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>
91 #include <linux/poison.h>
92 #include <linux/swap.h>
93 #include <linux/cache.h>
94 #include <linux/interrupt.h>
95 #include <linux/init.h>
96 #include <linux/compiler.h>
97 #include <linux/cpuset.h>
98 #include <linux/proc_fs.h>
99 #include <linux/seq_file.h>
100 #include <linux/notifier.h>
101 #include <linux/kallsyms.h>
102 #include <linux/cpu.h>
103 #include <linux/sysctl.h>
104 #include <linux/module.h>
105 #include <linux/rcupdate.h>
106 #include <linux/string.h>
107 #include <linux/uaccess.h>
108 #include <linux/nodemask.h>
109 #include <linux/kmemleak.h>
110 #include <linux/mempolicy.h>
111 #include <linux/mutex.h>
112 #include <linux/fault-inject.h>
113 #include <linux/rtmutex.h>
114 #include <linux/reciprocal_div.h>
115 #include <linux/debugobjects.h>
116 #include <linux/kmemcheck.h>
117 #include <linux/memory.h>
118 #include <linux/prefetch.h>
120 #include <net/sock.h>
122 #include <asm/cacheflush.h>
123 #include <asm/tlbflush.h>
124 #include <asm/page.h>
126 #include <trace/events/kmem.h>
128 #include "internal.h"
133 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
134 * 0 for faster, smaller code (especially in the critical paths).
136 * STATS - 1 to collect stats for /proc/slabinfo.
137 * 0 for faster, smaller code (especially in the critical paths).
139 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
142 #ifdef CONFIG_DEBUG_SLAB
145 #define FORCED_DEBUG 1
149 #define FORCED_DEBUG 0
152 /* Shouldn't this be in a header file somewhere? */
153 #define BYTES_PER_WORD sizeof(void *)
154 #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
156 #ifndef ARCH_KMALLOC_FLAGS
157 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
161 * true if a page was allocated from pfmemalloc reserves for network-based
164 static bool pfmemalloc_active __read_mostly;
169 * Bufctl's are used for linking objs within a slab
172 * This implementation relies on "struct page" for locating the cache &
173 * slab an object belongs to.
174 * This allows the bufctl structure to be small (one int), but limits
175 * the number of objects a slab (not a cache) can contain when off-slab
176 * bufctls are used. The limit is the size of the largest general cache
177 * that does not use off-slab slabs.
178 * For 32bit archs with 4 kB pages, is this 56.
179 * This is not serious, as it is only for large objects, when it is unwise
180 * to have too many per slab.
181 * Note: This limit can be raised by introducing a general cache whose size
182 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
185 typedef unsigned int kmem_bufctl_t;
186 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
187 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
188 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
189 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
194 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
195 * arrange for kmem_freepages to be called via RCU. This is useful if
196 * we need to approach a kernel structure obliquely, from its address
197 * obtained without the usual locking. We can lock the structure to
198 * stabilize it and check it's still at the given address, only if we
199 * can be sure that the memory has not been meanwhile reused for some
200 * other kind of object (which our subsystem's lock might corrupt).
202 * rcu_read_lock before reading the address, then rcu_read_unlock after
203 * taking the spinlock within the structure expected at that address.
206 struct rcu_head head;
207 struct kmem_cache *cachep;
214 * Manages the objs in a slab. Placed either at the beginning of mem allocated
215 * for a slab, or allocated from an general cache.
216 * Slabs are chained into three list: fully used, partial, fully free slabs.
221 struct list_head list;
222 unsigned long colouroff;
223 void *s_mem; /* including colour offset */
224 unsigned int inuse; /* num of objs active in slab */
226 unsigned short nodeid;
228 struct slab_rcu __slab_cover_slab_rcu;
236 * - LIFO ordering, to hand out cache-warm objects from _alloc
237 * - reduce the number of linked list operations
238 * - reduce spinlock operations
240 * The limit is stored in the per-cpu structure to reduce the data cache
247 unsigned int batchcount;
248 unsigned int touched;
251 * Must have this definition in here for the proper
252 * alignment of array_cache. Also simplifies accessing
255 * Entries should not be directly dereferenced as
256 * entries belonging to slabs marked pfmemalloc will
257 * have the lower bits set SLAB_OBJ_PFMEMALLOC
261 #define SLAB_OBJ_PFMEMALLOC 1
262 static inline bool is_obj_pfmemalloc(void *objp)
264 return (unsigned long)objp & SLAB_OBJ_PFMEMALLOC;
267 static inline void set_obj_pfmemalloc(void **objp)
269 *objp = (void *)((unsigned long)*objp | SLAB_OBJ_PFMEMALLOC);
273 static inline void clear_obj_pfmemalloc(void **objp)
275 *objp = (void *)((unsigned long)*objp & ~SLAB_OBJ_PFMEMALLOC);
279 * bootstrap: The caches do not work without cpuarrays anymore, but the
280 * cpuarrays are allocated from the generic caches...
282 #define BOOT_CPUCACHE_ENTRIES 1
283 struct arraycache_init {
284 struct array_cache cache;
285 void *entries[BOOT_CPUCACHE_ENTRIES];
289 * The slab lists for all objects.
291 struct kmem_cache_node {
292 struct list_head slabs_partial; /* partial list first, better asm code */
293 struct list_head slabs_full;
294 struct list_head slabs_free;
295 unsigned long free_objects;
296 unsigned int free_limit;
297 unsigned int colour_next; /* Per-node cache coloring */
298 spinlock_t list_lock;
299 struct array_cache *shared; /* shared per node */
300 struct array_cache **alien; /* on other nodes */
301 unsigned long next_reap; /* updated without locking */
302 int free_touched; /* updated without locking */
306 * Need this for bootstrapping a per node allocator.
308 #define NUM_INIT_LISTS (3 * MAX_NUMNODES)
309 static struct kmem_cache_node __initdata initkmem_list3[NUM_INIT_LISTS];
310 #define CACHE_CACHE 0
311 #define SIZE_AC MAX_NUMNODES
312 #define SIZE_L3 (2 * MAX_NUMNODES)
314 static int drain_freelist(struct kmem_cache *cache,
315 struct kmem_cache_node *l3, int tofree);
316 static void free_block(struct kmem_cache *cachep, void **objpp, int len,
318 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
319 static void cache_reap(struct work_struct *unused);
321 struct kmem_cache *kmalloc_caches[KMALLOC_SHIFT_HIGH + 1];
322 EXPORT_SYMBOL(kmalloc_caches);
324 #ifdef CONFIG_ZONE_DMA
325 struct kmem_cache *kmalloc_dma_caches[KMALLOC_SHIFT_HIGH + 1];
326 EXPORT_SYMBOL(kmalloc_dma_caches);
329 static int slab_early_init = 1;
331 #define INDEX_AC kmalloc_index(sizeof(struct arraycache_init))
332 #define INDEX_L3 kmalloc_index(sizeof(struct kmem_cache_node))
334 static void kmem_list3_init(struct kmem_cache_node *parent)
336 INIT_LIST_HEAD(&parent->slabs_full);
337 INIT_LIST_HEAD(&parent->slabs_partial);
338 INIT_LIST_HEAD(&parent->slabs_free);
339 parent->shared = NULL;
340 parent->alien = NULL;
341 parent->colour_next = 0;
342 spin_lock_init(&parent->list_lock);
343 parent->free_objects = 0;
344 parent->free_touched = 0;
347 #define MAKE_LIST(cachep, listp, slab, nodeid) \
349 INIT_LIST_HEAD(listp); \
350 list_splice(&(cachep->nodelists[nodeid]->slab), listp); \
353 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
355 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
356 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
357 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
360 #define CFLGS_OFF_SLAB (0x80000000UL)
361 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
363 #define BATCHREFILL_LIMIT 16
365 * Optimization question: fewer reaps means less probability for unnessary
366 * cpucache drain/refill cycles.
368 * OTOH the cpuarrays can contain lots of objects,
369 * which could lock up otherwise freeable slabs.
371 #define REAPTIMEOUT_CPUC (2*HZ)
372 #define REAPTIMEOUT_LIST3 (4*HZ)
375 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
376 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
377 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
378 #define STATS_INC_GROWN(x) ((x)->grown++)
379 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
380 #define STATS_SET_HIGH(x) \
382 if ((x)->num_active > (x)->high_mark) \
383 (x)->high_mark = (x)->num_active; \
385 #define STATS_INC_ERR(x) ((x)->errors++)
386 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
387 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
388 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
389 #define STATS_SET_FREEABLE(x, i) \
391 if ((x)->max_freeable < i) \
392 (x)->max_freeable = i; \
394 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
395 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
396 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
397 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
399 #define STATS_INC_ACTIVE(x) do { } while (0)
400 #define STATS_DEC_ACTIVE(x) do { } while (0)
401 #define STATS_INC_ALLOCED(x) do { } while (0)
402 #define STATS_INC_GROWN(x) do { } while (0)
403 #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
404 #define STATS_SET_HIGH(x) do { } while (0)
405 #define STATS_INC_ERR(x) do { } while (0)
406 #define STATS_INC_NODEALLOCS(x) do { } while (0)
407 #define STATS_INC_NODEFREES(x) do { } while (0)
408 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
409 #define STATS_SET_FREEABLE(x, i) do { } while (0)
410 #define STATS_INC_ALLOCHIT(x) do { } while (0)
411 #define STATS_INC_ALLOCMISS(x) do { } while (0)
412 #define STATS_INC_FREEHIT(x) do { } while (0)
413 #define STATS_INC_FREEMISS(x) do { } while (0)
419 * memory layout of objects:
421 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
422 * the end of an object is aligned with the end of the real
423 * allocation. Catches writes behind the end of the allocation.
424 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
426 * cachep->obj_offset: The real object.
427 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
428 * cachep->size - 1* BYTES_PER_WORD: last caller address
429 * [BYTES_PER_WORD long]
431 static int obj_offset(struct kmem_cache *cachep)
433 return cachep->obj_offset;
436 static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
438 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
439 return (unsigned long long*) (objp + obj_offset(cachep) -
440 sizeof(unsigned long long));
443 static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
445 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
446 if (cachep->flags & SLAB_STORE_USER)
447 return (unsigned long long *)(objp + cachep->size -
448 sizeof(unsigned long long) -
450 return (unsigned long long *) (objp + cachep->size -
451 sizeof(unsigned long long));
454 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
456 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
457 return (void **)(objp + cachep->size - BYTES_PER_WORD);
462 #define obj_offset(x) 0
463 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
464 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
465 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
470 * Do not go above this order unless 0 objects fit into the slab or
471 * overridden on the command line.
473 #define SLAB_MAX_ORDER_HI 1
474 #define SLAB_MAX_ORDER_LO 0
475 static int slab_max_order = SLAB_MAX_ORDER_LO;
476 static bool slab_max_order_set __initdata;
478 static inline struct kmem_cache *virt_to_cache(const void *obj)
480 struct page *page = virt_to_head_page(obj);
481 return page->slab_cache;
484 static inline struct slab *virt_to_slab(const void *obj)
486 struct page *page = virt_to_head_page(obj);
488 VM_BUG_ON(!PageSlab(page));
489 return page->slab_page;
492 static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
495 return slab->s_mem + cache->size * idx;
499 * We want to avoid an expensive divide : (offset / cache->size)
500 * Using the fact that size is a constant for a particular cache,
501 * we can replace (offset / cache->size) by
502 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
504 static inline unsigned int obj_to_index(const struct kmem_cache *cache,
505 const struct slab *slab, void *obj)
507 u32 offset = (obj - slab->s_mem);
508 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
511 static struct arraycache_init initarray_generic =
512 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
514 /* internal cache of cache description objs */
515 static struct kmem_cache kmem_cache_boot = {
517 .limit = BOOT_CPUCACHE_ENTRIES,
519 .size = sizeof(struct kmem_cache),
520 .name = "kmem_cache",
523 #define BAD_ALIEN_MAGIC 0x01020304ul
525 #ifdef CONFIG_LOCKDEP
528 * Slab sometimes uses the kmalloc slabs to store the slab headers
529 * for other slabs "off slab".
530 * The locking for this is tricky in that it nests within the locks
531 * of all other slabs in a few places; to deal with this special
532 * locking we put on-slab caches into a separate lock-class.
534 * We set lock class for alien array caches which are up during init.
535 * The lock annotation will be lost if all cpus of a node goes down and
536 * then comes back up during hotplug
538 static struct lock_class_key on_slab_l3_key;
539 static struct lock_class_key on_slab_alc_key;
541 static struct lock_class_key debugobj_l3_key;
542 static struct lock_class_key debugobj_alc_key;
544 static void slab_set_lock_classes(struct kmem_cache *cachep,
545 struct lock_class_key *l3_key, struct lock_class_key *alc_key,
548 struct array_cache **alc;
549 struct kmem_cache_node *l3;
552 l3 = cachep->nodelists[q];
556 lockdep_set_class(&l3->list_lock, l3_key);
559 * FIXME: This check for BAD_ALIEN_MAGIC
560 * should go away when common slab code is taught to
561 * work even without alien caches.
