3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
28 * This means, that your constructor is used only for newly allocated
29 * slabs and you must pass objects with the same initializations to
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 'slab_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>
92 #include <linux/poison.h>
93 #include <linux/swap.h>
94 #include <linux/cache.h>
95 #include <linux/interrupt.h>
96 #include <linux/init.h>
97 #include <linux/compiler.h>
98 #include <linux/cpuset.h>
99 #include <linux/proc_fs.h>
100 #include <linux/seq_file.h>
101 #include <linux/notifier.h>
102 #include <linux/kallsyms.h>
103 #include <linux/cpu.h>
104 #include <linux/sysctl.h>
105 #include <linux/module.h>
106 #include <linux/rcupdate.h>
107 #include <linux/string.h>
108 #include <linux/uaccess.h>
109 #include <linux/nodemask.h>
110 #include <linux/kmemleak.h>
111 #include <linux/mempolicy.h>
112 #include <linux/mutex.h>
113 #include <linux/fault-inject.h>
114 #include <linux/rtmutex.h>
115 #include <linux/reciprocal_div.h>
116 #include <linux/debugobjects.h>
117 #include <linux/kmemcheck.h>
118 #include <linux/memory.h>
119 #include <linux/prefetch.h>
121 #include <net/sock.h>
123 #include <asm/cacheflush.h>
124 #include <asm/tlbflush.h>
125 #include <asm/page.h>
127 #include <trace/events/kmem.h>
129 #include "internal.h"
132 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
133 * 0 for faster, smaller code (especially in the critical paths).
135 * STATS - 1 to collect stats for /proc/slabinfo.
136 * 0 for faster, smaller code (especially in the critical paths).
138 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
141 #ifdef CONFIG_DEBUG_SLAB
144 #define FORCED_DEBUG 1
148 #define FORCED_DEBUG 0
151 /* Shouldn't this be in a header file somewhere? */
152 #define BYTES_PER_WORD sizeof(void *)
153 #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
155 #ifndef ARCH_KMALLOC_FLAGS
156 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
160 * true if a page was allocated from pfmemalloc reserves for network-based
163 static bool pfmemalloc_active __read_mostly;
168 * Bufctl's are used for linking objs within a slab
171 * This implementation relies on "struct page" for locating the cache &
172 * slab an object belongs to.
173 * This allows the bufctl structure to be small (one int), but limits
174 * the number of objects a slab (not a cache) can contain when off-slab
175 * bufctls are used. The limit is the size of the largest general cache
176 * that does not use off-slab slabs.
177 * For 32bit archs with 4 kB pages, is this 56.
178 * This is not serious, as it is only for large objects, when it is unwise
179 * to have too many per slab.
180 * Note: This limit can be raised by introducing a general cache whose size
181 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
184 typedef unsigned int kmem_bufctl_t;
185 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
186 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
187 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
188 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
193 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
194 * arrange for kmem_freepages to be called via RCU. This is useful if
195 * we need to approach a kernel structure obliquely, from its address
196 * obtained without the usual locking. We can lock the structure to
197 * stabilize it and check it's still at the given address, only if we
198 * can be sure that the memory has not been meanwhile reused for some
199 * other kind of object (which our subsystem's lock might corrupt).
201 * rcu_read_lock before reading the address, then rcu_read_unlock after
202 * taking the spinlock within the structure expected at that address.
205 struct rcu_head head;
206 struct kmem_cache *cachep;
213 * Manages the objs in a slab. Placed either at the beginning of mem allocated
214 * for a slab, or allocated from an general cache.
215 * Slabs are chained into three list: fully used, partial, fully free slabs.
220 struct list_head list;
221 unsigned long colouroff;
222 void *s_mem; /* including colour offset */
223 unsigned int inuse; /* num of objs active in slab */
225 unsigned short nodeid;
227 struct slab_rcu __slab_cover_slab_rcu;
235 * - LIFO ordering, to hand out cache-warm objects from _alloc
236 * - reduce the number of linked list operations
237 * - reduce spinlock operations
239 * The limit is stored in the per-cpu structure to reduce the data cache
246 unsigned int batchcount;
247 unsigned int touched;
250 * Must have this definition in here for the proper
251 * alignment of array_cache. Also simplifies accessing
254 * Entries should not be directly dereferenced as
255 * entries belonging to slabs marked pfmemalloc will
256 * have the lower bits set SLAB_OBJ_PFMEMALLOC
260 #define SLAB_OBJ_PFMEMALLOC 1
261 static inline bool is_obj_pfmemalloc(void *objp)
263 return (unsigned long)objp & SLAB_OBJ_PFMEMALLOC;
266 static inline void set_obj_pfmemalloc(void **objp)
268 *objp = (void *)((unsigned long)*objp | SLAB_OBJ_PFMEMALLOC);
272 static inline void clear_obj_pfmemalloc(void **objp)
274 *objp = (void *)((unsigned long)*objp & ~SLAB_OBJ_PFMEMALLOC);
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 (3 * MAX_NUMNODES)
308 static struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
309 #define CACHE_CACHE 0
310 #define SIZE_AC MAX_NUMNODES
311 #define SIZE_L3 (2 * 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, gfp_t gfp);
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); \
375 #define CFLGS_OFF_SLAB (0x80000000UL)
376 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
378 #define BATCHREFILL_LIMIT 16
380 * Optimization question: fewer reaps means less probability for unnessary
381 * cpucache drain/refill cycles.
383 * OTOH the cpuarrays can contain lots of objects,
384 * which could lock up otherwise freeable slabs.
386 #define REAPTIMEOUT_CPUC (2*HZ)
387 #define REAPTIMEOUT_LIST3 (4*HZ)
390 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
391 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
392 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
393 #define STATS_INC_GROWN(x) ((x)->grown++)
394 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
395 #define STATS_SET_HIGH(x) \
397 if ((x)->num_active > (x)->high_mark) \
398 (x)->high_mark = (x)->num_active; \
400 #define STATS_INC_ERR(x) ((x)->errors++)
401 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
402 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
403 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
404 #define STATS_SET_FREEABLE(x, i) \
406 if ((x)->max_freeable < i) \
407 (x)->max_freeable = i; \
409 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
410 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
411 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
412 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
414 #define STATS_INC_ACTIVE(x) do { } while (0)
415 #define STATS_DEC_ACTIVE(x) do { } while (0)
416 #define STATS_INC_ALLOCED(x) do { } while (0)
417 #define STATS_INC_GROWN(x) do { } while (0)
418 #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
419 #define STATS_SET_HIGH(x) do { } while (0)
420 #define STATS_INC_ERR(x) do { } while (0)
421 #define STATS_INC_NODEALLOCS(x) do { } while (0)
422 #define STATS_INC_NODEFREES(x) do { } while (0)
423 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
424 #define STATS_SET_FREEABLE(x, i) do { } while (0)
425 #define STATS_INC_ALLOCHIT(x) do { } while (0)
426 #define STATS_INC_ALLOCMISS(x) do { } while (0)
427 #define STATS_INC_FREEHIT(x) do { } while (0)
428 #define STATS_INC_FREEMISS(x) do { } while (0)
434 * memory layout of objects:
436 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
437 * the end of an object is aligned with the end of the real
438 * allocation. Catches writes behind the end of the allocation.
439 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
441 * cachep->obj_offset: The real object.
442 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
443 * cachep->size - 1* BYTES_PER_WORD: last caller address
444 * [BYTES_PER_WORD long]
446 static int obj_offset(struct kmem_cache *cachep)
448 return cachep->obj_offset;
451 static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
453 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
454 return (unsigned long long*) (objp + obj_offset(cachep) -
455 sizeof(unsigned long long));
458 static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
460 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
461 if (cachep->flags & SLAB_STORE_USER)
462 return (unsigned long long *)(objp + cachep->size -
463 sizeof(unsigned long long) -
465 return (unsigned long long *) (objp + cachep->size -
466 sizeof(unsigned long long));
469 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
471 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
472 return (void **)(objp + cachep->size - BYTES_PER_WORD);
477 #define obj_offset(x) 0
478 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
479 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
480 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
485 * Do not go above this order unless 0 objects fit into the slab or
486 * overridden on the command line.
488 #define SLAB_MAX_ORDER_HI 1
489 #define SLAB_MAX_ORDER_LO 0
490 static int slab_max_order = SLAB_MAX_ORDER_LO;
491 static bool slab_max_order_set __initdata;
493 static inline struct kmem_cache *virt_to_cache(const void *obj)
495 struct page *page = virt_to_head_page(obj);
496 return page->slab_cache;
499 static inline struct slab *virt_to_slab(const void *obj)
501 struct page *page = virt_to_head_page(obj);
503 VM_BUG_ON(!PageSlab(page));
504 return page->slab_page;
507 static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
510 return slab->s_mem + cache->size * idx;
514 * We want to avoid an expensive divide : (offset / cache->size)
515 * Using the fact that size is a constant for a particular cache,
516 * we can replace (offset / cache->size) by
517 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
519 static inline unsigned int obj_to_index(const struct kmem_cache *cache,
520 const struct slab *slab, void *obj)
522 u32 offset = (obj - slab->s_mem);
523 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
527 * These are the default caches for kmalloc. Custom caches can have other sizes.
529 struct cache_sizes malloc_sizes[] = {
530 #define CACHE(x) { .cs_size = (x) },
531 #include <linux/kmalloc_sizes.h>
535 EXPORT_SYMBOL(malloc_sizes);
537 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
543 static struct cache_names __initdata cache_names[] = {
544 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
545 #include <linux/kmalloc_sizes.h>
550 static struct arraycache_init initarray_cache __initdata =
551 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
552 static struct arraycache_init initarray_generic =
553 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
555 /* internal cache of cache description objs */
556 static struct kmem_list3 *kmem_cache_nodelists[MAX_NUMNODES];
557 static struct kmem_cache kmem_cache_boot = {
558 .nodelists = kmem_cache_nodelists,
560 .limit = BOOT_CPUCACHE_ENTRIES,
562 .size = sizeof(struct kmem_cache),
563 .name = "kmem_cache",
566 #define BAD_ALIEN_MAGIC 0x01020304ul
568 #ifdef CONFIG_LOCKDEP
571 * Slab sometimes uses the kmalloc slabs to store the slab headers
572 * for other slabs "off slab".
573 * The locking for this is tricky in that it nests within the locks
574 * of all other slabs in a few places; to deal with this special
575 * locking we put on-slab caches into a separate lock-class.
577 * We set lock class for alien array caches which are up during init.
578 * The lock annotation will be lost if all cpus of a node goes down and
579 * then comes back up during hotplug
581 static struct lock_class_key on_slab_l3_key;
582 static struct lock_class_key on_slab_alc_key;
584 static struct lock_class_key debugobj_l3_key;
585 static struct lock_class_key debugobj_alc_key;
587 static void slab_set_lock_classes(struct kmem_cache *cachep,
588 struct lock_class_key *l3_key, struct lock_class_key *alc_key,
591 struct array_cache **alc;
592 struct kmem_list3 *l3;
595 l3 = cachep->nodelists[q];
599 lockdep_set_class(&l3->list_lock, l3_key);
602 * FIXME: This check for BAD_ALIEN_MAGIC
603 * should go away when common slab code is taught to
604 * work even without alien caches.
605 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
606 * for alloc_alien_cache,
608 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
612 lockdep_set_class(&alc[r]->lock, alc_key);
616 static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
618 slab_set_lock_classes(cachep, &debugobj_l3_key, &debugobj_alc_key, node);
621 static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
625 for_each_online_node(node)
626 slab_set_debugobj_lock_classes_node(cachep, node);
629 static void init_node_lock_keys(int q)
631 struct cache_sizes *s = malloc_sizes;
636 for (s = malloc_sizes; s->cs_size != ULONG_MAX; s++) {
637 struct kmem_list3 *l3;
639 l3 = s->cs_cachep->nodelists[q];
640 if (!l3 || OFF_SLAB(s->cs_cachep))
643 slab_set_lock_classes(s->cs_cachep, &on_slab_l3_key,
644 &on_slab_alc_key, q);
648 static inline void init_lock_keys(void)
653 init_node_lock_keys(node);
656 static void init_node_lock_keys(int q)
660 static inline void init_lock_keys(void)
664 static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
668 static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
673 static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
675 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
677 return cachep->array[smp_processor_id()];
680 static inline struct kmem_cache *__find_general_cachep(size_t size,
683 struct cache_sizes *csizep = malloc_sizes;
686 /* This happens if someone tries to call
687 * kmem_cache_create(), or __kmalloc(), before
688 * the generic caches are initialized.
690 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
693 return ZERO_SIZE_PTR;
695 while (size > csizep->cs_size)
699 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
700 * has cs_{dma,}cachep==NULL. Thus no special case
701 * for large kmalloc calls required.
703 #ifdef CONFIG_ZONE_DMA
704 if (unlikely(gfpflags & GFP_DMA))
705 return csizep->cs_dmacachep;
707 return csizep->cs_cachep;
710 static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
712 return __find_general_cachep(size, gfpflags);
715 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
717 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
721 * Calculate the number of objects and left-over bytes for a given buffer size.
