_id();
3409 if (unlikely(!cachep->nodelists[nodeid])) {
3410 /* Node not bootstrapped yet */
3411 ptr = fallback_alloc(cachep, flags);
3415 if (nodeid == numa_node_id()) {
3417 * Use the locally cached objects if possible.
3418 * However ____cache_alloc does not allow fallback
3419 * to other nodes. It may fail while we still have
3420 * objects on other nodes available.
3422 ptr = ____cache_alloc(cachep, flags);
3426 /* ___cache_alloc_node can fall back to other nodes */
3427 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3429 local_irq_restore(save_flags);
3430 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3435 static __always_inline void *
3436 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3440 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3441 objp = alternate_node_alloc(cache, flags);
3445 objp = ____cache_alloc(cache, flags);
3448 * We may just have run out of memory on the local node.
3449 * ____cache_alloc_node() knows how to locate memory on other nodes
3452 objp = ____cache_alloc_node(cache, flags, numa_node_id());
3459 static __always_inline void *
3460 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3462 return ____cache_alloc(cachep, flags);
3465 #endif /* CONFIG_NUMA */
3467 static __always_inline void *
3468 __cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller)
3470 unsigned long save_flags;
3473 cache_alloc_debugcheck_before(cachep, flags);
3474 local_irq_save(save_flags);
3475 objp = __do_cache_alloc(cachep, flags);
3476 local_irq_restore(save_flags);
3477 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3484 * Caller needs to acquire correct kmem_list's list_lock
3486 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3490 struct kmem_list3 *l3;
3492 for (i = 0; i < nr_objects; i++) {
3493 void *objp = objpp[i];
3496 slabp = virt_to_slab(objp);
3497 l3 = cachep->nodelists[node];
3498 list_del(&slabp->list);
3499 check_spinlock_acquired_node(cachep, node);
3500 check_slabp(cachep, slabp);
3501 slab_put_obj(cachep, slabp, objp, node);
3502 STATS_DEC_ACTIVE(cachep);
3504 check_slabp(cachep, slabp);
3506 /* fixup slab chains */
3507 if (slabp->inuse == 0) {
3508 if (l3->free_objects > l3->free_limit) {
3509 l3->free_objects -= cachep->num;
3510 /* No need to drop any previously held
3511 * lock here, even if we have a off-slab slab
3512 * descriptor it is guaranteed to come from
3513 * a different cache, refer to comments before
3516 slab_destroy(cachep, slabp);
3518 list_add(&slabp->list, &l3->slabs_free);
3521 /* Unconditionally move a slab to the end of the
3522 * partial list on free - maximum time for the
3523 * other objects to be freed, too.
3525 list_add_tail(&slabp->list, &l3->slabs_partial);
3530 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3533 struct kmem_list3 *l3;
3534 int node = numa_node_id();
3536 batchcount = ac->batchcount;
3538 BUG_ON(!batchcount || batchcount > ac->avail);
3541 l3 = cachep->nodelists[node];
3542 spin_lock(&l3->list_lock);
3544 struct array_cache *shared_array = l3->shared;
3545 int max = shared_array->limit - shared_array->avail;
3547 if (batchcount > max)
3549 memcpy(&(shared_array->entry[shared_array->avail]),
3550 ac->entry, sizeof(void *) * batchcount);
3551 shared_array->avail += batchcount;
3556 free_block(cachep, ac->entry, batchcount, node);
3561 struct list_head *p;
3563 p = l3->slabs_free.next;
3564 while (p != &(l3->slabs_free)) {
3567 slabp = list_entry(p, struct slab, list);
3568 BUG_ON(slabp->inuse);
3573 STATS_SET_FREEABLE(cachep, i);
3576 spin_unlock(&l3->list_lock);
3577 ac->avail -= batchcount;
3578 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3582 * Release an obj back to its cache. If the obj has a constructed state, it must
3583 * be in this state _before_ it is released. Called with disabled ints.
3585 static inline void __cache_free(struct kmem_cache *cachep, void *objp)
3587 struct array_cache *ac = cpu_cache_get(cachep);
3590 objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0));
3592 if (use_alien_caches && cache_free_alien(cachep, objp))
3595 if (likely(ac->avail < ac->limit)) {
3596 STATS_INC_FREEHIT(cachep);
3597 ac->entry[ac->avail++] = objp;
3600 STATS_INC_FREEMISS(cachep);
3601 cache_flusharray(cachep, ac);
3602 ac->entry[ac->avail++] = objp;
3607 * kmem_cache_alloc - Allocate an object
3608 * @cachep: The cache to allocate from.
