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;
165 /* Legal flag mask for kmem_cache_create(). */
167 # define CREATE_MASK (SLAB_RED_ZONE | \
168 SLAB_POISON | SLAB_HWCACHE_ALIGN | \
171 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
172 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
173 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
175 # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \
177 SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \
178 SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \
179 SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK)
185 * Bufctl's are used for linking objs within a slab
188 * This implementation relies on "struct page" for locating the cache &
189 * slab an object belongs to.
190 * This allows the bufctl structure to be small (one int), but limits
191 * the number of objects a slab (not a cache) can contain when off-slab
192 * bufctls are used. The limit is the size of the largest general cache
193 * that does not use off-slab slabs.
194 * For 32bit archs with 4 kB pages, is this 56.
195 * This is not serious, as it is only for large objects, when it is unwise
196 * to have too many per slab.
197 * Note: This limit can be raised by introducing a general cache whose size
198 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
201 typedef unsigned int kmem_bufctl_t;
202 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
203 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
204 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
205 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
210 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
211 * arrange for kmem_freepages to be called via RCU. This is useful if
212 * we need to approach a kernel structure obliquely, from its address
213 * obtained without the usual locking. We can lock the structure to
214 * stabilize it and check it's still at the given address, only if we
215 * can be sure that the memory has not been meanwhile reused for some
216 * other kind of object (which our subsystem's lock might corrupt).
218 * rcu_read_lock before reading the address, then rcu_read_unlock after
219 * taking the spinlock within the structure expected at that address.
222 struct rcu_head head;
223 struct kmem_cache *cachep;
230 * Manages the objs in a slab. Placed either at the beginning of mem allocated
231 * for a slab, or allocated from an general cache.
232 * Slabs are chained into three list: fully used, partial, fully free slabs.
237 struct list_head list;
238 unsigned long colouroff;
239 void *s_mem; /* including colour offset */
240 unsigned int inuse; /* num of objs active in slab */
242 unsigned short nodeid;
244 struct slab_rcu __slab_cover_slab_rcu;
252 * - LIFO ordering, to hand out cache-warm objects from _alloc
253 * - reduce the number of linked list operations
254 * - reduce spinlock operations
256 * The limit is stored in the per-cpu structure to reduce the data cache
263 unsigned int batchcount;
264 unsigned int touched;
267 * Must have this definition in here for the proper
268 * alignment of array_cache. Also simplifies accessing
271 * Entries should not be directly dereferenced as
272 * entries belonging to slabs marked pfmemalloc will
273 * have the lower bits set SLAB_OBJ_PFMEMALLOC
277 #define SLAB_OBJ_PFMEMALLOC 1
278 static inline bool is_obj_pfmemalloc(void *objp)
280 return (unsigned long)objp & SLAB_OBJ_PFMEMALLOC;
283 static inline void set_obj_pfmemalloc(void **objp)
285 *objp = (void *)((unsigned long)*objp | SLAB_OBJ_PFMEMALLOC);
289 static inline void clear_obj_pfmemalloc(void **objp)
291 *objp = (void *)((unsigned long)*objp & ~SLAB_OBJ_PFMEMALLOC);
295 * bootstrap: The caches do not work without cpuarrays anymore, but the
296 * cpuarrays are allocated from the generic caches...
298 #define BOOT_CPUCACHE_ENTRIES 1
299 struct arraycache_init {
300 struct array_cache cache;
301 void *entries[BOOT_CPUCACHE_ENTRIES];
305 * The slab lists for all objects.
308 struct list_head slabs_partial; /* partial list first, better asm code */
309 struct list_head slabs_full;
310 struct list_head slabs_free;
311 unsigned long free_objects;
312 unsigned int free_limit;
313 unsigned int colour_next; /* Per-node cache coloring */
314 spinlock_t list_lock;
315 struct array_cache *shared; /* shared per node */
316 struct array_cache **alien; /* on other nodes */
317 unsigned long next_reap; /* updated without locking */
318 int free_touched; /* updated without locking */
322 * Need this for bootstrapping a per node allocator.
324 #define NUM_INIT_LISTS (3 * MAX_NUMNODES)
325 static struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS];
326 #define CACHE_CACHE 0
327 #define SIZE_AC MAX_NUMNODES
328 #define SIZE_L3 (2 * MAX_NUMNODES)
330 static int drain_freelist(struct kmem_cache *cache,
331 struct kmem_list3 *l3, int tofree);
332 static void free_block(struct kmem_cache *cachep, void **objpp, int len,
334 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
335 static void cache_reap(struct work_struct *unused);
338 * This function must be completely optimized away if a constant is passed to
339 * it. Mostly the same as what is in linux/slab.h except it returns an index.
341 static __always_inline int index_of(const size_t size)
343 extern void __bad_size(void);
345 if (__builtin_constant_p(size)) {
353 #include <linux/kmalloc_sizes.h>
361 static int slab_early_init = 1;
363 #define INDEX_AC index_of(sizeof(struct arraycache_init))
364 #define INDEX_L3 index_of(sizeof(struct kmem_list3))
366 static void kmem_list3_init(struct kmem_list3 *parent)
368 INIT_LIST_HEAD(&parent->slabs_full);
369 INIT_LIST_HEAD(&parent->slabs_partial);
370 INIT_LIST_HEAD(&parent->slabs_free);
371 parent->shared = NULL;
372 parent->alien = NULL;
373 parent->colour_next = 0;
374 spin_lock_init(&parent->list_lock);
375 parent->free_objects = 0;
376 parent->free_touched = 0;
379 #define MAKE_LIST(cachep, listp, slab, nodeid) \
381 INIT_LIST_HEAD(listp); \
382 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
385 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
387 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
388 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
389 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
392 #define CFLGS_OFF_SLAB (0x80000000UL)
393 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
395 #define BATCHREFILL_LIMIT 16
397 * Optimization question: fewer reaps means less probability for unnessary
398 * cpucache drain/refill cycles.
400 * OTOH the cpuarrays can contain lots of objects,
401 * which could lock up otherwise freeable slabs.
403 #define REAPTIMEOUT_CPUC (2*HZ)
404 #define REAPTIMEOUT_LIST3 (4*HZ)
407 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
408 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
409 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
410 #define STATS_INC_GROWN(x) ((x)->grown++)
411 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
412 #define STATS_SET_HIGH(x) \
414 if ((x)->num_active > (x)->high_mark) \
415 (x)->high_mark = (x)->num_active; \
417 #define STATS_INC_ERR(x) ((x)->errors++)
418 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
419 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
420 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
421 #define STATS_SET_FREEABLE(x, i) \
423 if ((x)->max_freeable < i) \
424 (x)->max_freeable = i; \
426 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
427 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
428 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
429 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
431 #define STATS_INC_ACTIVE(x) do { } while (0)
432 #define STATS_DEC_ACTIVE(x) do { } while (0)
433 #define STATS_INC_ALLOCED(x) do { } while (0)
434 #define STATS_INC_GROWN(x) do { } while (0)
435 #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
436 #define STATS_SET_HIGH(x) do { } while (0)
437 #define STATS_INC_ERR(x) do { } while (0)
438 #define STATS_INC_NODEALLOCS(x) do { } while (0)
439 #define STATS_INC_NODEFREES(x) do { } while (0)
440 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
441 #define STATS_SET_FREEABLE(x, i) do { } while (0)
442 #define STATS_INC_ALLOCHIT(x) do { } while (0)
443 #define STATS_INC_ALLOCMISS(x) do { } while (0)
444 #define STATS_INC_FREEHIT(x) do { } while (0)
445 #define STATS_INC_FREEMISS(x) do { } while (0)
451 * memory layout of objects:
453 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
454 * the end of an object is aligned with the end of the real
455 * allocation. Catches writes behind the end of the allocation.
456 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
458 * cachep->obj_offset: The real object.
459 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
460 * cachep->size - 1* BYTES_PER_WORD: last caller address
461 * [BYTES_PER_WORD long]
463 static int obj_offset(struct kmem_cache *cachep)
465 return cachep->obj_offset;
468 static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
470 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
471 return (unsigned long long*) (objp + obj_offset(cachep) -
472 sizeof(unsigned long long));
475 static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
477 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
478 if (cachep->flags & SLAB_STORE_USER)
479 return (unsigned long long *)(objp + cachep->size -
480 sizeof(unsigned long long) -
482 return (unsigned long long *) (objp + cachep->size -
483 sizeof(unsigned long long));
486 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
488 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
489 return (void **)(objp + cachep->size - BYTES_PER_WORD);
494 #define obj_offset(x) 0
495 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
496 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
497 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
502 * Do not go above this order unless 0 objects fit into the slab or
503 * overridden on the command line.
505 #define SLAB_MAX_ORDER_HI 1
506 #define SLAB_MAX_ORDER_LO 0
507 static int slab_max_order = SLAB_MAX_ORDER_LO;
508 static bool slab_max_order_set __initdata;
510 static inline struct kmem_cache *virt_to_cache(const void *obj)
512 struct page *page = virt_to_head_page(obj);
513 return page->slab_cache;
516 static inline struct slab *virt_to_slab(const void *obj)
518 struct page *page = virt_to_head_page(obj);
520 VM_BUG_ON(!PageSlab(page));
521 return page->slab_page;
524 static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
527 return slab->s_mem + cache->size * idx;
531 * We want to avoid an expensive divide : (offset / cache->size)
532 * Using the fact that size is a constant for a particular cache,
533 * we can replace (offset / cache->size) by
534 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
536 static inline unsigned int obj_to_index(const struct kmem_cache *cache,
537 const struct slab *slab, void *obj)
539 u32 offset = (obj - slab->s_mem);
540 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
544 * These are the default caches for kmalloc. Custom caches can have other sizes.
546 struct cache_sizes malloc_sizes[] = {
547 #define CACHE(x) { .cs_size = (x) },
548 #include <linux/kmalloc_sizes.h>
552 EXPORT_SYMBOL(malloc_sizes);
554 /* Must match cache_sizes above. Out of line to keep cache footprint low. */
560 static struct cache_names __initdata cache_names[] = {
561 #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" },
562 #include <linux/kmalloc_sizes.h>
567 static struct arraycache_init initarray_cache __initdata =
568 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
569 static struct arraycache_init initarray_generic =
570 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
572 /* internal cache of cache description objs */
573 static struct kmem_list3 *kmem_cache_nodelists[MAX_NUMNODES];
574 static struct kmem_cache kmem_cache_boot = {
575 .nodelists = kmem_cache_nodelists,
577 .limit = BOOT_CPUCACHE_ENTRIES,
579 .size = sizeof(struct kmem_cache),
580 .name = "kmem_cache",
583 #define BAD_ALIEN_MAGIC 0x01020304ul
585 #ifdef CONFIG_LOCKDEP
588 * Slab sometimes uses the kmalloc slabs to store the slab headers
589 * for other slabs "off slab".
590 * The locking for this is tricky in that it nests within the locks
591 * of all other slabs in a few places; to deal with this special
592 * locking we put on-slab caches into a separate lock-class.
594 * We set lock class for alien array caches which are up during init.
595 * The lock annotation will be lost if all cpus of a node goes down and
596 * then comes back up during hotplug
598 static struct lock_class_key on_slab_l3_key;
599 static struct lock_class_key on_slab_alc_key;
601 static struct lock_class_key debugobj_l3_key;
602 static struct lock_class_key debugobj_alc_key;
604 static void slab_set_lock_classes(struct kmem_cache *cachep,
605 struct lock_class_key *l3_key, struct lock_class_key *alc_key,
608 struct array_cache **alc;
609 struct kmem_list3 *l3;
612 l3 = cachep->nodelists[q];
616 lockdep_set_class(&l3->list_lock, l3_key);
619 * FIXME: This check for BAD_ALIEN_MAGIC
620 * should go away when common slab code is taught to
621 * work even without alien caches.
622 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
623 * for alloc_alien_cache,
625 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
629 lockdep_set_class(&alc[r]->lock, alc_key);
633 static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
635 slab_set_lock_classes(cachep, &debugobj_l3_key, &debugobj_alc_key, node);
638 static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
642 for_each_online_node(node)
643 slab_set_debugobj_lock_classes_node(cachep, node);
646 static void init_node_lock_keys(int q)
648 struct cache_sizes *s = malloc_sizes;
653 for (s = malloc_sizes; s->cs_size != ULONG_MAX; s++) {
654 struct kmem_list3 *l3;
656 l3 = s->cs_cachep->nodelists[q];
657 if (!l3 || OFF_SLAB(s->cs_cachep))
660 slab_set_lock_classes(s->cs_cachep, &on_slab_l3_key,
661 &on_slab_alc_key, q);
665 static inline void init_lock_keys(void)
670 init_node_lock_keys(node);
673 static void init_node_lock_keys(int q)
677 static inline void init_lock_keys(void)
681 static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
685 static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
690 static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
692 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
694 return cachep->array[smp_processor_id()];
697 static inline struct kmem_cache *__find_general_cachep(size_t size,
700 struct cache_sizes *csizep = malloc_sizes;
703 /* This happens if someone tries to call
704 * kmem_cache_create(), or __kmalloc(), before
705 * the generic caches are initialized.
707 BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL);
710 return ZERO_SIZE_PTR;
712 while (size > csizep->cs_size)
716 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
717 * has cs_{dma,}cachep==NULL. Thus no special case
718 * for large kmalloc calls required.
