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.
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_list3 __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_list3 *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_list3))
334 static void kmem_list3_init(struct kmem_list3 *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_list3 *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_list3 *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)
611 struct kmem_list3 *l3;
612 l3 = cachep->nodelists[q];
616 slab_set_lock_classes(cachep, &on_slab_l3_key,
617 &on_slab_alc_key, q);
620 static inline void on_slab_lock_classes(struct kmem_cache *cachep)
624 VM_BUG_ON(OFF_SLAB(cachep));
626 on_slab_lock_classes_node(cachep, node);
629 static inline void init_lock_keys(void)
634 init_node_lock_keys(node);
637 static void init_node_lock_keys(int q)
641 static inline void init_lock_keys(void)
645 static inline void on_slab_lock_classes(struct kmem_cache *cachep)
649 static inline void on_slab_lock_classes_node(struct kmem_cache *cachep, int node)
653 static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
657 static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
662 static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
664 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
666 return cachep->array[smp_processor_id()];
669 static inline struct kmem_cache *__find_general_cachep(size_t size,
675 /* This happens if someone tries to call
676 * kmem_cache_create(), or __kmalloc(), before
677 * the generic caches are initialized.
679 BUG_ON(kmalloc_caches[INDEX_AC] == NULL);
682 return ZERO_SIZE_PTR;
684 i = kmalloc_index(size);
687 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
688 * has cs_{dma,}cachep==NULL. Thus no special case
689 * for large kmalloc calls required.
691 #ifdef CONFIG_ZONE_DMA
692 if (unlikely(gfpflags & GFP_DMA))
693 return kmalloc_dma_caches[i];
695 return kmalloc_caches[i];
698 static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
700 return __find_general_cachep(size, gfpflags);
703 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
705 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
709 * Calculate the number of objects and left-over bytes for a given buffer size.
711 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
712 size_t align, int flags, size_t *left_over,
717 size_t slab_size = PAGE_SIZE << gfporder;
720 * The slab management structure can be either off the slab or
721 * on it. For the latter case, the memory allocated for a
725 * - One kmem_bufctl_t for each object
726 * - Padding to respect alignment of @align
727 * - @buffer_size bytes for each object
729 * If the slab management structure is off the slab, then the
730 * alignment will already be calculated into the size. Because
731 * the slabs are all pages aligned, the objects will be at the
732 * correct alignment when allocated.
734 if (flags & CFLGS_OFF_SLAB) {
736 nr_objs = slab_size / buffer_size;
738 if (nr_objs > SLAB_LIMIT)
739 nr_objs = SLAB_LIMIT;
742 * Ignore padding for the initial guess. The padding
743 * is at most @align-1 bytes, and @buffer_size is at
744 * least @align. In the worst case, this result will
745 * be one greater than the number of objects that fit
746 * into the memory allocation when taking the padding
749 nr_objs = (slab_size - sizeof(struct slab)) /
750 (buffer_size + sizeof(kmem_bufctl_t));
753 * This calculated number will be either the right
754 * amount, or one greater than what we want.
756 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
760 if (nr_objs > SLAB_LIMIT)
761 nr_objs = SLAB_LIMIT;
763 mgmt_size = slab_mgmt_size(nr_objs, align);
766 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
770 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
772 static void __slab_error(const char *function, struct kmem_cache *cachep,
775 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
776 function, cachep->name, msg);
778 add_taint(TAINT_BAD_PAGE);
783 * By default on NUMA we use alien caches to stage the freeing of
784 * objects allocated from other nodes. This causes massive memory
785 * inefficiencies when using fake NUMA setup to split memory into a
786 * large number of small nodes, so it can be disabled on the command
790 static int use_alien_caches __read_mostly = 1;
791 static int __init noaliencache_setup(char *s)
793 use_alien_caches = 0;
796 __setup("noaliencache", noaliencache_setup);
798 static int __init slab_max_order_setup(char *str)
800 get_option(&str, &slab_max_order);
801 slab_max_order = slab_max_order < 0 ? 0 :
802 min(slab_max_order, MAX_ORDER - 1);
803 slab_max_order_set = true;
807 __setup("slab_max_order=", slab_max_order_setup);
811 * Special reaping functions for NUMA systems called from cache_reap().
812 * These take care of doing round robin flushing of alien caches (containing
813 * objects freed on different nodes from which they were allocated) and the
814 * flushing of remote pcps by calling drain_node_pages.
816 static DEFINE_PER_CPU(unsigned long, slab_reap_node);
818 static void init_reap_node(int cpu)
822 node = next_node(cpu_to_mem(cpu), node_online_map);
823 if (node == MAX_NUMNODES)
824 node = first_node(node_online_map);
826 per_cpu(slab_reap_node, cpu) = node;
829 static void next_reap_node(void)
831 int node = __this_cpu_read(slab_reap_node);
833 node = next_node(node, node_online_map);
834 if (unlikely(node >= MAX_NUMNODES))
835 node = first_node(node_online_map);
836 __this_cpu_write(slab_reap_node, node);
840 #define init_reap_node(cpu) do { } while (0)
841 #define next_reap_node(void) do { } while (0)
845 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
846 * via the workqueue/eventd.
847 * Add the CPU number into the expiration time to minimize the possibility of
848 * the CPUs getting into lockstep and contending for the global cache chain
851 static void __cpuinit start_cpu_timer(int cpu)
853 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
856 * When this gets called from do_initcalls via cpucache_init(),
857 * init_workqueues() has already run, so keventd will be setup
860 if (keventd_up() && reap_work->work.func == NULL) {
862 INIT_DEFERRABLE_WORK(reap_work, cache_reap);
863 schedule_delayed_work_on(cpu, reap_work,
864 __round_jiffies_relative(HZ, cpu));
868 static struct array_cache *alloc_arraycache(int node, int entries,
869 int batchcount, gfp_t gfp)
871 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
872 struct array_cache *nc = NULL;
874 nc = kmalloc_node(memsize, gfp, node);
876 * The array_cache structures contain pointers to free object.