3609 * @flags: See kmalloc().
3611 * Allocate an object from this cache. The flags are only relevant
3612 * if the cache has no available objects.
3614 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3616 return __cache_alloc(cachep, flags, __builtin_return_address(0));
3618 EXPORT_SYMBOL(kmem_cache_alloc);
3621 * kmem_cache_zalloc - Allocate an object. The memory is set to zero.
3622 * @cache: The cache to allocate from.
3623 * @flags: See kmalloc().
3625 * Allocate an object from this cache and set the allocated memory to zero.
3626 * The flags are only relevant if the cache has no available objects.
3628 void *kmem_cache_zalloc(struct kmem_cache *cache, gfp_t flags)
3630 void *ret = __cache_alloc(cache, flags, __builtin_return_address(0));
3632 memset(ret, 0, obj_size(cache));
3635 EXPORT_SYMBOL(kmem_cache_zalloc);
3638 * kmem_ptr_validate - check if an untrusted pointer might
3640 * @cachep: the cache we're checking against
3641 * @ptr: pointer to validate
3643 * This verifies that the untrusted pointer looks sane:
3644 * it is _not_ a guarantee that the pointer is actually
3645 * part of the slab cache in question, but it at least
3646 * validates that the pointer can be dereferenced and
3647 * looks half-way sane.
3649 * Currently only used for dentry validation.
3651 int kmem_ptr_validate(struct kmem_cache *cachep, const void *ptr)
3653 unsigned long addr = (unsigned long)ptr;
3654 unsigned long min_addr = PAGE_OFFSET;
3655 unsigned long align_mask = BYTES_PER_WORD - 1;
3656 unsigned long size = cachep->buffer_size;
3659 if (unlikely(addr < min_addr))
3661 if (unlikely(addr > (unsigned long)high_memory - size))
3663 if (unlikely(addr & align_mask))
3665 if (unlikely(!kern_addr_valid(addr)))
3667 if (unlikely(!kern_addr_valid(addr + size - 1)))
3669 page = virt_to_page(ptr);
3670 if (unlikely(!PageSlab(page)))
3672 if (unlikely(page_get_cache(page) != cachep))
3680 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3682 return __cache_alloc_node(cachep, flags, nodeid,
3683 __builtin_return_address(0));
3685 EXPORT_SYMBOL(kmem_cache_alloc_node);
3687 static __always_inline void *
3688 __do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller)
3690 struct kmem_cache *cachep;
3692 cachep = kmem_find_general_cachep(size, flags);
3693 if (unlikely(cachep == NULL))
3695 return kmem_cache_alloc_node(cachep, flags, node);
3698 #ifdef CONFIG_DEBUG_SLAB
3699 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3701 return __do_kmalloc_node(size, flags, node,
3702 __builtin_return_address(0));
3704 EXPORT_SYMBOL(__kmalloc_node);
3706 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3707 int node, void *caller)
3709 return __do_kmalloc_node(size, flags, node, caller);
3711 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3713 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3715 return __do_kmalloc_node(size, flags, node, NULL);
3717 EXPORT_SYMBOL(__kmalloc_node);
3718 #endif /* CONFIG_DEBUG_SLAB */
3719 #endif /* CONFIG_NUMA */
3722 * __do_kmalloc - allocate memory
3723 * @size: how many bytes of memory are required.
3724 * @flags: the type of memory to allocate (see kmalloc).
3725 * @caller: function caller for debug tracking of the caller
3727 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3730 struct kmem_cache *cachep;
3732 /* If you want to save a few bytes .text space: replace
3734 * Then kmalloc uses the uninlined functions instead of the inline
3737 cachep = __find_general_cachep(size, flags);
3738 if (unlikely(cachep == NULL))
3740 return __cache_alloc(cachep, flags, caller);
3744 #ifdef CONFIG_DEBUG_SLAB
3745 void *__kmalloc(size_t size, gfp_t flags)
3747 return __do_kmalloc(size, flags, __builtin_return_address(0));
3749 EXPORT_SYMBOL(__kmalloc);
3751 void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller)
3753 return __do_kmalloc(size, flags, caller);
3755 EXPORT_SYMBOL(__kmalloc_track_caller);
3758 void *__kmalloc(size_t size, gfp_t flags)
3760 return __do_kmalloc(size, flags, NULL);
3762 EXPORT_SYMBOL(__kmalloc);
3766 * krealloc - reallocate memory. The contents will remain unchanged.