720 #ifdef CONFIG_ZONE_DMA
721 if (unlikely(gfpflags & GFP_DMA))
722 return csizep->cs_dmacachep;
724 return csizep->cs_cachep;
727 static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
729 return __find_general_cachep(size, gfpflags);
732 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
734 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
738 * Calculate the number of objects and left-over bytes for a given buffer size.
740 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
741 size_t align, int flags, size_t *left_over,
746 size_t slab_size = PAGE_SIZE << gfporder;
749 * The slab management structure can be either off the slab or
750 * on it. For the latter case, the memory allocated for a
754 * - One kmem_bufctl_t for each object
755 * - Padding to respect alignment of @align
756 * - @buffer_size bytes for each object
758 * If the slab management structure is off the slab, then the
759 * alignment will already be calculated into the size. Because
760 * the slabs are all pages aligned, the objects will be at the
761 * correct alignment when allocated.
763 if (flags & CFLGS_OFF_SLAB) {
765 nr_objs = slab_size / buffer_size;
767 if (nr_objs > SLAB_LIMIT)
768 nr_objs = SLAB_LIMIT;
771 * Ignore padding for the initial guess. The padding
772 * is at most @align-1 bytes, and @buffer_size is at
773 * least @align. In the worst case, this result will
774 * be one greater than the number of objects that fit
775 * into the memory allocation when taking the padding
778 nr_objs = (slab_size - sizeof(struct slab)) /
779 (buffer_size + sizeof(kmem_bufctl_t));
782 * This calculated number will be either the right
783 * amount, or one greater than what we want.
785 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
789 if (nr_objs > SLAB_LIMIT)
790 nr_objs = SLAB_LIMIT;
792 mgmt_size = slab_mgmt_size(nr_objs, align);
795 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
799 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
801 static void __slab_error(const char *function, struct kmem_cache *cachep,
804 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
805 function, cachep->name, msg);
807 add_taint(TAINT_BAD_PAGE);
812 * By default on NUMA we use alien caches to stage the freeing of
813 * objects allocated from other nodes. This causes massive memory
814 * inefficiencies when using fake NUMA setup to split memory into a
815 * large number of small nodes, so it can be disabled on the command
819 static int use_alien_caches __read_mostly = 1;
820 static int __init noaliencache_setup(char *s)
822 use_alien_caches = 0;
825 __setup("noaliencache", noaliencache_setup);
827 static int __init slab_max_order_setup(char *str)
829 get_option(&str, &slab_max_order);
830 slab_max_order = slab_max_order < 0 ? 0 :
831 min(slab_max_order, MAX_ORDER - 1);
832 slab_max_order_set = true;
836 __setup("slab_max_order=", slab_max_order_setup);
840 * Special reaping functions for NUMA systems called from cache_reap().
841 * These take care of doing round robin flushing of alien caches (containing
842 * objects freed on different nodes from which they were allocated) and the
843 * flushing of remote pcps by calling drain_node_pages.
845 static DEFINE_PER_CPU(unsigned long, slab_reap_node);
847 static void init_reap_node(int cpu)
851 node = next_node(cpu_to_mem(cpu), node_online_map);
852 if (node == MAX_NUMNODES)
853 node = first_node(node_online_map);
855 per_cpu(slab_reap_node, cpu) = node;
858 static void next_reap_node(void)
860 int node = __this_cpu_read(slab_reap_node);
862 node = next_node(node, node_online_map);
863 if (unlikely(node >= MAX_NUMNODES))
864 node = first_node(node_online_map);
865 __this_cpu_write(slab_reap_node, node);
869 #define init_reap_node(cpu) do { } while (0)
870 #define next_reap_node(void) do { } while (0)
874 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
875 * via the workqueue/eventd.
876 * Add the CPU number into the expiration time to minimize the possibility of
877 * the CPUs getting into lockstep and contending for the global cache chain
880 static void __cpuinit start_cpu_timer(int cpu)
882 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
885 * When this gets called from do_initcalls via cpucache_init(),
886 * init_workqueues() has already run, so keventd will be setup
889 if (keventd_up() && reap_work->work.func == NULL) {
891 INIT_DEFERRABLE_WORK(reap_work, cache_reap);
892 schedule_delayed_work_on(cpu, reap_work,
893 __round_jiffies_relative(HZ, cpu));
897 static struct array_cache *alloc_arraycache(int node, int entries,
898 int batchcount, gfp_t gfp)
900 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
901 struct array_cache *nc = NULL;
903 nc = kmalloc_node(memsize, gfp, node);
905 * The array_cache structures contain pointers to free object.
906 * However, when such objects are allocated or transferred to another
907 * cache the pointers are not cleared and they could be counted as
908 * valid references during a kmemleak scan. Therefore, kmemleak must
909 * not scan such objects.
911 kmemleak_no_scan(nc);
915 nc->batchcount = batchcount;
917 spin_lock_init(&nc->lock);
922 static inline bool is_slab_pfmemalloc(struct slab *slabp)
924 struct page *page = virt_to_page(slabp->s_mem);
926 return PageSlabPfmemalloc(page);
929 /* Clears pfmemalloc_active if no slabs have pfmalloc set */
930 static void recheck_pfmemalloc_active(struct kmem_cache *cachep,
931 struct array_cache *ac)
933 struct kmem_list3 *l3 = cachep->nodelists[numa_mem_id()];
937 if (!pfmemalloc_active)
940 spin_lock_irqsave(&l3->list_lock, flags);
941 list_for_each_entry(slabp, &l3->slabs_full, list)
942 if (is_slab_pfmemalloc(slabp))
945 list_for_each_entry(slabp, &l3->slabs_partial, list)
946 if (is_slab_pfmemalloc(slabp))
949 list_for_each_entry(slabp, &l3->slabs_free, list)
950 if (is_slab_pfmemalloc(slabp))
953 pfmemalloc_active = false;
955 spin_unlock_irqrestore(&l3->list_lock, flags);
958 static void *__ac_get_obj(struct kmem_cache *cachep, struct array_cache *ac,
959 gfp_t flags, bool force_refill)
962 void *objp = ac->entry[--ac->avail];
964 /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
965 if (unlikely(is_obj_pfmemalloc(objp))) {
966 struct kmem_list3 *l3;
968 if (gfp_pfmemalloc_allowed(flags)) {
969 clear_obj_pfmemalloc(&objp);
973 /* The caller cannot use PFMEMALLOC objects, find another one */
974 for (i = 0; i < ac->avail; i++) {
975 /* If a !PFMEMALLOC object is found, swap them */
976 if (!is_obj_pfmemalloc(ac->entry[i])) {
978 ac->entry[i] = ac->entry[ac->avail];
979 ac->entry[ac->avail] = objp;
985 * If there are empty slabs on the slabs_free list and we are
986 * being forced to refill the cache, mark this one !pfmemalloc.
988 l3 = cachep->nodelists[numa_mem_id()];
989 if (!list_empty(&l3->slabs_free) && force_refill) {
990 struct slab *slabp = virt_to_slab(objp);
991 ClearPageSlabPfmemalloc(virt_to_head_page(slabp->s_mem));
992 clear_obj_pfmemalloc(&objp);
993 recheck_pfmemalloc_active(cachep, ac);
997 /* No !PFMEMALLOC objects available */
1005 static inline void *ac_get_obj(struct kmem_cache *cachep,
1006 struct array_cache *ac, gfp_t flags, bool force_refill)
1010 if (unlikely(sk_memalloc_socks()))
1011 objp = __ac_get_obj(cachep, ac, flags, force_refill);
1013 objp = ac->entry[--ac->avail];
1018 static void *__ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
1021 if (unlikely(pfmemalloc_active)) {
1022 /* Some pfmemalloc slabs exist, check if this is one */
1023 struct page *page = virt_to_head_page(objp);
1024 if (PageSlabPfmemalloc(page))
1025 set_obj_pfmemalloc(&objp);
1031 static inline void ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
1034 if (unlikely(sk_memalloc_socks()))
1035 objp = __ac_put_obj(cachep, ac, objp);
1037 ac->entry[ac->avail++] = objp;
1041 * Transfer objects in one arraycache to another.
1042 * Locking must be handled by the caller.
1044 * Return the number of entries transferred.
1046 static int transfer_objects(struct array_cache *to,
1047 struct array_cache *from, unsigned int max)
1049 /* Figure out how many entries to transfer */
1050 int nr = min3(from->avail, max, to->limit - to->avail);
1055 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
1056 sizeof(void *) *nr);
1065 #define drain_alien_cache(cachep, alien) do { } while (0)
1066 #define reap_alien(cachep, l3) do { } while (0)
1068 static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
1070 return (struct array_cache **)BAD_ALIEN_MAGIC;
1073 static inline void free_alien_cache(struct array_cache **ac_ptr)
1077 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1082 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
1088 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
1089 gfp_t flags, int nodeid)
1094 #else /* CONFIG_NUMA */
1096 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
1097 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
1099 static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
1101 struct array_cache **ac_ptr;
1102 int memsize = sizeof(void *) * nr_node_ids;
1107 ac_ptr = kzalloc_node(memsize, gfp, node);
1110 if (i == node || !node_online(i))
1112 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
1114 for (i--; i >= 0; i--)
1124 static void free_alien_cache(struct array_cache **ac_ptr)
1135 static void __drain_alien_cache(struct kmem_cache *cachep,
1136 struct array_cache *ac, int node)
1138 struct kmem_list3 *rl3 = cachep->nodelists[node];
1141 spin_lock(&rl3->list_lock);
1143 * Stuff objects into the remote nodes shared array first.
1144 * That way we could avoid the overhead of putting the objects
1145 * into the free lists and getting them back later.
1148 transfer_objects(rl3->shared, ac, ac->limit);
1150 free_block(cachep, ac->entry, ac->avail, node);
1152 spin_unlock(&rl3->list_lock);
1157 * Called from cache_reap() to regularly drain alien caches round robin.
1159 static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1161 int node = __this_cpu_read(slab_reap_node);
1164 struct array_cache *ac = l3->alien[node];
1166 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1167 __drain_alien_cache(cachep, ac, node);
1168 spin_unlock_irq(&ac->lock);
1173 static void drain_alien_cache(struct kmem_cache *cachep,
1174 struct array_cache **alien)
1177 struct array_cache *ac;
1178 unsigned long flags;
1180 for_each_online_node(i) {
1183 spin_lock_irqsave(&ac->lock, flags);
1184 __drain_alien_cache(cachep, ac, i);
1185 spin_unlock_irqrestore(&ac->lock, flags);
1190 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1192 struct slab *slabp = virt_to_slab(objp);
1193 int nodeid = slabp->nodeid;
1194 struct kmem_list3 *l3;
1195 struct array_cache *alien = NULL;
1198 node = numa_mem_id();
1201 * Make sure we are not freeing a object from another node to the array
1202 * cache on this cpu.
1204 if (likely(slabp->nodeid == node))
1207 l3 = cachep->nodelists[node];
1208 STATS_INC_NODEFREES(cachep);
1209 if (l3->alien && l3->alien[nodeid]) {
1210 alien = l3->alien[nodeid];
1211 spin_lock(&alien->lock);
1212 if (unlikely(alien->avail == alien->limit)) {
1213 STATS_INC_ACOVERFLOW(cachep);
1214 __drain_alien_cache(cachep, alien, nodeid);
1216 ac_put_obj(cachep, alien, objp);
1217 spin_unlock(&alien->lock);
1219 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1220 free_block(cachep, &objp, 1, nodeid);
1221 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1228 * Allocates and initializes nodelists for a node on each slab cache, used for
1229 * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3
1230 * will be allocated off-node since memory is not yet online for the new node.
1231 * When hotplugging memory or a cpu, existing nodelists are not replaced if
1234 * Must hold slab_mutex.
1236 static int init_cache_nodelists_node(int node)
1238 struct kmem_cache *cachep;
1239 struct kmem_list3 *l3;
1240 const int memsize = sizeof(struct kmem_list3);
1242 list_for_each_entry(cachep, &slab_caches, list) {
1244 * Set up the size64 kmemlist for cpu before we can
1245 * begin anything. Make sure some other cpu on this
1246 * node has not already allocated this
1248 if (!cachep->nodelists[node]) {
1249 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1252 kmem_list3_init(l3);
1253 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1254 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1257 * The l3s don't come and go as CPUs come and
1258 * go. slab_mutex is sufficient
1261 cachep->nodelists[node] = l3;
1264 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1265 cachep->nodelists[node]->free_limit =
1266 (1 + nr_cpus_node(node)) *
1267 cachep->batchcount + cachep->num;
1268 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1273 static void __cpuinit cpuup_canceled(long cpu)
1275 struct kmem_cache *cachep;
1276 struct kmem_list3 *l3 = NULL;
1277 int node = cpu_to_mem(cpu);
1278 const struct cpumask *mask = cpumask_of_node(node);
1280 list_for_each_entry(cachep, &slab_caches, list) {
1281 struct array_cache *nc;
1282 struct array_cache *shared;
1283 struct array_cache **alien;
1285 /* cpu is dead; no one can alloc from it. */
1286 nc = cachep->array[cpu];
1287 cachep->array[cpu] = NULL;
1288 l3 = cachep->nodelists[node];
1291 goto free_array_cache;
1293 spin_lock_irq(&l3->list_lock);
1295 /* Free limit for this kmem_list3 */
1296 l3->free_limit -= cachep->batchcount;
1298 free_block(cachep, nc->entry, nc->avail, node);
1300 if (!cpumask_empty(mask)) {
1301 spin_unlock_irq(&l3->list_lock);
1302 goto free_array_cache;
1305 shared = l3->shared;
1307 free_block(cachep, shared->entry,
1308 shared->avail, node);
1315 spin_unlock_irq(&l3->list_lock);
1319 drain_alien_cache(cachep, alien);
1320 free_alien_cache(alien);
1326 * In the previous loop, all the objects were freed to
1327 * the respective cache's slabs, now we can go ahead and
1328 * shrink each nodelist to its limit.