877 * However, when such objects are allocated or transferred to another
878 * cache the pointers are not cleared and they could be counted as
879 * valid references during a kmemleak scan. Therefore, kmemleak must
880 * not scan such objects.
882 kmemleak_no_scan(nc);
886 nc->batchcount = batchcount;
888 spin_lock_init(&nc->lock);
893 static inline bool is_slab_pfmemalloc(struct slab *slabp)
895 struct page *page = virt_to_page(slabp->s_mem);
897 return PageSlabPfmemalloc(page);
900 /* Clears pfmemalloc_active if no slabs have pfmalloc set */
901 static void recheck_pfmemalloc_active(struct kmem_cache *cachep,
902 struct array_cache *ac)
904 struct kmem_list3 *l3 = cachep->nodelists[numa_mem_id()];
908 if (!pfmemalloc_active)
911 spin_lock_irqsave(&l3->list_lock, flags);
912 list_for_each_entry(slabp, &l3->slabs_full, list)
913 if (is_slab_pfmemalloc(slabp))
916 list_for_each_entry(slabp, &l3->slabs_partial, list)
917 if (is_slab_pfmemalloc(slabp))
920 list_for_each_entry(slabp, &l3->slabs_free, list)
921 if (is_slab_pfmemalloc(slabp))
924 pfmemalloc_active = false;
926 spin_unlock_irqrestore(&l3->list_lock, flags);
929 static void *__ac_get_obj(struct kmem_cache *cachep, struct array_cache *ac,
930 gfp_t flags, bool force_refill)
933 void *objp = ac->entry[--ac->avail];
935 /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
936 if (unlikely(is_obj_pfmemalloc(objp))) {
937 struct kmem_list3 *l3;
939 if (gfp_pfmemalloc_allowed(flags)) {
940 clear_obj_pfmemalloc(&objp);
944 /* The caller cannot use PFMEMALLOC objects, find another one */
945 for (i = 0; i < ac->avail; i++) {
946 /* If a !PFMEMALLOC object is found, swap them */
947 if (!is_obj_pfmemalloc(ac->entry[i])) {
949 ac->entry[i] = ac->entry[ac->avail];
950 ac->entry[ac->avail] = objp;
956 * If there are empty slabs on the slabs_free list and we are
957 * being forced to refill the cache, mark this one !pfmemalloc.
959 l3 = cachep->nodelists[numa_mem_id()];
960 if (!list_empty(&l3->slabs_free) && force_refill) {
961 struct slab *slabp = virt_to_slab(objp);
962 ClearPageSlabPfmemalloc(virt_to_head_page(slabp->s_mem));
963 clear_obj_pfmemalloc(&objp);
964 recheck_pfmemalloc_active(cachep, ac);
968 /* No !PFMEMALLOC objects available */
976 static inline void *ac_get_obj(struct kmem_cache *cachep,
977 struct array_cache *ac, gfp_t flags, bool force_refill)
981 if (unlikely(sk_memalloc_socks()))
982 objp = __ac_get_obj(cachep, ac, flags, force_refill);
984 objp = ac->entry[--ac->avail];
989 static void *__ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
992 if (unlikely(pfmemalloc_active)) {
993 /* Some pfmemalloc slabs exist, check if this is one */
994 struct page *page = virt_to_head_page(objp);
995 if (PageSlabPfmemalloc(page))
996 set_obj_pfmemalloc(&objp);
1002 static inline void ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
1005 if (unlikely(sk_memalloc_socks()))
1006 objp = __ac_put_obj(cachep, ac, objp);
1008 ac->entry[ac->avail++] = objp;
1012 * Transfer objects in one arraycache to another.
1013 * Locking must be handled by the caller.
1015 * Return the number of entries transferred.
1017 static int transfer_objects(struct array_cache *to,
1018 struct array_cache *from, unsigned int max)
1020 /* Figure out how many entries to transfer */
1021 int nr = min3(from->avail, max, to->limit - to->avail);
1026 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
1027 sizeof(void *) *nr);
1036 #define drain_alien_cache(cachep, alien) do { } while (0)
1037 #define reap_alien(cachep, l3) do { } while (0)
1039 static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
1041 return (struct array_cache **)BAD_ALIEN_MAGIC;
1044 static inline void free_alien_cache(struct array_cache **ac_ptr)
1048 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1053 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
1059 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
1060 gfp_t flags, int nodeid)
1065 #else /* CONFIG_NUMA */
1067 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
1068 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
1070 static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
1072 struct array_cache **ac_ptr;
1073 int memsize = sizeof(void *) * nr_node_ids;
1078 ac_ptr = kzalloc_node(memsize, gfp, node);
1081 if (i == node || !node_online(i))
1083 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
1085 for (i--; i >= 0; i--)
1095 static void free_alien_cache(struct array_cache **ac_ptr)
1106 static void __drain_alien_cache(struct kmem_cache *cachep,
1107 struct array_cache *ac, int node)
1109 struct kmem_list3 *rl3 = cachep->nodelists[node];
1112 spin_lock(&rl3->list_lock);
1114 * Stuff objects into the remote nodes shared array first.
1115 * That way we could avoid the overhead of putting the objects
1116 * into the free lists and getting them back later.
1119 transfer_objects(rl3->shared, ac, ac->limit);
1121 free_block(cachep, ac->entry, ac->avail, node);
1123 spin_unlock(&rl3->list_lock);
1128 * Called from cache_reap() to regularly drain alien caches round robin.