3768 * @p: object to reallocate memory for.
3769 * @new_size: how many bytes of memory are required.
3770 * @flags: the type of memory to allocate.
3772 * The contents of the object pointed to are preserved up to the
3773 * lesser of the new and old sizes. If @p is %NULL, krealloc()
3774 * behaves exactly like kmalloc(). If @size is 0 and @p is not a
3775 * %NULL pointer, the object pointed to is freed.
3777 void *krealloc(const void *p, size_t new_size, gfp_t flags)
3779 struct kmem_cache *cache, *new_cache;
3783 return kmalloc_track_caller(new_size, flags);
3785 if (unlikely(!new_size)) {
3790 cache = virt_to_cache(p);
3791 new_cache = __find_general_cachep(new_size, flags);
3794 * If new size fits in the current cache, bail out.
3796 if (likely(cache == new_cache))
3800 * We are on the slow-path here so do not use __cache_alloc
3801 * because it bloats kernel text.
3803 ret = kmalloc_track_caller(new_size, flags);
3805 memcpy(ret, p, min(new_size, ksize(p)));
3810 EXPORT_SYMBOL(krealloc);
3813 * kmem_cache_free - Deallocate an object
3814 * @cachep: The cache the allocation was from.
3815 * @objp: The previously allocated object.
3817 * Free an object which was previously allocated from this
3820 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3822 unsigned long flags;
3824 BUG_ON(virt_to_cache(objp) != cachep);
3826 local_irq_save(flags);
3827 debug_check_no_locks_freed(objp, obj_size(cachep));
3828 __cache_free(cachep, objp);
3829 local_irq_restore(flags);
3831 EXPORT_SYMBOL(kmem_cache_free);
3834 * kfree - free previously allocated memory
3835 * @objp: pointer returned by kmalloc.
3837 * If @objp is NULL, no operation is performed.
3839 * Don't free memory not originally allocated by kmalloc()
3840 * or you will run into trouble.
3842 void kfree(const void *objp)
3844 struct kmem_cache *c;
3845 unsigned long flags;
3847 if (unlikely(!objp))
3849 local_irq_save(flags);
3850 kfree_debugcheck(objp);
3851 c = virt_to_cache(objp);
3852 debug_check_no_locks_freed(objp, obj_size(c));
3853 __cache_free(c, (void *)objp);
3854 local_irq_restore(flags);
3856 EXPORT_SYMBOL(kfree);
3858 unsigned int kmem_cache_size(struct kmem_cache *cachep)
3860 return obj_size(cachep);
3862 EXPORT_SYMBOL(kmem_cache_size);
3864 const char *kmem_cache_name(struct kmem_cache *cachep)
3866 return cachep->name;
3868 EXPORT_SYMBOL_GPL(kmem_cache_name);
3871 * This initializes kmem_list3 or resizes varioius caches for all nodes.