723 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
724 size_t align, int flags, size_t *left_over,
729 size_t slab_size = PAGE_SIZE << gfporder;
732 * The slab management structure can be either off the slab or
733 * on it. For the latter case, the memory allocated for a
737 * - One kmem_bufctl_t for each object
738 * - Padding to respect alignment of @align
739 * - @buffer_size bytes for each object
741 * If the slab management structure is off the slab, then the
742 * alignment will already be calculated into the size. Because
743 * the slabs are all pages aligned, the objects will be at the
744 * correct alignment when allocated.
746 if (flags & CFLGS_OFF_SLAB) {
748 nr_objs = slab_size / buffer_size;
750 if (nr_objs > SLAB_LIMIT)
751 nr_objs = SLAB_LIMIT;
754 * Ignore padding for the initial guess. The padding
755 * is at most @align-1 bytes, and @buffer_size is at
756 * least @align. In the worst case, this result will
757 * be one greater than the number of objects that fit
758 * into the memory allocation when taking the padding
761 nr_objs = (slab_size - sizeof(struct slab)) /
762 (buffer_size + sizeof(kmem_bufctl_t));
765 * This calculated number will be either the right
766 * amount, or one greater than what we want.
768 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
772 if (nr_objs > SLAB_LIMIT)
773 nr_objs = SLAB_LIMIT;
775 mgmt_size = slab_mgmt_size(nr_objs, align);
778 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
782 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
784 static void __slab_error(const char *function, struct kmem_cache *cachep,
787 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
788 function, cachep->name, msg);
790 add_taint(TAINT_BAD_PAGE);
795 * By default on NUMA we use alien caches to stage the freeing of
796 * objects allocated from other nodes. This causes massive memory
797 * inefficiencies when using fake NUMA setup to split memory into a
798 * large number of small nodes, so it can be disabled on the command
802 static int use_alien_caches __read_mostly = 1;
803 static int __init noaliencache_setup(char *s)
805 use_alien_caches = 0;
808 __setup("noaliencache", noaliencache_setup);
810 static int __init slab_max_order_setup(char *str)
812 get_option(&str, &slab_max_order);
813 slab_max_order = slab_max_order < 0 ? 0 :
814 min(slab_max_order, MAX_ORDER - 1);
815 slab_max_order_set = true;
819 __setup("slab_max_order=", slab_max_order_setup);
823 * Special reaping functions for NUMA systems called from cache_reap().
824 * These take care of doing round robin flushing of alien caches (containing
825 * objects freed on different nodes from which they were allocated) and the
826 * flushing of remote pcps by calling drain_node_pages.
828 static DEFINE_PER_CPU(unsigned long, slab_reap_node);
830 static void init_reap_node(int cpu)
834 node = next_node(cpu_to_mem(cpu), node_online_map);
835 if (node == MAX_NUMNODES)
836 node = first_node(node_online_map);
838 per_cpu(slab_reap_node, cpu) = node;
841 static void next_reap_node(void)
843 int node = __this_cpu_read(slab_reap_node);
845 node = next_node(node, node_online_map);
846 if (unlikely(node >= MAX_NUMNODES))
847 node = first_node(node_online_map);
848 __this_cpu_write(slab_reap_node, node);
852 #define init_reap_node(cpu) do { } while (0)
853 #define next_reap_node(void) do { } while (0)
857 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
858 * via the workqueue/eventd.
859 * Add the CPU number into the expiration time to minimize the possibility of
860 * the CPUs getting into lockstep and contending for the global cache chain
863 static void __cpuinit start_cpu_timer(int cpu)
865 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
868 * When this gets called from do_initcalls via cpucache_init(),
869 * init_workqueues() has already run, so keventd will be setup
872 if (keventd_up() && reap_work->work.func == NULL) {
874 INIT_DEFERRABLE_WORK(reap_work, cache_reap);
875 schedule_delayed_work_on(cpu, reap_work,
876 __round_jiffies_relative(HZ, cpu));
880 static struct array_cache *alloc_arraycache(int node, int entries,
881 int batchcount, gfp_t gfp)
883 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
884 struct array_cache *nc = NULL;
886 nc = kmalloc_node(memsize, gfp, node);
888 * The array_cache structures contain pointers to free object.
889 * However, when such objects are allocated or transferred to another
890 * cache the pointers are not cleared and they could be counted as
891 * valid references during a kmemleak scan. Therefore, kmemleak must
892 * not scan such objects.
894 kmemleak_no_scan(nc);
898 nc->batchcount = batchcount;
900 spin_lock_init(&nc->lock);
905 static inline bool is_slab_pfmemalloc(struct slab *slabp)
907 struct page *page = virt_to_page(slabp->s_mem);
909 return PageSlabPfmemalloc(page);
912 /* Clears pfmemalloc_active if no slabs have pfmalloc set */
913 static void recheck_pfmemalloc_active(struct kmem_cache *cachep,
914 struct array_cache *ac)
916 struct kmem_list3 *l3 = cachep->nodelists[numa_mem_id()];
920 if (!pfmemalloc_active)
923 spin_lock_irqsave(&l3->list_lock, flags);
924 list_for_each_entry(slabp, &l3->slabs_full, list)
925 if (is_slab_pfmemalloc(slabp))
928 list_for_each_entry(slabp, &l3->slabs_partial, list)
929 if (is_slab_pfmemalloc(slabp))
932 list_for_each_entry(slabp, &l3->slabs_free, list)
933 if (is_slab_pfmemalloc(slabp))
936 pfmemalloc_active = false;
938 spin_unlock_irqrestore(&l3->list_lock, flags);
941 static void *__ac_get_obj(struct kmem_cache *cachep, struct array_cache *ac,
942 gfp_t flags, bool force_refill)
945 void *objp = ac->entry[--ac->avail];
947 /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
948 if (unlikely(is_obj_pfmemalloc(objp))) {
949 struct kmem_list3 *l3;
951 if (gfp_pfmemalloc_allowed(flags)) {
952 clear_obj_pfmemalloc(&objp);
956 /* The caller cannot use PFMEMALLOC objects, find another one */
957 for (i = 0; i < ac->avail; i++) {
958 /* If a !PFMEMALLOC object is found, swap them */
959 if (!is_obj_pfmemalloc(ac->entry[i])) {
961 ac->entry[i] = ac->entry[ac->avail];
962 ac->entry[ac->avail] = objp;
968 * If there are empty slabs on the slabs_free list and we are
969 * being forced to refill the cache, mark this one !pfmemalloc.
971 l3 = cachep->nodelists[numa_mem_id()];
972 if (!list_empty(&l3->slabs_free) && force_refill) {
973 struct slab *slabp = virt_to_slab(objp);
974 ClearPageSlabPfmemalloc(virt_to_head_page(slabp->s_mem));
975 clear_obj_pfmemalloc(&objp);
976 recheck_pfmemalloc_active(cachep, ac);
980 /* No !PFMEMALLOC objects available */
988 static inline void *ac_get_obj(struct kmem_cache *cachep,
989 struct array_cache *ac, gfp_t flags, bool force_refill)
993 if (unlikely(sk_memalloc_socks()))
994 objp = __ac_get_obj(cachep, ac, flags, force_refill);
996 objp = ac->entry[--ac->avail];
1001 static void *__ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
1004 if (unlikely(pfmemalloc_active)) {
1005 /* Some pfmemalloc slabs exist, check if this is one */
1006 struct page *page = virt_to_head_page(objp);
1007 if (PageSlabPfmemalloc(page))
1008 set_obj_pfmemalloc(&objp);
1014 static inline void ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
1017 if (unlikely(sk_memalloc_socks()))
1018 objp = __ac_put_obj(cachep, ac, objp);
1020 ac->entry[ac->avail++] = objp;
1024 * Transfer objects in one arraycache to another.
1025 * Locking must be handled by the caller.
1027 * Return the number of entries transferred.
1029 static int transfer_objects(struct array_cache *to,
1030 struct array_cache *from, unsigned int max)
1032 /* Figure out how many entries to transfer */
1033 int nr = min3(from->avail, max, to->limit - to->avail);
1038 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
1039 sizeof(void *) *nr);
1048 #define drain_alien_cache(cachep, alien) do { } while (0)
1049 #define reap_alien(cachep, l3) do { } while (0)
1051 static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
1053 return (struct array_cache **)BAD_ALIEN_MAGIC;
1056 static inline void free_alien_cache(struct array_cache **ac_ptr)
1060 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1065 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
1071 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
1072 gfp_t flags, int nodeid)
1077 #else /* CONFIG_NUMA */
1079 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
1080 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
1082 static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
1084 struct array_cache **ac_ptr;
1085 int memsize = sizeof(void *) * nr_node_ids;
1090 ac_ptr = kzalloc_node(memsize, gfp, node);
1093 if (i == node || !node_online(i))
1095 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
1097 for (i--; i >= 0; i--)
1107 static void free_alien_cache(struct array_cache **ac_ptr)
1118 static void __drain_alien_cache(struct kmem_cache *cachep,
1119 struct array_cache *ac, int node)
1121 struct kmem_list3 *rl3 = cachep->nodelists[node];
1124 spin_lock(&rl3->list_lock);
1126 * Stuff objects into the remote nodes shared array first.
1127 * That way we could avoid the overhead of putting the objects
1128 * into the free lists and getting them back later.
1131 transfer_objects(rl3->shared, ac, ac->limit);
1133 free_block(cachep, ac->entry, ac->avail, node);
1135 spin_unlock(&rl3->list_lock);
1140 * Called from cache_reap() to regularly drain alien caches round robin.
1142 static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1144 int node = __this_cpu_read(slab_reap_node);
1147 struct array_cache *ac = l3->alien[node];
1149 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1150 __drain_alien_cache(cachep, ac, node);
1151 spin_unlock_irq(&ac->lock);
1156 static void drain_alien_cache(struct kmem_cache *cachep,
1157 struct array_cache **alien)
1160 struct array_cache *ac;
1161 unsigned long flags;
1163 for_each_online_node(i) {
1166 spin_lock_irqsave(&ac->lock, flags);
1167 __drain_alien_cache(cachep, ac, i);
1168 spin_unlock_irqrestore(&ac->lock, flags);
1173 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1175 struct slab *slabp = virt_to_slab(objp);
1176 int nodeid = slabp->nodeid;
1177 struct kmem_list3 *l3;
1178 struct array_cache *alien = NULL;
1181 node = numa_mem_id();
1184 * Make sure we are not freeing a object from another node to the array
1185 * cache on this cpu.
1187 if (likely(slabp->nodeid == node))
1190 l3 = cachep->nodelists[node];
1191 STATS_INC_NODEFREES(cachep);
1192 if (l3->alien && l3->alien[nodeid]) {
1193 alien = l3->alien[nodeid];
1194 spin_lock(&alien->lock);
1195 if (unlikely(alien->avail == alien->limit)) {
1196 STATS_INC_ACOVERFLOW(cachep);
1197 __drain_alien_cache(cachep, alien, nodeid);
1199 ac_put_obj(cachep, alien, objp);
1200 spin_unlock(&alien->lock);
1202 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1203 free_block(cachep, &objp, 1, nodeid);
1204 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1211 * Allocates and initializes nodelists for a node on each slab cache, used for
1212 * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3
1213 * will be allocated off-node since memory is not yet online for the new node.
1214 * When hotplugging memory or a cpu, existing nodelists are not replaced if
1217 * Must hold slab_mutex.
1219 static int init_cache_nodelists_node(int node)
1221 struct kmem_cache *cachep;
1222 struct kmem_list3 *l3;
1223 const int memsize = sizeof(struct kmem_list3);
1225 list_for_each_entry(cachep, &slab_caches, list) {
1227 * Set up the size64 kmemlist for cpu before we can
1228 * begin anything. Make sure some other cpu on this
1229 * node has not already allocated this
1231 if (!cachep->nodelists[node]) {
1232 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1235 kmem_list3_init(l3);
1236 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1237 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1240 * The l3s don't come and go as CPUs come and
1241 * go. slab_mutex is sufficient
1244 cachep->nodelists[node] = l3;
1247 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1248 cachep->nodelists[node]->free_limit =
1249 (1 + nr_cpus_node(node)) *
1250 cachep->batchcount + cachep->num;
1251 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1256 static void __cpuinit cpuup_canceled(long cpu)
1258 struct kmem_cache *cachep;
1259 struct kmem_list3 *l3 = NULL;
1260 int node = cpu_to_mem(cpu);
1261 const struct cpumask *mask = cpumask_of_node(node);
1263 list_for_each_entry(cachep, &slab_caches, list) {
1264 struct array_cache *nc;
1265 struct array_cache *shared;
1266 struct array_cache **alien;
1268 /* cpu is dead; no one can alloc from it. */
1269 nc = cachep->array[cpu];
1270 cachep->array[cpu] = NULL;
1271 l3 = cachep->nodelists[node];
1274 goto free_array_cache;
1276 spin_lock_irq(&l3->list_lock);
1278 /* Free limit for this kmem_list3 */
1279 l3->free_limit -= cachep->batchcount;
1281 free_block(cachep, nc->entry, nc->avail, node);
1283 if (!cpumask_empty(mask)) {
1284 spin_unlock_irq(&l3->list_lock);
1285 goto free_array_cache;
1288 shared = l3->shared;
1290 free_block(cachep, shared->entry,
1291 shared->avail, node);
1298 spin_unlock_irq(&l3->list_lock);
1302 drain_alien_cache(cachep, alien);
1303 free_alien_cache(alien);
1309 * In the previous loop, all the objects were freed to
1310 * the respective cache's slabs, now we can go ahead and
1311 * shrink each nodelist to its limit.