1130 static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3)
1132 int node = __this_cpu_read(slab_reap_node);
1135 struct array_cache *ac = l3->alien[node];
1137 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1138 __drain_alien_cache(cachep, ac, node);
1139 spin_unlock_irq(&ac->lock);
1144 static void drain_alien_cache(struct kmem_cache *cachep,
1145 struct array_cache **alien)
1148 struct array_cache *ac;
1149 unsigned long flags;
1151 for_each_online_node(i) {
1154 spin_lock_irqsave(&ac->lock, flags);
1155 __drain_alien_cache(cachep, ac, i);
1156 spin_unlock_irqrestore(&ac->lock, flags);
1161 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1163 struct slab *slabp = virt_to_slab(objp);
1164 int nodeid = slabp->nodeid;
1165 struct kmem_list3 *l3;
1166 struct array_cache *alien = NULL;
1169 node = numa_mem_id();
1172 * Make sure we are not freeing a object from another node to the array
1173 * cache on this cpu.
1175 if (likely(slabp->nodeid == node))
1178 l3 = cachep->nodelists[node];
1179 STATS_INC_NODEFREES(cachep);
1180 if (l3->alien && l3->alien[nodeid]) {
1181 alien = l3->alien[nodeid];
1182 spin_lock(&alien->lock);
1183 if (unlikely(alien->avail == alien->limit)) {
1184 STATS_INC_ACOVERFLOW(cachep);
1185 __drain_alien_cache(cachep, alien, nodeid);
1187 ac_put_obj(cachep, alien, objp);
1188 spin_unlock(&alien->lock);
1190 spin_lock(&(cachep->nodelists[nodeid])->list_lock);
1191 free_block(cachep, &objp, 1, nodeid);
1192 spin_unlock(&(cachep->nodelists[nodeid])->list_lock);
1199 * Allocates and initializes nodelists for a node on each slab cache, used for
1200 * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3
1201 * will be allocated off-node since memory is not yet online for the new node.
1202 * When hotplugging memory or a cpu, existing nodelists are not replaced if
1205 * Must hold slab_mutex.
1207 static int init_cache_nodelists_node(int node)
1209 struct kmem_cache *cachep;
1210 struct kmem_list3 *l3;
1211 const int memsize = sizeof(struct kmem_list3);
1213 list_for_each_entry(cachep, &slab_caches, list) {
1215 * Set up the size64 kmemlist for cpu before we can
1216 * begin anything. Make sure some other cpu on this
1217 * node has not already allocated this
1219 if (!cachep->nodelists[node]) {
1220 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1223 kmem_list3_init(l3);
1224 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1225 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1228 * The l3s don't come and go as CPUs come and
1229 * go. slab_mutex is sufficient
1232 cachep->nodelists[node] = l3;
1235 spin_lock_irq(&cachep->nodelists[node]->list_lock);
1236 cachep->nodelists[node]->free_limit =
1237 (1 + nr_cpus_node(node)) *
1238 cachep->batchcount + cachep->num;
1239 spin_unlock_irq(&cachep->nodelists[node]->list_lock);
1244 static void __cpuinit cpuup_canceled(long cpu)
1246 struct kmem_cache *cachep;
1247 struct kmem_list3 *l3 = NULL;
1248 int node = cpu_to_mem(cpu);
1249 const struct cpumask *mask = cpumask_of_node(node);
1251 list_for_each_entry(cachep, &slab_caches, list) {
1252 struct array_cache *nc;
1253 struct array_cache *shared;
1254 struct array_cache **alien;
1256 /* cpu is dead; no one can alloc from it. */
1257 nc = cachep->array[cpu];
1258 cachep->array[cpu] = NULL;
1259 l3 = cachep->nodelists[node];
1262 goto free_array_cache;
1264 spin_lock_irq(&l3->list_lock);
1266 /* Free limit for this kmem_list3 */
1267 l3->free_limit -= cachep->batchcount;
1269 free_block(cachep, nc->entry, nc->avail, node);
1271 if (!cpumask_empty(mask)) {
1272 spin_unlock_irq(&l3->list_lock);
1273 goto free_array_cache;
1276 shared = l3->shared;
1278 free_block(cachep, shared->entry,
1279 shared->avail, node);
1286 spin_unlock_irq(&l3->list_lock);
1290 drain_alien_cache(cachep, alien);
1291 free_alien_cache(alien);
1297 * In the previous loop, all the objects were freed to
1298 * the respective cache's slabs, now we can go ahead and
1299 * shrink each nodelist to its limit.
1301 list_for_each_entry(cachep, &slab_caches, list) {
1302 l3 = cachep->nodelists[node];
1305 drain_freelist(cachep, l3, l3->free_objects);
1309 static int __cpuinit cpuup_prepare(long cpu)
1311 struct kmem_cache *cachep;
1312 struct kmem_list3 *l3 = NULL;
1313 int node = cpu_to_mem(cpu);
1317 * We need to do this right in the beginning since
1318 * alloc_arraycache's are going to use this list.
1319 * kmalloc_node allows us to add the slab to the right
1320 * kmem_list3 and not this cpu's kmem_list3
1322 err = init_cache_nodelists_node(node);
1327 * Now we can go ahead with allocating the shared arrays and
1330 list_for_each_entry(cachep, &slab_caches, list) {
1331 struct array_cache *nc;
1332 struct array_cache *shared = NULL;
1333 struct array_cache **alien = NULL;
1335 nc = alloc_arraycache(node, cachep->limit,
1336 cachep->batchcount, GFP_KERNEL);
1339 if (cachep->shared) {
1340 shared = alloc_arraycache(node,
1341 cachep->shared * cachep->batchcount,
1342 0xbaadf00d, GFP_KERNEL);
1348 if (use_alien_caches) {
1349 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
1356 cachep->array[cpu] = nc;
1357 l3 = cachep->nodelists[node];
1360 spin_lock_irq(&l3->list_lock);
1363 * We are serialised from CPU_DEAD or
1364 * CPU_UP_CANCELLED by the cpucontrol lock
1366 l3->shared = shared;
1375 spin_unlock_irq(&l3->list_lock);
1377 free_alien_cache(alien);
1378 if (cachep->flags & SLAB_DEBUG_OBJECTS)
1379 slab_set_debugobj_lock_classes_node(cachep, node);
1380 else if (!OFF_SLAB(cachep) &&
1381 !(cachep->flags & SLAB_DESTROY_BY_RCU))
1382 on_slab_lock_classes_node(cachep, node);
1384 init_node_lock_keys(node);
1388 cpuup_canceled(cpu);
1392 static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1393 unsigned long action, void *hcpu)
1395 long cpu = (long)hcpu;
1399 case CPU_UP_PREPARE:
1400 case CPU_UP_PREPARE_FROZEN:
1401 mutex_lock(&slab_mutex);
1402 err = cpuup_prepare(cpu);
1403 mutex_unlock(&slab_mutex);
1406 case CPU_ONLINE_FROZEN:
1407 start_cpu_timer(cpu);
1409 #ifdef CONFIG_HOTPLUG_CPU
1410 case CPU_DOWN_PREPARE:
1411 case CPU_DOWN_PREPARE_FROZEN:
1413 * Shutdown cache reaper. Note that the slab_mutex is
1414 * held so that if cache_reap() is invoked it cannot do
1415 * anything expensive but will only modify reap_work
1416 * and reschedule the timer.