3873 static int alloc_kmemlist(struct kmem_cache *cachep)
3876 struct kmem_list3 *l3;
3877 struct array_cache *new_shared;
3878 struct array_cache **new_alien = NULL;
3880 for_each_online_node(node) {
3882 if (use_alien_caches) {
3883 new_alien = alloc_alien_cache(node, cachep->limit);
3889 if (cachep->shared) {
3890 new_shared = alloc_arraycache(node,
3891 cachep->shared*cachep->batchcount,
3894 free_alien_cache(new_alien);
3899 l3 = cachep->nodelists[node];
3901 struct array_cache *shared = l3->shared;
3903 spin_lock_irq(&l3->list_lock);
3906 free_block(cachep, shared->entry,
3907 shared->avail, node);
3909 l3->shared = new_shared;
3911 l3->alien = new_alien;
3914 l3->free_limit = (1 + nr_cpus_node(node)) *
3915 cachep->batchcount + cachep->num;
3916 spin_unlock_irq(&l3->list_lock);
3918 free_alien_cache(new_alien);
3921 l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node);
3923 free_alien_cache(new_alien);
3928 kmem_list3_init(l3);
3929 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3930 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3931 l3->shared = new_shared;
3932 l3->alien = new_alien;
3933 l3->free_limit = (1 + nr_cpus_node(node)) *
3934 cachep->batchcount + cachep->num;
3935 cachep->nodelists[node] = l3;
3940 if (!cachep->next.next) {
3941 /* Cache is not active yet. Roll back what we did */
3944 if (cachep->nodelists[node]) {
3945 l3 = cachep->nodelists[node];
3948 free_alien_cache(l3->alien);
3950 cachep->nodelists[node] = NULL;
3958 struct ccupdate_struct {
3959 struct kmem_cache *cachep;
3960 struct array_cache *new[NR_CPUS];
3963 static void do_ccupdate_local(void *info)
3965 struct ccupdate_struct *new = info;
3966 struct array_cache *old;
3969 old = cpu_cache_get(new->cachep);
3971 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3972 new->new[smp_processor_id()] = old;
3975 /* Always called with the cache_chain_mutex held */
3976 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3977 int batchcount, int shared)
3979 struct ccupdate_struct *new;
3982 new = kzalloc(sizeof(*new), GFP_KERNEL);
3986 for_each_online_cpu(i) {
3987 new->new[i] = alloc_arraycache(cpu_to_node(i), limit,
3990 for (i--; i >= 0; i--)
3996 new->cachep = cachep;
3998 on_each_cpu(do_ccupdate_local, (void *)new, 1, 1);
4001 cachep->batchcount = batchcount;
4002 cachep->limit = limit;
4003 cachep->shared = shared;
4005 for_each_online_cpu(i) {
4006 struct array_cache *ccold = new->new[i];
4009 spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
4010 free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i));
4011 spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock);
4015 return alloc_kmemlist(cachep);
4018 /* Called with cache_chain_mutex held always */
4019 static int enable_cpucache(struct kmem_cache *cachep)
4025 * The head array serves three purposes:
4026 * - create a LIFO ordering, i.e. return objects that are cache-warm
4027 * - reduce the number of spinlock operations.
4028 * - reduce the number of linked list operations on the slab and
4029 * bufctl chains: array operations are cheaper.
4030 * The numbers are guessed, we should auto-tune as described by
4033 if (cachep->buffer_size > 131072)
4035 else if (cachep->buffer_size > PAGE_SIZE)
4037 else if (cachep->buffer_size > 1024)
4039 else if (cachep->buffer_size > 256)
4045 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
4046 * allocation behaviour: Most allocs on one cpu, most free operations
4047 * on another cpu. For these cases, an efficient object passing between
4048 * cpus is necessary. This is provided by a shared array. The array
4049 * replaces Bonwick's magazine layer.
4050 * On uniprocessor, it's functionally equivalent (but less efficient)
4051 * to a larger limit. Thus disabled by default.
4054 if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1)
4059 * With debugging enabled, large batchcount lead to excessively long
4060 * periods with disabled local interrupts. Limit the batchcount
4065 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared);
4067 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
4068 cachep->name, -err);
4073 * Drain an array if it contains any elements taking the l3 lock only if
4074 * necessary. Note that the l3 listlock also protects the array_cache
4075 * if drain_array() is used on the shared array.
4077 void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
4078 struct array_cache *ac, int force, int node)
4082 if (!ac || !ac->avail)
4084 if (ac->touched && !force) {
4087 spin_lock_irq(&l3->list_lock);
4089 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4090 if (tofree > ac->avail)
4091 tofree = (ac->avail + 1) / 2;
4092 free_block(cachep, ac->entry, tofree, node);
4093 ac->avail -= tofree;
4094 memmove(ac->entry, &(ac->entry[tofree]),
4095 sizeof(void *) * ac->avail);
4097 spin_unlock_irq(&l3->list_lock);
4102 * cache_reap - Reclaim memory from caches.
4103 * @w: work descriptor
4105 * Called from workqueue/eventd every few seconds.
4107 * - clear the per-cpu caches for this CPU.
4108 * - return freeable pages to the main free memory pool.
4110 * If we cannot acquire the cache chain mutex then just give up - we'll try
4111 * again on the next iteration.
4113 static void cache_reap(struct work_struct *w)
4115 struct kmem_cache *searchp;
4116 struct kmem_list3 *l3;
4117 int node = numa_node_id();
4118 struct delayed_work *work =
4119 container_of(w, struct delayed_work, work);
4121 if (!mutex_trylock(&cache_chain_mutex))
4122 /* Give up. Setup the next iteration. */
4125 list_for_each_entry(searchp, &cache_chain, next) {
4129 * We only take the l3 lock if absolutely necessary and we
4130 * have established with reasonable certainty that
4131 * we can do some work if the lock was obtained.