562 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
563 * for alloc_alien_cache,
565 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
569 lockdep_set_class(&alc[r]->lock, alc_key);
573 static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
575 slab_set_lock_classes(cachep, &debugobj_l3_key, &debugobj_alc_key, node);
578 static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
582 for_each_online_node(node)
583 slab_set_debugobj_lock_classes_node(cachep, node);
586 static void init_node_lock_keys(int q)
593 for (i = 1; i < PAGE_SHIFT + MAX_ORDER; i++) {
594 struct kmem_cache_node *l3;
595 struct kmem_cache *cache = kmalloc_caches[i];
600 l3 = cache->nodelists[q];
601 if (!l3 || OFF_SLAB(cache))
604 slab_set_lock_classes(cache, &on_slab_l3_key,
605 &on_slab_alc_key, q);
609 static void on_slab_lock_classes_node(struct kmem_cache *cachep, int q)
612 if (!cachep->nodelists[q])
615 slab_set_lock_classes(cachep, &on_slab_l3_key,
616 &on_slab_alc_key, q);
619 static inline void on_slab_lock_classes(struct kmem_cache *cachep)
623 VM_BUG_ON(OFF_SLAB(cachep));
625 on_slab_lock_classes_node(cachep, node);
628 static inline void init_lock_keys(void)
633 init_node_lock_keys(node);
636 static void init_node_lock_keys(int q)
640 static inline void init_lock_keys(void)
644 static inline void on_slab_lock_classes(struct kmem_cache *cachep)
648 static inline void on_slab_lock_classes_node(struct kmem_cache *cachep, int node)
652 static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
656 static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
661 static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
663 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
665 return cachep->array[smp_processor_id()];
668 static inline struct kmem_cache *__find_general_cachep(size_t size,
674 /* This happens if someone tries to call
675 * kmem_cache_create(), or __kmalloc(), before
676 * the generic caches are initialized.
678 BUG_ON(kmalloc_caches[INDEX_AC] == NULL);
681 return ZERO_SIZE_PTR;
683 i = kmalloc_index(size);
686 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
687 * has cs_{dma,}cachep==NULL. Thus no special case
688 * for large kmalloc calls required.
690 #ifdef CONFIG_ZONE_DMA
691 if (unlikely(gfpflags & GFP_DMA))
692 return kmalloc_dma_caches[i];
694 return kmalloc_caches[i];
697 static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
699 return __find_general_cachep(size, gfpflags);
702 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
704 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
708 * Calculate the number of objects and left-over bytes for a given buffer size.
710 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
711 size_t align, int flags, size_t *left_over,
716 size_t slab_size = PAGE_SIZE << gfporder;
719 * The slab management structure can be either off the slab or
720 * on it. For the latter case, the memory allocated for a
724 * - One kmem_bufctl_t for each object
725 * - Padding to respect alignment of @align
726 * - @buffer_size bytes for each object
728 * If the slab management structure is off the slab, then the
729 * alignment will already be calculated into the size. Because
730 * the slabs are all pages aligned, the objects will be at the
731 * correct alignment when allocated.
733 if (flags & CFLGS_OFF_SLAB) {
735 nr_objs = slab_size / buffer_size;
737 if (nr_objs > SLAB_LIMIT)
738 nr_objs = SLAB_LIMIT;
741 * Ignore padding for the initial guess. The padding
742 * is at most @align-1 bytes, and @buffer_size is at
743 * least @align. In the worst case, this result will
744 * be one greater than the number of objects that fit
745 * into the memory allocation when taking the padding
748 nr_objs = (slab_size - sizeof(struct slab)) /
749 (buffer_size + sizeof(kmem_bufctl_t));
752 * This calculated number will be either the right
753 * amount, or one greater than what we want.
755 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
759 if (nr_objs > SLAB_LIMIT)
760 nr_objs = SLAB_LIMIT;
762 mgmt_size = slab_mgmt_size(nr_objs, align);
765 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
769 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
771 static void __slab_error(const char *function, struct kmem_cache *cachep,
774 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
775 function, cachep->name, msg);
777 add_taint(TAINT_BAD_PAGE);
782 * By default on NUMA we use alien caches to stage the freeing of
783 * objects allocated from other nodes. This causes massive memory
784 * inefficiencies when using fake NUMA setup to split memory into a
785 * large number of small nodes, so it can be disabled on the command
789 static int use_alien_caches __read_mostly = 1;
790 static int __init noaliencache_setup(char *s)
792 use_alien_caches = 0;
795 __setup("noaliencache", noaliencache_setup);
797 static int __init slab_max_order_setup(char *str)
799 get_option(&str, &slab_max_order);
800 slab_max_order = slab_max_order < 0 ? 0 :
801 min(slab_max_order, MAX_ORDER - 1);
802 slab_max_order_set = true;
806 __setup("slab_max_order=", slab_max_order_setup);
810 * Special reaping functions for NUMA systems called from cache_reap().
811 * These take care of doing round robin flushing of alien caches (containing
812 * objects freed on different nodes from which they were allocated) and the
813 * flushing of remote pcps by calling drain_node_pages.
815 static DEFINE_PER_CPU(unsigned long, slab_reap_node);
817 static void init_reap_node(int cpu)
821 node = next_node(cpu_to_mem(cpu), node_online_map);
822 if (node == MAX_NUMNODES)
823 node = first_node(node_online_map);
825 per_cpu(slab_reap_node, cpu) = node;
828 static void next_reap_node(void)
830 int node = __this_cpu_read(slab_reap_node);
832 node = next_node(node, node_online_map);
833 if (unlikely(node >= MAX_NUMNODES))
834 node = first_node(node_online_map);
835 __this_cpu_write(slab_reap_node, node);
839 #define init_reap_node(cpu) do { } while (0)
840 #define next_reap_node(void) do { } while (0)
844 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
845 * via the workqueue/eventd.
846 * Add the CPU number into the expiration time to minimize the possibility of
847 * the CPUs getting into lockstep and contending for the global cache chain
850 static void __cpuinit start_cpu_timer(int cpu)
852 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
855 * When this gets called from do_initcalls via cpucache_init(),
856 * init_workqueues() has already run, so keventd will be setup
859 if (keventd_up() && reap_work->work.func == NULL) {
861 INIT_DEFERRABLE_WORK(reap_work, cache_reap);
862 schedule_delayed_work_on(cpu, reap_work,
863 __round_jiffies_relative(HZ, cpu));
867 static struct array_cache *alloc_arraycache(int node, int entries,
868 int batchcount, gfp_t gfp)
870 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
871 struct array_cache *nc = NULL;
873 nc = kmalloc_node(memsize, gfp, node);
875 * The array_cache structures contain pointers to free object.
876 * However, when such objects are allocated or transferred to another
877 * cache the pointers are not cleared and they could be counted as
878 * valid references during a kmemleak scan. Therefore, kmemleak must
879 * not scan such objects.
881 kmemleak_no_scan(nc);
885 nc->batchcount = batchcount;
887 spin_lock_init(&nc->lock);
892 static inline bool is_slab_pfmemalloc(struct slab *slabp)
894 struct page *page = virt_to_page(slabp->s_mem);
896 return PageSlabPfmemalloc(page);
899 /* Clears pfmemalloc_active if no slabs have pfmalloc set */
900 static void recheck_pfmemalloc_active(struct kmem_cache *cachep,
901 struct array_cache *ac)
903 struct kmem_cache_node *l3 = cachep->nodelists[numa_mem_id()];
907 if (!pfmemalloc_active)
910 spin_lock_irqsave(&l3->list_lock, flags);
911 list_for_each_entry(slabp, &l3->slabs_full, list)
912 if (is_slab_pfmemalloc(slabp))
915 list_for_each_entry(slabp, &l3->slabs_partial, list)
916 if (is_slab_pfmemalloc(slabp))
919 list_for_each_entry(slabp, &l3->slabs_free, list)
920 if (is_slab_pfmemalloc(slabp))
923 pfmemalloc_active = false;
925 spin_unlock_irqrestore(&l3->list_lock, flags);
928 static void *__ac_get_obj(struct kmem_cache *cachep, struct array_cache *ac,
929 gfp_t flags, bool force_refill)
932 void *objp = ac->entry[--ac->avail];
934 /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
935 if (unlikely(is_obj_pfmemalloc(objp))) {
936 struct kmem_cache_node *l3;
938 if (gfp_pfmemalloc_allowed(flags)) {
939 clear_obj_pfmemalloc(&objp);
943 /* The caller cannot use PFMEMALLOC objects, find another one */
944 for (i = 0; i < ac->avail; i++) {
945 /* If a !PFMEMALLOC object is found, swap them */
946 if (!is_obj_pfmemalloc(ac->entry[i])) {
948 ac->entry[i] = ac->entry[ac->avail];
949 ac->entry[ac->avail] = objp;
955 * If there are empty slabs on the slabs_free list and we are
956 * being forced to refill the cache, mark this one !pfmemalloc.
958 l3 = cachep->nodelists[numa_mem_id()];
959 if (!list_empty(&l3->slabs_free) && force_refill) {
960 struct slab *slabp = virt_to_slab(objp);
961 ClearPageSlabPfmemalloc(virt_to_head_page(slabp->s_mem));
962 clear_obj_pfmemalloc(&objp);
963 recheck_pfmemalloc_active(cachep, ac);
967 /* No !PFMEMALLOC objects available */
975 static inline void *ac_get_obj(struct kmem_cache *cachep,
976 struct array_cache *ac, gfp_t flags, bool force_refill)
980 if (unlikely(sk_memalloc_socks()))
981 objp = __ac_get_obj(cachep, ac, flags, force_refill);
983 objp = ac->entry[--ac->avail];
988 static void *__ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
991 if (unlikely(pfmemalloc_active)) {
992 /* Some pfmemalloc slabs exist, check if this is one */
993 struct page *page = virt_to_head_page(objp);
994 if (PageSlabPfmemalloc(page))
995 set_obj_pfmemalloc(&objp);
1001 static inline void ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
1004 if (unlikely(sk_memalloc_socks()))
1005 objp = __ac_put_obj(cachep, ac, objp);
1007 ac->entry[ac->avail++] = objp;
1011 * Transfer objects in one arraycache to another.
1012 * Locking must be handled by the caller.
1014 * Return the number of entries transferred.
1016 static int transfer_objects(struct array_cache *to,
1017 struct array_cache *from, unsigned int max)
1019 /* Figure out how many entries to transfer */
1020 int nr = min3(from->avail, max, to->limit - to->avail);
1025 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
1026 sizeof(void *) *nr);
1035 #define drain_alien_cache(cachep, alien) do { } while (0)
1036 #define reap_alien(cachep, l3) do { } while (0)
1038 static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
1040 return (struct array_cache **)BAD_ALIEN_MAGIC;
1043 static inline void free_alien_cache(struct array_cache **ac_ptr)
1047 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1052 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
1058 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
1059 gfp_t flags, int nodeid)
1064 #else /* CONFIG_NUMA */
1066 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
1067 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
1069 static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
1071 struct array_cache **ac_ptr;
1072 int memsize = sizeof(void *) * nr_node_ids;
1077 ac_ptr = kzalloc_node(memsize, gfp, node);
1080 if (i == node || !node_online(i))
1082 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
1084 for (i--; i >= 0; i--)
1094 static void free_alien_cache(struct array_cache **ac_ptr)
1105 static void __drain_alien_cache(struct kmem_cache *cachep,
1106 struct array_cache *ac, int node)
1108 struct kmem_cache_node *rl3 = cachep->nodelists[node];
1111 spin_lock(&rl3->list_lock);
1113 * Stuff objects into the remote nodes shared array first.
1114 * That way we could avoid the overhead of putting the objects
1115 * into the free lists and getting them back later.
1118 transfer_objects(rl3->shared, ac, ac->limit);
1120 free_block(cachep, ac->entry, ac->avail, node);
1122 spin_unlock(&rl3->list_lock);
1127 * Called from cache_reap() to regularly drain alien caches round robin.
1129 static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *l3)
1131 int node = __this_cpu_read(slab_reap_node);
1134 struct array_cache *ac = l3->alien[node];
1136 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1137 __drain_alien_cache(cachep, ac, node);
1138 spin_unlock_irq(&ac->lock);
1143 static void drain_alien_cache(struct kmem_cache *cachep,
1144 struct array_cache **alien)
1147 struct array_cache *ac;
1148 unsigned long flags;
1150 for_each_online_node(i) {
1153 spin_lock_irqsave(&ac->lock, flags);
1154 __drain_alien_cache(cachep, ac, i);
1155 spin_unlock_irqrestore(&ac->lock, flags);
1160 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1162 struct slab *slabp = virt_to_slab(objp);
1163 int nodeid = slabp->nodeid;
1164 struct kmem_cache_node *l3;
1165 struct array_cache *alien = NULL;
1168 node = numa_mem_id();
1171 * Make sure we are not freeing a object from another node to the array
1172 * cache on this cpu.
1174 if (likely(slabp->nodeid == node))
1177 l3 = cachep->nodelists[node];
1178 STATS_INC_NODEFREES(cachep);
1179 if (l3->alien && l3->alien[nodeid]) {
1180 alien = l3->alien[nodeid];
1181 spin_lock(&alien->lock);
1182 if (unlikely(alien->avail == alien->limit)) {
1183 STATS_INC_ACOVERFLOW(cachep);
1184 __drain_alien_cache(cachep, alien, nodeid);
1186 ac_put_obj(cachep, alien, objp);
1187 spin_unlock(&alien->lock);
1189 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1190 free_block(cachep, &objp, 1, nodeid);
1191 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1198 * Allocates and initializes nodelists for a node on each slab cache, used for
1199 * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3
1200 * will be allocated off-node since memory is not yet online for the new node.