1418 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1419 /* Now the cache_reaper is guaranteed to be not running. */
1420 per_cpu(slab_reap_work, cpu).work.func = NULL;
1422 case CPU_DOWN_FAILED:
1423 case CPU_DOWN_FAILED_FROZEN:
1424 start_cpu_timer(cpu);
1427 case CPU_DEAD_FROZEN:
1429 * Even if all the cpus of a node are down, we don't free the
1430 * kmem_list3 of any cache. This to avoid a race between
1431 * cpu_down, and a kmalloc allocation from another cpu for
1432 * memory from the node of the cpu going down. The list3
1433 * structure is usually allocated from kmem_cache_create() and
1434 * gets destroyed at kmem_cache_destroy().
1438 case CPU_UP_CANCELED:
1439 case CPU_UP_CANCELED_FROZEN:
1440 mutex_lock(&slab_mutex);
1441 cpuup_canceled(cpu);
1442 mutex_unlock(&slab_mutex);
1445 return notifier_from_errno(err);
1448 static struct notifier_block __cpuinitdata cpucache_notifier = {
1449 &cpuup_callback, NULL, 0
1452 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1454 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1455 * Returns -EBUSY if all objects cannot be drained so that the node is not
1458 * Must hold slab_mutex.
1460 static int __meminit drain_cache_nodelists_node(int node)
1462 struct kmem_cache *cachep;
1465 list_for_each_entry(cachep, &slab_caches, list) {
1466 struct kmem_list3 *l3;
1468 l3 = cachep->nodelists[node];
1472 drain_freelist(cachep, l3, l3->free_objects);
1474 if (!list_empty(&l3->slabs_full) ||
1475 !list_empty(&l3->slabs_partial)) {
1483 static int __meminit slab_memory_callback(struct notifier_block *self,
1484 unsigned long action, void *arg)
1486 struct memory_notify *mnb = arg;
1490 nid = mnb->status_change_nid;
1495 case MEM_GOING_ONLINE:
1496 mutex_lock(&slab_mutex);
1497 ret = init_cache_nodelists_node(nid);
1498 mutex_unlock(&slab_mutex);
1500 case MEM_GOING_OFFLINE:
1501 mutex_lock(&slab_mutex);
1502 ret = drain_cache_nodelists_node(nid);
1503 mutex_unlock(&slab_mutex);
1507 case MEM_CANCEL_ONLINE:
1508 case MEM_CANCEL_OFFLINE:
1512 return notifier_from_errno(ret);
1514 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1517 * swap the static kmem_list3 with kmalloced memory
1519 static void __init init_list(struct kmem_cache *cachep, struct kmem_list3 *list,
1522 struct kmem_list3 *ptr;
1524 ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid);
1527 memcpy(ptr, list, sizeof(struct kmem_list3));
1529 * Do not assume that spinlocks can be initialized via memcpy:
1531 spin_lock_init(&ptr->list_lock);
1533 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1534 cachep->nodelists[nodeid] = ptr;
1538 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1539 * size of kmem_list3.
1541 static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1545 for_each_online_node(node) {
1546 cachep->nodelists[node] = &initkmem_list3[index + node];
1547 cachep->nodelists[node]->next_reap = jiffies +
1549 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1554 * The memory after the last cpu cache pointer is used for the
1555 * the nodelists pointer.
1557 static void setup_nodelists_pointer(struct kmem_cache *cachep)
1559 cachep->nodelists = (struct kmem_list3 **)&cachep->array[nr_cpu_ids];
1563 * Initialisation. Called after the page allocator have been initialised and
1564 * before smp_init().
1566 void __init kmem_cache_init(void)
1570 kmem_cache = &kmem_cache_boot;
1571 setup_nodelists_pointer(kmem_cache);
1573 if (num_possible_nodes() == 1)
1574 use_alien_caches = 0;
1576 for (i = 0; i < NUM_INIT_LISTS; i++)
1577 kmem_list3_init(&initkmem_list3[i]);
1579 set_up_list3s(kmem_cache, CACHE_CACHE);
1582 * Fragmentation resistance on low memory - only use bigger
1583 * page orders on machines with more than 32MB of memory if
1584 * not overridden on the command line.
1586 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
1587 slab_max_order = SLAB_MAX_ORDER_HI;
1589 /* Bootstrap is tricky, because several objects are allocated
1590 * from caches that do not exist yet:
1591 * 1) initialize the kmem_cache cache: it contains the struct
1592 * kmem_cache structures of all caches, except kmem_cache itself:
1593 * kmem_cache is statically allocated.
1594 * Initially an __init data area is used for the head array and the
1595 * kmem_list3 structures, it's replaced with a kmalloc allocated
1596 * array at the end of the bootstrap.