1313 list_for_each_entry(cachep, &slab_caches, list) {
1314 l3 = cachep->nodelists[node];
1317 drain_freelist(cachep, l3, l3->free_objects);
1321 static int __cpuinit cpuup_prepare(long cpu)
1323 struct kmem_cache *cachep;
1324 struct kmem_list3 *l3 = NULL;
1325 int node = cpu_to_mem(cpu);
1329 * We need to do this right in the beginning since
1330 * alloc_arraycache's are going to use this list.
1331 * kmalloc_node allows us to add the slab to the right
1332 * kmem_list3 and not this cpu's kmem_list3
1334 err = init_cache_nodelists_node(node);
1339 * Now we can go ahead with allocating the shared arrays and
1342 list_for_each_entry(cachep, &slab_caches, list) {
1343 struct array_cache *nc;
1344 struct array_cache *shared = NULL;
1345 struct array_cache **alien = NULL;
1347 nc = alloc_arraycache(node, cachep->limit,
1348 cachep->batchcount, GFP_KERNEL);
1351 if (cachep->shared) {
1352 shared = alloc_arraycache(node,
1353 cachep->shared * cachep->batchcount,
1354 0xbaadf00d, GFP_KERNEL);
1360 if (use_alien_caches) {
1361 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
1368 cachep->array[cpu] = nc;
1369 l3 = cachep->nodelists[node];
1372 spin_lock_irq(&l3->list_lock);
1375 * We are serialised from CPU_DEAD or
1376 * CPU_UP_CANCELLED by the cpucontrol lock
1378 l3->shared = shared;
1387 spin_unlock_irq(&l3->list_lock);
1389 free_alien_cache(alien);
1390 if (cachep->flags & SLAB_DEBUG_OBJECTS)
1391 slab_set_debugobj_lock_classes_node(cachep, node);
1393 init_node_lock_keys(node);
1397 cpuup_canceled(cpu);
1401 static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1402 unsigned long action, void *hcpu)
1404 long cpu = (long)hcpu;
1408 case CPU_UP_PREPARE:
1409 case CPU_UP_PREPARE_FROZEN:
1410 mutex_lock(&slab_mutex);
1411 err = cpuup_prepare(cpu);
1412 mutex_unlock(&slab_mutex);
1415 case CPU_ONLINE_FROZEN:
1416 start_cpu_timer(cpu);
1418 #ifdef CONFIG_HOTPLUG_CPU
1419 case CPU_DOWN_PREPARE:
1420 case CPU_DOWN_PREPARE_FROZEN:
1422 * Shutdown cache reaper. Note that the slab_mutex is
1423 * held so that if cache_reap() is invoked it cannot do
1424 * anything expensive but will only modify reap_work
1425 * and reschedule the timer.
1427 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1428 /* Now the cache_reaper is guaranteed to be not running. */
1429 per_cpu(slab_reap_work, cpu).work.func = NULL;
1431 case CPU_DOWN_FAILED:
1432 case CPU_DOWN_FAILED_FROZEN:
1433 start_cpu_timer(cpu);
1436 case CPU_DEAD_FROZEN:
1438 * Even if all the cpus of a node are down, we don't free the
1439 * kmem_list3 of any cache. This to avoid a race between
1440 * cpu_down, and a kmalloc allocation from another cpu for
1441 * memory from the node of the cpu going down. The list3
1442 * structure is usually allocated from kmem_cache_create() and
1443 * gets destroyed at kmem_cache_destroy().
1447 case CPU_UP_CANCELED:
1448 case CPU_UP_CANCELED_FROZEN:
1449 mutex_lock(&slab_mutex);
1450 cpuup_canceled(cpu);
1451 mutex_unlock(&slab_mutex);
1454 return notifier_from_errno(err);
1457 static struct notifier_block __cpuinitdata cpucache_notifier = {
1458 &cpuup_callback, NULL, 0
1461 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1463 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1464 * Returns -EBUSY if all objects cannot be drained so that the node is not
1467 * Must hold slab_mutex.
1469 static int __meminit drain_cache_nodelists_node(int node)
1471 struct kmem_cache *cachep;
1474 list_for_each_entry(cachep, &slab_caches, list) {
1475 struct kmem_list3 *l3;
1477 l3 = cachep->nodelists[node];
1481 drain_freelist(cachep, l3, l3->free_objects);
1483 if (!list_empty(&l3->slabs_full) ||
1484 !list_empty(&l3->slabs_partial)) {
1492 static int __meminit slab_memory_callback(struct notifier_block *self,
1493 unsigned long action, void *arg)
1495 struct memory_notify *mnb = arg;
1499 nid = mnb->status_change_nid;
1504 case MEM_GOING_ONLINE:
1505 mutex_lock(&slab_mutex);
1506 ret = init_cache_nodelists_node(nid);
1507 mutex_unlock(&slab_mutex);
1509 case MEM_GOING_OFFLINE:
1510 mutex_lock(&slab_mutex);
1511 ret = drain_cache_nodelists_node(nid);
1512 mutex_unlock(&slab_mutex);
1516 case MEM_CANCEL_ONLINE:
1517 case MEM_CANCEL_OFFLINE:
1521 return notifier_from_errno(ret);
1523 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1526 * swap the static kmem_list3 with kmalloced memory
1528 static void __init init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1531 struct kmem_list3 *ptr;
1533 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid);
1536 memcpy(ptr, list, sizeof(struct kmem_list3));
1538 * Do not assume that spinlocks can be initialized via memcpy:
1540 spin_lock_init(&ptr->list_lock);
1542 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1543 cachep->nodelists[nodeid] = ptr;
1547 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1548 * size of kmem_list3.
1550 static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1554 for_each_online_node(node) {
1555 cachep->nodelists[node] = &initkmem_list3[index + node];
1556 cachep->nodelists[node]->next_reap = jiffies +
1558 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1563 * Initialisation. Called after the page allocator have been initialised and
1564 * before smp_init().
1566 void __init kmem_cache_init(void)
1569 struct cache_sizes *sizes;
1570 struct cache_names *names;
1575 kmem_cache = &kmem_cache_boot;
1577 if (num_possible_nodes() == 1)
1578 use_alien_caches = 0;
1580 for (i = 0; i < NUM_INIT_LISTS; i++) {
1581 kmem_list3_init(&initkmem_list3[i]);
1582 if (i < MAX_NUMNODES)
1583 kmem_cache->nodelists[i] = NULL;
1585 set_up_list3s(kmem_cache, CACHE_CACHE);
1588 * Fragmentation resistance on low memory - only use bigger
1589 * page orders on machines with more than 32MB of memory if
1590 * not overridden on the command line.
1592 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
1593 slab_max_order = SLAB_MAX_ORDER_HI;
1595 /* Bootstrap is tricky, because several objects are allocated
1596 * from caches that do not exist yet:
1597 * 1) initialize the kmem_cache cache: it contains the struct
1598 * kmem_cache structures of all caches, except kmem_cache itself:
1599 * kmem_cache is statically allocated.
1600 * Initially an __init data area is used for the head array and the
1601 * kmem_list3 structures, it's replaced with a kmalloc allocated
1602 * array at the end of the bootstrap.
1603 * 2) Create the first kmalloc cache.
1604 * The struct kmem_cache for the new cache is allocated normally.
1605 * An __init data area is used for the head array.
1606 * 3) Create the remaining kmalloc caches, with minimally sized
1608 * 4) Replace the __init data head arrays for kmem_cache and the first
1609 * kmalloc cache with kmalloc allocated arrays.
1610 * 5) Replace the __init data for kmem_list3 for kmem_cache and
1611 * the other cache's with kmalloc allocated memory.
1612 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1615 node = numa_mem_id();
1617 /* 1) create the kmem_cache */
1618 INIT_LIST_HEAD(&slab_caches);
1619 list_add(&kmem_cache->list, &slab_caches);
1620 kmem_cache->colour_off = cache_line_size();
1621 kmem_cache->array[smp_processor_id()] = &initarray_cache.cache;
1622 kmem_cache->nodelists[node] = &initkmem_list3[CACHE_CACHE + node];
1625 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1627 kmem_cache->size = offsetof(struct kmem_cache, array[nr_cpu_ids]) +
1628 nr_node_ids * sizeof(struct kmem_list3 *);
1629 kmem_cache->object_size = kmem_cache->size;
1630 kmem_cache->size = ALIGN(kmem_cache->object_size,
1632 kmem_cache->reciprocal_buffer_size =
1633 reciprocal_value(kmem_cache->size);
1635 for (order = 0; order < MAX_ORDER; order++) {
1636 cache_estimate(order, kmem_cache->size,
1637 cache_line_size(), 0, &left_over, &kmem_cache->num);
1638 if (kmem_cache->num)
1641 BUG_ON(!kmem_cache->num);
1642 kmem_cache->gfporder = order;
1643 kmem_cache->colour = left_over / kmem_cache->colour_off;
1644 kmem_cache->slab_size = ALIGN(kmem_cache->num * sizeof(kmem_bufctl_t) +
1645 sizeof(struct slab), cache_line_size());
1647 /* 2+3) create the kmalloc caches */
1648 sizes = malloc_sizes;
1649 names = cache_names;
1652 * Initialize the caches that provide memory for the array cache and the
1653 * kmem_list3 structures first. Without this, further allocations will
1657 sizes[INDEX_AC].cs_cachep = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
1658 sizes[INDEX_AC].cs_cachep->name = names[INDEX_AC].name;
1659 sizes[INDEX_AC].cs_cachep->size = sizes[INDEX_AC].cs_size;
1660 sizes[INDEX_AC].cs_cachep->object_size = sizes[INDEX_AC].cs_size;
1661 sizes[INDEX_AC].cs_cachep->align = ARCH_KMALLOC_MINALIGN;
1662 __kmem_cache_create(sizes[INDEX_AC].cs_cachep, ARCH_KMALLOC_FLAGS|SLAB_PANIC);
1663 list_add(&sizes[INDEX_AC].cs_cachep->list, &slab_caches);
1665 if (INDEX_AC != INDEX_L3) {
1666 sizes[INDEX_L3].cs_cachep = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
1667 sizes[INDEX_L3].cs_cachep->name = names[INDEX_L3].name;
1668 sizes[INDEX_L3].cs_cachep->size = sizes[INDEX_L3].cs_size;
1669 sizes[INDEX_L3].cs_cachep->object_size = sizes[INDEX_L3].cs_size;
1670 sizes[INDEX_L3].cs_cachep->align = ARCH_KMALLOC_MINALIGN;
1671 __kmem_cache_create(sizes[INDEX_L3].cs_cachep, ARCH_KMALLOC_FLAGS|SLAB_PANIC);
1672 list_add(&sizes[INDEX_L3].cs_cachep->list, &slab_caches);
1675 slab_early_init = 0;
1677 while (sizes->cs_size != ULONG_MAX) {
1679 * For performance, all the general caches are L1 aligned.
1680 * This should be particularly beneficial on SMP boxes, as it
1681 * eliminates "false sharing".
1682 * Note for systems short on memory removing the alignment will
1683 * allow tighter packing of the smaller caches.
1685 if (!sizes->cs_cachep) {
1686 sizes->cs_cachep = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
1687 sizes->cs_cachep->name = names->name;
1688 sizes->cs_cachep->size = sizes->cs_size;
1689 sizes->cs_cachep->object_size = sizes->cs_size;
1690 sizes->cs_cachep->align = ARCH_KMALLOC_MINALIGN;
1691 __kmem_cache_create(sizes->cs_cachep, ARCH_KMALLOC_FLAGS|SLAB_PANIC);
1692 list_add(&sizes->cs_cachep->list, &slab_caches);
1694 #ifdef CONFIG_ZONE_DMA
1695 sizes->cs_dmacachep = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
1696 sizes->cs_dmacachep->name = names->name_dma;
1697 sizes->cs_dmacachep->size = sizes->cs_size;
1698 sizes->cs_dmacachep->object_size = sizes->cs_size;
1699 sizes->cs_dmacachep->align = ARCH_KMALLOC_MINALIGN;
1700 __kmem_cache_create(sizes->cs_dmacachep,
1701 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA| SLAB_PANIC);
1702 list_add(&sizes->cs_dmacachep->list, &slab_caches);
1707 /* 4) Replace the bootstrap head arrays */
1709 struct array_cache *ptr;
1711 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1713 BUG_ON(cpu_cache_get(kmem_cache) != &initarray_cache.cache);
1714 memcpy(ptr, cpu_cache_get(kmem_cache),
1715 sizeof(struct arraycache_init));
1717 * Do not assume that spinlocks can be initialized via memcpy:
1719 spin_lock_init(&ptr->lock);
1721 kmem_cache->array[smp_processor_id()] = ptr;
1723 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1725 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1726 != &initarray_generic.cache);
1727 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1728 sizeof(struct arraycache_init));
1730 * Do not assume that spinlocks can be initialized via memcpy:
1732 spin_lock_init(&ptr->lock);
1734 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1737 /* 5) Replace the bootstrap kmem_list3's */
1741 for_each_online_node(nid) {
1742 init_list(kmem_cache, &initkmem_list3[CACHE_CACHE + nid], nid);
1744 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1745 &initkmem_list3[SIZE_AC + nid], nid);
1747 if (INDEX_AC != INDEX_L3) {
1748 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1749 &initkmem_list3[SIZE_L3 + nid], nid);
1757 void __init kmem_cache_init_late(void)
1759 struct kmem_cache *cachep;
1763 /* 6) resize the head arrays to their final sizes */
1764 mutex_lock(&slab_mutex);
1765 list_for_each_entry(cachep, &slab_caches, list)
1766 if (enable_cpucache(cachep, GFP_NOWAIT))
1768 mutex_unlock(&slab_mutex);
1770 /* Annotate slab for lockdep -- annotate the malloc caches */
1777 * Register a cpu startup notifier callback that initializes
1778 * cpu_cache_get for all new cpus
1780 register_cpu_notifier(&cpucache_notifier);
1784 * Register a memory hotplug callback that initializes and frees
1787 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1791 * The reap timers are started later, with a module init call: That part
1792 * of the kernel is not yet operational.