1201 * When hotplugging memory or a cpu, existing nodelists are not replaced if
1204 * Must hold slab_mutex.
1206 static int init_cache_nodelists_node(int node)
1208 struct kmem_cache *cachep;
1209 struct kmem_cache_node *l3;
1210 const int memsize = sizeof(struct kmem_cache_node);
1212 list_for_each_entry(cachep, &slab_caches, list) {
1214 * Set up the size64 kmemlist for cpu before we can
1215 * begin anything. Make sure some other cpu on this
1216 * node has not already allocated this
1218 if (!cachep->nodelists[node]) {
1219 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1222 kmem_list3_init(l3);
1223 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1224 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1227 * The l3s don't come and go as CPUs come and
1228 * go. slab_mutex is sufficient
1231 cachep->nodelists[node] = l3;
1234 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1235 cachep->nodelists[node]->free_limit =
1236 (1 + nr_cpus_node(node)) *
1237 cachep->batchcount + cachep->num;
1238 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1243 static void __cpuinit cpuup_canceled(long cpu)
1245 struct kmem_cache *cachep;
1246 struct kmem_cache_node *l3 = NULL;
1247 int node = cpu_to_mem(cpu);
1248 const struct cpumask *mask = cpumask_of_node(node);
1250 list_for_each_entry(cachep, &slab_caches, list) {
1251 struct array_cache *nc;
1252 struct array_cache *shared;
1253 struct array_cache **alien;
1255 /* cpu is dead; no one can alloc from it. */
1256 nc = cachep->array[cpu];
1257 cachep->array[cpu] = NULL;
1258 l3 = cachep->nodelists[node];
1261 goto free_array_cache;
1263 spin_lock_irq(&l3->list_lock);
1265 /* Free limit for this kmem_list3 */
1266 l3->free_limit -= cachep->batchcount;
1268 free_block(cachep, nc->entry, nc->avail, node);
1270 if (!cpumask_empty(mask)) {
1271 spin_unlock_irq(&l3->list_lock);
1272 goto free_array_cache;
1275 shared = l3->shared;
1277 free_block(cachep, shared->entry,
1278 shared->avail, node);
1285 spin_unlock_irq(&l3->list_lock);
1289 drain_alien_cache(cachep, alien);
1290 free_alien_cache(alien);
1296 * In the previous loop, all the objects were freed to
1297 * the respective cache's slabs, now we can go ahead and
1298 * shrink each nodelist to its limit.
1300 list_for_each_entry(cachep, &slab_caches, list) {
1301 l3 = cachep->nodelists[node];
1304 drain_freelist(cachep, l3, l3->free_objects);
1308 static int __cpuinit cpuup_prepare(long cpu)
1310 struct kmem_cache *cachep;
1311 struct kmem_cache_node *l3 = NULL;
1312 int node = cpu_to_mem(cpu);
1316 * We need to do this right in the beginning since
1317 * alloc_arraycache's are going to use this list.
1318 * kmalloc_node allows us to add the slab to the right
1319 * kmem_list3 and not this cpu's kmem_list3
1321 err = init_cache_nodelists_node(node);
1326 * Now we can go ahead with allocating the shared arrays and
1329 list_for_each_entry(cachep, &slab_caches, list) {
1330 struct array_cache *nc;
1331 struct array_cache *shared = NULL;
1332 struct array_cache **alien = NULL;
1334 nc = alloc_arraycache(node, cachep->limit,
1335 cachep->batchcount, GFP_KERNEL);
1338 if (cachep->shared) {
1339 shared = alloc_arraycache(node,
1340 cachep->shared * cachep->batchcount,
1341 0xbaadf00d, GFP_KERNEL);
1347 if (use_alien_caches) {
1348 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
1355 cachep->array[cpu] = nc;
1356 l3 = cachep->nodelists[node];
1359 spin_lock_irq(&l3->list_lock);
1362 * We are serialised from CPU_DEAD or
1363 * CPU_UP_CANCELLED by the cpucontrol lock
1365 l3->shared = shared;
1374 spin_unlock_irq(&l3->list_lock);
1376 free_alien_cache(alien);
1377 if (cachep->flags & SLAB_DEBUG_OBJECTS)
1378 slab_set_debugobj_lock_classes_node(cachep, node);
1379 else if (!OFF_SLAB(cachep) &&
1380 !(cachep->flags & SLAB_DESTROY_BY_RCU))
1381 on_slab_lock_classes_node(cachep, node);
1383 init_node_lock_keys(node);
1387 cpuup_canceled(cpu);
1391 static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1392 unsigned long action, void *hcpu)
1394 long cpu = (long)hcpu;
1398 case CPU_UP_PREPARE:
1399 case CPU_UP_PREPARE_FROZEN:
1400 mutex_lock(&slab_mutex);
1401 err = cpuup_prepare(cpu);
1402 mutex_unlock(&slab_mutex);
1405 case CPU_ONLINE_FROZEN:
1406 start_cpu_timer(cpu);
1408 #ifdef CONFIG_HOTPLUG_CPU
1409 case CPU_DOWN_PREPARE:
1410 case CPU_DOWN_PREPARE_FROZEN:
1412 * Shutdown cache reaper. Note that the slab_mutex is
1413 * held so that if cache_reap() is invoked it cannot do
1414 * anything expensive but will only modify reap_work
1415 * and reschedule the timer.
1417 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1418 /* Now the cache_reaper is guaranteed to be not running. */
1419 per_cpu(slab_reap_work, cpu).work.func = NULL;
1421 case CPU_DOWN_FAILED:
1422 case CPU_DOWN_FAILED_FROZEN:
1423 start_cpu_timer(cpu);
1426 case CPU_DEAD_FROZEN:
1428 * Even if all the cpus of a node are down, we don't free the
1429 * kmem_list3 of any cache. This to avoid a race between
1430 * cpu_down, and a kmalloc allocation from another cpu for
1431 * memory from the node of the cpu going down. The list3
1432 * structure is usually allocated from kmem_cache_create() and
1433 * gets destroyed at kmem_cache_destroy().
1437 case CPU_UP_CANCELED:
1438 case CPU_UP_CANCELED_FROZEN:
1439 mutex_lock(&slab_mutex);
1440 cpuup_canceled(cpu);
1441 mutex_unlock(&slab_mutex);
1444 return notifier_from_errno(err);
1447 static struct notifier_block __cpuinitdata cpucache_notifier = {
1448 &cpuup_callback, NULL, 0
1451 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1453 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1454 * Returns -EBUSY if all objects cannot be drained so that the node is not
1457 * Must hold slab_mutex.
1459 static int __meminit drain_cache_nodelists_node(int node)
1461 struct kmem_cache *cachep;
1464 list_for_each_entry(cachep, &slab_caches, list) {
1465 struct kmem_cache_node *l3;
1467 l3 = cachep->nodelists[node];
1471 drain_freelist(cachep, l3, l3->free_objects);
1473 if (!list_empty(&l3->slabs_full) ||
1474 !list_empty(&l3->slabs_partial)) {
1482 static int __meminit slab_memory_callback(struct notifier_block *self,
1483 unsigned long action, void *arg)
1485 struct memory_notify *mnb = arg;
1489 nid = mnb->status_change_nid;
1494 case MEM_GOING_ONLINE:
1495 mutex_lock(&slab_mutex);
1496 ret = init_cache_nodelists_node(nid);
1497 mutex_unlock(&slab_mutex);
1499 case MEM_GOING_OFFLINE:
1500 mutex_lock(&slab_mutex);
1501 ret = drain_cache_nodelists_node(nid);
1502 mutex_unlock(&slab_mutex);
1506 case MEM_CANCEL_ONLINE:
1507 case MEM_CANCEL_OFFLINE:
1511 return notifier_from_errno(ret);
1513 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1516 * swap the static kmem_list3 with kmalloced memory
1518 static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list,
1521 struct kmem_cache_node *ptr;
1523 ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid);
1526 memcpy(ptr, list, sizeof(struct kmem_cache_node));
1528 * Do not assume that spinlocks can be initialized via memcpy:
1530 spin_lock_init(&ptr->list_lock);
1532 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1533 cachep->nodelists[nodeid] = ptr;
1537 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1538 * size of kmem_list3.
1540 static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1544 for_each_online_node(node) {
1545 cachep->nodelists[node] = &initkmem_list3[index + node];
1546 cachep->nodelists[node]->next_reap = jiffies +
1548 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1553 * The memory after the last cpu cache pointer is used for the
1554 * the nodelists pointer.
1556 static void setup_nodelists_pointer(struct kmem_cache *cachep)
1558 cachep->nodelists = (struct kmem_cache_node **)&cachep->array[nr_cpu_ids];
1562 * Initialisation. Called after the page allocator have been initialised and
1563 * before smp_init().
1565 void __init kmem_cache_init(void)
1569 kmem_cache = &kmem_cache_boot;
1570 setup_nodelists_pointer(kmem_cache);
1572 if (num_possible_nodes() == 1)
1573 use_alien_caches = 0;
1575 for (i = 0; i < NUM_INIT_LISTS; i++)
1576 kmem_list3_init(&initkmem_list3[i]);
1578 set_up_list3s(kmem_cache, CACHE_CACHE);
1581 * Fragmentation resistance on low memory - only use bigger
1582 * page orders on machines with more than 32MB of memory if
1583 * not overridden on the command line.
1585 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
1586 slab_max_order = SLAB_MAX_ORDER_HI;
1588 /* Bootstrap is tricky, because several objects are allocated
1589 * from caches that do not exist yet:
1590 * 1) initialize the kmem_cache cache: it contains the struct
1591 * kmem_cache structures of all caches, except kmem_cache itself:
1592 * kmem_cache is statically allocated.
1593 * Initially an __init data area is used for the head array and the
1594 * kmem_list3 structures, it's replaced with a kmalloc allocated
1595 * array at the end of the bootstrap.
1596 * 2) Create the first kmalloc cache.
1597 * The struct kmem_cache for the new cache is allocated normally.
1598 * An __init data area is used for the head array.
1599 * 3) Create the remaining kmalloc caches, with minimally sized
1601 * 4) Replace the __init data head arrays for kmem_cache and the first
1602 * kmalloc cache with kmalloc allocated arrays.
1603 * 5) Replace the __init data for kmem_list3 for kmem_cache and
1604 * the other cache's with kmalloc allocated memory.
1605 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1608 /* 1) create the kmem_cache */
1611 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1613 create_boot_cache(kmem_cache, "kmem_cache",
1614 offsetof(struct kmem_cache, array[nr_cpu_ids]) +
1615 nr_node_ids * sizeof(struct kmem_cache_node *),
1616 SLAB_HWCACHE_ALIGN);
1617 list_add(&kmem_cache->list, &slab_caches);
1619 /* 2+3) create the kmalloc caches */
1622 * Initialize the caches that provide memory for the array cache and the
1623 * kmem_list3 structures first. Without this, further allocations will
1627 kmalloc_caches[INDEX_AC] = create_kmalloc_cache("kmalloc-ac",
1628 kmalloc_size(INDEX_AC), ARCH_KMALLOC_FLAGS);
1630 if (INDEX_AC != INDEX_L3)
1631 kmalloc_caches[INDEX_L3] =
1632 create_kmalloc_cache("kmalloc-l3",
1633 kmalloc_size(INDEX_L3), ARCH_KMALLOC_FLAGS);
1635 slab_early_init = 0;
1637 for (i = 1; i < PAGE_SHIFT + MAX_ORDER; i++) {
1638 size_t cs_size = kmalloc_size(i);
1640 if (cs_size < KMALLOC_MIN_SIZE)
1643 if (!kmalloc_caches[i]) {
1645 * For performance, all the general caches are L1 aligned.
1646 * This should be particularly beneficial on SMP boxes, as it
1647 * eliminates "false sharing".
1648 * Note for systems short on memory removing the alignment will
1649 * allow tighter packing of the smaller caches.