1330 list_for_each_entry(cachep, &slab_caches, list) {
1331 l3 = cachep->nodelists[node];
1334 drain_freelist(cachep, l3, l3->free_objects);
1338 static int __cpuinit cpuup_prepare(long cpu)
1340 struct kmem_cache *cachep;
1341 struct kmem_list3 *l3 = NULL;
1342 int node = cpu_to_mem(cpu);
1346 * We need to do this right in the beginning since
1347 * alloc_arraycache's are going to use this list.
1348 * kmalloc_node allows us to add the slab to the right
1349 * kmem_list3 and not this cpu's kmem_list3
1351 err = init_cache_nodelists_node(node);
1356 * Now we can go ahead with allocating the shared arrays and
1359 list_for_each_entry(cachep, &slab_caches, list) {
1360 struct array_cache *nc;
1361 struct array_cache *shared = NULL;
1362 struct array_cache **alien = NULL;
1364 nc = alloc_arraycache(node, cachep->limit,
1365 cachep->batchcount, GFP_KERNEL);
1368 if (cachep->shared) {
1369 shared = alloc_arraycache(node,
1370 cachep->shared * cachep->batchcount,
1371 0xbaadf00d, GFP_KERNEL);
1377 if (use_alien_caches) {
1378 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
1385 cachep->array[cpu] = nc;
1386 l3 = cachep->nodelists[node];
1389 spin_lock_irq(&l3->list_lock);
1392 * We are serialised from CPU_DEAD or
1393 * CPU_UP_CANCELLED by the cpucontrol lock
1395 l3->shared = shared;
1404 spin_unlock_irq(&l3->list_lock);
1406 free_alien_cache(alien);
1407 if (cachep->flags & SLAB_DEBUG_OBJECTS)
1408 slab_set_debugobj_lock_classes_node(cachep, node);
1410 init_node_lock_keys(node);
1414 cpuup_canceled(cpu);
1418 static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1419 unsigned long action, void *hcpu)
1421 long cpu = (long)hcpu;
1425 case CPU_UP_PREPARE:
1426 case CPU_UP_PREPARE_FROZEN:
1427 mutex_lock(&slab_mutex);
1428 err = cpuup_prepare(cpu);
1429 mutex_unlock(&slab_mutex);
1432 case CPU_ONLINE_FROZEN:
1433 start_cpu_timer(cpu);
1435 #ifdef CONFIG_HOTPLUG_CPU
1436 case CPU_DOWN_PREPARE:
1437 case CPU_DOWN_PREPARE_FROZEN:
1439 * Shutdown cache reaper. Note that the slab_mutex is
1440 * held so that if cache_reap() is invoked it cannot do
1441 * anything expensive but will only modify reap_work
1442 * and reschedule the timer.
1444 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1445 /* Now the cache_reaper is guaranteed to be not running. */
1446 per_cpu(slab_reap_work, cpu).work.func = NULL;
1448 case CPU_DOWN_FAILED:
1449 case CPU_DOWN_FAILED_FROZEN:
1450 start_cpu_timer(cpu);
1453 case CPU_DEAD_FROZEN:
1455 * Even if all the cpus of a node are down, we don't free the
1456 * kmem_list3 of any cache. This to avoid a race between
1457 * cpu_down, and a kmalloc allocation from another cpu for
1458 * memory from the node of the cpu going down. The list3
1459 * structure is usually allocated from kmem_cache_create() and
1460 * gets destroyed at kmem_cache_destroy().
1464 case CPU_UP_CANCELED:
1465 case CPU_UP_CANCELED_FROZEN:
1466 mutex_lock(&slab_mutex);
1467 cpuup_canceled(cpu);
1468 mutex_unlock(&slab_mutex);
1471 return notifier_from_errno(err);
1474 static struct notifier_block __cpuinitdata cpucache_notifier = {
1475 &cpuup_callback, NULL, 0
1478 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1480 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1481 * Returns -EBUSY if all objects cannot be drained so that the node is not
1484 * Must hold slab_mutex.
1486 static int __meminit drain_cache_nodelists_node(int node)
1488 struct kmem_cache *cachep;
1491 list_for_each_entry(cachep, &slab_caches, list) {
1492 struct kmem_list3 *l3;
1494 l3 = cachep->nodelists[node];
1498 drain_freelist(cachep, l3, l3->free_objects);
1500 if (!list_empty(&l3->slabs_full) ||
1501 !list_empty(&l3->slabs_partial)) {
1509 static int __meminit slab_memory_callback(struct notifier_block *self,
1510 unsigned long action, void *arg)
1512 struct memory_notify *mnb = arg;
1516 nid = mnb->status_change_nid;
1521 case MEM_GOING_ONLINE:
1522 mutex_lock(&slab_mutex);
1523 ret = init_cache_nodelists_node(nid);
1524 mutex_unlock(&slab_mutex);
1526 case MEM_GOING_OFFLINE:
1527 mutex_lock(&slab_mutex);
1528 ret = drain_cache_nodelists_node(nid);
1529 mutex_unlock(&slab_mutex);
1533 case MEM_CANCEL_ONLINE:
1534 case MEM_CANCEL_OFFLINE:
1538 return notifier_from_errno(ret);
1540 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1543 * swap the static kmem_list3 with kmalloced memory
1545 static void __init init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1548 struct kmem_list3 *ptr;
1550 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid);
1553 memcpy(ptr, list, sizeof(struct kmem_list3));
1555 * Do not assume that spinlocks can be initialized via memcpy:
1557 spin_lock_init(&ptr->list_lock);
1559 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1560 cachep->nodelists[nodeid] = ptr;
1564 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1565 * size of kmem_list3.
1567 static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1571 for_each_online_node(node) {
1572 cachep->nodelists[node] = &initkmem_list3[index + node];
1573 cachep->nodelists[node]->next_reap = jiffies +
1575 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1580 * Initialisation. Called after the page allocator have been initialised and
1581 * before smp_init().
1583 void __init kmem_cache_init(void)
1586 struct cache_sizes *sizes;
1587 struct cache_names *names;
1592 kmem_cache = &kmem_cache_boot;
1594 if (num_possible_nodes() == 1)
1595 use_alien_caches = 0;
1597 for (i = 0; i < NUM_INIT_LISTS; i++) {
1598 kmem_list3_init(&initkmem_list3[i]);
1599 if (i < MAX_NUMNODES)
1600 kmem_cache->nodelists[i] = NULL;
1602 set_up_list3s(kmem_cache, CACHE_CACHE);
1605 * Fragmentation resistance on low memory - only use bigger
1606 * page orders on machines with more than 32MB of memory if
1607 * not overridden on the command line.
1609 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
1610 slab_max_order = SLAB_MAX_ORDER_HI;
1612 /* Bootstrap is tricky, because several objects are allocated
1613 * from caches that do not exist yet:
1614 * 1) initialize the kmem_cache cache: it contains the struct
1615 * kmem_cache structures of all caches, except kmem_cache itself:
1616 * kmem_cache is statically allocated.
1617 * Initially an __init data area is used for the head array and the
1618 * kmem_list3 structures, it's replaced with a kmalloc allocated
1619 * array at the end of the bootstrap.
1620 * 2) Create the first kmalloc cache.
1621 * The struct kmem_cache for the new cache is allocated normally.
1622 * An __init data area is used for the head array.
1623 * 3) Create the remaining kmalloc caches, with minimally sized
1625 * 4) Replace the __init data head arrays for kmem_cache and the first
1626 * kmalloc cache with kmalloc allocated arrays.
1627 * 5) Replace the __init data for kmem_list3 for kmem_cache and
1628 * the other cache's with kmalloc allocated memory.
1629 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1632 node = numa_mem_id();
1634 /* 1) create the kmem_cache */
1635 INIT_LIST_HEAD(&slab_caches);
1636 list_add(&kmem_cache->list, &slab_caches);
1637 kmem_cache->colour_off = cache_line_size();
1638 kmem_cache->array[smp_processor_id()] = &initarray_cache.cache;
1639 kmem_cache->nodelists[node] = &initkmem_list3[CACHE_CACHE + node];
1642 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1644 kmem_cache->size = offsetof(struct kmem_cache, array[nr_cpu_ids]) +
1645 nr_node_ids * sizeof(struct kmem_list3 *);
1646 kmem_cache->object_size = kmem_cache->size;
1647 kmem_cache->size = ALIGN(kmem_cache->object_size,
1649 kmem_cache->reciprocal_buffer_size =
1650 reciprocal_value(kmem_cache->size);
1652 for (order = 0; order < MAX_ORDER; order++) {
1653 cache_estimate(order, kmem_cache->size,
1654 cache_line_size(), 0, &left_over, &kmem_cache->num);
1655 if (kmem_cache->num)
1658 BUG_ON(!kmem_cache->num);
1659 kmem_cache->gfporder = order;
1660 kmem_cache->colour = left_over / kmem_cache->colour_off;
1661 kmem_cache->slab_size = ALIGN(kmem_cache->num * sizeof(kmem_bufctl_t) +
1662 sizeof(struct slab), cache_line_size());
1664 /* 2+3) create the kmalloc caches */
1665 sizes = malloc_sizes;
1666 names = cache_names;
1669 * Initialize the caches that provide memory for the array cache and the
1670 * kmem_list3 structures first. Without this, further allocations will
1674 sizes[INDEX_AC].cs_cachep = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
1675 sizes[INDEX_AC].cs_cachep->name = names[INDEX_AC].name;
1676 sizes[INDEX_AC].cs_cachep->size = sizes[INDEX_AC].cs_size;
1677 sizes[INDEX_AC].cs_cachep->object_size = sizes[INDEX_AC].cs_size;
1678 sizes[INDEX_AC].cs_cachep->align = ARCH_KMALLOC_MINALIGN;
1679 __kmem_cache_create(sizes[INDEX_AC].cs_cachep, ARCH_KMALLOC_FLAGS|SLAB_PANIC);
1680 list_add(&sizes[INDEX_AC].cs_cachep->list, &slab_caches);
1682 if (INDEX_AC != INDEX_L3) {
1683 sizes[INDEX_L3].cs_cachep = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
1684 sizes[INDEX_L3].cs_cachep->name = names[INDEX_L3].name;
1685 sizes[INDEX_L3].cs_cachep->size = sizes[INDEX_L3].cs_size;
1686 sizes[INDEX_L3].cs_cachep->object_size = sizes[INDEX_L3].cs_size;
1687 sizes[INDEX_L3].cs_cachep->align = ARCH_KMALLOC_MINALIGN;
1688 __kmem_cache_create(sizes[INDEX_L3].cs_cachep, ARCH_KMALLOC_FLAGS|SLAB_PANIC);
1689 list_add(&sizes[INDEX_L3].cs_cachep->list, &slab_caches);
1692 slab_early_init = 0;
1694 while (sizes->cs_size != ULONG_MAX) {
1696 * For performance, all the general caches are L1 aligned.
1697 * This should be particularly beneficial on SMP boxes, as it
1698 * eliminates "false sharing".
1699 * Note for systems short on memory removing the alignment will
1700 * allow tighter packing of the smaller caches.
1702 if (!sizes->cs_cachep) {
1703 sizes->cs_cachep = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
1704 sizes->cs_cachep->name = names->name;
1705 sizes->cs_cachep->size = sizes->cs_size;
1706 sizes->cs_cachep->object_size = sizes->cs_size;
1707 sizes->cs_cachep->align = ARCH_KMALLOC_MINALIGN;
1708 __kmem_cache_create(sizes->cs_cachep, ARCH_KMALLOC_FLAGS|SLAB_PANIC);
1709 list_add(&sizes->cs_cachep->list, &slab_caches);
1711 #ifdef CONFIG_ZONE_DMA
1712 sizes->cs_dmacachep = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
1713 sizes->cs_dmacachep->name = names->name_dma;
1714 sizes->cs_dmacachep->size = sizes->cs_size;
1715 sizes->cs_dmacachep->object_size = sizes->cs_size;
1716 sizes->cs_dmacachep->align = ARCH_KMALLOC_MINALIGN;
1717 __kmem_cache_create(sizes->cs_dmacachep,
1718 ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA| SLAB_PANIC);
1719 list_add(&sizes->cs_dmacachep->list, &slab_caches);
1724 /* 4) Replace the bootstrap head arrays */
1726 struct array_cache *ptr;
1728 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1730 BUG_ON(cpu_cache_get(kmem_cache) != &initarray_cache.cache);
1731 memcpy(ptr, cpu_cache_get(kmem_cache),
1732 sizeof(struct arraycache_init));
1734 * Do not assume that spinlocks can be initialized via memcpy:
1736 spin_lock_init(&ptr->lock);
1738 kmem_cache->array[smp_processor_id()] = ptr;
1740 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1742 BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep)
1743 != &initarray_generic.cache);
1744 memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep),
1745 sizeof(struct arraycache_init));
1747 * Do not assume that spinlocks can be initialized via memcpy:
1749 spin_lock_init(&ptr->lock);
1751 malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] =
1754 /* 5) Replace the bootstrap kmem_list3's */
1758 for_each_online_node(nid) {
1759 init_list(kmem_cache, &initkmem_list3[CACHE_CACHE + nid], nid);
1761 init_list(malloc_sizes[INDEX_AC].cs_cachep,
1762 &initkmem_list3[SIZE_AC + nid], nid);
1764 if (INDEX_AC != INDEX_L3) {
1765 init_list(malloc_sizes[INDEX_L3].cs_cachep,
1766 &initkmem_list3[SIZE_L3 + nid], nid);
1774 void __init kmem_cache_init_late(void)
1776 struct kmem_cache *cachep;
1780 /* 6) resize the head arrays to their final sizes */
1781 mutex_lock(&slab_mutex);
1782 list_for_each_entry(cachep, &slab_caches, list)
1783 if (enable_cpucache(cachep, GFP_NOWAIT))
1785 mutex_unlock(&slab_mutex);
1787 /* Annotate slab for lockdep -- annotate the malloc caches */
1794 * Register a cpu startup notifier callback that initializes
1795 * cpu_cache_get for all new cpus
1797 register_cpu_notifier(&cpucache_notifier);
1801 * Register a memory hotplug callback that initializes and frees
1804 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1808 * The reap timers are started later, with a module init call: That part
1809 * of the kernel is not yet operational.