4133 l3 = searchp->nodelists[node];
4135 reap_alien(searchp, l3);
4137 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
4140 * These are racy checks but it does not matter
4141 * if we skip one check or scan twice.
4143 if (time_after(l3->next_reap, jiffies))
4146 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
4148 drain_array(searchp, l3, l3->shared, 0, node);
4150 if (l3->free_touched)
4151 l3->free_touched = 0;
4155 freed = drain_freelist(searchp, l3, (l3->free_limit +
4156 5 * searchp->num - 1) / (5 * searchp->num));
4157 STATS_ADD_REAPED(searchp, freed);
4163 mutex_unlock(&cache_chain_mutex);
4165 refresh_cpu_vm_stats(smp_processor_id());
4167 /* Set up the next iteration */
4168 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
4171 #ifdef CONFIG_PROC_FS
4173 static void print_slabinfo_header(struct seq_file *m)
4176 * Output format version, so at least we can change it
4177 * without _too_ many complaints.
4180 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4182 seq_puts(m, "slabinfo - version: 2.1\n");
4184 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4185 "<objperslab> <pagesperslab>");
4186 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4187 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4189 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
4190 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
4191 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4196 static void *s_start(struct seq_file *m, loff_t *pos)
4199 struct list_head *p;
4201 mutex_lock(&cache_chain_mutex);
4203 print_slabinfo_header(m);
4204 p = cache_chain.next;
4207 if (p == &cache_chain)
4210 return list_entry(p, struct kmem_cache, next);
4213 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4215 struct kmem_cache *cachep = p;
4217 return cachep->next.next == &cache_chain ?
4218 NULL : list_entry(cachep->next.next, struct kmem_cache, next);
4221 static void s_stop(struct seq_file *m, void *p)
4223 mutex_unlock(&cache_chain_mutex);
4226 static int s_show(struct seq_file *m, void *p)
4228 struct kmem_cache *cachep = p;
4230 unsigned long active_objs;
4231 unsigned long num_objs;
4232 unsigned long active_slabs = 0;
4233 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4237 struct kmem_list3 *l3;
4241 for_each_online_node(node) {
4242 l3 = cachep->nodelists[node];
4247 spin_lock_irq(&l3->list_lock);
4249 list_for_each_entry(slabp, &l3->slabs_full, list) {
4250 if (slabp->inuse != cachep->num && !error)
4251 error = "slabs_full accounting error";
4252 active_objs += cachep->num;
4255 list_for_each_entry(slabp, &l3->slabs_partial, list) {
4256 if (slabp->inuse == cachep->num && !error)
4257 error = "slabs_partial inuse accounting error";
4258 if (!slabp->inuse && !error)
4259 error = "slabs_partial/inuse accounting error";
4260 active_objs += slabp->inuse;
4263 list_for_each_entry(slabp, &l3->slabs_free, list) {
4264 if (slabp->inuse && !error)
4265 error = "slabs_free/inuse accounting error";
4268 free_objects += l3->free_objects;
4270 shared_avail += l3->shared->avail;
4272 spin_unlock_irq(&l3->list_lock);
4274 num_slabs += active_slabs;
4275 num_objs = num_slabs * cachep->num;
4276 if (num_objs - active_objs != free_objects && !error)
4277 error = "free_objects accounting error";
4279 name = cachep->name;
4281 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4283 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
4284 name, active_objs, num_objs, cachep->buffer_size,
4285 cachep->num, (1 << cachep->gfporder));
4286 seq_printf(m, " : tunables %4u %4u %4u",
4287 cachep->limit, cachep->batchcount, cachep->shared);
4288 seq_printf(m, " : slabdata %6lu %6lu %6lu",
4289 active_slabs, num_slabs, shared_avail);
4292 unsigned long high = cachep->high_mark;
4293 unsigned long allocs = cachep->num_allocations;
4294 unsigned long grown = cachep->grown;
4295 unsigned long reaped = cachep->reaped;
4296 unsigned long errors = cachep->errors;
4297 unsigned long max_freeable = cachep->max_freeable;
4298 unsigned long node_allocs = cachep->node_allocs;
4299 unsigned long node_frees = cachep->node_frees;