1597 * 2) Create the first kmalloc cache.
1598 * The struct kmem_cache for the new cache is allocated normally.
1599 * An __init data area is used for the head array.
1600 * 3) Create the remaining kmalloc caches, with minimally sized
1602 * 4) Replace the __init data head arrays for kmem_cache and the first
1603 * kmalloc cache with kmalloc allocated arrays.
1604 * 5) Replace the __init data for kmem_list3 for kmem_cache and
1605 * the other cache's with kmalloc allocated memory.
1606 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1609 /* 1) create the kmem_cache */
1612 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1614 create_boot_cache(kmem_cache, "kmem_cache",
1615 offsetof(struct kmem_cache, array[nr_cpu_ids]) +
1616 nr_node_ids * sizeof(struct kmem_list3 *),
1617 SLAB_HWCACHE_ALIGN);
1618 list_add(&kmem_cache->list, &slab_caches);
1620 /* 2+3) create the kmalloc caches */
1623 * Initialize the caches that provide memory for the array cache and the
1624 * kmem_list3 structures first. Without this, further allocations will
1628 kmalloc_caches[INDEX_AC] = create_kmalloc_cache("kmalloc-ac",
1629 kmalloc_size(INDEX_AC), ARCH_KMALLOC_FLAGS);
1631 if (INDEX_AC != INDEX_L3)
1632 kmalloc_caches[INDEX_L3] =
1633 create_kmalloc_cache("kmalloc-l3",
1634 kmalloc_size(INDEX_L3), ARCH_KMALLOC_FLAGS);
1636 slab_early_init = 0;
1638 for (i = 1; i < PAGE_SHIFT + MAX_ORDER; i++) {
1639 size_t cs_size = kmalloc_size(i);
1641 if (cs_size < KMALLOC_MIN_SIZE)
1644 if (!kmalloc_caches[i]) {
1646 * For performance, all the general caches are L1 aligned.
1647 * This should be particularly beneficial on SMP boxes, as it
1648 * eliminates "false sharing".
1649 * Note for systems short on memory removing the alignment will
1650 * allow tighter packing of the smaller caches.
1652 kmalloc_caches[i] = create_kmalloc_cache("kmalloc",
1653 cs_size, ARCH_KMALLOC_FLAGS);
1656 #ifdef CONFIG_ZONE_DMA
1657 kmalloc_dma_caches[i] = create_kmalloc_cache(
1658 "kmalloc-dma", cs_size,
1659 SLAB_CACHE_DMA|ARCH_KMALLOC_FLAGS);
1662 /* 4) Replace the bootstrap head arrays */
1664 struct array_cache *ptr;
1666 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1668 memcpy(ptr, cpu_cache_get(kmem_cache),
1669 sizeof(struct arraycache_init));
1671 * Do not assume that spinlocks can be initialized via memcpy:
1673 spin_lock_init(&ptr->lock);
1675 kmem_cache->array[smp_processor_id()] = ptr;
1677 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1679 BUG_ON(cpu_cache_get(kmalloc_caches[INDEX_AC])
1680 != &initarray_generic.cache);
1681 memcpy(ptr, cpu_cache_get(kmalloc_caches[INDEX_AC]),
1682 sizeof(struct arraycache_init));
1684 * Do not assume that spinlocks can be initialized via memcpy:
1686 spin_lock_init(&ptr->lock);
1688 kmalloc_caches[INDEX_AC]->array[smp_processor_id()] = ptr;
1690 /* 5) Replace the bootstrap kmem_list3's */
1694 for_each_online_node(nid) {
1695 init_list(kmem_cache, &initkmem_list3[CACHE_CACHE + nid], nid);
1697 init_list(kmalloc_caches[INDEX_AC],
1698 &initkmem_list3[SIZE_AC + nid], nid);
1700 if (INDEX_AC != INDEX_L3) {
1701 init_list(kmalloc_caches[INDEX_L3],
1702 &initkmem_list3[SIZE_L3 + nid], nid);
1709 /* Create the proper names */
1710 for (i = 1; i < PAGE_SHIFT + MAX_ORDER; i++) {
1712 struct kmem_cache *c = kmalloc_caches[i];
1717 s = kasprintf(GFP_NOWAIT, "kmalloc-%d", kmalloc_size(i));
1722 #ifdef CONFIG_ZONE_DMA
1723 c = kmalloc_dma_caches[i];
1725 s = kasprintf(GFP_NOWAIT, "dma-kmalloc-%d", kmalloc_size(i));
1732 void __init kmem_cache_init_late(void)
1734 struct kmem_cache *cachep;
1738 /* 6) resize the head arrays to their final sizes */
1739 mutex_lock(&slab_mutex);
1740 list_for_each_entry(cachep, &slab_caches, list)
1741 if (enable_cpucache(cachep, GFP_NOWAIT))
1743 mutex_unlock(&slab_mutex);
1745 /* Annotate slab for lockdep -- annotate the malloc caches */
1752 * Register a cpu startup notifier callback that initializes
1753 * cpu_cache_get for all new cpus
1755 register_cpu_notifier(&cpucache_notifier);
1759 * Register a memory hotplug callback that initializes and frees
1762 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1766 * The reap timers are started later, with a module init call: That part
1767 * of the kernel is not yet operational.