1651 kmalloc_caches[i] = create_kmalloc_cache("kmalloc",
1652 cs_size, ARCH_KMALLOC_FLAGS);
1655 #ifdef CONFIG_ZONE_DMA
1656 kmalloc_dma_caches[i] = create_kmalloc_cache(
1657 "kmalloc-dma", cs_size,
1658 SLAB_CACHE_DMA|ARCH_KMALLOC_FLAGS);
1661 /* 4) Replace the bootstrap head arrays */
1663 struct array_cache *ptr;
1665 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1667 memcpy(ptr, cpu_cache_get(kmem_cache),
1668 sizeof(struct arraycache_init));
1670 * Do not assume that spinlocks can be initialized via memcpy:
1672 spin_lock_init(&ptr->lock);
1674 kmem_cache->array[smp_processor_id()] = ptr;
1676 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1678 BUG_ON(cpu_cache_get(kmalloc_caches[INDEX_AC])
1679 != &initarray_generic.cache);
1680 memcpy(ptr, cpu_cache_get(kmalloc_caches[INDEX_AC]),
1681 sizeof(struct arraycache_init));
1683 * Do not assume that spinlocks can be initialized via memcpy:
1685 spin_lock_init(&ptr->lock);
1687 kmalloc_caches[INDEX_AC]->array[smp_processor_id()] = ptr;
1689 /* 5) Replace the bootstrap kmem_list3's */
1693 for_each_online_node(nid) {
1694 init_list(kmem_cache, &initkmem_list3[CACHE_CACHE + nid], nid);
1696 init_list(kmalloc_caches[INDEX_AC],
1697 &initkmem_list3[SIZE_AC + nid], nid);
1699 if (INDEX_AC != INDEX_L3) {
1700 init_list(kmalloc_caches[INDEX_L3],
1701 &initkmem_list3[SIZE_L3 + nid], nid);
1708 /* Create the proper names */
1709 for (i = 1; i < PAGE_SHIFT + MAX_ORDER; i++) {
1711 struct kmem_cache *c = kmalloc_caches[i];
1716 s = kasprintf(GFP_NOWAIT, "kmalloc-%d", kmalloc_size(i));
1721 #ifdef CONFIG_ZONE_DMA
1722 c = kmalloc_dma_caches[i];
1724 s = kasprintf(GFP_NOWAIT, "dma-kmalloc-%d", kmalloc_size(i));
1731 void __init kmem_cache_init_late(void)
1733 struct kmem_cache *cachep;
1737 /* 6) resize the head arrays to their final sizes */
1738 mutex_lock(&slab_mutex);
1739 list_for_each_entry(cachep, &slab_caches, list)
1740 if (enable_cpucache(cachep, GFP_NOWAIT))
1742 mutex_unlock(&slab_mutex);
1744 /* Annotate slab for lockdep -- annotate the malloc caches */
1751 * Register a cpu startup notifier callback that initializes
1752 * cpu_cache_get for all new cpus
1754 register_cpu_notifier(&cpucache_notifier);
1758 * Register a memory hotplug callback that initializes and frees
1761 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1765 * The reap timers are started later, with a module init call: That part
1766 * of the kernel is not yet operational.
1770 static int __init cpucache_init(void)
1775 * Register the timers that return unneeded pages to the page allocator
1777 for_each_online_cpu(cpu)
1778 start_cpu_timer(cpu);
1784 __initcall(cpucache_init);
1786 static noinline void
1787 slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1789 struct kmem_cache_node *l3;
1791 unsigned long flags;
1795 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1797 printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n",
1798 cachep->name, cachep->size, cachep->gfporder);
1800 for_each_online_node(node) {
1801 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1802 unsigned long active_slabs = 0, num_slabs = 0;
1804 l3 = cachep->nodelists[node];
1808 spin_lock_irqsave(&l3->list_lock, flags);
1809 list_for_each_entry(slabp, &l3->slabs_full, list) {
1810 active_objs += cachep->num;
1813 list_for_each_entry(slabp, &l3->slabs_partial, list) {
1814 active_objs += slabp->inuse;
1817 list_for_each_entry(slabp, &l3->slabs_free, list)
1820 free_objects += l3->free_objects;
1821 spin_unlock_irqrestore(&l3->list_lock, flags);
1823 num_slabs += active_slabs;
1824 num_objs = num_slabs * cachep->num;
1826 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1827 node, active_slabs, num_slabs, active_objs, num_objs,
1833 * Interface to system's page allocator. No need to hold the cache-lock.
1835 * If we requested dmaable memory, we will get it. Even if we
1836 * did not request dmaable memory, we might get it, but that
1837 * would be relatively rare and ignorable.
1839 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1847 * Nommu uses slab's for process anonymous memory allocations, and thus
1848 * requires __GFP_COMP to properly refcount higher order allocations
1850 flags |= __GFP_COMP;
1853 flags |= cachep->allocflags;
1854 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1855 flags |= __GFP_RECLAIMABLE;
1857 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1859 if (!(flags & __GFP_NOWARN) && printk_ratelimit())
1860 slab_out_of_memory(cachep, flags, nodeid);
1864 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1865 if (unlikely(page->pfmemalloc))
1866 pfmemalloc_active = true;
1868 nr_pages = (1 << cachep->gfporder);
1869 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1870 add_zone_page_state(page_zone(page),
1871 NR_SLAB_RECLAIMABLE, nr_pages);
1873 add_zone_page_state(page_zone(page),
1874 NR_SLAB_UNRECLAIMABLE, nr_pages);
1875 for (i = 0; i < nr_pages; i++) {
1876 __SetPageSlab(page + i);
1878 if (page->pfmemalloc)
1879 SetPageSlabPfmemalloc(page + i);
1881 memcg_bind_pages(cachep, cachep->gfporder);
1883 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1884 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1887 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1889 kmemcheck_mark_unallocated_pages(page, nr_pages);
1892 return page_address(page);
1896 * Interface to system's page release.
1898 static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1900 unsigned long i = (1 << cachep->gfporder);
1901 struct page *page = virt_to_page(addr);
1902 const unsigned long nr_freed = i;
1904 kmemcheck_free_shadow(page, cachep->gfporder);
1906 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1907 sub_zone_page_state(page_zone(page),
1908 NR_SLAB_RECLAIMABLE, nr_freed);
1910 sub_zone_page_state(page_zone(page),
1911 NR_SLAB_UNRECLAIMABLE, nr_freed);
1913 BUG_ON(!PageSlab(page));
1914 __ClearPageSlabPfmemalloc(page);
1915 __ClearPageSlab(page);
1919 memcg_release_pages(cachep, cachep->gfporder);
1920 if (current->reclaim_state)
1921 current->reclaim_state->reclaimed_slab += nr_freed;
1922 free_memcg_kmem_pages((unsigned long)addr, cachep->gfporder);
1925 static void kmem_rcu_free(struct rcu_head *head)
1927 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1928 struct kmem_cache *cachep = slab_rcu->cachep;
1930 kmem_freepages(cachep, slab_rcu->addr);
1931 if (OFF_SLAB(cachep))
1932 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1937 #ifdef CONFIG_DEBUG_PAGEALLOC
1938 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1939 unsigned long caller)
1941 int size = cachep->object_size;
1943 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1945 if (size < 5 * sizeof(unsigned long))
1948 *addr++ = 0x12345678;
1950 *addr++ = smp_processor_id();
1951 size -= 3 * sizeof(unsigned long);
1953 unsigned long *sptr = &caller;
1954 unsigned long svalue;
1956 while (!kstack_end(sptr)) {
1958 if (kernel_text_address(svalue)) {
1960 size -= sizeof(unsigned long);
1961 if (size <= sizeof(unsigned long))
1967 *addr++ = 0x87654321;
1971 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1973 int size = cachep->object_size;
1974 addr = &((char *)addr)[obj_offset(cachep)];
1976 memset(addr, val, size);
1977 *(unsigned char *)(addr + size - 1) = POISON_END;
1980 static void dump_line(char *data, int offset, int limit)
1983 unsigned char error = 0;
1986 printk(KERN_ERR "%03x: ", offset);
1987 for (i = 0; i < limit; i++) {
1988 if (data[offset + i] != POISON_FREE) {
1989 error = data[offset + i];
1993 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1994 &data[offset], limit, 1);
1996 if (bad_count == 1) {
1997 error ^= POISON_FREE;
1998 if (!(error & (error - 1))) {
1999 printk(KERN_ERR "Single bit error detected. Probably "
2002 printk(KERN_ERR "Run memtest86+ or a similar memory "
2005 printk(KERN_ERR "Run a memory test tool.\n");
2014 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
2019 if (cachep->flags & SLAB_RED_ZONE) {
2020 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
2021 *dbg_redzone1(cachep, objp),
2022 *dbg_redzone2(cachep, objp));
2025 if (cachep->flags & SLAB_STORE_USER) {
2026 printk(KERN_ERR "Last user: [<%p>]",
2027 *dbg_userword(cachep, objp));
2028 print_symbol("(%s)",
2029 (unsigned long)*dbg_userword(cachep, objp));
2032 realobj = (char *)objp + obj_offset(cachep);
2033 size = cachep->object_size;
2034 for (i = 0; i < size && lines; i += 16, lines--) {
2037 if (i + limit > size)
2039 dump_line(realobj, i, limit);
2043 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
2049 realobj = (char *)objp + obj_offset(cachep);
2050 size = cachep->object_size;
2052 for (i = 0; i < size; i++) {
2053 char exp = POISON_FREE;
2056 if (realobj[i] != exp) {
2062 "Slab corruption (%s): %s start=%p, len=%d\n",
2063 print_tainted(), cachep->name, realobj, size);
2064 print_objinfo(cachep, objp, 0);
2066 /* Hexdump the affected line */
2069 if (i + limit > size)
2071 dump_line(realobj, i, limit);
2074 /* Limit to 5 lines */
2080 /* Print some data about the neighboring objects, if they
2083 struct slab *slabp = virt_to_slab(objp);
2086 objnr = obj_to_index(cachep, slabp, objp);
2088 objp = index_to_obj(cachep, slabp, objnr - 1);
2089 realobj = (char *)objp + obj_offset(cachep);
2090 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
2092 print_objinfo(cachep, objp, 2);
2094 if (objnr + 1 < cachep->num) {
2095 objp = index_to_obj(cachep, slabp, objnr + 1);
2096 realobj = (char *)objp + obj_offset(cachep);
2097 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
2099 print_objinfo(cachep, objp, 2);
2106 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
2109 for (i = 0; i < cachep->num; i++) {
2110 void *objp = index_to_obj(cachep, slabp, i);
2112 if (cachep->flags & SLAB_POISON) {
2113 #ifdef CONFIG_DEBUG_PAGEALLOC
2114 if (cachep->size % PAGE_SIZE == 0 &&
2116 kernel_map_pages(virt_to_page(objp),
2117 cachep->size / PAGE_SIZE, 1);
2119 check_poison_obj(cachep, objp);
2121 check_poison_obj(cachep, objp);
2124 if (cachep->flags & SLAB_RED_ZONE) {
2125 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2126 slab_error(cachep, "start of a freed object "
2128 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2129 slab_error(cachep, "end of a freed object "
2135 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
2141 * slab_destroy - destroy and release all objects in a slab
2142 * @cachep: cache pointer being destroyed
2143 * @slabp: slab pointer being destroyed
2145 * Destroy all the objs in a slab, and release the mem back to the system.
2146 * Before calling the slab must have been unlinked from the cache. The
2147 * cache-lock is not held/needed.
2149 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
2151 void *addr = slabp->s_mem - slabp->colouroff;
2153 slab_destroy_debugcheck(cachep, slabp);
2154 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
2155 struct slab_rcu *slab_rcu;
2157 slab_rcu = (struct slab_rcu *)slabp;
2158 slab_rcu->cachep = cachep;
2159 slab_rcu->addr = addr;
2160 call_rcu(&slab_rcu->head, kmem_rcu_free);
2162 kmem_freepages(cachep, addr);
2163 if (OFF_SLAB(cachep))
2164 kmem_cache_free(cachep->slabp_cache, slabp);
2169 * calculate_slab_order - calculate size (page order) of slabs
2170 * @cachep: pointer to the cache that is being created
2171 * @size: size of objects to be created in this cache.
2172 * @align: required alignment for the objects.
2173 * @flags: slab allocation flags
2175 * Also calculates the number of objects per slab.
2177 * This could be made much more intelligent. For now, try to avoid using
2178 * high order pages for slabs. When the gfp() functions are more friendly
2179 * towards high-order requests, this should be changed.
2181 static size_t calculate_slab_order(struct kmem_cache *cachep,
2182 size_t size, size_t align, unsigned long flags)
2184 unsigned long offslab_limit;
2185 size_t left_over = 0;
2188 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
2192 cache_estimate(gfporder, size, align, flags, &remainder, &num);
2196 if (flags & CFLGS_OFF_SLAB) {
2198 * Max number of objs-per-slab for caches which
2199 * use off-slab slabs. Needed to avoid a possible
2200 * looping condition in cache_grow().
2202 offslab_limit = size - sizeof(struct slab);
2203 offslab_limit /= sizeof(kmem_bufctl_t);
2205 if (num > offslab_limit)
2209 /* Found something acceptable - save it away */
2211 cachep->gfporder = gfporder;
2212 left_over = remainder;
2215 * A VFS-reclaimable slab tends to have most allocations
2216 * as GFP_NOFS and we really don't want to have to be allocating
2217 * higher-order pages when we are unable to shrink dcache.
2219 if (flags & SLAB_RECLAIM_ACCOUNT)
2223 * Large number of objects is good, but very large slabs are
2224 * currently bad for the gfp()s.
2226 if (gfporder >= slab_max_order)
2230 * Acceptable internal fragmentation?
2232 if (left_over * 8 <= (PAGE_SIZE << gfporder))
2238 static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
2240 if (slab_state >= FULL)
2241 return enable_cpucache(cachep, gfp);
2243 if (slab_state == DOWN) {
2245 * Note: Creation of first cache (kmem_cache).
2246 * The setup_list3s is taken care
2247 * of by the caller of __kmem_cache_create
2249 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2250 slab_state = PARTIAL;
2251 } else if (slab_state == PARTIAL) {
2253 * Note: the second kmem_cache_create must create the cache
2254 * that's used by kmalloc(24), otherwise the creation of
2255 * further caches will BUG().
2257 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2260 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2261 * the second cache, then we need to set up all its list3s,
2262 * otherwise the creation of further caches will BUG().