4300 unsigned long overflows = cachep->node_overflow;
4302 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \
4303 %4lu %4lu %4lu %4lu %4lu", allocs, high, grown,
4304 reaped, errors, max_freeable, node_allocs,
4305 node_frees, overflows);
4309 unsigned long allochit = atomic_read(&cachep->allochit);
4310 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4311 unsigned long freehit = atomic_read(&cachep->freehit);
4312 unsigned long freemiss = atomic_read(&cachep->freemiss);
4314 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4315 allochit, allocmiss, freehit, freemiss);
4323 * slabinfo_op - iterator that generates /proc/slabinfo
4332 * num-pages-per-slab
4333 * + further values on SMP and with statistics enabled
4336 const struct seq_operations slabinfo_op = {
4343 #define MAX_SLABINFO_WRITE 128
4345 * slabinfo_write - Tuning for the slab allocator
4347 * @buffer: user buffer
4348 * @count: data length
4351 ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4352 size_t count, loff_t *ppos)
4354 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4355 int limit, batchcount, shared, res;
4356 struct kmem_cache *cachep;
4358 if (count > MAX_SLABINFO_WRITE)
4360 if (copy_from_user(&kbuf, buffer, count))
4362 kbuf[MAX_SLABINFO_WRITE] = '\0';
4364 tmp = strchr(kbuf, ' ');
4369 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4372 /* Find the cache in the chain of caches. */
4373 mutex_lock(&cache_chain_mutex);
4375 list_for_each_entry(cachep, &cache_chain, next) {
4376 if (!strcmp(cachep->name, kbuf)) {
4377 if (limit < 1 || batchcount < 1 ||
4378 batchcount > limit || shared < 0) {
4381 res = do_tune_cpucache(cachep, limit,
4382 batchcount, shared);
4387 mutex_unlock(&cache_chain_mutex);
4393 #ifdef CONFIG_DEBUG_SLAB_LEAK
4395 static void *leaks_start(struct seq_file *m, loff_t *pos)
4398 struct list_head *p;
4400 mutex_lock(&cache_chain_mutex);
4401 p = cache_chain.next;
4404 if (p == &cache_chain)
4407 return list_entry(p, struct kmem_cache, next);
4410 static inline int add_caller(unsigned long *n, unsigned long v)
4420 unsigned long *q = p + 2 * i;
4434 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4440 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4446 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) {
4447 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4449 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4454 static void show_symbol(struct seq_file *m, unsigned long address)
4456 #ifdef CONFIG_KALLSYMS
4459 unsigned long offset, size;
4460 char namebuf[KSYM_NAME_LEN+1];
4462 name = kallsyms_lookup(address, &size, &offset, &modname, namebuf);
4465 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4467 seq_printf(m, " [%s]", modname);
4471 seq_printf(m, "%p", (void *)address);
4474 static int leaks_show(struct seq_file *m, void *p)
4476 struct kmem_cache *cachep = p;
4478 struct kmem_list3 *l3;
4480 unsigned long *n = m->private;
4484 if (!(cachep->flags & SLAB_STORE_USER))
4486 if (!(cachep->flags & SLAB_RED_ZONE))
4489 /* OK, we can do it */
4493 for_each_online_node(node) {
4494 l3 = cachep->nodelists[node];
4499 spin_lock_irq(&l3->list_lock);
4501 list_for_each_entry(slabp, &l3->slabs_full, list)
4502 handle_slab(n, cachep, slabp);
4503 list_for_each_entry(slabp, &l3->slabs_partial, list)
4504 handle_slab(n, cachep, slabp);
4505 spin_unlock_irq(&l3->list_lock);
4507 name = cachep->name;
4509 /* Increase the buffer size */
4510 mutex_unlock(&cache_chain_mutex);
4511 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4513 /* Too bad, we are really out */
4515 mutex_lock(&cache_chain_mutex);
4518 *(unsigned long *)m->private = n[0] * 2;
4520 mutex_lock(&cache_chain_mutex);
4521 /* Now make sure this entry will be retried */
4525 for (i = 0; i < n[1]; i++) {
4526 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4527 show_symbol(m, n[2*i+2]);
4534 const struct seq_operations slabstats_op = {
4535 .start = leaks_start,
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 if (unlikely(objp == NULL))
4560 return obj_size(virt_to_cache(objp));