1813 static int __init cpucache_init(void)
1818 * Register the timers that return unneeded pages to the page allocator
1820 for_each_online_cpu(cpu)
1821 start_cpu_timer(cpu);
1827 __initcall(cpucache_init);
1829 static noinline void
1830 slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1832 struct kmem_list3 *l3;
1834 unsigned long flags;
1838 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1840 printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n",
1841 cachep->name, cachep->size, cachep->gfporder);
1843 for_each_online_node(node) {
1844 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1845 unsigned long active_slabs = 0, num_slabs = 0;
1847 l3 = cachep->nodelists[node];
1851 spin_lock_irqsave(&l3->list_lock, flags);
1852 list_for_each_entry(slabp, &l3->slabs_full, list) {
1853 active_objs += cachep->num;
1856 list_for_each_entry(slabp, &l3->slabs_partial, list) {
1857 active_objs += slabp->inuse;
1860 list_for_each_entry(slabp, &l3->slabs_free, list)
1863 free_objects += l3->free_objects;
1864 spin_unlock_irqrestore(&l3->list_lock, flags);
1866 num_slabs += active_slabs;
1867 num_objs = num_slabs * cachep->num;
1869 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1870 node, active_slabs, num_slabs, active_objs, num_objs,
1876 * Interface to system's page allocator. No need to hold the cache-lock.
1878 * If we requested dmaable memory, we will get it. Even if we
1879 * did not request dmaable memory, we might get it, but that
1880 * would be relatively rare and ignorable.
1882 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1890 * Nommu uses slab's for process anonymous memory allocations, and thus
1891 * requires __GFP_COMP to properly refcount higher order allocations
1893 flags |= __GFP_COMP;
1896 flags |= cachep->allocflags;
1897 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1898 flags |= __GFP_RECLAIMABLE;
1900 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1902 if (!(flags & __GFP_NOWARN) && printk_ratelimit())
1903 slab_out_of_memory(cachep, flags, nodeid);
1907 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1908 if (unlikely(page->pfmemalloc))
1909 pfmemalloc_active = true;
1911 nr_pages = (1 << cachep->gfporder);
1912 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1913 add_zone_page_state(page_zone(page),
1914 NR_SLAB_RECLAIMABLE, nr_pages);
1916 add_zone_page_state(page_zone(page),
1917 NR_SLAB_UNRECLAIMABLE, nr_pages);
1918 for (i = 0; i < nr_pages; i++) {
1919 __SetPageSlab(page + i);
1921 if (page->pfmemalloc)
1922 SetPageSlabPfmemalloc(page + i);
1925 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1926 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1929 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1931 kmemcheck_mark_unallocated_pages(page, nr_pages);
1934 return page_address(page);
1938 * Interface to system's page release.
1940 static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1942 unsigned long i = (1 << cachep->gfporder);
1943 struct page *page = virt_to_page(addr);
1944 const unsigned long nr_freed = i;
1946 kmemcheck_free_shadow(page, cachep->gfporder);
1948 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1949 sub_zone_page_state(page_zone(page),
1950 NR_SLAB_RECLAIMABLE, nr_freed);
1952 sub_zone_page_state(page_zone(page),
1953 NR_SLAB_UNRECLAIMABLE, nr_freed);
1955 BUG_ON(!PageSlab(page));
1956 __ClearPageSlabPfmemalloc(page);
1957 __ClearPageSlab(page);
1960 if (current->reclaim_state)
1961 current->reclaim_state->reclaimed_slab += nr_freed;
1962 free_pages((unsigned long)addr, cachep->gfporder);
1965 static void kmem_rcu_free(struct rcu_head *head)
1967 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1968 struct kmem_cache *cachep = slab_rcu->cachep;
1970 kmem_freepages(cachep, slab_rcu->addr);
1971 if (OFF_SLAB(cachep))
1972 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1977 #ifdef CONFIG_DEBUG_PAGEALLOC
1978 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1979 unsigned long caller)
1981 int size = cachep->object_size;
1983 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1985 if (size < 5 * sizeof(unsigned long))
1988 *addr++ = 0x12345678;
1990 *addr++ = smp_processor_id();
1991 size -= 3 * sizeof(unsigned long);
1993 unsigned long *sptr = &caller;
1994 unsigned long svalue;
1996 while (!kstack_end(sptr)) {
1998 if (kernel_text_address(svalue)) {
2000 size -= sizeof(unsigned long);
2001 if (size <= sizeof(unsigned long))
2007 *addr++ = 0x87654321;
2011 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
2013 int size = cachep->object_size;
2014 addr = &((char *)addr)[obj_offset(cachep)];
2016 memset(addr, val, size);
2017 *(unsigned char *)(addr + size - 1) = POISON_END;
2020 static void dump_line(char *data, int offset, int limit)
2023 unsigned char error = 0;
2026 printk(KERN_ERR "%03x: ", offset);
2027 for (i = 0; i < limit; i++) {
2028 if (data[offset + i] != POISON_FREE) {
2029 error = data[offset + i];
2033 print_hex_dump(KERN_CONT, "", 0, 16, 1,
2034 &data[offset], limit, 1);
2036 if (bad_count == 1) {
2037 error ^= POISON_FREE;
2038 if (!(error & (error - 1))) {
2039 printk(KERN_ERR "Single bit error detected. Probably "
2042 printk(KERN_ERR "Run memtest86+ or a similar memory "
2045 printk(KERN_ERR "Run a memory test tool.\n");
2054 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
2059 if (cachep->flags & SLAB_RED_ZONE) {
2060 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
2061 *dbg_redzone1(cachep, objp),
2062 *dbg_redzone2(cachep, objp));
2065 if (cachep->flags & SLAB_STORE_USER) {
2066 printk(KERN_ERR "Last user: [<%p>]",
2067 *dbg_userword(cachep, objp));
2068 print_symbol("(%s)",
2069 (unsigned long)*dbg_userword(cachep, objp));
2072 realobj = (char *)objp + obj_offset(cachep);
2073 size = cachep->object_size;
2074 for (i = 0; i < size && lines; i += 16, lines--) {
2077 if (i + limit > size)
2079 dump_line(realobj, i, limit);
2083 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
2089 realobj = (char *)objp + obj_offset(cachep);
2090 size = cachep->object_size;
2092 for (i = 0; i < size; i++) {
2093 char exp = POISON_FREE;
2096 if (realobj[i] != exp) {
2102 "Slab corruption (%s): %s start=%p, len=%d\n",
2103 print_tainted(), cachep->name, realobj, size);
2104 print_objinfo(cachep, objp, 0);
2106 /* Hexdump the affected line */
2109 if (i + limit > size)
2111 dump_line(realobj, i, limit);
2114 /* Limit to 5 lines */
2120 /* Print some data about the neighboring objects, if they
2123 struct slab *slabp = virt_to_slab(objp);
2126 objnr = obj_to_index(cachep, slabp, objp);
2128 objp = index_to_obj(cachep, slabp, objnr - 1);
2129 realobj = (char *)objp + obj_offset(cachep);
2130 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
2132 print_objinfo(cachep, objp, 2);
2134 if (objnr + 1 < cachep->num) {
2135 objp = index_to_obj(cachep, slabp, objnr + 1);
2136 realobj = (char *)objp + obj_offset(cachep);
2137 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
2139 print_objinfo(cachep, objp, 2);
2146 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
2149 for (i = 0; i < cachep->num; i++) {
2150 void *objp = index_to_obj(cachep, slabp, i);
2152 if (cachep->flags & SLAB_POISON) {
2153 #ifdef CONFIG_DEBUG_PAGEALLOC
2154 if (cachep->size % PAGE_SIZE == 0 &&
2156 kernel_map_pages(virt_to_page(objp),
2157 cachep->size / PAGE_SIZE, 1);
2159 check_poison_obj(cachep, objp);
2161 check_poison_obj(cachep, objp);
2164 if (cachep->flags & SLAB_RED_ZONE) {
2165 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2166 slab_error(cachep, "start of a freed object "
2168 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2169 slab_error(cachep, "end of a freed object "
2175 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
2181 * slab_destroy - destroy and release all objects in a slab
2182 * @cachep: cache pointer being destroyed
2183 * @slabp: slab pointer being destroyed
2185 * Destroy all the objs in a slab, and release the mem back to the system.
2186 * Before calling the slab must have been unlinked from the cache. The
2187 * cache-lock is not held/needed.
2189 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
2191 void *addr = slabp->s_mem - slabp->colouroff;
2193 slab_destroy_debugcheck(cachep, slabp);
2194 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
2195 struct slab_rcu *slab_rcu;
2197 slab_rcu = (struct slab_rcu *)slabp;
2198 slab_rcu->cachep = cachep;
2199 slab_rcu->addr = addr;
2200 call_rcu(&slab_rcu->head, kmem_rcu_free);
2202 kmem_freepages(cachep, addr);
2203 if (OFF_SLAB(cachep))
2204 kmem_cache_free(cachep->slabp_cache, slabp);
2209 * calculate_slab_order - calculate size (page order) of slabs
2210 * @cachep: pointer to the cache that is being created
2211 * @size: size of objects to be created in this cache.
2212 * @align: required alignment for the objects.
2213 * @flags: slab allocation flags
2215 * Also calculates the number of objects per slab.
2217 * This could be made much more intelligent. For now, try to avoid using
2218 * high order pages for slabs. When the gfp() functions are more friendly
2219 * towards high-order requests, this should be changed.
2221 static size_t calculate_slab_order(struct kmem_cache *cachep,
2222 size_t size, size_t align, unsigned long flags)
2224 unsigned long offslab_limit;
2225 size_t left_over = 0;
2228 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
2232 cache_estimate(gfporder, size, align, flags, &remainder, &num);
2236 if (flags & CFLGS_OFF_SLAB) {
2238 * Max number of objs-per-slab for caches which
2239 * use off-slab slabs. Needed to avoid a possible
2240 * looping condition in cache_grow().
2242 offslab_limit = size - sizeof(struct slab);
2243 offslab_limit /= sizeof(kmem_bufctl_t);
2245 if (num > offslab_limit)
2249 /* Found something acceptable - save it away */
2251 cachep->gfporder = gfporder;
2252 left_over = remainder;
2255 * A VFS-reclaimable slab tends to have most allocations
2256 * as GFP_NOFS and we really don't want to have to be allocating
2257 * higher-order pages when we are unable to shrink dcache.
2259 if (flags & SLAB_RECLAIM_ACCOUNT)
2263 * Large number of objects is good, but very large slabs are
2264 * currently bad for the gfp()s.
2266 if (gfporder >= slab_max_order)
2270 * Acceptable internal fragmentation?
2272 if (left_over * 8 <= (PAGE_SIZE << gfporder))
2278 static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
2280 if (slab_state >= FULL)
2281 return enable_cpucache(cachep, gfp);
2283 if (slab_state == DOWN) {
2285 * Note: the first kmem_cache_create must create the cache
2286 * that's used by kmalloc(24), otherwise the creation of
2287 * further caches will BUG().
2289 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2292 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2293 * the first cache, then we need to set up all its list3s,
2294 * otherwise the creation of further caches will BUG().
2296 set_up_list3s(cachep, SIZE_AC);
2297 if (INDEX_AC == INDEX_L3)
2298 slab_state = PARTIAL_L3;
2300 slab_state = PARTIAL_ARRAYCACHE;
2302 cachep->array[smp_processor_id()] =
2303 kmalloc(sizeof(struct arraycache_init), gfp);
2305 if (slab_state == PARTIAL_ARRAYCACHE) {
2306 set_up_list3s(cachep, SIZE_L3);
2307 slab_state = PARTIAL_L3;
2310 for_each_online_node(node) {
2311 cachep->nodelists[node] =
2312 kmalloc_node(sizeof(struct kmem_list3),
2314 BUG_ON(!cachep->nodelists[node]);
2315 kmem_list3_init(cachep->nodelists[node]);
2319 cachep->nodelists[numa_mem_id()]->next_reap =
2320 jiffies + REAPTIMEOUT_LIST3 +
2321 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2323 cpu_cache_get(cachep)->avail = 0;
2324 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2325 cpu_cache_get(cachep)->batchcount = 1;
2326 cpu_cache_get(cachep)->touched = 0;
2327 cachep->batchcount = 1;
2328 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2333 * __kmem_cache_create - Create a cache.