2264 set_up_list3s(cachep, SIZE_AC);
2265 if (INDEX_AC == INDEX_L3)
2266 slab_state = PARTIAL_L3;
2268 slab_state = PARTIAL_ARRAYCACHE;
2270 /* Remaining boot caches */
2271 cachep->array[smp_processor_id()] =
2272 kmalloc(sizeof(struct arraycache_init), gfp);
2274 if (slab_state == PARTIAL_ARRAYCACHE) {
2275 set_up_list3s(cachep, SIZE_L3);
2276 slab_state = PARTIAL_L3;
2279 for_each_online_node(node) {
2280 cachep->nodelists[node] =
2281 kmalloc_node(sizeof(struct kmem_cache_node),
2283 BUG_ON(!cachep->nodelists[node]);
2284 kmem_list3_init(cachep->nodelists[node]);
2288 cachep->nodelists[numa_mem_id()]->next_reap =
2289 jiffies + REAPTIMEOUT_LIST3 +
2290 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2292 cpu_cache_get(cachep)->avail = 0;
2293 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2294 cpu_cache_get(cachep)->batchcount = 1;
2295 cpu_cache_get(cachep)->touched = 0;
2296 cachep->batchcount = 1;
2297 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2302 * __kmem_cache_create - Create a cache.
2303 * @cachep: cache management descriptor
2304 * @flags: SLAB flags
2306 * Returns a ptr to the cache on success, NULL on failure.
2307 * Cannot be called within a int, but can be interrupted.
2308 * The @ctor is run when new pages are allocated by the cache.
2312 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2313 * to catch references to uninitialised memory.
2315 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2316 * for buffer overruns.
2318 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2319 * cacheline. This can be beneficial if you're counting cycles as closely
2323 __kmem_cache_create (struct kmem_cache *cachep, unsigned long flags)
2325 size_t left_over, slab_size, ralign;
2328 size_t size = cachep->size;
2333 * Enable redzoning and last user accounting, except for caches with
2334 * large objects, if the increased size would increase the object size
2335 * above the next power of two: caches with object sizes just above a
2336 * power of two have a significant amount of internal fragmentation.
2338 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2339 2 * sizeof(unsigned long long)))
2340 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2341 if (!(flags & SLAB_DESTROY_BY_RCU))
2342 flags |= SLAB_POISON;
2344 if (flags & SLAB_DESTROY_BY_RCU)
2345 BUG_ON(flags & SLAB_POISON);
2349 * Check that size is in terms of words. This is needed to avoid
2350 * unaligned accesses for some archs when redzoning is used, and makes
2351 * sure any on-slab bufctl's are also correctly aligned.
2353 if (size & (BYTES_PER_WORD - 1)) {
2354 size += (BYTES_PER_WORD - 1);
2355 size &= ~(BYTES_PER_WORD - 1);
2359 * Redzoning and user store require word alignment or possibly larger.
2360 * Note this will be overridden by architecture or caller mandated
2361 * alignment if either is greater than BYTES_PER_WORD.
2363 if (flags & SLAB_STORE_USER)
2364 ralign = BYTES_PER_WORD;
2366 if (flags & SLAB_RED_ZONE) {
2367 ralign = REDZONE_ALIGN;
2368 /* If redzoning, ensure that the second redzone is suitably
2369 * aligned, by adjusting the object size accordingly. */
2370 size += REDZONE_ALIGN - 1;
2371 size &= ~(REDZONE_ALIGN - 1);
2374 /* 3) caller mandated alignment */
2375 if (ralign < cachep->align) {
2376 ralign = cachep->align;
2378 /* disable debug if necessary */
2379 if (ralign > __alignof__(unsigned long long))
2380 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2384 cachep->align = ralign;
2386 if (slab_is_available())
2391 setup_nodelists_pointer(cachep);
2395 * Both debugging options require word-alignment which is calculated
2398 if (flags & SLAB_RED_ZONE) {
2399 /* add space for red zone words */
2400 cachep->obj_offset += sizeof(unsigned long long);
2401 size += 2 * sizeof(unsigned long long);
2403 if (flags & SLAB_STORE_USER) {
2404 /* user store requires one word storage behind the end of
2405 * the real object. But if the second red zone needs to be
2406 * aligned to 64 bits, we must allow that much space.
2408 if (flags & SLAB_RED_ZONE)
2409 size += REDZONE_ALIGN;
2411 size += BYTES_PER_WORD;
2413 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2414 if (size >= kmalloc_size(INDEX_L3 + 1)
2415 && cachep->object_size > cache_line_size() && ALIGN(size, align) < PAGE_SIZE) {
2416 cachep->obj_offset += PAGE_SIZE - ALIGN(size, align);
2423 * Determine if the slab management is 'on' or 'off' slab.
2424 * (bootstrapping cannot cope with offslab caches so don't do
2425 * it too early on. Always use on-slab management when
2426 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2428 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
2429 !(flags & SLAB_NOLEAKTRACE))
2431 * Size is large, assume best to place the slab management obj
2432 * off-slab (should allow better packing of objs).
2434 flags |= CFLGS_OFF_SLAB;
2436 size = ALIGN(size, cachep->align);
2438 left_over = calculate_slab_order(cachep, size, cachep->align, flags);
2443 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2444 + sizeof(struct slab), cachep->align);
2447 * If the slab has been placed off-slab, and we have enough space then
2448 * move it on-slab. This is at the expense of any extra colouring.
2450 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2451 flags &= ~CFLGS_OFF_SLAB;
2452 left_over -= slab_size;
2455 if (flags & CFLGS_OFF_SLAB) {
2456 /* really off slab. No need for manual alignment */
2458 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2460 #ifdef CONFIG_PAGE_POISONING
2461 /* If we're going to use the generic kernel_map_pages()
2462 * poisoning, then it's going to smash the contents of
2463 * the redzone and userword anyhow, so switch them off.
2465 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2466 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2470 cachep->colour_off = cache_line_size();
2471 /* Offset must be a multiple of the alignment. */
2472 if (cachep->colour_off < cachep->align)
2473 cachep->colour_off = cachep->align;
2474 cachep->colour = left_over / cachep->colour_off;
2475 cachep->slab_size = slab_size;
2476 cachep->flags = flags;
2477 cachep->allocflags = 0;
2478 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2479 cachep->allocflags |= GFP_DMA;
2480 cachep->size = size;
2481 cachep->reciprocal_buffer_size = reciprocal_value(size);
2483 if (flags & CFLGS_OFF_SLAB) {
2484 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2486 * This is a possibility for one of the malloc_sizes caches.
2487 * But since we go off slab only for object size greater than
2488 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2489 * this should not happen at all.
2490 * But leave a BUG_ON for some lucky dude.
2492 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
2495 err = setup_cpu_cache(cachep, gfp);
2497 __kmem_cache_shutdown(cachep);
2501 if (flags & SLAB_DEBUG_OBJECTS) {
2503 * Would deadlock through slab_destroy()->call_rcu()->
2504 * debug_object_activate()->kmem_cache_alloc().
2506 WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU);
2508 slab_set_debugobj_lock_classes(cachep);
2509 } else if (!OFF_SLAB(cachep) && !(flags & SLAB_DESTROY_BY_RCU))
2510 on_slab_lock_classes(cachep);
2516 static void check_irq_off(void)
2518 BUG_ON(!irqs_disabled());
2521 static void check_irq_on(void)
2523 BUG_ON(irqs_disabled());
2526 static void check_spinlock_acquired(struct kmem_cache *cachep)
2530 assert_spin_locked(&cachep->nodelists[numa_mem_id()]->list_lock);
2534 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2538 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2543 #define check_irq_off() do { } while(0)
2544 #define check_irq_on() do { } while(0)
2545 #define check_spinlock_acquired(x) do { } while(0)
2546 #define check_spinlock_acquired_node(x, y) do { } while(0)
2549 static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *l3,
2550 struct array_cache *ac,
2551 int force, int node);
2553 static void do_drain(void *arg)
2555 struct kmem_cache *cachep = arg;
2556 struct array_cache *ac;
2557 int node = numa_mem_id();
2560 ac = cpu_cache_get(cachep);
2561 spin_lock(&cachep->nodelists[node]->list_lock);
2562 free_block(cachep, ac->entry, ac->avail, node);
2563 spin_unlock(&cachep->nodelists[node]->list_lock);
2567 static void drain_cpu_caches(struct kmem_cache *cachep)
2569 struct kmem_cache_node *l3;
2572 on_each_cpu(do_drain, cachep, 1);
2574 for_each_online_node(node) {
2575 l3 = cachep->nodelists[node];
2576 if (l3 && l3->alien)
2577 drain_alien_cache(cachep, l3->alien);
2580 for_each_online_node(node) {
2581 l3 = cachep->nodelists[node];
2583 drain_array(cachep, l3, l3->shared, 1, node);
2588 * Remove slabs from the list of free slabs.
2589 * Specify the number of slabs to drain in tofree.
2591 * Returns the actual number of slabs released.
2593 static int drain_freelist(struct kmem_cache *cache,
2594 struct kmem_cache_node *l3, int tofree)
2596 struct list_head *p;
2601 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
2603 spin_lock_irq(&l3->list_lock);
2604 p = l3->slabs_free.prev;
2605 if (p == &l3->slabs_free) {
2606 spin_unlock_irq(&l3->list_lock);
2610 slabp = list_entry(p, struct slab, list);
2612 BUG_ON(slabp->inuse);
2614 list_del(&slabp->list);
2616 * Safe to drop the lock. The slab is no longer linked
2619 l3->free_objects -= cache->num;
2620 spin_unlock_irq(&l3->list_lock);
2621 slab_destroy(cache, slabp);
2628 /* Called with slab_mutex held to protect against cpu hotplug */
2629 static int __cache_shrink(struct kmem_cache *cachep)
2632 struct kmem_cache_node *l3;
2634 drain_cpu_caches(cachep);
2637 for_each_online_node(i) {
2638 l3 = cachep->nodelists[i];
2642 drain_freelist(cachep, l3, l3->free_objects);
2644 ret += !list_empty(&l3->slabs_full) ||
2645 !list_empty(&l3->slabs_partial);
2647 return (ret ? 1 : 0);
2651 * kmem_cache_shrink - Shrink a cache.
2652 * @cachep: The cache to shrink.
2654 * Releases as many slabs as possible for a cache.
2655 * To help debugging, a zero exit status indicates all slabs were released.
2657 int kmem_cache_shrink(struct kmem_cache *cachep)
2660 BUG_ON(!cachep || in_interrupt());
2663 mutex_lock(&slab_mutex);
2664 ret = __cache_shrink(cachep);
2665 mutex_unlock(&slab_mutex);
2669 EXPORT_SYMBOL(kmem_cache_shrink);
2671 int __kmem_cache_shutdown(struct kmem_cache *cachep)
2674 struct kmem_cache_node *l3;
2675 int rc = __cache_shrink(cachep);
2680 for_each_online_cpu(i)
2681 kfree(cachep->array[i]);
2683 /* NUMA: free the list3 structures */
2684 for_each_online_node(i) {
2685 l3 = cachep->nodelists[i];
2688 free_alien_cache(l3->alien);
2696 * Get the memory for a slab management obj.
2697 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2698 * always come from malloc_sizes caches. The slab descriptor cannot
2699 * come from the same cache which is getting created because,
2700 * when we are searching for an appropriate cache for these
2701 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2702 * If we are creating a malloc_sizes cache here it would not be visible to
2703 * kmem_find_general_cachep till the initialization is complete.
2704 * Hence we cannot have slabp_cache same as the original cache.
2706 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2707 int colour_off, gfp_t local_flags,
2712 if (OFF_SLAB(cachep)) {
2713 /* Slab management obj is off-slab. */
2714 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2715 local_flags, nodeid);
2717 * If the first object in the slab is leaked (it's allocated
2718 * but no one has a reference to it), we want to make sure
2719 * kmemleak does not treat the ->s_mem pointer as a reference
2720 * to the object. Otherwise we will not report the leak.
2722 kmemleak_scan_area(&slabp->list, sizeof(struct list_head),
2727 slabp = objp + colour_off;
2728 colour_off += cachep->slab_size;
2731 slabp->colouroff = colour_off;
2732 slabp->s_mem = objp + colour_off;
2733 slabp->nodeid = nodeid;
2738 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2740 return (kmem_bufctl_t *) (slabp + 1);
2743 static void cache_init_objs(struct kmem_cache *cachep,
2748 for (i = 0; i < cachep->num; i++) {
2749 void *objp = index_to_obj(cachep, slabp, i);
2751 /* need to poison the objs? */
2752 if (cachep->flags & SLAB_POISON)
2753 poison_obj(cachep, objp, POISON_FREE);
2754 if (cachep->flags & SLAB_STORE_USER)
2755 *dbg_userword(cachep, objp) = NULL;
2757 if (cachep->flags & SLAB_RED_ZONE) {
2758 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2759 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2762 * Constructors are not allowed to allocate memory from the same
2763 * cache which they are a constructor for. Otherwise, deadlock.
2764 * They must also be threaded.