1771 static int __init cpucache_init(void)
1776 * Register the timers that return unneeded pages to the page allocator
1778 for_each_online_cpu(cpu)
1779 start_cpu_timer(cpu);
1785 __initcall(cpucache_init);
1787 static noinline void
1788 slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1790 struct kmem_list3 *l3;
1792 unsigned long flags;
1796 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1798 printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n",
1799 cachep->name, cachep->size, cachep->gfporder);
1801 for_each_online_node(node) {
1802 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1803 unsigned long active_slabs = 0, num_slabs = 0;
1805 l3 = cachep->nodelists[node];
1809 spin_lock_irqsave(&l3->list_lock, flags);
1810 list_for_each_entry(slabp, &l3->slabs_full, list) {
1811 active_objs += cachep->num;
1814 list_for_each_entry(slabp, &l3->slabs_partial, list) {
1815 active_objs += slabp->inuse;
1818 list_for_each_entry(slabp, &l3->slabs_free, list)
1821 free_objects += l3->free_objects;
1822 spin_unlock_irqrestore(&l3->list_lock, flags);
1824 num_slabs += active_slabs;
1825 num_objs = num_slabs * cachep->num;
1827 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1828 node, active_slabs, num_slabs, active_objs, num_objs,
1834 * Interface to system's page allocator. No need to hold the cache-lock.
1836 * If we requested dmaable memory, we will get it. Even if we
1837 * did not request dmaable memory, we might get it, but that
1838 * would be relatively rare and ignorable.
1840 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1848 * Nommu uses slab's for process anonymous memory allocations, and thus
1849 * requires __GFP_COMP to properly refcount higher order allocations
1851 flags |= __GFP_COMP;
1854 flags |= cachep->allocflags;
1855 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1856 flags |= __GFP_RECLAIMABLE;
1858 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1860 if (!(flags & __GFP_NOWARN) && printk_ratelimit())
1861 slab_out_of_memory(cachep, flags, nodeid);
1865 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1866 if (unlikely(page->pfmemalloc))
1867 pfmemalloc_active = true;
1869 nr_pages = (1 << cachep->gfporder);
1870 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1871 add_zone_page_state(page_zone(page),
1872 NR_SLAB_RECLAIMABLE, nr_pages);
1874 add_zone_page_state(page_zone(page),
1875 NR_SLAB_UNRECLAIMABLE, nr_pages);
1876 for (i = 0; i < nr_pages; i++) {
1877 __SetPageSlab(page + i);
1879 if (page->pfmemalloc)
1880 SetPageSlabPfmemalloc(page + i);
1882 memcg_bind_pages(cachep, cachep->gfporder);
1884 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1885 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1888 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1890 kmemcheck_mark_unallocated_pages(page, nr_pages);
1893 return page_address(page);
1897 * Interface to system's page release.
1899 static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1901 unsigned long i = (1 << cachep->gfporder);
1902 struct page *page = virt_to_page(addr);
1903 const unsigned long nr_freed = i;
1905 kmemcheck_free_shadow(page, cachep->gfporder);
1907 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1908 sub_zone_page_state(page_zone(page),
1909 NR_SLAB_RECLAIMABLE, nr_freed);
1911 sub_zone_page_state(page_zone(page),
1912 NR_SLAB_UNRECLAIMABLE, nr_freed);
1914 BUG_ON(!PageSlab(page));
1915 __ClearPageSlabPfmemalloc(page);
1916 __ClearPageSlab(page);
1920 memcg_release_pages(cachep, cachep->gfporder);
1921 if (current->reclaim_state)
1922 current->reclaim_state->reclaimed_slab += nr_freed;
1923 free_memcg_kmem_pages((unsigned long)addr, cachep->gfporder);
1926 static void kmem_rcu_free(struct rcu_head *head)
1928 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1929 struct kmem_cache *cachep = slab_rcu->cachep;
1931 kmem_freepages(cachep, slab_rcu->addr);
1932 if (OFF_SLAB(cachep))
1933 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1938 #ifdef CONFIG_DEBUG_PAGEALLOC
1939 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1940 unsigned long caller)
1942 int size = cachep->object_size;
1944 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1946 if (size < 5 * sizeof(unsigned long))
1949 *addr++ = 0x12345678;
1951 *addr++ = smp_processor_id();
1952 size -= 3 * sizeof(unsigned long);
1954 unsigned long *sptr = &caller;
1955 unsigned long svalue;
1957 while (!kstack_end(sptr)) {
1959 if (kernel_text_address(svalue)) {
1961 size -= sizeof(unsigned long);
1962 if (size <= sizeof(unsigned long))
1968 *addr++ = 0x87654321;
1972 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1974 int size = cachep->object_size;
1975 addr = &((char *)addr)[obj_offset(cachep)];
1977 memset(addr, val, size);
1978 *(unsigned char *)(addr + size - 1) = POISON_END;
1981 static void dump_line(char *data, int offset, int limit)
1984 unsigned char error = 0;
1987 printk(KERN_ERR "%03x: ", offset);
1988 for (i = 0; i < limit; i++) {
1989 if (data[offset + i] != POISON_FREE) {
1990 error = data[offset + i];
1994 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1995 &data[offset], limit, 1);
1997 if (bad_count == 1) {
1998 error ^= POISON_FREE;
1999 if (!(error & (error - 1))) {
2000 printk(KERN_ERR "Single bit error detected. Probably "
2003 printk(KERN_ERR "Run memtest86+ or a similar memory "
2006 printk(KERN_ERR "Run a memory test tool.\n");
2015 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
2020 if (cachep->flags & SLAB_RED_ZONE) {
2021 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
2022 *dbg_redzone1(cachep, objp),
2023 *dbg_redzone2(cachep, objp));
2026 if (cachep->flags & SLAB_STORE_USER) {
2027 printk(KERN_ERR "Last user: [<%p>]",
2028 *dbg_userword(cachep, objp));
2029 print_symbol("(%s)",
2030 (unsigned long)*dbg_userword(cachep, objp));
2033 realobj = (char *)objp + obj_offset(cachep);
2034 size = cachep->object_size;
2035 for (i = 0; i < size && lines; i += 16, lines--) {
2038 if (i + limit > size)
2040 dump_line(realobj, i, limit);
2044 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
2050 realobj = (char *)objp + obj_offset(cachep);