2334 * @name: A string which is used in /proc/slabinfo to identify this cache.
2335 * @size: The size of objects to be created in this cache.
2336 * @align: The required alignment for the objects.
2337 * @flags: SLAB flags
2338 * @ctor: A constructor for the objects.
2340 * Returns a ptr to the cache on success, NULL on failure.
2341 * Cannot be called within a int, but can be interrupted.
2342 * The @ctor is run when new pages are allocated by the cache.
2346 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2347 * to catch references to uninitialised memory.
2349 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2350 * for buffer overruns.
2352 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2353 * cacheline. This can be beneficial if you're counting cycles as closely
2357 __kmem_cache_create (struct kmem_cache *cachep, unsigned long flags)
2359 size_t left_over, slab_size, ralign;
2362 size_t size = cachep->size;
2367 * Enable redzoning and last user accounting, except for caches with
2368 * large objects, if the increased size would increase the object size
2369 * above the next power of two: caches with object sizes just above a
2370 * power of two have a significant amount of internal fragmentation.
2372 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2373 2 * sizeof(unsigned long long)))
2374 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2375 if (!(flags & SLAB_DESTROY_BY_RCU))
2376 flags |= SLAB_POISON;
2378 if (flags & SLAB_DESTROY_BY_RCU)
2379 BUG_ON(flags & SLAB_POISON);
2382 * Always checks flags, a caller might be expecting debug support which
2385 BUG_ON(flags & ~CREATE_MASK);
2388 * Check that size is in terms of words. This is needed to avoid
2389 * unaligned accesses for some archs when redzoning is used, and makes
2390 * sure any on-slab bufctl's are also correctly aligned.
2392 if (size & (BYTES_PER_WORD - 1)) {
2393 size += (BYTES_PER_WORD - 1);
2394 size &= ~(BYTES_PER_WORD - 1);
2397 /* calculate the final buffer alignment: */
2399 /* 1) arch recommendation: can be overridden for debug */
2400 if (flags & SLAB_HWCACHE_ALIGN) {
2402 * Default alignment: as specified by the arch code. Except if
2403 * an object is really small, then squeeze multiple objects into
2406 ralign = cache_line_size();
2407 while (size <= ralign / 2)
2410 ralign = BYTES_PER_WORD;
2414 * Redzoning and user store require word alignment or possibly larger.
2415 * Note this will be overridden by architecture or caller mandated
2416 * alignment if either is greater than BYTES_PER_WORD.
2418 if (flags & SLAB_STORE_USER)
2419 ralign = BYTES_PER_WORD;
2421 if (flags & SLAB_RED_ZONE) {
2422 ralign = REDZONE_ALIGN;
2423 /* If redzoning, ensure that the second redzone is suitably
2424 * aligned, by adjusting the object size accordingly. */
2425 size += REDZONE_ALIGN - 1;
2426 size &= ~(REDZONE_ALIGN - 1);
2429 /* 2) arch mandated alignment */
2430 if (ralign < ARCH_SLAB_MINALIGN) {
2431 ralign = ARCH_SLAB_MINALIGN;
2433 /* 3) caller mandated alignment */
2434 if (ralign < cachep->align) {
2435 ralign = cachep->align;
2437 /* disable debug if necessary */
2438 if (ralign > __alignof__(unsigned long long))
2439 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2443 cachep->align = ralign;
2445 if (slab_is_available())
2450 cachep->nodelists = (struct kmem_list3 **)&cachep->array[nr_cpu_ids];
2454 * Both debugging options require word-alignment which is calculated
2457 if (flags & SLAB_RED_ZONE) {
2458 /* add space for red zone words */
2459 cachep->obj_offset += sizeof(unsigned long long);
2460 size += 2 * sizeof(unsigned long long);
2462 if (flags & SLAB_STORE_USER) {
2463 /* user store requires one word storage behind the end of
2464 * the real object. But if the second red zone needs to be
2465 * aligned to 64 bits, we must allow that much space.
2467 if (flags & SLAB_RED_ZONE)
2468 size += REDZONE_ALIGN;
2470 size += BYTES_PER_WORD;
2472 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2473 if (size >= malloc_sizes[INDEX_L3 + 1].cs_size
2474 && cachep->object_size > cache_line_size()
2475 && ALIGN(size, cachep->align) < PAGE_SIZE) {
2476 cachep->obj_offset += PAGE_SIZE - ALIGN(size, cachep->align);
2483 * Determine if the slab management is 'on' or 'off' slab.
2484 * (bootstrapping cannot cope with offslab caches so don't do
2485 * it too early on. Always use on-slab management when
2486 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2488 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
2489 !(flags & SLAB_NOLEAKTRACE))
2491 * Size is large, assume best to place the slab management obj
2492 * off-slab (should allow better packing of objs).
2494 flags |= CFLGS_OFF_SLAB;
2496 size = ALIGN(size, cachep->align);
2498 left_over = calculate_slab_order(cachep, size, cachep->align, flags);
2503 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2504 + sizeof(struct slab), cachep->align);
2507 * If the slab has been placed off-slab, and we have enough space then
2508 * move it on-slab. This is at the expense of any extra colouring.
2510 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2511 flags &= ~CFLGS_OFF_SLAB;
2512 left_over -= slab_size;
2515 if (flags & CFLGS_OFF_SLAB) {
2516 /* really off slab. No need for manual alignment */
2518 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2520 #ifdef CONFIG_PAGE_POISONING
2521 /* If we're going to use the generic kernel_map_pages()
2522 * poisoning, then it's going to smash the contents of
2523 * the redzone and userword anyhow, so switch them off.
2525 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2526 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2530 cachep->colour_off = cache_line_size();
2531 /* Offset must be a multiple of the alignment. */
2532 if (cachep->colour_off < cachep->align)
2533 cachep->colour_off = cachep->align;
2534 cachep->colour = left_over / cachep->colour_off;
2535 cachep->slab_size = slab_size;
2536 cachep->flags = flags;
2537 cachep->allocflags = 0;
2538 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2539 cachep->allocflags |= GFP_DMA;
2540 cachep->size = size;
2541 cachep->reciprocal_buffer_size = reciprocal_value(size);
2543 if (flags & CFLGS_OFF_SLAB) {
2544 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2546 * This is a possibility for one of the malloc_sizes caches.
2547 * But since we go off slab only for object size greater than
2548 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2549 * this should not happen at all.
2550 * But leave a BUG_ON for some lucky dude.
2552 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
2555 err = setup_cpu_cache(cachep, gfp);
2557 __kmem_cache_shutdown(cachep);
2561 if (flags & SLAB_DEBUG_OBJECTS) {
2563 * Would deadlock through slab_destroy()->call_rcu()->
2564 * debug_object_activate()->kmem_cache_alloc().
2566 WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU);
2568 slab_set_debugobj_lock_classes(cachep);
2575 static void check_irq_off(void)
2577 BUG_ON(!irqs_disabled());
2580 static void check_irq_on(void)
2582 BUG_ON(irqs_disabled());
2585 static void check_spinlock_acquired(struct kmem_cache *cachep)
2589 assert_spin_locked(&cachep->nodelists[numa_mem_id()]->list_lock);
2593 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2597 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2602 #define check_irq_off() do { } while(0)
2603 #define check_irq_on() do { } while(0)
2604 #define check_spinlock_acquired(x) do { } while(0)
2605 #define check_spinlock_acquired_node(x, y) do { } while(0)
2608 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2609 struct array_cache *ac,
2610 int force, int node);
2612 static void do_drain(void *arg)
2614 struct kmem_cache *cachep = arg;
2615 struct array_cache *ac;
2616 int node = numa_mem_id();
2619 ac = cpu_cache_get(cachep);
2620 spin_lock(&cachep->nodelists[node]->list_lock);
2621 free_block(cachep, ac->entry, ac->avail, node);
2622 spin_unlock(&cachep->nodelists[node]->list_lock);
2626 static void drain_cpu_caches(struct kmem_cache *cachep)
2628 struct kmem_list3 *l3;
2631 on_each_cpu(do_drain, cachep, 1);
2633 for_each_online_node(node) {
2634 l3 = cachep->nodelists[node];
2635 if (l3 && l3->alien)
2636 drain_alien_cache(cachep, l3->alien);
2639 for_each_online_node(node) {
2640 l3 = cachep->nodelists[node];
2642 drain_array(cachep, l3, l3->shared, 1, node);
2647 * Remove slabs from the list of free slabs.
2648 * Specify the number of slabs to drain in tofree.
2650 * Returns the actual number of slabs released.
2652 static int drain_freelist(struct kmem_cache *cache,
2653 struct kmem_list3 *l3, int tofree)
2655 struct list_head *p;
2660 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
2662 spin_lock_irq(&l3->list_lock);
2663 p = l3->slabs_free.prev;
2664 if (p == &l3->slabs_free) {
2665 spin_unlock_irq(&l3->list_lock);
2669 slabp = list_entry(p, struct slab, list);
2671 BUG_ON(slabp->inuse);
2673 list_del(&slabp->list);
2675 * Safe to drop the lock. The slab is no longer linked
2678 l3->free_objects -= cache->num;
2679 spin_unlock_irq(&l3->list_lock);
2680 slab_destroy(cache, slabp);
2687 /* Called with slab_mutex held to protect against cpu hotplug */
2688 static int __cache_shrink(struct kmem_cache *cachep)
2691 struct kmem_list3 *l3;
2693 drain_cpu_caches(cachep);
2696 for_each_online_node(i) {
2697 l3 = cachep->nodelists[i];
2701 drain_freelist(cachep, l3, l3->free_objects);
2703 ret += !list_empty(&l3->slabs_full) ||
2704 !list_empty(&l3->slabs_partial);
2706 return (ret ? 1 : 0);
2710 * kmem_cache_shrink - Shrink a cache.
2711 * @cachep: The cache to shrink.
2713 * Releases as many slabs as possible for a cache.
2714 * To help debugging, a zero exit status indicates all slabs were released.
2716 int kmem_cache_shrink(struct kmem_cache *cachep)
2719 BUG_ON(!cachep || in_interrupt());
2722 mutex_lock(&slab_mutex);
2723 ret = __cache_shrink(cachep);
2724 mutex_unlock(&slab_mutex);
2728 EXPORT_SYMBOL(kmem_cache_shrink);
2730 int __kmem_cache_shutdown(struct kmem_cache *cachep)
2733 struct kmem_list3 *l3;
2734 int rc = __cache_shrink(cachep);
2739 for_each_online_cpu(i)
2740 kfree(cachep->array[i]);
2742 /* NUMA: free the list3 structures */
2743 for_each_online_node(i) {
2744 l3 = cachep->nodelists[i];
2747 free_alien_cache(l3->alien);
2755 * Get the memory for a slab management obj.
2756 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2757 * always come from malloc_sizes caches. The slab descriptor cannot
2758 * come from the same cache which is getting created because,
2759 * when we are searching for an appropriate cache for these
2760 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2761 * If we are creating a malloc_sizes cache here it would not be visible to
2762 * kmem_find_general_cachep till the initialization is complete.
2763 * Hence we cannot have slabp_cache same as the original cache.
2765 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2766 int colour_off, gfp_t local_flags,
2771 if (OFF_SLAB(cachep)) {
2772 /* Slab management obj is off-slab. */
2773 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2774 local_flags, nodeid);
2776 * If the first object in the slab is leaked (it's allocated
2777 * but no one has a reference to it), we want to make sure
2778 * kmemleak does not treat the ->s_mem pointer as a reference
2779 * to the object. Otherwise we will not report the leak.
2781 kmemleak_scan_area(&slabp->list, sizeof(struct list_head),
2786 slabp = objp + colour_off;
2787 colour_off += cachep->slab_size;
2790 slabp->colouroff = colour_off;
2791 slabp->s_mem = objp + colour_off;
2792 slabp->nodeid = nodeid;
2797 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2799 return (kmem_bufctl_t *) (slabp + 1);
2802 static void cache_init_objs(struct kmem_cache *cachep,
2807 for (i = 0; i < cachep->num; i++) {
2808 void *objp = index_to_obj(cachep, slabp, i);
2810 /* need to poison the objs? */
2811 if (cachep->flags & SLAB_POISON)
2812 poison_obj(cachep, objp, POISON_FREE);
2813 if (cachep->flags & SLAB_STORE_USER)
2814 *dbg_userword(cachep, objp) = NULL;
2816 if (cachep->flags & SLAB_RED_ZONE) {
2817 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2818 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2821 * Constructors are not allowed to allocate memory from the same
2822 * cache which they are a constructor for. Otherwise, deadlock.
2823 * They must also be threaded.