2766 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2767 cachep->ctor(objp + obj_offset(cachep));
2769 if (cachep->flags & SLAB_RED_ZONE) {
2770 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2771 slab_error(cachep, "constructor overwrote the"
2772 " end of an object");
2773 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2774 slab_error(cachep, "constructor overwrote the"
2775 " start of an object");
2777 if ((cachep->size % PAGE_SIZE) == 0 &&
2778 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2779 kernel_map_pages(virt_to_page(objp),
2780 cachep->size / PAGE_SIZE, 0);
2785 slab_bufctl(slabp)[i] = i + 1;
2787 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2790 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2792 if (CONFIG_ZONE_DMA_FLAG) {
2793 if (flags & GFP_DMA)
2794 BUG_ON(!(cachep->allocflags & GFP_DMA));
2796 BUG_ON(cachep->allocflags & GFP_DMA);
2800 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2803 void *objp = index_to_obj(cachep, slabp, slabp->free);
2807 next = slab_bufctl(slabp)[slabp->free];
2809 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2810 WARN_ON(slabp->nodeid != nodeid);
2817 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2818 void *objp, int nodeid)
2820 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2823 /* Verify that the slab belongs to the intended node */
2824 WARN_ON(slabp->nodeid != nodeid);
2826 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2827 printk(KERN_ERR "slab: double free detected in cache "
2828 "'%s', objp %p\n", cachep->name, objp);
2832 slab_bufctl(slabp)[objnr] = slabp->free;
2833 slabp->free = objnr;
2838 * Map pages beginning at addr to the given cache and slab. This is required
2839 * for the slab allocator to be able to lookup the cache and slab of a
2840 * virtual address for kfree, ksize, and slab debugging.
2842 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2848 page = virt_to_page(addr);
2851 if (likely(!PageCompound(page)))
2852 nr_pages <<= cache->gfporder;
2855 page->slab_cache = cache;
2856 page->slab_page = slab;
2858 } while (--nr_pages);
2862 * Grow (by 1) the number of slabs within a cache. This is called by
2863 * kmem_cache_alloc() when there are no active objs left in a cache.
2865 static int cache_grow(struct kmem_cache *cachep,
2866 gfp_t flags, int nodeid, void *objp)
2871 struct kmem_cache_node *l3;
2874 * Be lazy and only check for valid flags here, keeping it out of the
2875 * critical path in kmem_cache_alloc().
2877 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2878 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2880 /* Take the l3 list lock to change the colour_next on this node */
2882 l3 = cachep->nodelists[nodeid];
2883 spin_lock(&l3->list_lock);
2885 /* Get colour for the slab, and cal the next value. */
2886 offset = l3->colour_next;
2888 if (l3->colour_next >= cachep->colour)
2889 l3->colour_next = 0;
2890 spin_unlock(&l3->list_lock);
2892 offset *= cachep->colour_off;
2894 if (local_flags & __GFP_WAIT)
2898 * The test for missing atomic flag is performed here, rather than
2899 * the more obvious place, simply to reduce the critical path length
2900 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2901 * will eventually be caught here (where it matters).
2903 kmem_flagcheck(cachep, flags);
2906 * Get mem for the objs. Attempt to allocate a physical page from
2910 objp = kmem_getpages(cachep, local_flags, nodeid);
2914 /* Get slab management. */
2915 slabp = alloc_slabmgmt(cachep, objp, offset,
2916 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2920 slab_map_pages(cachep, slabp, objp);
2922 cache_init_objs(cachep, slabp);
2924 if (local_flags & __GFP_WAIT)
2925 local_irq_disable();
2927 spin_lock(&l3->list_lock);
2929 /* Make slab active. */
2930 list_add_tail(&slabp->list, &(l3->slabs_free));
2931 STATS_INC_GROWN(cachep);
2932 l3->free_objects += cachep->num;
2933 spin_unlock(&l3->list_lock);
2936 kmem_freepages(cachep, objp);
2938 if (local_flags & __GFP_WAIT)
2939 local_irq_disable();
2946 * Perform extra freeing checks:
2947 * - detect bad pointers.
2948 * - POISON/RED_ZONE checking
2950 static void kfree_debugcheck(const void *objp)
2952 if (!virt_addr_valid(objp)) {
2953 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2954 (unsigned long)objp);
2959 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2961 unsigned long long redzone1, redzone2;
2963 redzone1 = *dbg_redzone1(cache, obj);
2964 redzone2 = *dbg_redzone2(cache, obj);
2969 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2972 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2973 slab_error(cache, "double free detected");
2975 slab_error(cache, "memory outside object was overwritten");
2977 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2978 obj, redzone1, redzone2);
2981 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2982 unsigned long caller)
2988 BUG_ON(virt_to_cache(objp) != cachep);
2990 objp -= obj_offset(cachep);
2991 kfree_debugcheck(objp);
2992 page = virt_to_head_page(objp);
2994 slabp = page->slab_page;
2996 if (cachep->flags & SLAB_RED_ZONE) {
2997 verify_redzone_free(cachep, objp);
2998 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2999 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
3001 if (cachep->flags & SLAB_STORE_USER)
3002 *dbg_userword(cachep, objp) = (void *)caller;
3004 objnr = obj_to_index(cachep, slabp, objp);
3006 BUG_ON(objnr >= cachep->num);
3007 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
3009 #ifdef CONFIG_DEBUG_SLAB_LEAK
3010 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
3012 if (cachep->flags & SLAB_POISON) {
3013 #ifdef CONFIG_DEBUG_PAGEALLOC
3014 if ((cachep->size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
3015 store_stackinfo(cachep, objp, caller);
3016 kernel_map_pages(virt_to_page(objp),
3017 cachep->size / PAGE_SIZE, 0);
3019 poison_obj(cachep, objp, POISON_FREE);
3022 poison_obj(cachep, objp, POISON_FREE);
3028 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
3033 /* Check slab's freelist to see if this obj is there. */
3034 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
3036 if (entries > cachep->num || i >= cachep->num)
3039 if (entries != cachep->num - slabp->inuse) {
3041 printk(KERN_ERR "slab: Internal list corruption detected in "
3042 "cache '%s'(%d), slabp %p(%d). Tainted(%s). Hexdump:\n",
3043 cachep->name, cachep->num, slabp, slabp->inuse,
3045 print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 16, 1, slabp,
3046 sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t),
3052 #define kfree_debugcheck(x) do { } while(0)
3053 #define cache_free_debugcheck(x,objp,z) (objp)
3054 #define check_slabp(x,y) do { } while(0)
3057 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags,
3061 struct kmem_cache_node *l3;
3062 struct array_cache *ac;
3066 node = numa_mem_id();
3067 if (unlikely(force_refill))
3070 ac = cpu_cache_get(cachep);
3071 batchcount = ac->batchcount;
3072 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
3074 * If there was little recent activity on this cache, then
3075 * perform only a partial refill. Otherwise we could generate
3078 batchcount = BATCHREFILL_LIMIT;
3080 l3 = cachep->nodelists[node];
3082 BUG_ON(ac->avail > 0 || !l3);
3083 spin_lock(&l3->list_lock);
3085 /* See if we can refill from the shared array */
3086 if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) {
3087 l3->shared->touched = 1;
3091 while (batchcount > 0) {
3092 struct list_head *entry;
3094 /* Get slab alloc is to come from. */
3095 entry = l3->slabs_partial.next;
3096 if (entry == &l3->slabs_partial) {
3097 l3->free_touched = 1;
3098 entry = l3->slabs_free.next;
3099 if (entry == &l3->slabs_free)
3103 slabp = list_entry(entry, struct slab, list);
3104 check_slabp(cachep, slabp);
3105 check_spinlock_acquired(cachep);
3108 * The slab was either on partial or free list so
3109 * there must be at least one object available for
3112 BUG_ON(slabp->inuse >= cachep->num);
3114 while (slabp->inuse < cachep->num && batchcount--) {
3115 STATS_INC_ALLOCED(cachep);
3116 STATS_INC_ACTIVE(cachep);
3117 STATS_SET_HIGH(cachep);
3119 ac_put_obj(cachep, ac, slab_get_obj(cachep, slabp,
3122 check_slabp(cachep, slabp);
3124 /* move slabp to correct slabp list: */
3125 list_del(&slabp->list);
3126 if (slabp->free == BUFCTL_END)
3127 list_add(&slabp->list, &l3->slabs_full);
3129 list_add(&slabp->list, &l3->slabs_partial);
3133 l3->free_objects -= ac->avail;
3135 spin_unlock(&l3->list_lock);
3137 if (unlikely(!ac->avail)) {
3140 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
3142 /* cache_grow can reenable interrupts, then ac could change. */
3143 ac = cpu_cache_get(cachep);
3144 node = numa_mem_id();
3146 /* no objects in sight? abort */
3147 if (!x && (ac->avail == 0 || force_refill))
3150 if (!ac->avail) /* objects refilled by interrupt? */
3155 return ac_get_obj(cachep, ac, flags, force_refill);
3158 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3161 might_sleep_if(flags & __GFP_WAIT);
3163 kmem_flagcheck(cachep, flags);
3168 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3169 gfp_t flags, void *objp, unsigned long caller)
3173 if (cachep->flags & SLAB_POISON) {
3174 #ifdef CONFIG_DEBUG_PAGEALLOC
3175 if ((cachep->size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
3176 kernel_map_pages(virt_to_page(objp),
3177 cachep->size / PAGE_SIZE, 1);
3179 check_poison_obj(cachep, objp);
3181 check_poison_obj(cachep, objp);
3183 poison_obj(cachep, objp, POISON_INUSE);
3185 if (cachep->flags & SLAB_STORE_USER)
3186 *dbg_userword(cachep, objp) = (void *)caller;
3188 if (cachep->flags & SLAB_RED_ZONE) {
3189 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3190 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3191 slab_error(cachep, "double free, or memory outside"
3192 " object was overwritten");
3194 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3195 objp, *dbg_redzone1(cachep, objp),
3196 *dbg_redzone2(cachep, objp));
3198 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3199 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3201 #ifdef CONFIG_DEBUG_SLAB_LEAK
3206 slabp = virt_to_head_page(objp)->slab_page;
3207 objnr = (unsigned)(objp - slabp->s_mem) / cachep->size;
3208 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3211 objp += obj_offset(cachep);
3212 if (cachep->ctor && cachep->flags & SLAB_POISON)
3214 if (ARCH_SLAB_MINALIGN &&
3215 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
3216 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3217 objp, (int)ARCH_SLAB_MINALIGN);
3222 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3225 static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
3227 if (cachep == kmem_cache)
3230 return should_failslab(cachep->object_size, flags, cachep->flags);
3233 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3236 struct array_cache *ac;
3237 bool force_refill = false;
3241 ac = cpu_cache_get(cachep);
3242 if (likely(ac->avail)) {
3244 objp = ac_get_obj(cachep, ac, flags, false);
3247 * Allow for the possibility all avail objects are not allowed
3248 * by the current flags
3251 STATS_INC_ALLOCHIT(cachep);
3254 force_refill = true;
3257 STATS_INC_ALLOCMISS(cachep);
3258 objp = cache_alloc_refill(cachep, flags, force_refill);
3260 * the 'ac' may be updated by cache_alloc_refill(),
3261 * and kmemleak_erase() requires its correct value.
3263 ac = cpu_cache_get(cachep);
3267 * To avoid a false negative, if an object that is in one of the
3268 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3269 * treat the array pointers as a reference to the object.
3272 kmemleak_erase(&ac->entry[ac->avail]);
3278 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3280 * If we are in_interrupt, then process context, including cpusets and
3281 * mempolicy, may not apply and should not be used for allocation policy.
3283 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3285 int nid_alloc, nid_here;
3287 if (in_interrupt() || (flags & __GFP_THISNODE))
3289 nid_alloc = nid_here = numa_mem_id();
3290 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3291 nid_alloc = cpuset_slab_spread_node();
3292 else if (current->mempolicy)
3293 nid_alloc = slab_node();
3294 if (nid_alloc != nid_here)
3295 return ____cache_alloc_node(cachep, flags, nid_alloc);
3300 * Fallback function if there was no memory available and no objects on a
3301 * certain node and fall back is permitted. First we scan all the
3302 * available nodelists for available objects. If that fails then we
3303 * perform an allocation without specifying a node. This allows the page
3304 * allocator to do its reclaim / fallback magic. We then insert the
3305 * slab into the proper nodelist and then allocate from it.
3307 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3309 struct zonelist *zonelist;
3313 enum zone_type high_zoneidx = gfp_zone(flags);
3316 unsigned int cpuset_mems_cookie;
3318 if (flags & __GFP_THISNODE)
3321 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3324 cpuset_mems_cookie = get_mems_allowed();
3325 zonelist = node_zonelist(slab_node(), flags);
3329 * Look through allowed nodes for objects available
3330 * from existing per node queues.
3332 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3333 nid = zone_to_nid(zone);
3335 if (cpuset_zone_allowed_hardwall(zone, flags) &&
3336 cache->nodelists[nid] &&
3337 cache->nodelists[nid]->free_objects) {
3338 obj = ____cache_alloc_node(cache,
3339 flags | GFP_THISNODE, nid);
3347 * This allocation will be performed within the constraints
3348 * of the current cpuset / memory policy requirements.
3349 * We may trigger various forms of reclaim on the allowed
3350 * set and go into memory reserves if necessary.
3352 if (local_flags & __GFP_WAIT)
3354 kmem_flagcheck(cache, flags);
3355 obj = kmem_getpages(cache, local_flags, numa_mem_id());
3356 if (local_flags & __GFP_WAIT)
3357 local_irq_disable();
3360 * Insert into the appropriate per node queues
3362 nid = page_to_nid(virt_to_page(obj));
3363 if (cache_grow(cache, flags, nid, obj)) {
3364 obj = ____cache_alloc_node(cache,
3365 flags | GFP_THISNODE, nid);
3368 * Another processor may allocate the
3369 * objects in the slab since we are
3370 * not holding any locks.