1796 static int __init cpucache_init(void)
1801 * Register the timers that return unneeded pages to the page allocator
1803 for_each_online_cpu(cpu)
1804 start_cpu_timer(cpu);
1810 __initcall(cpucache_init);
1812 static noinline void
1813 slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1815 struct kmem_list3 *l3;
1817 unsigned long flags;
1821 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1823 printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n",
1824 cachep->name, cachep->size, cachep->gfporder);
1826 for_each_online_node(node) {
1827 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1828 unsigned long active_slabs = 0, num_slabs = 0;
1830 l3 = cachep->nodelists[node];
1834 spin_lock_irqsave(&l3->list_lock, flags);
1835 list_for_each_entry(slabp, &l3->slabs_full, list) {
1836 active_objs += cachep->num;
1839 list_for_each_entry(slabp, &l3->slabs_partial, list) {
1840 active_objs += slabp->inuse;
1843 list_for_each_entry(slabp, &l3->slabs_free, list)
1846 free_objects += l3->free_objects;
1847 spin_unlock_irqrestore(&l3->list_lock, flags);
1849 num_slabs += active_slabs;
1850 num_objs = num_slabs * cachep->num;
1852 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1853 node, active_slabs, num_slabs, active_objs, num_objs,
1859 * Interface to system's page allocator. No need to hold the cache-lock.
1861 * If we requested dmaable memory, we will get it. Even if we
1862 * did not request dmaable memory, we might get it, but that
1863 * would be relatively rare and ignorable.
1865 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1873 * Nommu uses slab's for process anonymous memory allocations, and thus
1874 * requires __GFP_COMP to properly refcount higher order allocations
1876 flags |= __GFP_COMP;
1879 flags |= cachep->allocflags;
1880 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1881 flags |= __GFP_RECLAIMABLE;
1883 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1885 if (!(flags & __GFP_NOWARN) && printk_ratelimit())
1886 slab_out_of_memory(cachep, flags, nodeid);
1890 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1891 if (unlikely(page->pfmemalloc))
1892 pfmemalloc_active = true;
1894 nr_pages = (1 << cachep->gfporder);
1895 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1896 add_zone_page_state(page_zone(page),
1897 NR_SLAB_RECLAIMABLE, nr_pages);
1899 add_zone_page_state(page_zone(page),
1900 NR_SLAB_UNRECLAIMABLE, nr_pages);
1901 for (i = 0; i < nr_pages; i++) {
1902 __SetPageSlab(page + i);
1904 if (page->pfmemalloc)
1905 SetPageSlabPfmemalloc(page + i);
1908 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1909 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1912 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1914 kmemcheck_mark_unallocated_pages(page, nr_pages);
1917 return page_address(page);
1921 * Interface to system's page release.
1923 static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1925 unsigned long i = (1 << cachep->gfporder);
1926 struct page *page = virt_to_page(addr);
1927 const unsigned long nr_freed = i;
1929 kmemcheck_free_shadow(page, cachep->gfporder);
1931 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1932 sub_zone_page_state(page_zone(page),
1933 NR_SLAB_RECLAIMABLE, nr_freed);
1935 sub_zone_page_state(page_zone(page),
1936 NR_SLAB_UNRECLAIMABLE, nr_freed);
1938 BUG_ON(!PageSlab(page));
1939 __ClearPageSlabPfmemalloc(page);
1940 __ClearPageSlab(page);
1943 if (current->reclaim_state)
1944 current->reclaim_state->reclaimed_slab += nr_freed;
1945 free_pages((unsigned long)addr, cachep->gfporder);
1948 static void kmem_rcu_free(struct rcu_head *head)
1950 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1951 struct kmem_cache *cachep = slab_rcu->cachep;
1953 kmem_freepages(cachep, slab_rcu->addr);
1954 if (OFF_SLAB(cachep))
1955 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1960 #ifdef CONFIG_DEBUG_PAGEALLOC
1961 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1962 unsigned long caller)
1964 int size = cachep->object_size;
1966 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1968 if (size < 5 * sizeof(unsigned long))
1971 *addr++ = 0x12345678;
1973 *addr++ = smp_processor_id();
1974 size -= 3 * sizeof(unsigned long);
1976 unsigned long *sptr = &caller;
1977 unsigned long svalue;
1979 while (!kstack_end(sptr)) {
1981 if (kernel_text_address(svalue)) {
1983 size -= sizeof(unsigned long);
1984 if (size <= sizeof(unsigned long))
1990 *addr++ = 0x87654321;
1994 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1996 int size = cachep->object_size;
1997 addr = &((char *)addr)[obj_offset(cachep)];
1999 memset(addr, val, size);
2000 *(unsigned char *)(addr + size - 1) = POISON_END;
2003 static void dump_line(char *data, int offset, int limit)
2006 unsigned char error = 0;
2009 printk(KERN_ERR "%03x: ", offset);
2010 for (i = 0; i < limit; i++) {
2011 if (data[offset + i] != POISON_FREE) {
2012 error = data[offset + i];
2016 print_hex_dump(KERN_CONT, "", 0, 16, 1,
2017 &data[offset], limit, 1);
2019 if (bad_count == 1) {
2020 error ^= POISON_FREE;
2021 if (!(error & (error - 1))) {
2022 printk(KERN_ERR "Single bit error detected. Probably "
2025 printk(KERN_ERR "Run memtest86+ or a similar memory "
2028 printk(KERN_ERR "Run a memory test tool.\n");
2037 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
2042 if (cachep->flags & SLAB_RED_ZONE) {
2043 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
2044 *dbg_redzone1(cachep, objp),
2045 *dbg_redzone2(cachep, objp));
2048 if (cachep->flags & SLAB_STORE_USER) {
2049 printk(KERN_ERR "Last user: [<%p>]",
2050 *dbg_userword(cachep, objp));
2051 print_symbol("(%s)",
2052 (unsigned long)*dbg_userword(cachep, objp));
2055 realobj = (char *)objp + obj_offset(cachep);
2056 size = cachep->object_size;
2057 for (i = 0; i < size && lines; i += 16, lines--) {
2060 if (i + limit > size)
2062 dump_line(realobj, i, limit);
2066 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
2072 realobj = (char *)objp + obj_offset(cachep);
2073 size = cachep->object_size;
2075 for (i = 0; i < size; i++) {
2076 char exp = POISON_FREE;
2079 if (realobj[i] != exp) {
2085 "Slab corruption (%s): %s start=%p, len=%d\n",
2086 print_tainted(), cachep->name, realobj, size);
2087 print_objinfo(cachep, objp, 0);
2089 /* Hexdump the affected line */
2092 if (i + limit > size)
2094 dump_line(realobj, i, limit);
2097 /* Limit to 5 lines */
2103 /* Print some data about the neighboring objects, if they
2106 struct slab *slabp = virt_to_slab(objp);
2109 objnr = obj_to_index(cachep, slabp, objp);
2111 objp = index_to_obj(cachep, slabp, objnr - 1);
2112 realobj = (char *)objp + obj_offset(cachep);
2113 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
2115 print_objinfo(cachep, objp, 2);
2117 if (objnr + 1 < cachep->num) {
2118 objp = index_to_obj(cachep, slabp, objnr + 1);
2119 realobj = (char *)objp + obj_offset(cachep);
2120 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
2122 print_objinfo(cachep, objp, 2);
2129 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
2132 for (i = 0; i < cachep->num; i++) {
2133 void *objp = index_to_obj(cachep, slabp, i);
2135 if (cachep->flags & SLAB_POISON) {
2136 #ifdef CONFIG_DEBUG_PAGEALLOC
2137 if (cachep->size % PAGE_SIZE == 0 &&
2139 kernel_map_pages(virt_to_page(objp),
2140 cachep->size / PAGE_SIZE, 1);
2142 check_poison_obj(cachep, objp);
2144 check_poison_obj(cachep, objp);
2147 if (cachep->flags & SLAB_RED_ZONE) {
2148 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2149 slab_error(cachep, "start of a freed object "
2151 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2152 slab_error(cachep, "end of a freed object "
2158 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
2164 * slab_destroy - destroy and release all objects in a slab
2165 * @cachep: cache pointer being destroyed
2166 * @slabp: slab pointer being destroyed
2168 * Destroy all the objs in a slab, and release the mem back to the system.
2169 * Before calling the slab must have been unlinked from the cache. The
2170 * cache-lock is not held/needed.
2172 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
2174 void *addr = slabp->s_mem - slabp->colouroff;
2176 slab_destroy_debugcheck(cachep, slabp);
2177 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
2178 struct slab_rcu *slab_rcu;
2180 slab_rcu = (struct slab_rcu *)slabp;
2181 slab_rcu->cachep = cachep;
2182 slab_rcu->addr = addr;
2183 call_rcu(&slab_rcu->head, kmem_rcu_free);
2185 kmem_freepages(cachep, addr);
2186 if (OFF_SLAB(cachep))
2187 kmem_cache_free(cachep->slabp_cache, slabp);
2192 * calculate_slab_order - calculate size (page order) of slabs
2193 * @cachep: pointer to the cache that is being created
2194 * @size: size of objects to be created in this cache.
2195 * @align: required alignment for the objects.
2196 * @flags: slab allocation flags
2198 * Also calculates the number of objects per slab.
2200 * This could be made much more intelligent. For now, try to avoid using
2201 * high order pages for slabs. When the gfp() functions are more friendly
2202 * towards high-order requests, this should be changed.
2204 static size_t calculate_slab_order(struct kmem_cache *cachep,
2205 size_t size, size_t align, unsigned long flags)
2207 unsigned long offslab_limit;
2208 size_t left_over = 0;
2211 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
2215 cache_estimate(gfporder, size, align, flags, &remainder, &num);
2219 if (flags & CFLGS_OFF_SLAB) {
2221 * Max number of objs-per-slab for caches which
2222 * use off-slab slabs. Needed to avoid a possible
2223 * looping condition in cache_grow().
2225 offslab_limit = size - sizeof(struct slab);
2226 offslab_limit /= sizeof(kmem_bufctl_t);
2228 if (num > offslab_limit)
2232 /* Found something acceptable - save it away */
2234 cachep->gfporder = gfporder;
2235 left_over = remainder;
2238 * A VFS-reclaimable slab tends to have most allocations
2239 * as GFP_NOFS and we really don't want to have to be allocating
2240 * higher-order pages when we are unable to shrink dcache.
2242 if (flags & SLAB_RECLAIM_ACCOUNT)
2246 * Large number of objects is good, but very large slabs are
2247 * currently bad for the gfp()s.
2249 if (gfporder >= slab_max_order)
2253 * Acceptable internal fragmentation?
2255 if (left_over * 8 <= (PAGE_SIZE << gfporder))
2261 static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
2263 if (slab_state >= FULL)
2264 return enable_cpucache(cachep, gfp);
2266 if (slab_state == DOWN) {
2268 * Note: the first kmem_cache_create must create the cache
2269 * that's used by kmalloc(24), otherwise the creation of
2270 * further caches will BUG().
2272 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2275 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2276 * the first cache, then we need to set up all its list3s,
2277 * otherwise the creation of further caches will BUG().
2279 set_up_list3s(cachep, SIZE_AC);
2280 if (INDEX_AC == INDEX_L3)
2281 slab_state = PARTIAL_L3;
2283 slab_state = PARTIAL_ARRAYCACHE;
2285 cachep->array[smp_processor_id()] =
2286 kmalloc(sizeof(struct arraycache_init), gfp);
2288 if (slab_state == PARTIAL_ARRAYCACHE) {
2289 set_up_list3s(cachep, SIZE_L3);
2290 slab_state = PARTIAL_L3;
2293 for_each_online_node(node) {
2294 cachep->nodelists[node] =
2295 kmalloc_node(sizeof(struct kmem_list3),
2297 BUG_ON(!cachep->nodelists[node]);
2298 kmem_list3_init(cachep->nodelists[node]);
2302 cachep->nodelists[numa_mem_id()]->next_reap =
2303 jiffies + REAPTIMEOUT_LIST3 +
2304 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2306 cpu_cache_get(cachep)->avail = 0;
2307 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2308 cpu_cache_get(cachep)->batchcount = 1;
2309 cpu_cache_get(cachep)->touched = 0;
2310 cachep->batchcount = 1;
2311 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2316 * __kmem_cache_create - Create a cache.