2825 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2826 cachep->ctor(objp + obj_offset(cachep));
2828 if (cachep->flags & SLAB_RED_ZONE) {
2829 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2830 slab_error(cachep, "constructor overwrote the"
2831 " end of an object");
2832 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2833 slab_error(cachep, "constructor overwrote the"
2834 " start of an object");
2836 if ((cachep->size % PAGE_SIZE) == 0 &&
2837 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2838 kernel_map_pages(virt_to_page(objp),
2839 cachep->size / PAGE_SIZE, 0);
2844 slab_bufctl(slabp)[i] = i + 1;
2846 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2849 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2851 if (CONFIG_ZONE_DMA_FLAG) {
2852 if (flags & GFP_DMA)
2853 BUG_ON(!(cachep->allocflags & GFP_DMA));
2855 BUG_ON(cachep->allocflags & GFP_DMA);
2859 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2862 void *objp = index_to_obj(cachep, slabp, slabp->free);
2866 next = slab_bufctl(slabp)[slabp->free];
2868 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2869 WARN_ON(slabp->nodeid != nodeid);
2876 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2877 void *objp, int nodeid)
2879 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2882 /* Verify that the slab belongs to the intended node */
2883 WARN_ON(slabp->nodeid != nodeid);
2885 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2886 printk(KERN_ERR "slab: double free detected in cache "
2887 "'%s', objp %p\n", cachep->name, objp);
2891 slab_bufctl(slabp)[objnr] = slabp->free;
2892 slabp->free = objnr;
2897 * Map pages beginning at addr to the given cache and slab. This is required
2898 * for the slab allocator to be able to lookup the cache and slab of a
2899 * virtual address for kfree, ksize, and slab debugging.
2901 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2907 page = virt_to_page(addr);
2910 if (likely(!PageCompound(page)))
2911 nr_pages <<= cache->gfporder;
2914 page->slab_cache = cache;
2915 page->slab_page = slab;
2917 } while (--nr_pages);
2921 * Grow (by 1) the number of slabs within a cache. This is called by
2922 * kmem_cache_alloc() when there are no active objs left in a cache.
2924 static int cache_grow(struct kmem_cache *cachep,
2925 gfp_t flags, int nodeid, void *objp)
2930 struct kmem_list3 *l3;
2933 * Be lazy and only check for valid flags here, keeping it out of the
2934 * critical path in kmem_cache_alloc().
2936 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2937 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2939 /* Take the l3 list lock to change the colour_next on this node */
2941 l3 = cachep->nodelists[nodeid];
2942 spin_lock(&l3->list_lock);
2944 /* Get colour for the slab, and cal the next value. */
2945 offset = l3->colour_next;
2947 if (l3->colour_next >= cachep->colour)
2948 l3->colour_next = 0;
2949 spin_unlock(&l3->list_lock);
2951 offset *= cachep->colour_off;
2953 if (local_flags & __GFP_WAIT)
2957 * The test for missing atomic flag is performed here, rather than
2958 * the more obvious place, simply to reduce the critical path length
2959 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2960 * will eventually be caught here (where it matters).
2962 kmem_flagcheck(cachep, flags);
2965 * Get mem for the objs. Attempt to allocate a physical page from
2969 objp = kmem_getpages(cachep, local_flags, nodeid);
2973 /* Get slab management. */
2974 slabp = alloc_slabmgmt(cachep, objp, offset,
2975 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2979 slab_map_pages(cachep, slabp, objp);
2981 cache_init_objs(cachep, slabp);
2983 if (local_flags & __GFP_WAIT)
2984 local_irq_disable();
2986 spin_lock(&l3->list_lock);
2988 /* Make slab active. */
2989 list_add_tail(&slabp->list, &(l3->slabs_free));
2990 STATS_INC_GROWN(cachep);
2991 l3->free_objects += cachep->num;
2992 spin_unlock(&l3->list_lock);
2995 kmem_freepages(cachep, objp);
2997 if (local_flags & __GFP_WAIT)
2998 local_irq_disable();
3005 * Perform extra freeing checks:
3006 * - detect bad pointers.
3007 * - POISON/RED_ZONE checking
3009 static void kfree_debugcheck(const void *objp)
3011 if (!virt_addr_valid(objp)) {
3012 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
3013 (unsigned long)objp);
3018 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
3020 unsigned long long redzone1, redzone2;
3022 redzone1 = *dbg_redzone1(cache, obj);
3023 redzone2 = *dbg_redzone2(cache, obj);
3028 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
3031 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
3032 slab_error(cache, "double free detected");
3034 slab_error(cache, "memory outside object was overwritten");
3036 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
3037 obj, redzone1, redzone2);
3040 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
3041 unsigned long caller)
3047 BUG_ON(virt_to_cache(objp) != cachep);
3049 objp -= obj_offset(cachep);
3050 kfree_debugcheck(objp);
3051 page = virt_to_head_page(objp);
3053 slabp = page->slab_page;
3055 if (cachep->flags & SLAB_RED_ZONE) {
3056 verify_redzone_free(cachep, objp);
3057 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
3058 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
3060 if (cachep->flags & SLAB_STORE_USER)
3061 *dbg_userword(cachep, objp) = (void *)caller;
3063 objnr = obj_to_index(cachep, slabp, objp);
3065 BUG_ON(objnr >= cachep->num);
3066 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
3068 #ifdef CONFIG_DEBUG_SLAB_LEAK
3069 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
3071 if (cachep->flags & SLAB_POISON) {
3072 #ifdef CONFIG_DEBUG_PAGEALLOC
3073 if ((cachep->size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
3074 store_stackinfo(cachep, objp, caller);
3075 kernel_map_pages(virt_to_page(objp),
3076 cachep->size / PAGE_SIZE, 0);
3078 poison_obj(cachep, objp, POISON_FREE);
3081 poison_obj(cachep, objp, POISON_FREE);
3087 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
3092 /* Check slab's freelist to see if this obj is there. */
3093 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
3095 if (entries > cachep->num || i >= cachep->num)
3098 if (entries != cachep->num - slabp->inuse) {
3100 printk(KERN_ERR "slab: Internal list corruption detected in "
3101 "cache '%s'(%d), slabp %p(%d). Tainted(%s). Hexdump:\n",
3102 cachep->name, cachep->num, slabp, slabp->inuse,
3104 print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 16, 1, slabp,
3105 sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t),
3111 #define kfree_debugcheck(x) do { } while(0)
3112 #define cache_free_debugcheck(x,objp,z) (objp)
3113 #define check_slabp(x,y) do { } while(0)
3116 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags,
3120 struct kmem_list3 *l3;
3121 struct array_cache *ac;
3125 node = numa_mem_id();
3126 if (unlikely(force_refill))
3129 ac = cpu_cache_get(cachep);
3130 batchcount = ac->batchcount;
3131 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
3133 * If there was little recent activity on this cache, then
3134 * perform only a partial refill. Otherwise we could generate
3137 batchcount = BATCHREFILL_LIMIT;
3139 l3 = cachep->nodelists[node];
3141 BUG_ON(ac->avail > 0 || !l3);
3142 spin_lock(&l3->list_lock);
3144 /* See if we can refill from the shared array */
3145 if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) {
3146 l3->shared->touched = 1;
3150 while (batchcount > 0) {
3151 struct list_head *entry;
3153 /* Get slab alloc is to come from. */
3154 entry = l3->slabs_partial.next;
3155 if (entry == &l3->slabs_partial) {
3156 l3->free_touched = 1;
3157 entry = l3->slabs_free.next;
3158 if (entry == &l3->slabs_free)
3162 slabp = list_entry(entry, struct slab, list);
3163 check_slabp(cachep, slabp);
3164 check_spinlock_acquired(cachep);
3167 * The slab was either on partial or free list so
3168 * there must be at least one object available for
3171 BUG_ON(slabp->inuse >= cachep->num);
3173 while (slabp->inuse < cachep->num && batchcount--) {
3174 STATS_INC_ALLOCED(cachep);
3175 STATS_INC_ACTIVE(cachep);
3176 STATS_SET_HIGH(cachep);
3178 ac_put_obj(cachep, ac, slab_get_obj(cachep, slabp,
3181 check_slabp(cachep, slabp);
3183 /* move slabp to correct slabp list: */
3184 list_del(&slabp->list);
3185 if (slabp->free == BUFCTL_END)
3186 list_add(&slabp->list, &l3->slabs_full);
3188 list_add(&slabp->list, &l3->slabs_partial);
3192 l3->free_objects -= ac->avail;
3194 spin_unlock(&l3->list_lock);
3196 if (unlikely(!ac->avail)) {
3199 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
3201 /* cache_grow can reenable interrupts, then ac could change. */
3202 ac = cpu_cache_get(cachep);
3203 node = numa_mem_id();
3205 /* no objects in sight? abort */
3206 if (!x && (ac->avail == 0 || force_refill))
3209 if (!ac->avail) /* objects refilled by interrupt? */
3214 return ac_get_obj(cachep, ac, flags, force_refill);
3217 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3220 might_sleep_if(flags & __GFP_WAIT);
3222 kmem_flagcheck(cachep, flags);
3227 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3228 gfp_t flags, void *objp, unsigned long caller)
3232 if (cachep->flags & SLAB_POISON) {
3233 #ifdef CONFIG_DEBUG_PAGEALLOC
3234 if ((cachep->size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
3235 kernel_map_pages(virt_to_page(objp),
3236 cachep->size / PAGE_SIZE, 1);
3238 check_poison_obj(cachep, objp);
3240 check_poison_obj(cachep, objp);
3242 poison_obj(cachep, objp, POISON_INUSE);
3244 if (cachep->flags & SLAB_STORE_USER)
3245 *dbg_userword(cachep, objp) = (void *)caller;
3247 if (cachep->flags & SLAB_RED_ZONE) {
3248 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3249 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3250 slab_error(cachep, "double free, or memory outside"
3251 " object was overwritten");
3253 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3254 objp, *dbg_redzone1(cachep, objp),
3255 *dbg_redzone2(cachep, objp));
3257 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3258 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3260 #ifdef CONFIG_DEBUG_SLAB_LEAK
3265 slabp = virt_to_head_page(objp)->slab_page;
3266 objnr = (unsigned)(objp - slabp->s_mem) / cachep->size;
3267 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3270 objp += obj_offset(cachep);
3271 if (cachep->ctor && cachep->flags & SLAB_POISON)
3273 if (ARCH_SLAB_MINALIGN &&
3274 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
3275 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3276 objp, (int)ARCH_SLAB_MINALIGN);
3281 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3284 static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
3286 if (cachep == kmem_cache)
3289 return should_failslab(cachep->object_size, flags, cachep->flags);
3292 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3295 struct array_cache *ac;
3296 bool force_refill = false;
3300 ac = cpu_cache_get(cachep);
3301 if (likely(ac->avail)) {
3303 objp = ac_get_obj(cachep, ac, flags, false);
3306 * Allow for the possibility all avail objects are not allowed
3307 * by the current flags
3310 STATS_INC_ALLOCHIT(cachep);
3313 force_refill = true;
3316 STATS_INC_ALLOCMISS(cachep);
3317 objp = cache_alloc_refill(cachep, flags, force_refill);
3319 * the 'ac' may be updated by cache_alloc_refill(),
3320 * and kmemleak_erase() requires its correct value.
3322 ac = cpu_cache_get(cachep);
3326 * To avoid a false negative, if an object that is in one of the
3327 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3328 * treat the array pointers as a reference to the object.
3331 kmemleak_erase(&ac->entry[ac->avail]);
3337 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3339 * If we are in_interrupt, then process context, including cpusets and
3340 * mempolicy, may not apply and should not be used for allocation policy.
3342 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3344 int nid_alloc, nid_here;
3346 if (in_interrupt() || (flags & __GFP_THISNODE))
3348 nid_alloc = nid_here = numa_mem_id();
3349 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3350 nid_alloc = cpuset_slab_spread_node();
3351 else if (current->mempolicy)
3352 nid_alloc = slab_node();
3353 if (nid_alloc != nid_here)
3354 return ____cache_alloc_node(cachep, flags, nid_alloc);
3359 * Fallback function if there was no memory available and no objects on a
3360 * certain node and fall back is permitted. First we scan all the
3361 * available nodelists for available objects. If that fails then we
3362 * perform an allocation without specifying a node. This allows the page
3363 * allocator to do its reclaim / fallback magic. We then insert the
3364 * slab into the proper nodelist and then allocate from it.
3366 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3368 struct zonelist *zonelist;
3372 enum zone_type high_zoneidx = gfp_zone(flags);
3375 unsigned int cpuset_mems_cookie;
3377 if (flags & __GFP_THISNODE)
3380 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3383 cpuset_mems_cookie = get_mems_allowed();
3384 zonelist = node_zonelist(slab_node(), flags);
3388 * Look through allowed nodes for objects available
3389 * from existing per node queues.
3391 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3392 nid = zone_to_nid(zone);
3394 if (cpuset_zone_allowed_hardwall(zone, flags) &&
3395 cache->nodelists[nid] &&
3396 cache->nodelists[nid]->free_objects) {
3397 obj = ____cache_alloc_node(cache,
3398 flags | GFP_THISNODE, nid);
3406 * This allocation will be performed within the constraints
3407 * of the current cpuset / memory policy requirements.
3408 * We may trigger various forms of reclaim on the allowed
3409 * set and go into memory reserves if necessary.
3411 if (local_flags & __GFP_WAIT)
3413 kmem_flagcheck(cache, flags);
3414 obj = kmem_getpages(cache, local_flags, numa_mem_id());
3415 if (local_flags & __GFP_WAIT)
3416 local_irq_disable();
3419 * Insert into the appropriate per node queues
3421 nid = page_to_nid(virt_to_page(obj));
3422 if (cache_grow(cache, flags, nid, obj)) {
3423 obj = ____cache_alloc_node(cache,
3424 flags | GFP_THISNODE, nid);
3427 * Another processor may allocate the
3428 * objects in the slab since we are
3429 * not holding any locks.