3374 /* cache_grow already freed obj */
3380 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !obj))
3386 * A interface to enable slab creation on nodeid
3388 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3391 struct list_head *entry;
3393 struct kmem_cache_node *l3;
3397 l3 = cachep->nodelists[nodeid];
3402 spin_lock(&l3->list_lock);
3403 entry = l3->slabs_partial.next;
3404 if (entry == &l3->slabs_partial) {
3405 l3->free_touched = 1;
3406 entry = l3->slabs_free.next;
3407 if (entry == &l3->slabs_free)
3411 slabp = list_entry(entry, struct slab, list);
3412 check_spinlock_acquired_node(cachep, nodeid);
3413 check_slabp(cachep, slabp);
3415 STATS_INC_NODEALLOCS(cachep);
3416 STATS_INC_ACTIVE(cachep);
3417 STATS_SET_HIGH(cachep);
3419 BUG_ON(slabp->inuse == cachep->num);
3421 obj = slab_get_obj(cachep, slabp, nodeid);
3422 check_slabp(cachep, slabp);
3424 /* move slabp to correct slabp list: */
3425 list_del(&slabp->list);
3427 if (slabp->free == BUFCTL_END)
3428 list_add(&slabp->list, &l3->slabs_full);
3430 list_add(&slabp->list, &l3->slabs_partial);
3432 spin_unlock(&l3->list_lock);
3436 spin_unlock(&l3->list_lock);
3437 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
3441 return fallback_alloc(cachep, flags);
3448 * kmem_cache_alloc_node - Allocate an object on the specified node
3449 * @cachep: The cache to allocate from.
3450 * @flags: See kmalloc().
3451 * @nodeid: node number of the target node.
3452 * @caller: return address of caller, used for debug information
3454 * Identical to kmem_cache_alloc but it will allocate memory on the given
3455 * node, which can improve the performance for cpu bound structures.
3457 * Fallback to other node is possible if __GFP_THISNODE is not set.
3459 static __always_inline void *
3460 slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3461 unsigned long caller)
3463 unsigned long save_flags;
3465 int slab_node = numa_mem_id();
3467 flags &= gfp_allowed_mask;
3469 lockdep_trace_alloc(flags);
3471 if (slab_should_failslab(cachep, flags))
3474 cachep = memcg_kmem_get_cache(cachep, flags);
3476 cache_alloc_debugcheck_before(cachep, flags);
3477 local_irq_save(save_flags);
3479 if (nodeid == NUMA_NO_NODE)
3482 if (unlikely(!cachep->nodelists[nodeid])) {
3483 /* Node not bootstrapped yet */
3484 ptr = fallback_alloc(cachep, flags);
3488 if (nodeid == slab_node) {
3490 * Use the locally cached objects if possible.
3491 * However ____cache_alloc does not allow fallback
3492 * to other nodes. It may fail while we still have
3493 * objects on other nodes available.
3495 ptr = ____cache_alloc(cachep, flags);
3499 /* ___cache_alloc_node can fall back to other nodes */
3500 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3502 local_irq_restore(save_flags);
3503 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3504 kmemleak_alloc_recursive(ptr, cachep->object_size, 1, cachep->flags,
3508 kmemcheck_slab_alloc(cachep, flags, ptr, cachep->object_size);
3510 if (unlikely((flags & __GFP_ZERO) && ptr))
3511 memset(ptr, 0, cachep->object_size);
3516 static __always_inline void *
3517 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3521 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3522 objp = alternate_node_alloc(cache, flags);
3526 objp = ____cache_alloc(cache, flags);
3529 * We may just have run out of memory on the local node.
3530 * ____cache_alloc_node() knows how to locate memory on other nodes
3533 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3540 static __always_inline void *
3541 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3543 return ____cache_alloc(cachep, flags);
3546 #endif /* CONFIG_NUMA */
3548 static __always_inline void *
3549 slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
3551 unsigned long save_flags;
3554 flags &= gfp_allowed_mask;
3556 lockdep_trace_alloc(flags);
3558 if (slab_should_failslab(cachep, flags))
3561 cachep = memcg_kmem_get_cache(cachep, flags);
3563 cache_alloc_debugcheck_before(cachep, flags);
3564 local_irq_save(save_flags);
3565 objp = __do_cache_alloc(cachep, flags);
3566 local_irq_restore(save_flags);
3567 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3568 kmemleak_alloc_recursive(objp, cachep->object_size, 1, cachep->flags,
3573 kmemcheck_slab_alloc(cachep, flags, objp, cachep->object_size);
3575 if (unlikely((flags & __GFP_ZERO) && objp))
3576 memset(objp, 0, cachep->object_size);
3582 * Caller needs to acquire correct kmem_list's list_lock
3584 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3588 struct kmem_cache_node *l3;
3590 for (i = 0; i < nr_objects; i++) {
3594 clear_obj_pfmemalloc(&objpp[i]);
3597 slabp = virt_to_slab(objp);
3598 l3 = cachep->nodelists[node];
3599 list_del(&slabp->list);
3600 check_spinlock_acquired_node(cachep, node);
3601 check_slabp(cachep, slabp);
3602 slab_put_obj(cachep, slabp, objp, node);
3603 STATS_DEC_ACTIVE(cachep);
3605 check_slabp(cachep, slabp);
3607 /* fixup slab chains */
3608 if (slabp->inuse == 0) {
3609 if (l3->free_objects > l3->free_limit) {
3610 l3->free_objects -= cachep->num;
3611 /* No need to drop any previously held
3612 * lock here, even if we have a off-slab slab
3613 * descriptor it is guaranteed to come from
3614 * a different cache, refer to comments before
3617 slab_destroy(cachep, slabp);
3619 list_add(&slabp->list, &l3->slabs_free);
3622 /* Unconditionally move a slab to the end of the
3623 * partial list on free - maximum time for the
3624 * other objects to be freed, too.
3626 list_add_tail(&slabp->list, &l3->slabs_partial);
3631 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3634 struct kmem_cache_node *l3;
3635 int node = numa_mem_id();
3637 batchcount = ac->batchcount;
3639 BUG_ON(!batchcount || batchcount > ac->avail);
3642 l3 = cachep->nodelists[node];
3643 spin_lock(&l3->list_lock);
3645 struct array_cache *shared_array = l3->shared;
3646 int max = shared_array->limit - shared_array->avail;
3648 if (batchcount > max)
3650 memcpy(&(shared_array->entry[shared_array->avail]),
3651 ac->entry, sizeof(void *) * batchcount);
3652 shared_array->avail += batchcount;
3657 free_block(cachep, ac->entry, batchcount, node);
3662 struct list_head *p;
3664 p = l3->slabs_free.next;
3665 while (p != &(l3->slabs_free)) {
3668 slabp = list_entry(p, struct slab, list);
3669 BUG_ON(slabp->inuse);
3674 STATS_SET_FREEABLE(cachep, i);
3677 spin_unlock(&l3->list_lock);
3678 ac->avail -= batchcount;
3679 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3683 * Release an obj back to its cache. If the obj has a constructed state, it must
3684 * be in this state _before_ it is released. Called with disabled ints.
3686 static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3687 unsigned long caller)
3689 struct array_cache *ac = cpu_cache_get(cachep);
3692 kmemleak_free_recursive(objp, cachep->flags);
3693 objp = cache_free_debugcheck(cachep, objp, caller);
3695 kmemcheck_slab_free(cachep, objp, cachep->object_size);
3698 * Skip calling cache_free_alien() when the platform is not numa.
3699 * This will avoid cache misses that happen while accessing slabp (which
3700 * is per page memory reference) to get nodeid. Instead use a global
3701 * variable to skip the call, which is mostly likely to be present in
3704 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3707 if (likely(ac->avail < ac->limit)) {
3708 STATS_INC_FREEHIT(cachep);
3710 STATS_INC_FREEMISS(cachep);
3711 cache_flusharray(cachep, ac);
3714 ac_put_obj(cachep, ac, objp);
3718 * kmem_cache_alloc - Allocate an object
3719 * @cachep: The cache to allocate from.
3720 * @flags: See kmalloc().
3722 * Allocate an object from this cache. The flags are only relevant
3723 * if the cache has no available objects.
3725 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3727 void *ret = slab_alloc(cachep, flags, _RET_IP_);
3729 trace_kmem_cache_alloc(_RET_IP_, ret,
3730 cachep->object_size, cachep->size, flags);
3734 EXPORT_SYMBOL(kmem_cache_alloc);
3736 #ifdef CONFIG_TRACING
3738 kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
3742 ret = slab_alloc(cachep, flags, _RET_IP_);
3744 trace_kmalloc(_RET_IP_, ret,
3745 size, cachep->size, flags);
3748 EXPORT_SYMBOL(kmem_cache_alloc_trace);
3752 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3754 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3756 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3757 cachep->object_size, cachep->size,
3762 EXPORT_SYMBOL(kmem_cache_alloc_node);
3764 #ifdef CONFIG_TRACING
3765 void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
3772 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3774 trace_kmalloc_node(_RET_IP_, ret,
3779 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3782 static __always_inline void *
3783 __do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
3785 struct kmem_cache *cachep;
3787 cachep = kmem_find_general_cachep(size, flags);
3788 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3790 return kmem_cache_alloc_node_trace(cachep, flags, node, size);
3793 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3794 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3796 return __do_kmalloc_node(size, flags, node, _RET_IP_);
3798 EXPORT_SYMBOL(__kmalloc_node);
3800 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3801 int node, unsigned long caller)
3803 return __do_kmalloc_node(size, flags, node, caller);
3805 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3807 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3809 return __do_kmalloc_node(size, flags, node, 0);
3811 EXPORT_SYMBOL(__kmalloc_node);
3812 #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3813 #endif /* CONFIG_NUMA */
3816 * __do_kmalloc - allocate memory
3817 * @size: how many bytes of memory are required.
3818 * @flags: the type of memory to allocate (see kmalloc).
3819 * @caller: function caller for debug tracking of the caller
3821 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3822 unsigned long caller)
3824 struct kmem_cache *cachep;
3827 /* If you want to save a few bytes .text space: replace
3829 * Then kmalloc uses the uninlined functions instead of the inline
3832 cachep = __find_general_cachep(size, flags);
3833 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3835 ret = slab_alloc(cachep, flags, caller);
3837 trace_kmalloc(caller, ret,
3838 size, cachep->size, flags);
3844 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3845 void *__kmalloc(size_t size, gfp_t flags)
3847 return __do_kmalloc(size, flags, _RET_IP_);
3849 EXPORT_SYMBOL(__kmalloc);
3851 void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3853 return __do_kmalloc(size, flags, caller);
3855 EXPORT_SYMBOL(__kmalloc_track_caller);
3858 void *__kmalloc(size_t size, gfp_t flags)
3860 return __do_kmalloc(size, flags, 0);
3862 EXPORT_SYMBOL(__kmalloc);
3866 * kmem_cache_free - Deallocate an object
3867 * @cachep: The cache the allocation was from.
3868 * @objp: The previously allocated object.
3870 * Free an object which was previously allocated from this
3873 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3875 unsigned long flags;
3876 cachep = cache_from_obj(cachep, objp);
3880 local_irq_save(flags);
3881 debug_check_no_locks_freed(objp, cachep->object_size);
3882 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3883 debug_check_no_obj_freed(objp, cachep->object_size);
3884 __cache_free(cachep, objp, _RET_IP_);
3885 local_irq_restore(flags);
3887 trace_kmem_cache_free(_RET_IP_, objp);
3889 EXPORT_SYMBOL(kmem_cache_free);
3892 * kfree - free previously allocated memory
3893 * @objp: pointer returned by kmalloc.
3895 * If @objp is NULL, no operation is performed.
3897 * Don't free memory not originally allocated by kmalloc()
3898 * or you will run into trouble.
3900 void kfree(const void *objp)
3902 struct kmem_cache *c;
3903 unsigned long flags;
3905 trace_kfree(_RET_IP_, objp);
3907 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3909 local_irq_save(flags);
3910 kfree_debugcheck(objp);
3911 c = virt_to_cache(objp);
3912 debug_check_no_locks_freed(objp, c->object_size);
3914 debug_check_no_obj_freed(objp, c->object_size);
3915 __cache_free(c, (void *)objp, _RET_IP_);
3916 local_irq_restore(flags);
3918 EXPORT_SYMBOL(kfree);
3921 * This initializes kmem_list3 or resizes various caches for all nodes.