2317 * @name: A string which is used in /proc/slabinfo to identify this cache.
2318 * @size: The size of objects to be created in this cache.
2319 * @align: The required alignment for the objects.
2320 * @flags: SLAB flags
2321 * @ctor: A constructor for the objects.
2323 * Returns a ptr to the cache on success, NULL on failure.
2324 * Cannot be called within a int, but can be interrupted.
2325 * The @ctor is run when new pages are allocated by the cache.
2329 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2330 * to catch references to uninitialised memory.
2332 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2333 * for buffer overruns.
2335 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2336 * cacheline. This can be beneficial if you're counting cycles as closely
2340 __kmem_cache_create (struct kmem_cache *cachep, unsigned long flags)
2342 size_t left_over, slab_size, ralign;
2345 size_t size = cachep->size;
2350 * Enable redzoning and last user accounting, except for caches with
2351 * large objects, if the increased size would increase the object size
2352 * above the next power of two: caches with object sizes just above a
2353 * power of two have a significant amount of internal fragmentation.
2355 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2356 2 * sizeof(unsigned long long)))
2357 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2358 if (!(flags & SLAB_DESTROY_BY_RCU))
2359 flags |= SLAB_POISON;
2361 if (flags & SLAB_DESTROY_BY_RCU)
2362 BUG_ON(flags & SLAB_POISON);
2366 * Check that size is in terms of words. This is needed to avoid
2367 * unaligned accesses for some archs when redzoning is used, and makes
2368 * sure any on-slab bufctl's are also correctly aligned.
2370 if (size & (BYTES_PER_WORD - 1)) {
2371 size += (BYTES_PER_WORD - 1);
2372 size &= ~(BYTES_PER_WORD - 1);
2375 /* calculate the final buffer alignment: */
2377 /* 1) arch recommendation: can be overridden for debug */
2378 if (flags & SLAB_HWCACHE_ALIGN) {
2380 * Default alignment: as specified by the arch code. Except if
2381 * an object is really small, then squeeze multiple objects into
2384 ralign = cache_line_size();
2385 while (size <= ralign / 2)
2388 ralign = BYTES_PER_WORD;
2392 * Redzoning and user store require word alignment or possibly larger.
2393 * Note this will be overridden by architecture or caller mandated
2394 * alignment if either is greater than BYTES_PER_WORD.
2396 if (flags & SLAB_STORE_USER)
2397 ralign = BYTES_PER_WORD;
2399 if (flags & SLAB_RED_ZONE) {
2400 ralign = REDZONE_ALIGN;
2401 /* If redzoning, ensure that the second redzone is suitably
2402 * aligned, by adjusting the object size accordingly. */
2403 size += REDZONE_ALIGN - 1;
2404 size &= ~(REDZONE_ALIGN - 1);
2407 /* 2) arch mandated alignment */
2408 if (ralign < ARCH_SLAB_MINALIGN) {
2409 ralign = ARCH_SLAB_MINALIGN;
2411 /* 3) caller mandated alignment */
2412 if (ralign < cachep->align) {
2413 ralign = cachep->align;
2415 /* disable debug if necessary */
2416 if (ralign > __alignof__(unsigned long long))
2417 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2421 cachep->align = ralign;
2423 if (slab_is_available())
2428 cachep->nodelists = (struct kmem_list3 **)&cachep->array[nr_cpu_ids];
2432 * Both debugging options require word-alignment which is calculated
2435 if (flags & SLAB_RED_ZONE) {
2436 /* add space for red zone words */
2437 cachep->obj_offset += sizeof(unsigned long long);
2438 size += 2 * sizeof(unsigned long long);
2440 if (flags & SLAB_STORE_USER) {
2441 /* user store requires one word storage behind the end of
2442 * the real object. But if the second red zone needs to be
2443 * aligned to 64 bits, we must allow that much space.
2445 if (flags & SLAB_RED_ZONE)
2446 size += REDZONE_ALIGN;
2448 size += BYTES_PER_WORD;
2450 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2451 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
2452 && cachep->object_size > cache_line_size()
2453 && ALIGN(size, cachep->align) < PAGE_SIZE) {
2454 cachep->obj_offset += PAGE_SIZE - ALIGN(size, cachep->align);
2461 * Determine if the slab management is 'on' or 'off' slab.
2462 * (bootstrapping cannot cope with offslab caches so don't do
2463 * it too early on. Always use on-slab management when
2464 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2466 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
2467 !(flags & SLAB_NOLEAKTRACE))
2469 * Size is large, assume best to place the slab management obj
2470 * off-slab (should allow better packing of objs).
2472 flags |= CFLGS_OFF_SLAB;
2474 size = ALIGN(size, cachep->align);
2476 left_over = calculate_slab_order(cachep, size, cachep->align, flags);
2481 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2482 + sizeof(struct slab), cachep->align);
2485 * If the slab has been placed off-slab, and we have enough space then
2486 * move it on-slab. This is at the expense of any extra colouring.
2488 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2489 flags &= ~CFLGS_OFF_SLAB;
2490 left_over -= slab_size;
2493 if (flags & CFLGS_OFF_SLAB) {
2494 /* really off slab. No need for manual alignment */
2496 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2498 #ifdef CONFIG_PAGE_POISONING
2499 /* If we're going to use the generic kernel_map_pages()
2500 * poisoning, then it's going to smash the contents of
2501 * the redzone and userword anyhow, so switch them off.
2503 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2504 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2508 cachep->colour_off = cache_line_size();
2509 /* Offset must be a multiple of the alignment. */
2510 if (cachep->colour_off < cachep->align)
2511 cachep->colour_off = cachep->align;
2512 cachep->colour = left_over / cachep->colour_off;
2513 cachep->slab_size = slab_size;
2514 cachep->flags = flags;
2515 cachep->allocflags = 0;
2516 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2517 cachep->allocflags |= GFP_DMA;
2518 cachep->size = size;
2519 cachep->reciprocal_buffer_size = reciprocal_value(size);
2521 if (flags & CFLGS_OFF_SLAB) {
2522 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2524 * This is a possibility for one of the malloc_sizes caches.
2525 * But since we go off slab only for object size greater than
2526 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2527 * this should not happen at all.
2528 * But leave a BUG_ON for some lucky dude.
2530 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
2533 err = setup_cpu_cache(cachep, gfp);
2535 __kmem_cache_shutdown(cachep);
2539 if (flags & SLAB_DEBUG_OBJECTS) {
2541 * Would deadlock through slab_destroy()->call_rcu()->
2542 * debug_object_activate()->kmem_cache_alloc().
2544 WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU);
2546 slab_set_debugobj_lock_classes(cachep);
2553 static void check_irq_off(void)
2555 BUG_ON(!irqs_disabled());
2558 static void check_irq_on(void)
2560 BUG_ON(irqs_disabled());
2563 static void check_spinlock_acquired(struct kmem_cache *cachep)
2567 assert_spin_locked(&cachep->nodelists[numa_mem_id()]->list_lock);
2571 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2575 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2580 #define check_irq_off() do { } while(0)
2581 #define check_irq_on() do { } while(0)
2582 #define check_spinlock_acquired(x) do { } while(0)
2583 #define check_spinlock_acquired_node(x, y) do { } while(0)
2586 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2587 struct array_cache *ac,
2588 int force, int node);
2590 static void do_drain(void *arg)
2592 struct kmem_cache *cachep = arg;
2593 struct array_cache *ac;
2594 int node = numa_mem_id();
2597 ac = cpu_cache_get(cachep);
2598 spin_lock(&cachep->nodelists[node]->list_lock);
2599 free_block(cachep, ac->entry, ac->avail, node);
2600 spin_unlock(&cachep->nodelists[node]->list_lock);
2604 static void drain_cpu_caches(struct kmem_cache *cachep)
2606 struct kmem_list3 *l3;
2609 on_each_cpu(do_drain, cachep, 1);
2611 for_each_online_node(node) {
2612 l3 = cachep->nodelists[node];
2613 if (l3 && l3->alien)
2614 drain_alien_cache(cachep, l3->alien);
2617 for_each_online_node(node) {
2618 l3 = cachep->nodelists[node];
2620 drain_array(cachep, l3, l3->shared, 1, node);
2625 * Remove slabs from the list of free slabs.
2626 * Specify the number of slabs to drain in tofree.
2628 * Returns the actual number of slabs released.
2630 static int drain_freelist(struct kmem_cache *cache,
2631 struct kmem_list3 *l3, int tofree)
2633 struct list_head *p;
2638 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
2640 spin_lock_irq(&l3->list_lock);
2641 p = l3->slabs_free.prev;
2642 if (p == &l3->slabs_free) {
2643 spin_unlock_irq(&l3->list_lock);
2647 slabp = list_entry(p, struct slab, list);
2649 BUG_ON(slabp->inuse);
2651 list_del(&slabp->list);
2653 * Safe to drop the lock. The slab is no longer linked
2656 l3->free_objects -= cache->num;
2657 spin_unlock_irq(&l3->list_lock);
2658 slab_destroy(cache, slabp);
2665 /* Called with slab_mutex held to protect against cpu hotplug */
2666 static int __cache_shrink(struct kmem_cache *cachep)
2669 struct kmem_list3 *l3;
2671 drain_cpu_caches(cachep);
2674 for_each_online_node(i) {
2675 l3 = cachep->nodelists[i];
2679 drain_freelist(cachep, l3, l3->free_objects);
2681 ret += !list_empty(&l3->slabs_full) ||
2682 !list_empty(&l3->slabs_partial);
2684 return (ret ? 1 : 0);
2688 * kmem_cache_shrink - Shrink a cache.
2689 * @cachep: The cache to shrink.
2691 * Releases as many slabs as possible for a cache.
2692 * To help debugging, a zero exit status indicates all slabs were released.
2694 int kmem_cache_shrink(struct kmem_cache *cachep)
2697 BUG_ON(!cachep || in_interrupt());
2700 mutex_lock(&slab_mutex);
2701 ret = __cache_shrink(cachep);
2702 mutex_unlock(&slab_mutex);
2706 EXPORT_SYMBOL(kmem_cache_shrink);
2708 int __kmem_cache_shutdown(struct kmem_cache *cachep)
2711 struct kmem_list3 *l3;
2712 int rc = __cache_shrink(cachep);
2717 for_each_online_cpu(i)
2718 kfree(cachep->array[i]);
2720 /* NUMA: free the list3 structures */
2721 for_each_online_node(i) {
2722 l3 = cachep->nodelists[i];
2725 free_alien_cache(l3->alien);
2733 * Get the memory for a slab management obj.
2734 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2735 * always come from malloc_sizes caches. The slab descriptor cannot
2736 * come from the same cache which is getting created because,
2737 * when we are searching for an appropriate cache for these
2738 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2739 * If we are creating a malloc_sizes cache here it would not be visible to
2740 * kmem_find_general_cachep till the initialization is complete.
2741 * Hence we cannot have slabp_cache same as the original cache.
2743 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2744 int colour_off, gfp_t local_flags,
2749 if (OFF_SLAB(cachep)) {
2750 /* Slab management obj is off-slab. */
2751 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2752 local_flags, nodeid);
2754 * If the first object in the slab is leaked (it's allocated
2755 * but no one has a reference to it), we want to make sure
2756 * kmemleak does not treat the ->s_mem pointer as a reference
2757 * to the object. Otherwise we will not report the leak.
2759 kmemleak_scan_area(&slabp->list, sizeof(struct list_head),
2764 slabp = objp + colour_off;
2765 colour_off += cachep->slab_size;
2768 slabp->colouroff = colour_off;
2769 slabp->s_mem = objp + colour_off;
2770 slabp->nodeid = nodeid;
2775 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2777 return (kmem_bufctl_t *) (slabp + 1);
2780 static void cache_init_objs(struct kmem_cache *cachep,
2785 for (i = 0; i < cachep->num; i++) {
2786 void *objp = index_to_obj(cachep, slabp, i);
2788 /* need to poison the objs? */
2789 if (cachep->flags & SLAB_POISON)
2790 poison_obj(cachep, objp, POISON_FREE);
2791 if (cachep->flags & SLAB_STORE_USER)
2792 *dbg_userword(cachep, objp) = NULL;
2794 if (cachep->flags & SLAB_RED_ZONE) {
2795 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2796 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2799 * Constructors are not allowed to allocate memory from the same
2800 * cache which they are a constructor for. Otherwise, deadlock.
2801 * They must also be threaded.