2051 size = cachep->object_size;
2053 for (i = 0; i < size; i++) {
2054 char exp = POISON_FREE;
2057 if (realobj[i] != exp) {
2063 "Slab corruption (%s): %s start=%p, len=%d\n",
2064 print_tainted(), cachep->name, realobj, size);
2065 print_objinfo(cachep, objp, 0);
2067 /* Hexdump the affected line */
2070 if (i + limit > size)
2072 dump_line(realobj, i, limit);
2075 /* Limit to 5 lines */
2081 /* Print some data about the neighboring objects, if they
2084 struct slab *slabp = virt_to_slab(objp);
2087 objnr = obj_to_index(cachep, slabp, objp);
2089 objp = index_to_obj(cachep, slabp, objnr - 1);
2090 realobj = (char *)objp + obj_offset(cachep);
2091 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
2093 print_objinfo(cachep, objp, 2);
2095 if (objnr + 1 < cachep->num) {
2096 objp = index_to_obj(cachep, slabp, objnr + 1);
2097 realobj = (char *)objp + obj_offset(cachep);
2098 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
2100 print_objinfo(cachep, objp, 2);
2107 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
2110 for (i = 0; i < cachep->num; i++) {
2111 void *objp = index_to_obj(cachep, slabp, i);
2113 if (cachep->flags & SLAB_POISON) {
2114 #ifdef CONFIG_DEBUG_PAGEALLOC
2115 if (cachep->size % PAGE_SIZE == 0 &&
2117 kernel_map_pages(virt_to_page(objp),
2118 cachep->size / PAGE_SIZE, 1);
2120 check_poison_obj(cachep, objp);
2122 check_poison_obj(cachep, objp);
2125 if (cachep->flags & SLAB_RED_ZONE) {
2126 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2127 slab_error(cachep, "start of a freed object "
2129 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2130 slab_error(cachep, "end of a freed object "
2136 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
2142 * slab_destroy - destroy and release all objects in a slab
2143 * @cachep: cache pointer being destroyed
2144 * @slabp: slab pointer being destroyed
2146 * Destroy all the objs in a slab, and release the mem back to the system.
2147 * Before calling the slab must have been unlinked from the cache. The
2148 * cache-lock is not held/needed.
2150 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
2152 void *addr = slabp->s_mem - slabp->colouroff;
2154 slab_destroy_debugcheck(cachep, slabp);
2155 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
2156 struct slab_rcu *slab_rcu;
2158 slab_rcu = (struct slab_rcu *)slabp;
2159 slab_rcu->cachep = cachep;
2160 slab_rcu->addr = addr;
2161 call_rcu(&slab_rcu->head, kmem_rcu_free);
2163 kmem_freepages(cachep, addr);
2164 if (OFF_SLAB(cachep))
2165 kmem_cache_free(cachep->slabp_cache, slabp);
2170 * calculate_slab_order - calculate size (page order) of slabs
2171 * @cachep: pointer to the cache that is being created
2172 * @size: size of objects to be created in this cache.
2173 * @align: required alignment for the objects.
2174 * @flags: slab allocation flags
2176 * Also calculates the number of objects per slab.
2178 * This could be made much more intelligent. For now, try to avoid using
2179 * high order pages for slabs. When the gfp() functions are more friendly
2180 * towards high-order requests, this should be changed.
2182 static size_t calculate_slab_order(struct kmem_cache *cachep,
2183 size_t size, size_t align, unsigned long flags)
2185 unsigned long offslab_limit;
2186 size_t left_over = 0;
2189 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
2193 cache_estimate(gfporder, size, align, flags, &remainder, &num);
2197 if (flags & CFLGS_OFF_SLAB) {
2199 * Max number of objs-per-slab for caches which
2200 * use off-slab slabs. Needed to avoid a possible
2201 * looping condition in cache_grow().
2203 offslab_limit = size - sizeof(struct slab);
2204 offslab_limit /= sizeof(kmem_bufctl_t);
2206 if (num > offslab_limit)
2210 /* Found something acceptable - save it away */
2212 cachep->gfporder = gfporder;
2213 left_over = remainder;
2216 * A VFS-reclaimable slab tends to have most allocations
2217 * as GFP_NOFS and we really don't want to have to be allocating
2218 * higher-order pages when we are unable to shrink dcache.
2220 if (flags & SLAB_RECLAIM_ACCOUNT)
2224 * Large number of objects is good, but very large slabs are
2225 * currently bad for the gfp()s.
2227 if (gfporder >= slab_max_order)
2231 * Acceptable internal fragmentation?
2233 if (left_over * 8 <= (PAGE_SIZE << gfporder))
2239 static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
2241 if (slab_state >= FULL)
2242 return enable_cpucache(cachep, gfp);
2244 if (slab_state == DOWN) {
2246 * Note: Creation of first cache (kmem_cache).
2247 * The setup_list3s is taken care
2248 * of by the caller of __kmem_cache_create
2250 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2251 slab_state = PARTIAL;
2252 } else if (slab_state == PARTIAL) {
2254 * Note: the second kmem_cache_create must create the cache
2255 * that's used by kmalloc(24), otherwise the creation of
2256 * further caches will BUG().
2258 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2261 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2262 * the second cache, then we need to set up all its list3s,
2263 * otherwise the creation of further caches will BUG().
2265 set_up_list3s(cachep, SIZE_AC);
2266 if (INDEX_AC == INDEX_L3)
2267 slab_state = PARTIAL_L3;
2269 slab_state = PARTIAL_ARRAYCACHE;
2271 /* Remaining boot caches */
2272 cachep->array[smp_processor_id()] =
2273 kmalloc(sizeof(struct arraycache_init), gfp);
2275 if (slab_state == PARTIAL_ARRAYCACHE) {
2276 set_up_list3s(cachep, SIZE_L3);
2277 slab_state = PARTIAL_L3;
2280 for_each_online_node(node) {
2281 cachep->nodelists[node] =
2282 kmalloc_node(sizeof(struct kmem_list3),
2284 BUG_ON(!cachep->nodelists[node]);
2285 kmem_list3_init(cachep->nodelists[node]);
2289 cachep->nodelists[numa_mem_id()]->next_reap =
2290 jiffies + REAPTIMEOUT_LIST3 +
2291 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2293 cpu_cache_get(cachep)->avail = 0;
2294 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2295 cpu_cache_get(cachep)->batchcount = 1;
2296 cpu_cache_get(cachep)->touched = 0;
2297 cachep->batchcount = 1;
2298 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2303 * __kmem_cache_create - Create a cache.