3433 /* cache_grow already freed obj */
3439 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !obj))
3445 * A interface to enable slab creation on nodeid
3447 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3450 struct list_head *entry;
3452 struct kmem_list3 *l3;
3456 l3 = cachep->nodelists[nodeid];
3461 spin_lock(&l3->list_lock);
3462 entry = l3->slabs_partial.next;
3463 if (entry == &l3->slabs_partial) {
3464 l3->free_touched = 1;
3465 entry = l3->slabs_free.next;
3466 if (entry == &l3->slabs_free)
3470 slabp = list_entry(entry, struct slab, list);
3471 check_spinlock_acquired_node(cachep, nodeid);
3472 check_slabp(cachep, slabp);
3474 STATS_INC_NODEALLOCS(cachep);
3475 STATS_INC_ACTIVE(cachep);
3476 STATS_SET_HIGH(cachep);
3478 BUG_ON(slabp->inuse == cachep->num);
3480 obj = slab_get_obj(cachep, slabp, nodeid);
3481 check_slabp(cachep, slabp);
3483 /* move slabp to correct slabp list: */
3484 list_del(&slabp->list);
3486 if (slabp->free == BUFCTL_END)
3487 list_add(&slabp->list, &l3->slabs_full);
3489 list_add(&slabp->list, &l3->slabs_partial);
3491 spin_unlock(&l3->list_lock);
3495 spin_unlock(&l3->list_lock);
3496 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
3500 return fallback_alloc(cachep, flags);
3507 * kmem_cache_alloc_node - Allocate an object on the specified node
3508 * @cachep: The cache to allocate from.
3509 * @flags: See kmalloc().
3510 * @nodeid: node number of the target node.
3511 * @caller: return address of caller, used for debug information
3513 * Identical to kmem_cache_alloc but it will allocate memory on the given
3514 * node, which can improve the performance for cpu bound structures.
3516 * Fallback to other node is possible if __GFP_THISNODE is not set.
3518 static __always_inline void *
3519 slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3520 unsigned long caller)
3522 unsigned long save_flags;
3524 int slab_node = numa_mem_id();
3526 flags &= gfp_allowed_mask;
3528 lockdep_trace_alloc(flags);
3530 if (slab_should_failslab(cachep, flags))
3533 cache_alloc_debugcheck_before(cachep, flags);
3534 local_irq_save(save_flags);
3536 if (nodeid == NUMA_NO_NODE)
3539 if (unlikely(!cachep->nodelists[nodeid])) {
3540 /* Node not bootstrapped yet */
3541 ptr = fallback_alloc(cachep, flags);
3545 if (nodeid == slab_node) {
3547 * Use the locally cached objects if possible.
3548 * However ____cache_alloc does not allow fallback
3549 * to other nodes. It may fail while we still have
3550 * objects on other nodes available.
3552 ptr = ____cache_alloc(cachep, flags);
3556 /* ___cache_alloc_node can fall back to other nodes */
3557 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3559 local_irq_restore(save_flags);
3560 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3561 kmemleak_alloc_recursive(ptr, cachep->object_size, 1, cachep->flags,
3565 kmemcheck_slab_alloc(cachep, flags, ptr, cachep->object_size);
3567 if (unlikely((flags & __GFP_ZERO) && ptr))
3568 memset(ptr, 0, cachep->object_size);
3573 static __always_inline void *
3574 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3578 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3579 objp = alternate_node_alloc(cache, flags);
3583 objp = ____cache_alloc(cache, flags);
3586 * We may just have run out of memory on the local node.
3587 * ____cache_alloc_node() knows how to locate memory on other nodes
3590 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3597 static __always_inline void *
3598 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3600 return ____cache_alloc(cachep, flags);
3603 #endif /* CONFIG_NUMA */
3605 static __always_inline void *
3606 slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
3608 unsigned long save_flags;
3611 flags &= gfp_allowed_mask;
3613 lockdep_trace_alloc(flags);
3615 if (slab_should_failslab(cachep, flags))
3618 cache_alloc_debugcheck_before(cachep, flags);
3619 local_irq_save(save_flags);
3620 objp = __do_cache_alloc(cachep, flags);
3621 local_irq_restore(save_flags);
3622 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3623 kmemleak_alloc_recursive(objp, cachep->object_size, 1, cachep->flags,
3628 kmemcheck_slab_alloc(cachep, flags, objp, cachep->object_size);
3630 if (unlikely((flags & __GFP_ZERO) && objp))
3631 memset(objp, 0, cachep->object_size);
3637 * Caller needs to acquire correct kmem_list's list_lock
3639 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3643 struct kmem_list3 *l3;
3645 for (i = 0; i < nr_objects; i++) {
3649 clear_obj_pfmemalloc(&objpp[i]);
3652 slabp = virt_to_slab(objp);
3653 l3 = cachep->nodelists[node];
3654 list_del(&slabp->list);
3655 check_spinlock_acquired_node(cachep, node);
3656 check_slabp(cachep, slabp);
3657 slab_put_obj(cachep, slabp, objp, node);
3658 STATS_DEC_ACTIVE(cachep);
3660 check_slabp(cachep, slabp);
3662 /* fixup slab chains */
3663 if (slabp->inuse == 0) {
3664 if (l3->free_objects > l3->free_limit) {
3665 l3->free_objects -= cachep->num;
3666 /* No need to drop any previously held
3667 * lock here, even if we have a off-slab slab
3668 * descriptor it is guaranteed to come from
3669 * a different cache, refer to comments before
3672 slab_destroy(cachep, slabp);
3674 list_add(&slabp->list, &l3->slabs_free);
3677 /* Unconditionally move a slab to the end of the
3678 * partial list on free - maximum time for the
3679 * other objects to be freed, too.
3681 list_add_tail(&slabp->list, &l3->slabs_partial);
3686 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3689 struct kmem_list3 *l3;
3690 int node = numa_mem_id();
3692 batchcount = ac->batchcount;
3694 BUG_ON(!batchcount || batchcount > ac->avail);
3697 l3 = cachep->nodelists[node];
3698 spin_lock(&l3->list_lock);
3700 struct array_cache *shared_array = l3->shared;
3701 int max = shared_array->limit - shared_array->avail;
3703 if (batchcount > max)
3705 memcpy(&(shared_array->entry[shared_array->avail]),
3706 ac->entry, sizeof(void *) * batchcount);
3707 shared_array->avail += batchcount;
3712 free_block(cachep, ac->entry, batchcount, node);
3717 struct list_head *p;
3719 p = l3->slabs_free.next;
3720 while (p != &(l3->slabs_free)) {
3723 slabp = list_entry(p, struct slab, list);
3724 BUG_ON(slabp->inuse);
3729 STATS_SET_FREEABLE(cachep, i);
3732 spin_unlock(&l3->list_lock);
3733 ac->avail -= batchcount;
3734 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3738 * Release an obj back to its cache. If the obj has a constructed state, it must
3739 * be in this state _before_ it is released. Called with disabled ints.
3741 static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3742 unsigned long caller)
3744 struct array_cache *ac = cpu_cache_get(cachep);
3747 kmemleak_free_recursive(objp, cachep->flags);
3748 objp = cache_free_debugcheck(cachep, objp, caller);
3750 kmemcheck_slab_free(cachep, objp, cachep->object_size);
3753 * Skip calling cache_free_alien() when the platform is not numa.
3754 * This will avoid cache misses that happen while accessing slabp (which
3755 * is per page memory reference) to get nodeid. Instead use a global
3756 * variable to skip the call, which is mostly likely to be present in
3759 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3762 if (likely(ac->avail < ac->limit)) {
3763 STATS_INC_FREEHIT(cachep);
3765 STATS_INC_FREEMISS(cachep);
3766 cache_flusharray(cachep, ac);
3769 ac_put_obj(cachep, ac, objp);
3773 * kmem_cache_alloc - Allocate an object
3774 * @cachep: The cache to allocate from.
3775 * @flags: See kmalloc().
3777 * Allocate an object from this cache. The flags are only relevant
3778 * if the cache has no available objects.
3780 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3782 void *ret = slab_alloc(cachep, flags, _RET_IP_);
3784 trace_kmem_cache_alloc(_RET_IP_, ret,
3785 cachep->object_size, cachep->size, flags);
3789 EXPORT_SYMBOL(kmem_cache_alloc);
3791 #ifdef CONFIG_TRACING
3793 kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
3797 ret = slab_alloc(cachep, flags, _RET_IP_);
3799 trace_kmalloc(_RET_IP_, ret,
3800 size, cachep->size, flags);
3803 EXPORT_SYMBOL(kmem_cache_alloc_trace);
3807 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3809 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3811 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3812 cachep->object_size, cachep->size,
3817 EXPORT_SYMBOL(kmem_cache_alloc_node);
3819 #ifdef CONFIG_TRACING
3820 void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
3827 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3829 trace_kmalloc_node(_RET_IP_, ret,
3834 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3837 static __always_inline void *
3838 __do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
3840 struct kmem_cache *cachep;
3842 cachep = kmem_find_general_cachep(size, flags);
3843 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3845 return kmem_cache_alloc_node_trace(cachep, flags, node, size);
3848 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3849 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3851 return __do_kmalloc_node(size, flags, node, _RET_IP_);
3853 EXPORT_SYMBOL(__kmalloc_node);
3855 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3856 int node, unsigned long caller)
3858 return __do_kmalloc_node(size, flags, node, caller);
3860 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3862 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3864 return __do_kmalloc_node(size, flags, node, 0);
3866 EXPORT_SYMBOL(__kmalloc_node);
3867 #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3868 #endif /* CONFIG_NUMA */
3871 * __do_kmalloc - allocate memory
3872 * @size: how many bytes of memory are required.
3873 * @flags: the type of memory to allocate (see kmalloc).
3874 * @caller: function caller for debug tracking of the caller
3876 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3877 unsigned long caller)
3879 struct kmem_cache *cachep;
3882 /* If you want to save a few bytes .text space: replace
3884 * Then kmalloc uses the uninlined functions instead of the inline
3887 cachep = __find_general_cachep(size, flags);
3888 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3890 ret = slab_alloc(cachep, flags, caller);
3892 trace_kmalloc(caller, ret,
3893 size, cachep->size, flags);
3899 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3900 void *__kmalloc(size_t size, gfp_t flags)
3902 return __do_kmalloc(size, flags, _RET_IP_);
3904 EXPORT_SYMBOL(__kmalloc);
3906 void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3908 return __do_kmalloc(size, flags, caller);
3910 EXPORT_SYMBOL(__kmalloc_track_caller);
3913 void *__kmalloc(size_t size, gfp_t flags)
3915 return __do_kmalloc(size, flags, 0);
3917 EXPORT_SYMBOL(__kmalloc);
3921 * kmem_cache_free - Deallocate an object
3922 * @cachep: The cache the allocation was from.
3923 * @objp: The previously allocated object.
3925 * Free an object which was previously allocated from this
3928 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3930 unsigned long flags;
3932 local_irq_save(flags);
3933 debug_check_no_locks_freed(objp, cachep->object_size);
3934 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3935 debug_check_no_obj_freed(objp, cachep->object_size);
3936 __cache_free(cachep, objp, _RET_IP_);
3937 local_irq_restore(flags);
3939 trace_kmem_cache_free(_RET_IP_, objp);
3941 EXPORT_SYMBOL(kmem_cache_free);
3944 * kfree - free previously allocated memory
3945 * @objp: pointer returned by kmalloc.
3947 * If @objp is NULL, no operation is performed.
3949 * Don't free memory not originally allocated by kmalloc()
3950 * or you will run into trouble.
3952 void kfree(const void *objp)
3954 struct kmem_cache *c;
3955 unsigned long flags;
3957 trace_kfree(_RET_IP_, objp);
3959 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3961 local_irq_save(flags);
3962 kfree_debugcheck(objp);
3963 c = virt_to_cache(objp);
3964 debug_check_no_locks_freed(objp, c->object_size);
3966 debug_check_no_obj_freed(objp, c->object_size);
3967 __cache_free(c, (void *)objp, _RET_IP_);
3968 local_irq_restore(flags);
3970 EXPORT_SYMBOL(kfree);
3972 unsigned int kmem_cache_size(struct kmem_cache *cachep)
3974 return cachep->object_size;
3976 EXPORT_SYMBOL(kmem_cache_size);
3979 * This initializes kmem_list3 or resizes various caches for all nodes.