2803 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2804 cachep->ctor(objp + obj_offset(cachep));
2806 if (cachep->flags & SLAB_RED_ZONE) {
2807 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2808 slab_error(cachep, "constructor overwrote the"
2809 " end of an object");
2810 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2811 slab_error(cachep, "constructor overwrote the"
2812 " start of an object");
2814 if ((cachep->size % PAGE_SIZE) == 0 &&
2815 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2816 kernel_map_pages(virt_to_page(objp),
2817 cachep->size / PAGE_SIZE, 0);
2822 slab_bufctl(slabp)[i] = i + 1;
2824 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2827 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2829 if (CONFIG_ZONE_DMA_FLAG) {
2830 if (flags & GFP_DMA)
2831 BUG_ON(!(cachep->allocflags & GFP_DMA));
2833 BUG_ON(cachep->allocflags & GFP_DMA);
2837 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2840 void *objp = index_to_obj(cachep, slabp, slabp->free);
2844 next = slab_bufctl(slabp)[slabp->free];
2846 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2847 WARN_ON(slabp->nodeid != nodeid);
2854 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2855 void *objp, int nodeid)
2857 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2860 /* Verify that the slab belongs to the intended node */
2861 WARN_ON(slabp->nodeid != nodeid);
2863 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2864 printk(KERN_ERR "slab: double free detected in cache "
2865 "'%s', objp %p\n", cachep->name, objp);
2869 slab_bufctl(slabp)[objnr] = slabp->free;
2870 slabp->free = objnr;
2875 * Map pages beginning at addr to the given cache and slab. This is required
2876 * for the slab allocator to be able to lookup the cache and slab of a
2877 * virtual address for kfree, ksize, and slab debugging.
2879 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2885 page = virt_to_page(addr);
2888 if (likely(!PageCompound(page)))
2889 nr_pages <<= cache->gfporder;
2892 page->slab_cache = cache;
2893 page->slab_page = slab;
2895 } while (--nr_pages);
2899 * Grow (by 1) the number of slabs within a cache. This is called by
2900 * kmem_cache_alloc() when there are no active objs left in a cache.
2902 static int cache_grow(struct kmem_cache *cachep,
2903 gfp_t flags, int nodeid, void *objp)
2908 struct kmem_list3 *l3;
2911 * Be lazy and only check for valid flags here, keeping it out of the
2912 * critical path in kmem_cache_alloc().
2914 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2915 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2917 /* Take the l3 list lock to change the colour_next on this node */
2919 l3 = cachep->nodelists[nodeid];
2920 spin_lock(&l3->list_lock);
2922 /* Get colour for the slab, and cal the next value. */
2923 offset = l3->colour_next;
2925 if (l3->colour_next >= cachep->colour)
2926 l3->colour_next = 0;
2927 spin_unlock(&l3->list_lock);
2929 offset *= cachep->colour_off;
2931 if (local_flags & __GFP_WAIT)
2935 * The test for missing atomic flag is performed here, rather than
2936 * the more obvious place, simply to reduce the critical path length
2937 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2938 * will eventually be caught here (where it matters).
2940 kmem_flagcheck(cachep, flags);
2943 * Get mem for the objs. Attempt to allocate a physical page from
2947 objp = kmem_getpages(cachep, local_flags, nodeid);
2951 /* Get slab management. */
2952 slabp = alloc_slabmgmt(cachep, objp, offset,
2953 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2957 slab_map_pages(cachep, slabp, objp);
2959 cache_init_objs(cachep, slabp);
2961 if (local_flags & __GFP_WAIT)
2962 local_irq_disable();
2964 spin_lock(&l3->list_lock);
2966 /* Make slab active. */
2967 list_add_tail(&slabp->list, &(l3->slabs_free));
2968 STATS_INC_GROWN(cachep);
2969 l3->free_objects += cachep->num;
2970 spin_unlock(&l3->list_lock);
2973 kmem_freepages(cachep, objp);
2975 if (local_flags & __GFP_WAIT)
2976 local_irq_disable();
2983 * Perform extra freeing checks:
2984 * - detect bad pointers.
2985 * - POISON/RED_ZONE checking
2987 static void kfree_debugcheck(const void *objp)
2989 if (!virt_addr_valid(objp)) {
2990 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2991 (unsigned long)objp);
2996 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2998 unsigned long long redzone1, redzone2;
3000 redzone1 = *dbg_redzone1(cache, obj);
3001 redzone2 = *dbg_redzone2(cache, obj);
3006 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
3009 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
3010 slab_error(cache, "double free detected");
3012 slab_error(cache, "memory outside object was overwritten");
3014 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
3015 obj, redzone1, redzone2);
3018 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
3019 unsigned long caller)
3025 BUG_ON(virt_to_cache(objp) != cachep);
3027 objp -= obj_offset(cachep);
3028 kfree_debugcheck(objp);
3029 page = virt_to_head_page(objp);
3031 slabp = page->slab_page;
3033 if (cachep->flags & SLAB_RED_ZONE) {
3034 verify_redzone_free(cachep, objp);
3035 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
3036 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
3038 if (cachep->flags & SLAB_STORE_USER)
3039 *dbg_userword(cachep, objp) = (void *)caller;
3041 objnr = obj_to_index(cachep, slabp, objp);
3043 BUG_ON(objnr >= cachep->num);
3044 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
3046 #ifdef CONFIG_DEBUG_SLAB_LEAK
3047 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
3049 if (cachep->flags & SLAB_POISON) {
3050 #ifdef CONFIG_DEBUG_PAGEALLOC
3051 if ((cachep->size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
3052 store_stackinfo(cachep, objp, caller);
3053 kernel_map_pages(virt_to_page(objp),
3054 cachep->size / PAGE_SIZE, 0);
3056 poison_obj(cachep, objp, POISON_FREE);
3059 poison_obj(cachep, objp, POISON_FREE);
3065 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
3070 /* Check slab's freelist to see if this obj is there. */
3071 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
3073 if (entries > cachep->num || i >= cachep->num)
3076 if (entries != cachep->num - slabp->inuse) {
3078 printk(KERN_ERR "slab: Internal list corruption detected in "
3079 "cache '%s'(%d), slabp %p(%d). Tainted(%s). Hexdump:\n",
3080 cachep->name, cachep->num, slabp, slabp->inuse,
3082 print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 16, 1, slabp,
3083 sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t),
3089 #define kfree_debugcheck(x) do { } while(0)
3090 #define cache_free_debugcheck(x,objp,z) (objp)
3091 #define check_slabp(x,y) do { } while(0)
3094 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags,
3098 struct kmem_list3 *l3;
3099 struct array_cache *ac;
3103 node = numa_mem_id();
3104 if (unlikely(force_refill))
3107 ac = cpu_cache_get(cachep);
3108 batchcount = ac->batchcount;
3109 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
3111 * If there was little recent activity on this cache, then
3112 * perform only a partial refill. Otherwise we could generate
3115 batchcount = BATCHREFILL_LIMIT;
3117 l3 = cachep->nodelists[node];
3119 BUG_ON(ac->avail > 0 || !l3);
3120 spin_lock(&l3->list_lock);
3122 /* See if we can refill from the shared array */
3123 if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) {
3124 l3->shared->touched = 1;
3128 while (batchcount > 0) {
3129 struct list_head *entry;
3131 /* Get slab alloc is to come from. */
3132 entry = l3->slabs_partial.next;
3133 if (entry == &l3->slabs_partial) {
3134 l3->free_touched = 1;
3135 entry = l3->slabs_free.next;
3136 if (entry == &l3->slabs_free)
3140 slabp = list_entry(entry, struct slab, list);
3141 check_slabp(cachep, slabp);
3142 check_spinlock_acquired(cachep);
3145 * The slab was either on partial or free list so
3146 * there must be at least one object available for
3149 BUG_ON(slabp->inuse >= cachep->num);
3151 while (slabp->inuse < cachep->num && batchcount--) {
3152 STATS_INC_ALLOCED(cachep);
3153 STATS_INC_ACTIVE(cachep);
3154 STATS_SET_HIGH(cachep);
3156 ac_put_obj(cachep, ac, slab_get_obj(cachep, slabp,
3159 check_slabp(cachep, slabp);
3161 /* move slabp to correct slabp list: */
3162 list_del(&slabp->list);
3163 if (slabp->free == BUFCTL_END)
3164 list_add(&slabp->list, &l3->slabs_full);
3166 list_add(&slabp->list, &l3->slabs_partial);
3170 l3->free_objects -= ac->avail;
3172 spin_unlock(&l3->list_lock);
3174 if (unlikely(!ac->avail)) {
3177 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
3179 /* cache_grow can reenable interrupts, then ac could change. */
3180 ac = cpu_cache_get(cachep);
3181 node = numa_mem_id();
3183 /* no objects in sight? abort */
3184 if (!x && (ac->avail == 0 || force_refill))
3187 if (!ac->avail) /* objects refilled by interrupt? */
3192 return ac_get_obj(cachep, ac, flags, force_refill);
3195 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3198 might_sleep_if(flags & __GFP_WAIT);
3200 kmem_flagcheck(cachep, flags);
3205 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3206 gfp_t flags, void *objp, unsigned long caller)
3210 if (cachep->flags & SLAB_POISON) {
3211 #ifdef CONFIG_DEBUG_PAGEALLOC
3212 if ((cachep->size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
3213 kernel_map_pages(virt_to_page(objp),
3214 cachep->size / PAGE_SIZE, 1);
3216 check_poison_obj(cachep, objp);
3218 check_poison_obj(cachep, objp);
3220 poison_obj(cachep, objp, POISON_INUSE);
3222 if (cachep->flags & SLAB_STORE_USER)
3223 *dbg_userword(cachep, objp) = (void *)caller;
3225 if (cachep->flags & SLAB_RED_ZONE) {
3226 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3227 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3228 slab_error(cachep, "double free, or memory outside"
3229 " object was overwritten");
3231 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3232 objp, *dbg_redzone1(cachep, objp),
3233 *dbg_redzone2(cachep, objp));
3235 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3236 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3238 #ifdef CONFIG_DEBUG_SLAB_LEAK
3243 slabp = virt_to_head_page(objp)->slab_page;
3244 objnr = (unsigned)(objp - slabp->s_mem) / cachep->size;
3245 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3248 objp += obj_offset(cachep);
3249 if (cachep->ctor && cachep->flags & SLAB_POISON)
3251 if (ARCH_SLAB_MINALIGN &&
3252 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
3253 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3254 objp, (int)ARCH_SLAB_MINALIGN);
3259 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3262 static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
3264 if (cachep == kmem_cache)
3267 return should_failslab(cachep->object_size, flags, cachep->flags);
3270 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3273 struct array_cache *ac;
3274 bool force_refill = false;
3278 ac = cpu_cache_get(cachep);
3279 if (likely(ac->avail)) {
3281 objp = ac_get_obj(cachep, ac, flags, false);
3284 * Allow for the possibility all avail objects are not allowed
3285 * by the current flags
3288 STATS_INC_ALLOCHIT(cachep);
3291 force_refill = true;
3294 STATS_INC_ALLOCMISS(cachep);
3295 objp = cache_alloc_refill(cachep, flags, force_refill);
3297 * the 'ac' may be updated by cache_alloc_refill(),
3298 * and kmemleak_erase() requires its correct value.
3300 ac = cpu_cache_get(cachep);
3304 * To avoid a false negative, if an object that is in one of the
3305 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3306 * treat the array pointers as a reference to the object.
3309 kmemleak_erase(&ac->entry[ac->avail]);
3315 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3317 * If we are in_interrupt, then process context, including cpusets and
3318 * mempolicy, may not apply and should not be used for allocation policy.
3320 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3322 int nid_alloc, nid_here;
3324 if (in_interrupt() || (flags & __GFP_THISNODE))
3326 nid_alloc = nid_here = numa_mem_id();
3327 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3328 nid_alloc = cpuset_slab_spread_node();
3329 else if (current->mempolicy)
3330 nid_alloc = slab_node();
3331 if (nid_alloc != nid_here)
3332 return ____cache_alloc_node(cachep, flags, nid_alloc);
3337 * Fallback function if there was no memory available and no objects on a
3338 * certain node and fall back is permitted. First we scan all the
3339 * available nodelists for available objects. If that fails then we
3340 * perform an allocation without specifying a node. This allows the page
3341 * allocator to do its reclaim / fallback magic. We then insert the
3342 * slab into the proper nodelist and then allocate from it.
3344 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3346 struct zonelist *zonelist;
3350 enum zone_type high_zoneidx = gfp_zone(flags);
3353 unsigned int cpuset_mems_cookie;
3355 if (flags & __GFP_THISNODE)
3358 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3361 cpuset_mems_cookie = get_mems_allowed();
3362 zonelist = node_zonelist(slab_node(), flags);
3366 * Look through allowed nodes for objects available
3367 * from existing per node queues.
3369 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3370 nid = zone_to_nid(zone);
3372 if (cpuset_zone_allowed_hardwall(zone, flags) &&
3373 cache->nodelists[nid] &&
3374 cache->nodelists[nid]->free_objects) {
3375 obj = ____cache_alloc_node(cache,
3376 flags | GFP_THISNODE, nid);
3384 * This allocation will be performed within the constraints
3385 * of the current cpuset / memory policy requirements.
3386 * We may trigger various forms of reclaim on the allowed
3387 * set and go into memory reserves if necessary.