3923 static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
3926 struct kmem_cache_node *l3;
3927 struct array_cache *new_shared;
3928 struct array_cache **new_alien = NULL;
3930 for_each_online_node(node) {
3932 if (use_alien_caches) {
3933 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
3939 if (cachep->shared) {
3940 new_shared = alloc_arraycache(node,
3941 cachep->shared*cachep->batchcount,
3944 free_alien_cache(new_alien);
3949 l3 = cachep->nodelists[node];
3951 struct array_cache *shared = l3->shared;
3953 spin_lock_irq(&l3->list_lock);
3956 free_block(cachep, shared->entry,
3957 shared->avail, node);
3959 l3->shared = new_shared;
3961 l3->alien = new_alien;
3964 l3->free_limit = (1 + nr_cpus_node(node)) *
3965 cachep->batchcount + cachep->num;
3966 spin_unlock_irq(&l3->list_lock);
3968 free_alien_cache(new_alien);
3971 l3 = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
3973 free_alien_cache(new_alien);
3978 kmem_list3_init(l3);
3979 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3980 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3981 l3->shared = new_shared;
3982 l3->alien = new_alien;
3983 l3->free_limit = (1 + nr_cpus_node(node)) *
3984 cachep->batchcount + cachep->num;
3985 cachep->nodelists[node] = l3;
3990 if (!cachep->list.next) {
3991 /* Cache is not active yet. Roll back what we did */
3994 if (cachep->nodelists[node]) {
3995 l3 = cachep->nodelists[node];
3998 free_alien_cache(l3->alien);
4000 cachep->nodelists[node] = NULL;
4008 struct ccupdate_struct {
4009 struct kmem_cache *cachep;
4010 struct array_cache *new[0];
4013 static void do_ccupdate_local(void *info)
4015 struct ccupdate_struct *new = info;
4016 struct array_cache *old;
4019 old = cpu_cache_get(new->cachep);
4021 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
4022 new->new[smp_processor_id()] = old;
4025 /* Always called with the slab_mutex held */
4026 static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
4027 int batchcount, int shared, gfp_t gfp)
4029 struct ccupdate_struct *new;
4032 new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *),
4037 for_each_online_cpu(i) {
4038 new->new[i] = alloc_arraycache(cpu_to_mem(i), limit,
4041 for (i--; i >= 0; i--)
4047 new->cachep = cachep;
4049 on_each_cpu(do_ccupdate_local, (void *)new, 1);
4052 cachep->batchcount = batchcount;
4053 cachep->limit = limit;
4054 cachep->shared = shared;
4056 for_each_online_cpu(i) {
4057 struct array_cache *ccold = new->new[i];
4060 spin_lock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
4061 free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i));
4062 spin_unlock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
4066 return alloc_kmemlist(cachep, gfp);
4069 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
4070 int batchcount, int shared, gfp_t gfp)
4073 struct kmem_cache *c = NULL;
4076 ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
4078 if (slab_state < FULL)
4081 if ((ret < 0) || !is_root_cache(cachep))
4084 VM_BUG_ON(!mutex_is_locked(&slab_mutex));
4085 for_each_memcg_cache_index(i) {
4086 c = cache_from_memcg(cachep, i);
4088 /* return value determined by the parent cache only */
4089 __do_tune_cpucache(c, limit, batchcount, shared, gfp);
4095 /* Called with slab_mutex held always */
4096 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
4103 if (!is_root_cache(cachep)) {
4104 struct kmem_cache *root = memcg_root_cache(cachep);
4105 limit = root->limit;
4106 shared = root->shared;
4107 batchcount = root->batchcount;
4110 if (limit && shared && batchcount)
4113 * The head array serves three purposes:
4114 * - create a LIFO ordering, i.e. return objects that are cache-warm
4115 * - reduce the number of spinlock operations.
4116 * - reduce the number of linked list operations on the slab and
4117 * bufctl chains: array operations are cheaper.
4118 * The numbers are guessed, we should auto-tune as described by
4121 if (cachep->size > 131072)
4123 else if (cachep->size > PAGE_SIZE)
4125 else if (cachep->size > 1024)
4127 else if (cachep->size > 256)
4133 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
4134 * allocation behaviour: Most allocs on one cpu, most free operations
4135 * on another cpu. For these cases, an efficient object passing between
4136 * cpus is necessary. This is provided by a shared array. The array
4137 * replaces Bonwick's magazine layer.
4138 * On uniprocessor, it's functionally equivalent (but less efficient)
4139 * to a larger limit. Thus disabled by default.
4142 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
4147 * With debugging enabled, large batchcount lead to excessively long
4148 * periods with disabled local interrupts. Limit the batchcount
4153 batchcount = (limit + 1) / 2;
4155 err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
4157 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
4158 cachep->name, -err);
4163 * Drain an array if it contains any elements taking the l3 lock only if
4164 * necessary. Note that the l3 listlock also protects the array_cache
4165 * if drain_array() is used on the shared array.
4167 static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *l3,
4168 struct array_cache *ac, int force, int node)
4172 if (!ac || !ac->avail)
4174 if (ac->touched && !force) {
4177 spin_lock_irq(&l3->list_lock);
4179 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4180 if (tofree > ac->avail)
4181 tofree = (ac->avail + 1) / 2;
4182 free_block(cachep, ac->entry, tofree, node);
4183 ac->avail -= tofree;
4184 memmove(ac->entry, &(ac->entry[tofree]),
4185 sizeof(void *) * ac->avail);
4187 spin_unlock_irq(&l3->list_lock);
4192 * cache_reap - Reclaim memory from caches.
4193 * @w: work descriptor
4195 * Called from workqueue/eventd every few seconds.
4197 * - clear the per-cpu caches for this CPU.
4198 * - return freeable pages to the main free memory pool.
4200 * If we cannot acquire the cache chain mutex then just give up - we'll try
4201 * again on the next iteration.
4203 static void cache_reap(struct work_struct *w)
4205 struct kmem_cache *searchp;
4206 struct kmem_cache_node *l3;
4207 int node = numa_mem_id();
4208 struct delayed_work *work = to_delayed_work(w);
4210 if (!mutex_trylock(&slab_mutex))
4211 /* Give up. Setup the next iteration. */
4214 list_for_each_entry(searchp, &slab_caches, list) {
4218 * We only take the l3 lock if absolutely necessary and we
4219 * have established with reasonable certainty that
4220 * we can do some work if the lock was obtained.
4222 l3 = searchp->nodelists[node];
4224 reap_alien(searchp, l3);
4226 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
4229 * These are racy checks but it does not matter
4230 * if we skip one check or scan twice.
4232 if (time_after(l3->next_reap, jiffies))
4235 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
4237 drain_array(searchp, l3, l3->shared, 0, node);
4239 if (l3->free_touched)
4240 l3->free_touched = 0;
4244 freed = drain_freelist(searchp, l3, (l3->free_limit +
4245 5 * searchp->num - 1) / (5 * searchp->num));
4246 STATS_ADD_REAPED(searchp, freed);
4252 mutex_unlock(&slab_mutex);
4255 /* Set up the next iteration */
4256 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
4259 #ifdef CONFIG_SLABINFO
4260 void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
4263 unsigned long active_objs;
4264 unsigned long num_objs;
4265 unsigned long active_slabs = 0;
4266 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4270 struct kmem_cache_node *l3;
4274 for_each_online_node(node) {
4275 l3 = cachep->nodelists[node];
4280 spin_lock_irq(&l3->list_lock);
4282 list_for_each_entry(slabp, &l3->slabs_full, list) {
4283 if (slabp->inuse != cachep->num && !error)
4284 error = "slabs_full accounting error";
4285 active_objs += cachep->num;
4288 list_for_each_entry(slabp, &l3->slabs_partial, list) {
4289 if (slabp->inuse == cachep->num && !error)
4290 error = "slabs_partial inuse accounting error";
4291 if (!slabp->inuse && !error)
4292 error = "slabs_partial/inuse accounting error";
4293 active_objs += slabp->inuse;
4296 list_for_each_entry(slabp, &l3->slabs_free, list) {
4297 if (slabp->inuse && !error)
4298 error = "slabs_free/inuse accounting error";
4301 free_objects += l3->free_objects;
4303 shared_avail += l3->shared->avail;
4305 spin_unlock_irq(&l3->list_lock);
4307 num_slabs += active_slabs;
4308 num_objs = num_slabs * cachep->num;
4309 if (num_objs - active_objs != free_objects && !error)
4310 error = "free_objects accounting error";
4312 name = cachep->name;
4314 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4316 sinfo->active_objs = active_objs;
4317 sinfo->num_objs = num_objs;
4318 sinfo->active_slabs = active_slabs;
4319 sinfo->num_slabs = num_slabs;
4320 sinfo->shared_avail = shared_avail;
4321 sinfo->limit = cachep->limit;
4322 sinfo->batchcount = cachep->batchcount;
4323 sinfo->shared = cachep->shared;
4324 sinfo->objects_per_slab = cachep->num;
4325 sinfo->cache_order = cachep->gfporder;
4328 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
4332 unsigned long high = cachep->high_mark;
4333 unsigned long allocs = cachep->num_allocations;
4334 unsigned long grown = cachep->grown;
4335 unsigned long reaped = cachep->reaped;
4336 unsigned long errors = cachep->errors;
4337 unsigned long max_freeable = cachep->max_freeable;
4338 unsigned long node_allocs = cachep->node_allocs;
4339 unsigned long node_frees = cachep->node_frees;
4340 unsigned long overflows = cachep->node_overflow;
4342 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4343 "%4lu %4lu %4lu %4lu %4lu",
4344 allocs, high, grown,
4345 reaped, errors, max_freeable, node_allocs,
4346 node_frees, overflows);
4350 unsigned long allochit = atomic_read(&cachep->allochit);
4351 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4352 unsigned long freehit = atomic_read(&cachep->freehit);
4353 unsigned long freemiss = atomic_read(&cachep->freemiss);
4355 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4356 allochit, allocmiss, freehit, freemiss);
4361 #define MAX_SLABINFO_WRITE 128
4363 * slabinfo_write - Tuning for the slab allocator
4365 * @buffer: user buffer
4366 * @count: data length
4369 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4370 size_t count, loff_t *ppos)
4372 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4373 int limit, batchcount, shared, res;
4374 struct kmem_cache *cachep;
4376 if (count > MAX_SLABINFO_WRITE)
4378 if (copy_from_user(&kbuf, buffer, count))
4380 kbuf[MAX_SLABINFO_WRITE] = '\0';
4382 tmp = strchr(kbuf, ' ');
4387 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4390 /* Find the cache in the chain of caches. */
4391 mutex_lock(&slab_mutex);
4393 list_for_each_entry(cachep, &slab_caches, list) {
4394 if (!strcmp(cachep->name, kbuf)) {
4395 if (limit < 1 || batchcount < 1 ||
4396 batchcount > limit || shared < 0) {
4399 res = do_tune_cpucache(cachep, limit,
4406 mutex_unlock(&slab_mutex);
4412 #ifdef CONFIG_DEBUG_SLAB_LEAK
4414 static void *leaks_start(struct seq_file *m, loff_t *pos)
4416 mutex_lock(&slab_mutex);
4417 return seq_list_start(&slab_caches, *pos);
4420 static inline int add_caller(unsigned long *n, unsigned long v)
4430 unsigned long *q = p + 2 * i;
4444 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4450 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4456 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->size) {
4457 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4459 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4464 static void show_symbol(struct seq_file *m, unsigned long address)
4466 #ifdef CONFIG_KALLSYMS
4467 unsigned long offset, size;
4468 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4470 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4471 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4473 seq_printf(m, " [%s]", modname);
4477 seq_printf(m, "%p", (void *)address);
4480 static int leaks_show(struct seq_file *m, void *p)
4482 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
4484 struct kmem_cache_node *l3;
4486 unsigned long *n = m->private;
4490 if (!(cachep->flags & SLAB_STORE_USER))
4492 if (!(cachep->flags & SLAB_RED_ZONE))
4495 /* OK, we can do it */
4499 for_each_online_node(node) {
4500 l3 = cachep->nodelists[node];
4505 spin_lock_irq(&l3->list_lock);
4507 list_for_each_entry(slabp, &l3->slabs_full, list)
4508 handle_slab(n, cachep, slabp);
4509 list_for_each_entry(slabp, &l3->slabs_partial, list)
4510 handle_slab(n, cachep, slabp);
4511 spin_unlock_irq(&l3->list_lock);
4513 name = cachep->name;
4515 /* Increase the buffer size */
4516 mutex_unlock(&slab_mutex);
4517 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4519 /* Too bad, we are really out */
4521 mutex_lock(&slab_mutex);
4524 *(unsigned long *)m->private = n[0] * 2;
4526 mutex_lock(&slab_mutex);
4527 /* Now make sure this entry will be retried */
4531 for (i = 0; i < n[1]; i++) {
4532 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4533 show_symbol(m, n[2*i+2]);
4540 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4542 return seq_list_next(p, &slab_caches, pos);
4545 static void s_stop(struct seq_file *m, void *p)
4547 mutex_unlock(&slab_mutex);
4550 static const struct seq_operations slabstats_op = {
4551 .start = leaks_start,
4557 static int slabstats_open(struct inode *inode, struct file *file)
4559 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4562 ret = seq_open(file, &slabstats_op);
4564 struct seq_file *m = file->private_data;
4565 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4574 static const struct file_operations proc_slabstats_operations = {
4575 .open = slabstats_open,
4577 .llseek = seq_lseek,
4578 .release = seq_release_private,
4582 static int __init slab_proc_init(void)
4584 #ifdef CONFIG_DEBUG_SLAB_LEAK
4585 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4589 module_init(slab_proc_init);
4593 * ksize - get the actual amount of memory allocated for a given object
4594 * @objp: Pointer to the object
4596 * kmalloc may internally round up allocations and return more memory
4597 * than requested. ksize() can be used to determine the actual amount of
4598 * memory allocated. The caller may use this additional memory, even though
4599 * a smaller amount of memory was initially specified with the kmalloc call.
4600 * The caller must guarantee that objp points to a valid object previously
4601 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4602 * must not be freed during the duration of the call.
4604 size_t ksize(const void *objp)
4607 if (unlikely(objp == ZERO_SIZE_PTR))
4610 return virt_to_cache(objp)->object_size;
4612 EXPORT_SYMBOL(ksize);