3981 static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
3984 struct kmem_list3 *l3;
3985 struct array_cache *new_shared;
3986 struct array_cache **new_alien = NULL;
3988 for_each_online_node(node) {
3990 if (use_alien_caches) {
3991 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
3997 if (cachep->shared) {
3998 new_shared = alloc_arraycache(node,
3999 cachep->shared*cachep->batchcount,
4002 free_alien_cache(new_alien);
4007 l3 = cachep->nodelists[node];
4009 struct array_cache *shared = l3->shared;
4011 spin_lock_irq(&l3->list_lock);
4014 free_block(cachep, shared->entry,
4015 shared->avail, node);
4017 l3->shared = new_shared;
4019 l3->alien = new_alien;
4022 l3->free_limit = (1 + nr_cpus_node(node)) *
4023 cachep->batchcount + cachep->num;
4024 spin_unlock_irq(&l3->list_lock);
4026 free_alien_cache(new_alien);
4029 l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node);
4031 free_alien_cache(new_alien);
4036 kmem_list3_init(l3);
4037 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
4038 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
4039 l3->shared = new_shared;
4040 l3->alien = new_alien;
4041 l3->free_limit = (1 + nr_cpus_node(node)) *
4042 cachep->batchcount + cachep->num;
4043 cachep->nodelists[node] = l3;
4048 if (!cachep->list.next) {
4049 /* Cache is not active yet. Roll back what we did */
4052 if (cachep->nodelists[node]) {
4053 l3 = cachep->nodelists[node];
4056 free_alien_cache(l3->alien);
4058 cachep->nodelists[node] = NULL;
4066 struct ccupdate_struct {
4067 struct kmem_cache *cachep;
4068 struct array_cache *new[0];
4071 static void do_ccupdate_local(void *info)
4073 struct ccupdate_struct *new = info;
4074 struct array_cache *old;
4077 old = cpu_cache_get(new->cachep);
4079 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
4080 new->new[smp_processor_id()] = old;
4083 /* Always called with the slab_mutex held */
4084 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
4085 int batchcount, int shared, gfp_t gfp)
4087 struct ccupdate_struct *new;
4090 new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *),
4095 for_each_online_cpu(i) {
4096 new->new[i] = alloc_arraycache(cpu_to_mem(i), limit,
4099 for (i--; i >= 0; i--)
4105 new->cachep = cachep;
4107 on_each_cpu(do_ccupdate_local, (void *)new, 1);
4110 cachep->batchcount = batchcount;
4111 cachep->limit = limit;
4112 cachep->shared = shared;
4114 for_each_online_cpu(i) {
4115 struct array_cache *ccold = new->new[i];
4118 spin_lock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
4119 free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i));
4120 spin_unlock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
4124 return alloc_kmemlist(cachep, gfp);
4127 /* Called with slab_mutex held always */
4128 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
4134 * The head array serves three purposes:
4135 * - create a LIFO ordering, i.e. return objects that are cache-warm
4136 * - reduce the number of spinlock operations.
4137 * - reduce the number of linked list operations on the slab and
4138 * bufctl chains: array operations are cheaper.
4139 * The numbers are guessed, we should auto-tune as described by
4142 if (cachep->size > 131072)
4144 else if (cachep->size > PAGE_SIZE)
4146 else if (cachep->size > 1024)
4148 else if (cachep->size > 256)
4154 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
4155 * allocation behaviour: Most allocs on one cpu, most free operations
4156 * on another cpu. For these cases, an efficient object passing between
4157 * cpus is necessary. This is provided by a shared array. The array
4158 * replaces Bonwick's magazine layer.
4159 * On uniprocessor, it's functionally equivalent (but less efficient)
4160 * to a larger limit. Thus disabled by default.
4163 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
4168 * With debugging enabled, large batchcount lead to excessively long
4169 * periods with disabled local interrupts. Limit the batchcount
4174 err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp);
4176 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
4177 cachep->name, -err);
4182 * Drain an array if it contains any elements taking the l3 lock only if
4183 * necessary. Note that the l3 listlock also protects the array_cache
4184 * if drain_array() is used on the shared array.
4186 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
4187 struct array_cache *ac, int force, int node)
4191 if (!ac || !ac->avail)
4193 if (ac->touched && !force) {
4196 spin_lock_irq(&l3->list_lock);
4198 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4199 if (tofree > ac->avail)
4200 tofree = (ac->avail + 1) / 2;
4201 free_block(cachep, ac->entry, tofree, node);
4202 ac->avail -= tofree;
4203 memmove(ac->entry, &(ac->entry[tofree]),
4204 sizeof(void *) * ac->avail);
4206 spin_unlock_irq(&l3->list_lock);
4211 * cache_reap - Reclaim memory from caches.
4212 * @w: work descriptor
4214 * Called from workqueue/eventd every few seconds.
4216 * - clear the per-cpu caches for this CPU.
4217 * - return freeable pages to the main free memory pool.
4219 * If we cannot acquire the cache chain mutex then just give up - we'll try
4220 * again on the next iteration.
4222 static void cache_reap(struct work_struct *w)
4224 struct kmem_cache *searchp;
4225 struct kmem_list3 *l3;
4226 int node = numa_mem_id();
4227 struct delayed_work *work = to_delayed_work(w);
4229 if (!mutex_trylock(&slab_mutex))
4230 /* Give up. Setup the next iteration. */
4233 list_for_each_entry(searchp, &slab_caches, list) {
4237 * We only take the l3 lock if absolutely necessary and we
4238 * have established with reasonable certainty that
4239 * we can do some work if the lock was obtained.
4241 l3 = searchp->nodelists[node];
4243 reap_alien(searchp, l3);
4245 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
4248 * These are racy checks but it does not matter
4249 * if we skip one check or scan twice.
4251 if (time_after(l3->next_reap, jiffies))
4254 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
4256 drain_array(searchp, l3, l3->shared, 0, node);
4258 if (l3->free_touched)
4259 l3->free_touched = 0;
4263 freed = drain_freelist(searchp, l3, (l3->free_limit +
4264 5 * searchp->num - 1) / (5 * searchp->num));
4265 STATS_ADD_REAPED(searchp, freed);
4271 mutex_unlock(&slab_mutex);
4274 /* Set up the next iteration */
4275 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
4278 #ifdef CONFIG_SLABINFO
4280 static void print_slabinfo_header(struct seq_file *m)
4283 * Output format version, so at least we can change it
4284 * without _too_ many complaints.
4287 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
4289 seq_puts(m, "slabinfo - version: 2.1\n");
4291 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4292 "<objperslab> <pagesperslab>");
4293 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4294 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4296 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> "
4297 "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
4298 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
4303 static void *s_start(struct seq_file *m, loff_t *pos)
4307 mutex_lock(&slab_mutex);
4309 print_slabinfo_header(m);
4311 return seq_list_start(&slab_caches, *pos);
4314 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4316 return seq_list_next(p, &slab_caches, pos);
4319 static void s_stop(struct seq_file *m, void *p)
4321 mutex_unlock(&slab_mutex);
4324 static int s_show(struct seq_file *m, void *p)
4326 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
4328 unsigned long active_objs;
4329 unsigned long num_objs;
4330 unsigned long active_slabs = 0;
4331 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4335 struct kmem_list3 *l3;
4339 for_each_online_node(node) {
4340 l3 = cachep->nodelists[node];
4345 spin_lock_irq(&l3->list_lock);
4347 list_for_each_entry(slabp, &l3->slabs_full, list) {
4348 if (slabp->inuse != cachep->num && !error)
4349 error = "slabs_full accounting error";
4350 active_objs += cachep->num;
4353 list_for_each_entry(slabp, &l3->slabs_partial, list) {
4354 if (slabp->inuse == cachep->num && !error)
4355 error = "slabs_partial inuse accounting error";
4356 if (!slabp->inuse && !error)
4357 error = "slabs_partial/inuse accounting error";
4358 active_objs += slabp->inuse;
4361 list_for_each_entry(slabp, &l3->slabs_free, list) {
4362 if (slabp->inuse && !error)
4363 error = "slabs_free/inuse accounting error";
4366 free_objects += l3->free_objects;
4368 shared_avail += l3->shared->avail;
4370 spin_unlock_irq(&l3->list_lock);
4372 num_slabs += active_slabs;
4373 num_objs = num_slabs * cachep->num;
4374 if (num_objs - active_objs != free_objects && !error)
4375 error = "free_objects accounting error";
4377 name = cachep->name;
4379 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4381 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
4382 name, active_objs, num_objs, cachep->size,
4383 cachep->num, (1 << cachep->gfporder));
4384 seq_printf(m, " : tunables %4u %4u %4u",
4385 cachep->limit, cachep->batchcount, cachep->shared);
4386 seq_printf(m, " : slabdata %6lu %6lu %6lu",
4387 active_slabs, num_slabs, shared_avail);
4390 unsigned long high = cachep->high_mark;
4391 unsigned long allocs = cachep->num_allocations;
4392 unsigned long grown = cachep->grown;
4393 unsigned long reaped = cachep->reaped;
4394 unsigned long errors = cachep->errors;
4395 unsigned long max_freeable = cachep->max_freeable;
4396 unsigned long node_allocs = cachep->node_allocs;
4397 unsigned long node_frees = cachep->node_frees;
4398 unsigned long overflows = cachep->node_overflow;
4400 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4401 "%4lu %4lu %4lu %4lu %4lu",
4402 allocs, high, grown,
4403 reaped, errors, max_freeable, node_allocs,
4404 node_frees, overflows);
4408 unsigned long allochit = atomic_read(&cachep->allochit);
4409 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4410 unsigned long freehit = atomic_read(&cachep->freehit);
4411 unsigned long freemiss = atomic_read(&cachep->freemiss);
4413 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4414 allochit, allocmiss, freehit, freemiss);
4422 * slabinfo_op - iterator that generates /proc/slabinfo
4431 * num-pages-per-slab
4432 * + further values on SMP and with statistics enabled
4435 static const struct seq_operations slabinfo_op = {
4442 #define MAX_SLABINFO_WRITE 128
4444 * slabinfo_write - Tuning for the slab allocator
4446 * @buffer: user buffer
4447 * @count: data length
4450 static ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4451 size_t count, loff_t *ppos)
4453 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4454 int limit, batchcount, shared, res;
4455 struct kmem_cache *cachep;
4457 if (count > MAX_SLABINFO_WRITE)
4459 if (copy_from_user(&kbuf, buffer, count))
4461 kbuf[MAX_SLABINFO_WRITE] = '\0';
4463 tmp = strchr(kbuf, ' ');
4468 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4471 /* Find the cache in the chain of caches. */
4472 mutex_lock(&slab_mutex);
4474 list_for_each_entry(cachep, &slab_caches, list) {
4475 if (!strcmp(cachep->name, kbuf)) {
4476 if (limit < 1 || batchcount < 1 ||
4477 batchcount > limit || shared < 0) {
4480 res = do_tune_cpucache(cachep, limit,
4487 mutex_unlock(&slab_mutex);
4493 static int slabinfo_open(struct inode *inode, struct file *file)
4495 return seq_open(file, &slabinfo_op);
4498 static const struct file_operations proc_slabinfo_operations = {
4499 .open = slabinfo_open,
4501 .write = slabinfo_write,
4502 .llseek = seq_lseek,
4503 .release = seq_release,
4506 #ifdef CONFIG_DEBUG_SLAB_LEAK
4508 static void *leaks_start(struct seq_file *m, loff_t *pos)
4510 mutex_lock(&slab_mutex);
4511 return seq_list_start(&slab_caches, *pos);
4514 static inline int add_caller(unsigned long *n, unsigned long v)
4524 unsigned long *q = p + 2 * i;
4538 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4544 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4550 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->size) {
4551 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4553 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4558 static void show_symbol(struct seq_file *m, unsigned long address)
4560 #ifdef CONFIG_KALLSYMS
4561 unsigned long offset, size;
4562 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4564 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4565 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4567 seq_printf(m, " [%s]", modname);
4571 seq_printf(m, "%p", (void *)address);
4574 static int leaks_show(struct seq_file *m, void *p)
4576 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
4578 struct kmem_list3 *l3;
4580 unsigned long *n = m->private;
4584 if (!(cachep->flags & SLAB_STORE_USER))
4586 if (!(cachep->flags & SLAB_RED_ZONE))
4589 /* OK, we can do it */
4593 for_each_online_node(node) {
4594 l3 = cachep->nodelists[node];
4599 spin_lock_irq(&l3->list_lock);
4601 list_for_each_entry(slabp, &l3->slabs_full, list)
4602 handle_slab(n, cachep, slabp);
4603 list_for_each_entry(slabp, &l3->slabs_partial, list)
4604 handle_slab(n, cachep, slabp);
4605 spin_unlock_irq(&l3->list_lock);
4607 name = cachep->name;
4609 /* Increase the buffer size */
4610 mutex_unlock(&slab_mutex);
4611 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4613 /* Too bad, we are really out */
4615 mutex_lock(&slab_mutex);
4618 *(unsigned long *)m->private = n[0] * 2;
4620 mutex_lock(&slab_mutex);
4621 /* Now make sure this entry will be retried */
4625 for (i = 0; i < n[1]; i++) {
4626 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4627 show_symbol(m, n[2*i+2]);
4634 static const struct seq_operations slabstats_op = {
4635 .start = leaks_start,
4641 static int slabstats_open(struct inode *inode, struct file *file)
4643 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4646 ret = seq_open(file, &slabstats_op);
4648 struct seq_file *m = file->private_data;
4649 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4658 static const struct file_operations proc_slabstats_operations = {
4659 .open = slabstats_open,
4661 .llseek = seq_lseek,
4662 .release = seq_release_private,
4666 static int __init slab_proc_init(void)
4668 proc_create("slabinfo",S_IWUSR|S_IRUSR,NULL,&proc_slabinfo_operations);
4669 #ifdef CONFIG_DEBUG_SLAB_LEAK
4670 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4674 module_init(slab_proc_init);
4678 * ksize - get the actual amount of memory allocated for a given object
4679 * @objp: Pointer to the object
4681 * kmalloc may internally round up allocations and return more memory
4682 * than requested. ksize() can be used to determine the actual amount of
4683 * memory allocated. The caller may use this additional memory, even though
4684 * a smaller amount of memory was initially specified with the kmalloc call.
4685 * The caller must guarantee that objp points to a valid object previously
4686 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4687 * must not be freed during the duration of the call.
4689 size_t ksize(const void *objp)
4692 if (unlikely(objp == ZERO_SIZE_PTR))
4695 return virt_to_cache(objp)->object_size;
4697 EXPORT_SYMBOL(ksize);