3389 if (local_flags & __GFP_WAIT)
3391 kmem_flagcheck(cache, flags);
3392 obj = kmem_getpages(cache, local_flags, numa_mem_id());
3393 if (local_flags & __GFP_WAIT)
3394 local_irq_disable();
3397 * Insert into the appropriate per node queues
3399 nid = page_to_nid(virt_to_page(obj));
3400 if (cache_grow(cache, flags, nid, obj)) {
3401 obj = ____cache_alloc_node(cache,
3402 flags | GFP_THISNODE, nid);
3405 * Another processor may allocate the
3406 * objects in the slab since we are
3407 * not holding any locks.
3411 /* cache_grow already freed obj */
3417 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !obj))
3423 * A interface to enable slab creation on nodeid
3425 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3428 struct list_head *entry;
3430 struct kmem_list3 *l3;
3434 l3 = cachep->nodelists[nodeid];
3439 spin_lock(&l3->list_lock);
3440 entry = l3->slabs_partial.next;
3441 if (entry == &l3->slabs_partial) {
3442 l3->free_touched = 1;
3443 entry = l3->slabs_free.next;
3444 if (entry == &l3->slabs_free)
3448 slabp = list_entry(entry, struct slab, list);
3449 check_spinlock_acquired_node(cachep, nodeid);
3450 check_slabp(cachep, slabp);
3452 STATS_INC_NODEALLOCS(cachep);
3453 STATS_INC_ACTIVE(cachep);
3454 STATS_SET_HIGH(cachep);
3456 BUG_ON(slabp->inuse == cachep->num);
3458 obj = slab_get_obj(cachep, slabp, nodeid);
3459 check_slabp(cachep, slabp);
3461 /* move slabp to correct slabp list: */
3462 list_del(&slabp->list);
3464 if (slabp->free == BUFCTL_END)
3465 list_add(&slabp->list, &l3->slabs_full);
3467 list_add(&slabp->list, &l3->slabs_partial);
3469 spin_unlock(&l3->list_lock);
3473 spin_unlock(&l3->list_lock);
3474 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
3478 return fallback_alloc(cachep, flags);
3485 * kmem_cache_alloc_node - Allocate an object on the specified node
3486 * @cachep: The cache to allocate from.
3487 * @flags: See kmalloc().
3488 * @nodeid: node number of the target node.
3489 * @caller: return address of caller, used for debug information
3491 * Identical to kmem_cache_alloc but it will allocate memory on the given
3492 * node, which can improve the performance for cpu bound structures.
3494 * Fallback to other node is possible if __GFP_THISNODE is not set.
3496 static __always_inline void *
3497 slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3498 unsigned long caller)
3500 unsigned long save_flags;
3502 int slab_node = numa_mem_id();
3504 flags &= gfp_allowed_mask;
3506 lockdep_trace_alloc(flags);
3508 if (slab_should_failslab(cachep, flags))
3511 cache_alloc_debugcheck_before(cachep, flags);
3512 local_irq_save(save_flags);
3514 if (nodeid == NUMA_NO_NODE)
3517 if (unlikely(!cachep->nodelists[nodeid])) {
3518 /* Node not bootstrapped yet */
3519 ptr = fallback_alloc(cachep, flags);
3523 if (nodeid == slab_node) {
3525 * Use the locally cached objects if possible.
3526 * However ____cache_alloc does not allow fallback
3527 * to other nodes. It may fail while we still have
3528 * objects on other nodes available.
3530 ptr = ____cache_alloc(cachep, flags);
3534 /* ___cache_alloc_node can fall back to other nodes */
3535 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3537 local_irq_restore(save_flags);
3538 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3539 kmemleak_alloc_recursive(ptr, cachep->object_size, 1, cachep->flags,
3543 kmemcheck_slab_alloc(cachep, flags, ptr, cachep->object_size);
3545 if (unlikely((flags & __GFP_ZERO) && ptr))
3546 memset(ptr, 0, cachep->object_size);
3551 static __always_inline void *
3552 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3556 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3557 objp = alternate_node_alloc(cache, flags);
3561 objp = ____cache_alloc(cache, flags);
3564 * We may just have run out of memory on the local node.
3565 * ____cache_alloc_node() knows how to locate memory on other nodes
3568 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3575 static __always_inline void *
3576 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3578 return ____cache_alloc(cachep, flags);
3581 #endif /* CONFIG_NUMA */
3583 static __always_inline void *
3584 slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
3586 unsigned long save_flags;
3589 flags &= gfp_allowed_mask;
3591 lockdep_trace_alloc(flags);
3593 if (slab_should_failslab(cachep, flags))
3596 cache_alloc_debugcheck_before(cachep, flags);
3597 local_irq_save(save_flags);
3598 objp = __do_cache_alloc(cachep, flags);
3599 local_irq_restore(save_flags);
3600 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3601 kmemleak_alloc_recursive(objp, cachep->object_size, 1, cachep->flags,
3606 kmemcheck_slab_alloc(cachep, flags, objp, cachep->object_size);
3608 if (unlikely((flags & __GFP_ZERO) && objp))
3609 memset(objp, 0, cachep->object_size);
3615 * Caller needs to acquire correct kmem_list's list_lock
3617 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3621 struct kmem_list3 *l3;
3623 for (i = 0; i < nr_objects; i++) {
3627 clear_obj_pfmemalloc(&objpp[i]);
3630 slabp = virt_to_slab(objp);
3631 l3 = cachep->nodelists[node];
3632 list_del(&slabp->list);
3633 check_spinlock_acquired_node(cachep, node);
3634 check_slabp(cachep, slabp);
3635 slab_put_obj(cachep, slabp, objp, node);
3636 STATS_DEC_ACTIVE(cachep);
3638 check_slabp(cachep, slabp);
3640 /* fixup slab chains */
3641 if (slabp->inuse == 0) {
3642 if (l3->free_objects > l3->free_limit) {
3643 l3->free_objects -= cachep->num;
3644 /* No need to drop any previously held
3645 * lock here, even if we have a off-slab slab
3646 * descriptor it is guaranteed to come from
3647 * a different cache, refer to comments before
3650 slab_destroy(cachep, slabp);
3652 list_add(&slabp->list, &l3->slabs_free);
3655 /* Unconditionally move a slab to the end of the
3656 * partial list on free - maximum time for the
3657 * other objects to be freed, too.
3659 list_add_tail(&slabp->list, &l3->slabs_partial);
3664 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3667 struct kmem_list3 *l3;
3668 int node = numa_mem_id();
3670 batchcount = ac->batchcount;
3672 BUG_ON(!batchcount || batchcount > ac->avail);
3675 l3 = cachep->nodelists[node];
3676 spin_lock(&l3->list_lock);
3678 struct array_cache *shared_array = l3->shared;
3679 int max = shared_array->limit - shared_array->avail;
3681 if (batchcount > max)
3683 memcpy(&(shared_array->entry[shared_array->avail]),
3684 ac->entry, sizeof(void *) * batchcount);
3685 shared_array->avail += batchcount;
3690 free_block(cachep, ac->entry, batchcount, node);
3695 struct list_head *p;
3697 p = l3->slabs_free.next;
3698 while (p != &(l3->slabs_free)) {
3701 slabp = list_entry(p, struct slab, list);
3702 BUG_ON(slabp->inuse);
3707 STATS_SET_FREEABLE(cachep, i);
3710 spin_unlock(&l3->list_lock);
3711 ac->avail -= batchcount;
3712 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3716 * Release an obj back to its cache. If the obj has a constructed state, it must
3717 * be in this state _before_ it is released. Called with disabled ints.
3719 static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3720 unsigned long caller)
3722 struct array_cache *ac = cpu_cache_get(cachep);
3725 kmemleak_free_recursive(objp, cachep->flags);
3726 objp = cache_free_debugcheck(cachep, objp, caller);
3728 kmemcheck_slab_free(cachep, objp, cachep->object_size);
3731 * Skip calling cache_free_alien() when the platform is not numa.
3732 * This will avoid cache misses that happen while accessing slabp (which
3733 * is per page memory reference) to get nodeid. Instead use a global
3734 * variable to skip the call, which is mostly likely to be present in
3737 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3740 if (likely(ac->avail < ac->limit)) {
3741 STATS_INC_FREEHIT(cachep);
3743 STATS_INC_FREEMISS(cachep);
3744 cache_flusharray(cachep, ac);
3747 ac_put_obj(cachep, ac, objp);
3751 * kmem_cache_alloc - Allocate an object
3752 * @cachep: The cache to allocate from.
3753 * @flags: See kmalloc().
3755 * Allocate an object from this cache. The flags are only relevant
3756 * if the cache has no available objects.
3758 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3760 void *ret = slab_alloc(cachep, flags, _RET_IP_);
3762 trace_kmem_cache_alloc(_RET_IP_, ret,
3763 cachep->object_size, cachep->size, flags);
3767 EXPORT_SYMBOL(kmem_cache_alloc);
3769 #ifdef CONFIG_TRACING
3771 kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
3775 ret = slab_alloc(cachep, flags, _RET_IP_);
3777 trace_kmalloc(_RET_IP_, ret,
3778 size, cachep->size, flags);
3781 EXPORT_SYMBOL(kmem_cache_alloc_trace);
3785 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3787 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3789 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3790 cachep->object_size, cachep->size,
3795 EXPORT_SYMBOL(kmem_cache_alloc_node);
3797 #ifdef CONFIG_TRACING
3798 void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
3805 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3807 trace_kmalloc_node(_RET_IP_, ret,
3812 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3815 static __always_inline void *
3816 __do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
3818 struct kmem_cache *cachep;
3820 cachep = kmem_find_general_cachep(size, flags);
3821 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3823 return kmem_cache_alloc_node_trace(cachep, flags, node, size);
3826 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3827 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3829 return __do_kmalloc_node(size, flags, node, _RET_IP_);
3831 EXPORT_SYMBOL(__kmalloc_node);
3833 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3834 int node, unsigned long caller)
3836 return __do_kmalloc_node(size, flags, node, caller);
3838 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3840 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3842 return __do_kmalloc_node(size, flags, node, 0);
3844 EXPORT_SYMBOL(__kmalloc_node);
3845 #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3846 #endif /* CONFIG_NUMA */
3849 * __do_kmalloc - allocate memory
3850 * @size: how many bytes of memory are required.
3851 * @flags: the type of memory to allocate (see kmalloc).
3852 * @caller: function caller for debug tracking of the caller
3854 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3855 unsigned long caller)
3857 struct kmem_cache *cachep;
3860 /* If you want to save a few bytes .text space: replace
3862 * Then kmalloc uses the uninlined functions instead of the inline
3865 cachep = __find_general_cachep(size, flags);
3866 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3868 ret = slab_alloc(cachep, flags, caller);
3870 trace_kmalloc(caller, ret,
3871 size, cachep->size, flags);
3877 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3878 void *__kmalloc(size_t size, gfp_t flags)
3880 return __do_kmalloc(size, flags, _RET_IP_);
3882 EXPORT_SYMBOL(__kmalloc);
3884 void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3886 return __do_kmalloc(size, flags, caller);
3888 EXPORT_SYMBOL(__kmalloc_track_caller);
3891 void *__kmalloc(size_t size, gfp_t flags)
3893 return __do_kmalloc(size, flags, 0);
3895 EXPORT_SYMBOL(__kmalloc);
3899 * kmem_cache_free - Deallocate an object
3900 * @cachep: The cache the allocation was from.
3901 * @objp: The previously allocated object.
3903 * Free an object which was previously allocated from this
3906 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3908 unsigned long flags;
3910 local_irq_save(flags);
3911 debug_check_no_locks_freed(objp, cachep->object_size);
3912 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3913 debug_check_no_obj_freed(objp, cachep->object_size);
3914 __cache_free(cachep, objp, _RET_IP_);
3915 local_irq_restore(flags);
3917 trace_kmem_cache_free(_RET_IP_, objp);
3919 EXPORT_SYMBOL(kmem_cache_free);
3922 * kfree - free previously allocated memory
3923 * @objp: pointer returned by kmalloc.
3925 * If @objp is NULL, no operation is performed.
3927 * Don't free memory not originally allocated by kmalloc()
3928 * or you will run into trouble.
3930 void kfree(const void *objp)
3932 struct kmem_cache *c;
3933 unsigned long flags;
3935 trace_kfree(_RET_IP_, objp);
3937 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3939 local_irq_save(flags);
3940 kfree_debugcheck(objp);
3941 c = virt_to_cache(objp);
3942 debug_check_no_locks_freed(objp, c->object_size);
3944 debug_check_no_obj_freed(objp, c->object_size);
3945 __cache_free(c, (void *)objp, _RET_IP_);
3946 local_irq_restore(flags);
3948 EXPORT_SYMBOL(kfree);
3951 * This initializes kmem_list3 or resizes various caches for all nodes.
3953 static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
3956 struct kmem_list3 *l3;
3957 struct array_cache *new_shared;
3958 struct array_cache **new_alien = NULL;
3960 for_each_online_node(node) {
3962 if (use_alien_caches) {
3963 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
3969 if (cachep->shared) {
3970 new_shared = alloc_arraycache(node,
3971 cachep->shared*cachep->batchcount,
3974 free_alien_cache(new_alien);
3979 l3 = cachep->nodelists[node];
3981 struct array_cache *shared = l3->shared;
3983 spin_lock_irq(&l3->list_lock);
3986 free_block(cachep, shared->entry,
3987 shared->avail, node);
3989 l3->shared = new_shared;
3991 l3->alien = new_alien;
3994 l3->free_limit = (1 + nr_cpus_node(node)) *
3995 cachep->batchcount + cachep->num;
3996 spin_unlock_irq(&l3->list_lock);
3998 free_alien_cache(new_alien);
4001 l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node);
4003 free_alien_cache(new_alien);
4008 kmem_list3_init(l3);
4009 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
4010 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
4011 l3->shared = new_shared;
4012 l3->alien = new_alien;
4013 l3->free_limit = (1 + nr_cpus_node(node)) *
4014 cachep->batchcount + cachep->num;
4015 cachep->nodelists[node] = l3;
4020 if (!cachep->list.next) {
4021 /* Cache is not active yet. Roll back what we did */
4024 if (cachep->nodelists[node]) {
4025 l3 = cachep->nodelists[node];
4028 free_alien_cache(l3->alien);
4030 cachep->nodelists[node] = NULL;
4038 struct ccupdate_struct {
4039 struct kmem_cache *cachep;
4040 struct array_cache *new[0];
4043 static void do_ccupdate_local(void *info)
4045 struct ccupdate_struct *new = info;
4046 struct array_cache *old;
4049 old = cpu_cache_get(new->cachep);
4051 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
4052 new->new[smp_processor_id()] = old;
4055 /* Always called with the slab_mutex held */
4056 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
4057 int batchcount, int shared, gfp_t gfp)
4059 struct ccupdate_struct *new;
4062 new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *),
4067 for_each_online_cpu(i) {
4068 new->new[i] = alloc_arraycache(cpu_to_mem(i), limit,
4071 for (i--; i >= 0; i--)
4077 new->cachep = cachep;
4079 on_each_cpu(do_ccupdate_local, (void *)new, 1);
4082 cachep->batchcount = batchcount;
4083 cachep->limit = limit;
4084 cachep->shared = shared;
4086 for_each_online_cpu(i) {
4087 struct array_cache *ccold = new->new[i];
4090 spin_lock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
4091 free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i));
4092 spin_unlock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
4096 return alloc_kmemlist(cachep, gfp);
4099 /* Called with slab_mutex held always */
4100 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
4106 * The head array serves three purposes:
4107 * - create a LIFO ordering, i.e. return objects that are cache-warm
4108 * - reduce the number of spinlock operations.
4109 * - reduce the number of linked list operations on the slab and
4110 * bufctl chains: array operations are cheaper.
4111 * The numbers are guessed, we should auto-tune as described by
4114 if (cachep->size > 131072)
4116 else if (cachep->size > PAGE_SIZE)
4118 else if (cachep->size > 1024)
4120 else if (cachep->size > 256)
4126 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
4127 * allocation behaviour: Most allocs on one cpu, most free operations
4128 * on another cpu. For these cases, an efficient object passing between
4129 * cpus is necessary. This is provided by a shared array. The array
4130 * replaces Bonwick's magazine layer.
4131 * On uniprocessor, it's functionally equivalent (but less efficient)
4132 * to a larger limit. Thus disabled by default.
4135 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
4140 * With debugging enabled, large batchcount lead to excessively long
4141 * periods with disabled local interrupts. Limit the batchcount
4146 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp);
4148 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
4149 cachep->name, -err);
4154 * Drain an array if it contains any elements taking the l3 lock only if
4155 * necessary. Note that the l3 listlock also protects the array_cache
4156 * if drain_array() is used on the shared array.
4158 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
4159 struct array_cache *ac, int force, int node)
4163 if (!ac || !ac->avail)
4165 if (ac->touched && !force) {
4168 spin_lock_irq(&l3->list_lock);
4170 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4171 if (tofree > ac->avail)
4172 tofree = (ac->avail + 1) / 2;
4173 free_block(cachep, ac->entry, tofree, node);
4174 ac->avail -= tofree;
4175 memmove(ac->entry, &(ac->entry[tofree]),
4176 sizeof(void *) * ac->avail);
4178 spin_unlock_irq(&l3->list_lock);
4183 * cache_reap - Reclaim memory from caches.
4184 * @w: work descriptor
4186 * Called from workqueue/eventd every few seconds.
4188 * - clear the per-cpu caches for this CPU.
4189 * - return freeable pages to the main free memory pool.
4191 * If we cannot acquire the cache chain mutex then just give up - we'll try
4192 * again on the next iteration.
4194 static void cache_reap(struct work_struct *w)
4196 struct kmem_cache *searchp;
4197 struct kmem_list3 *l3;
4198 int node = numa_mem_id();
4199 struct delayed_work *work = to_delayed_work(w);
4201 if (!mutex_trylock(&slab_mutex))
4202 /* Give up. Setup the next iteration. */
4205 list_for_each_entry(searchp, &slab_caches, list) {
4209 * We only take the l3 lock if absolutely necessary and we
4210 * have established with reasonable certainty that
4211 * we can do some work if the lock was obtained.
4213 l3 = searchp->nodelists[node];
4215 reap_alien(searchp, l3);
4217 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
4220 * These are racy checks but it does not matter
4221 * if we skip one check or scan twice.
4223 if (time_after(l3->next_reap, jiffies))
4226 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
4228 drain_array(searchp, l3, l3->shared, 0, node);
4230 if (l3->free_touched)
4231 l3->free_touched = 0;
4235 freed = drain_freelist(searchp, l3, (l3->free_limit +
4236 5 * searchp->num - 1) / (5 * searchp->num));
4237 STATS_ADD_REAPED(searchp, freed);
4243 mutex_unlock(&slab_mutex);
4246 /* Set up the next iteration */
4247 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
4250 #ifdef CONFIG_SLABINFO
4251 void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
4254 unsigned long active_objs;
4255 unsigned long num_objs;
4256 unsigned long active_slabs = 0;
4257 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4261 struct kmem_list3 *l3;
4265 for_each_online_node(node) {
4266 l3 = cachep->nodelists[node];
4271 spin_lock_irq(&l3->list_lock);
4273 list_for_each_entry(slabp, &l3->slabs_full, list) {
4274 if (slabp->inuse != cachep->num && !error)
4275 error = "slabs_full accounting error";
4276 active_objs += cachep->num;
4279 list_for_each_entry(slabp, &l3->slabs_partial, list) {
4280 if (slabp->inuse == cachep->num && !error)
4281 error = "slabs_partial inuse accounting error";
4282 if (!slabp->inuse && !error)
4283 error = "slabs_partial/inuse accounting error";
4284 active_objs += slabp->inuse;
4287 list_for_each_entry(slabp, &l3->slabs_free, list) {
4288 if (slabp->inuse && !error)
4289 error = "slabs_free/inuse accounting error";
4292 free_objects += l3->free_objects;
4294 shared_avail += l3->shared->avail;
4296 spin_unlock_irq(&l3->list_lock);
4298 num_slabs += active_slabs;
4299 num_objs = num_slabs * cachep->num;
4300 if (num_objs - active_objs != free_objects && !error)
4301 error = "free_objects accounting error";
4303 name = cachep->name;
4305 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4307 sinfo->active_objs = active_objs;
4308 sinfo->num_objs = num_objs;
4309 sinfo->active_slabs = active_slabs;
4310 sinfo->num_slabs = num_slabs;
4311 sinfo->shared_avail = shared_avail;
4312 sinfo->limit = cachep->limit;
4313 sinfo->batchcount = cachep->batchcount;
4314 sinfo->shared = cachep->shared;
4315 sinfo->objects_per_slab = cachep->num;
4316 sinfo->cache_order = cachep->gfporder;
4319 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
4323 unsigned long high = cachep->high_mark;
4324 unsigned long allocs = cachep->num_allocations;
4325 unsigned long grown = cachep->grown;
4326 unsigned long reaped = cachep->reaped;
4327 unsigned long errors = cachep->errors;
4328 unsigned long max_freeable = cachep->max_freeable;
4329 unsigned long node_allocs = cachep->node_allocs;
4330 unsigned long node_frees = cachep->node_frees;
4331 unsigned long overflows = cachep->node_overflow;
4333 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4334 "%4lu %4lu %4lu %4lu %4lu",
4335 allocs, high, grown,
4336 reaped, errors, max_freeable, node_allocs,
4337 node_frees, overflows);
4341 unsigned long allochit = atomic_read(&cachep->allochit);
4342 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4343 unsigned long freehit = atomic_read(&cachep->freehit);
4344 unsigned long freemiss = atomic_read(&cachep->freemiss);
4346 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4347 allochit, allocmiss, freehit, freemiss);
4352 #define MAX_SLABINFO_WRITE 128
4354 * slabinfo_write - Tuning for the slab allocator
4356 * @buffer: user buffer
4357 * @count: data length
4360 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4361 size_t count, loff_t *ppos)
4363 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4364 int limit, batchcount, shared, res;
4365 struct kmem_cache *cachep;
4367 if (count > MAX_SLABINFO_WRITE)
4369 if (copy_from_user(&kbuf, buffer, count))
4371 kbuf[MAX_SLABINFO_WRITE] = '\0';
4373 tmp = strchr(kbuf, ' ');
4378 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4381 /* Find the cache in the chain of caches. */
4382 mutex_lock(&slab_mutex);
4384 list_for_each_entry(cachep, &slab_caches, list) {
4385 if (!strcmp(cachep->name, kbuf)) {
4386 if (limit < 1 || batchcount < 1 ||
4387 batchcount > limit || shared < 0) {
4390 res = do_tune_cpucache(cachep, limit,
4397 mutex_unlock(&slab_mutex);
4403 #ifdef CONFIG_DEBUG_SLAB_LEAK
4405 static void *leaks_start(struct seq_file *m, loff_t *pos)
4407 mutex_lock(&slab_mutex);
4408 return seq_list_start(&slab_caches, *pos);
4411 static inline int add_caller(unsigned long *n, unsigned long v)
4421 unsigned long *q = p + 2 * i;
4435 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4441 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4447 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->size) {
4448 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4450 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4455 static void show_symbol(struct seq_file *m, unsigned long address)
4457 #ifdef CONFIG_KALLSYMS
4458 unsigned long offset, size;
4459 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4461 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4462 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4464 seq_printf(m, " [%s]", modname);
4468 seq_printf(m, "%p", (void *)address);
4471 static int leaks_show(struct seq_file *m, void *p)
4473 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
4475 struct kmem_list3 *l3;
4477 unsigned long *n = m->private;
4481 if (!(cachep->flags & SLAB_STORE_USER))
4483 if (!(cachep->flags & SLAB_RED_ZONE))
4486 /* OK, we can do it */
4490 for_each_online_node(node) {
4491 l3 = cachep->nodelists[node];
4496 spin_lock_irq(&l3->list_lock);
4498 list_for_each_entry(slabp, &l3->slabs_full, list)
4499 handle_slab(n, cachep, slabp);
4500 list_for_each_entry(slabp, &l3->slabs_partial, list)
4501 handle_slab(n, cachep, slabp);
4502 spin_unlock_irq(&l3->list_lock);
4504 name = cachep->name;
4506 /* Increase the buffer size */
4507 mutex_unlock(&slab_mutex);
4508 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4510 /* Too bad, we are really out */
4512 mutex_lock(&slab_mutex);
4515 *(unsigned long *)m->private = n[0] * 2;
4517 mutex_lock(&slab_mutex);
4518 /* Now make sure this entry will be retried */
4522 for (i = 0; i < n[1]; i++) {
4523 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4524 show_symbol(m, n[2*i+2]);
4531 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4533 return seq_list_next(p, &slab_caches, pos);
4536 static void s_stop(struct seq_file *m, void *p)
4538 mutex_unlock(&slab_mutex);
4541 static const struct seq_operations slabstats_op = {
4542 .start = leaks_start,
4548 static int slabstats_open(struct inode *inode, struct file *file)
4550 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4553 ret = seq_open(file, &slabstats_op);
4555 struct seq_file *m = file->private_data;
4556 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4565 static const struct file_operations proc_slabstats_operations = {
4566 .open = slabstats_open,
4568 .llseek = seq_lseek,
4569 .release = seq_release_private,
4573 static int __init slab_proc_init(void)
4575 #ifdef CONFIG_DEBUG_SLAB_LEAK
4576 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4580 module_init(slab_proc_init);
4584 * ksize - get the actual amount of memory allocated for a given object
4585 * @objp: Pointer to the object
4587 * kmalloc may internally round up allocations and return more memory
4588 * than requested. ksize() can be used to determine the actual amount of
4589 * memory allocated. The caller may use this additional memory, even though
4590 * a smaller amount of memory was initially specified with the kmalloc call.
4591 * The caller must guarantee that objp points to a valid object previously
4592 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4593 * must not be freed during the duration of the call.
4595 size_t ksize(const void *objp)
4598 if (unlikely(objp == ZERO_SIZE_PTR))
4601 return virt_to_cache(objp)->object_size;
4603 EXPORT_SYMBOL(ksize);