2304 * @cachep: cache management descriptor
2305 * @flags: SLAB flags
2307 * Returns a ptr to the cache on success, NULL on failure.
2308 * Cannot be called within a int, but can be interrupted.
2309 * The @ctor is run when new pages are allocated by the cache.
2313 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2314 * to catch references to uninitialised memory.
2316 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2317 * for buffer overruns.
2319 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2320 * cacheline. This can be beneficial if you're counting cycles as closely
2324 __kmem_cache_create (struct kmem_cache *cachep, unsigned long flags)
2326 size_t left_over, slab_size, ralign;
2329 size_t size = cachep->size;
2334 * Enable redzoning and last user accounting, except for caches with
2335 * large objects, if the increased size would increase the object size
2336 * above the next power of two: caches with object sizes just above a
2337 * power of two have a significant amount of internal fragmentation.
2339 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2340 2 * sizeof(unsigned long long)))
2341 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2342 if (!(flags & SLAB_DESTROY_BY_RCU))
2343 flags |= SLAB_POISON;
2345 if (flags & SLAB_DESTROY_BY_RCU)
2346 BUG_ON(flags & SLAB_POISON);
2350 * Check that size is in terms of words. This is needed to avoid
2351 * unaligned accesses for some archs when redzoning is used, and makes
2352 * sure any on-slab bufctl's are also correctly aligned.
2354 if (size & (BYTES_PER_WORD - 1)) {
2355 size += (BYTES_PER_WORD - 1);
2356 size &= ~(BYTES_PER_WORD - 1);
2360 * Redzoning and user store require word alignment or possibly larger.
2361 * Note this will be overridden by architecture or caller mandated
2362 * alignment if either is greater than BYTES_PER_WORD.
2364 if (flags & SLAB_STORE_USER)
2365 ralign = BYTES_PER_WORD;
2367 if (flags & SLAB_RED_ZONE) {
2368 ralign = REDZONE_ALIGN;
2369 /* If redzoning, ensure that the second redzone is suitably
2370 * aligned, by adjusting the object size accordingly. */
2371 size += REDZONE_ALIGN - 1;
2372 size &= ~(REDZONE_ALIGN - 1);
2375 /* 3) caller mandated alignment */
2376 if (ralign < cachep->align) {
2377 ralign = cachep->align;
2379 /* disable debug if necessary */
2380 if (ralign > __alignof__(unsigned long long))
2381 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2385 cachep->align = ralign;
2387 if (slab_is_available())
2392 setup_nodelists_pointer(cachep);
2396 * Both debugging options require word-alignment which is calculated
2399 if (flags & SLAB_RED_ZONE) {
2400 /* add space for red zone words */
2401 cachep->obj_offset += sizeof(unsigned long long);
2402 size += 2 * sizeof(unsigned long long);
2404 if (flags & SLAB_STORE_USER) {
2405 /* user store requires one word storage behind the end of
2406 * the real object. But if the second red zone needs to be
2407 * aligned to 64 bits, we must allow that much space.
2409 if (flags & SLAB_RED_ZONE)
2410 size += REDZONE_ALIGN;
2412 size += BYTES_PER_WORD;
2414 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2415 if (size >= kmalloc_size(INDEX_L3 + 1)
2416 && cachep->object_size > cache_line_size() && ALIGN(size, align) < PAGE_SIZE) {
2417 cachep->obj_offset += PAGE_SIZE - ALIGN(size, align);
2424 * Determine if the slab management is 'on' or 'off' slab.
2425 * (bootstrapping cannot cope with offslab caches so don't do
2426 * it too early on. Always use on-slab management when
2427 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2429 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
2430 !(flags & SLAB_NOLEAKTRACE))
2432 * Size is large, assume best to place the slab management obj
2433 * off-slab (should allow better packing of objs).
2435 flags |= CFLGS_OFF_SLAB;
2437 size = ALIGN(size, cachep->align);
2439 left_over = calculate_slab_order(cachep, size, cachep->align, flags);
2444 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2445 + sizeof(struct slab), cachep->align);
2448 * If the slab has been placed off-slab, and we have enough space then
2449 * move it on-slab. This is at the expense of any extra colouring.
2451 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2452 flags &= ~CFLGS_OFF_SLAB;
2453 left_over -= slab_size;
2456 if (flags & CFLGS_OFF_SLAB) {
2457 /* really off slab. No need for manual alignment */
2459 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2461 #ifdef CONFIG_PAGE_POISONING
2462 /* If we're going to use the generic kernel_map_pages()
2463 * poisoning, then it's going to smash the contents of
2464 * the redzone and userword anyhow, so switch them off.
2466 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2467 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2471 cachep->colour_off = cache_line_size();
2472 /* Offset must be a multiple of the alignment. */
2473 if (cachep->colour_off < cachep->align)
2474 cachep->colour_off = cachep->align;
2475 cachep->colour = left_over / cachep->colour_off;
2476 cachep->slab_size = slab_size;
2477 cachep->flags = flags;
2478 cachep->allocflags = 0;
2479 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2480 cachep->allocflags |= GFP_DMA;
2481 cachep->size = size;
2482 cachep->reciprocal_buffer_size = reciprocal_value(size);
2484 if (flags & CFLGS_OFF_SLAB) {
2485 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2487 * This is a possibility for one of the malloc_sizes caches.
2488 * But since we go off slab only for object size greater than
2489 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2490 * this should not happen at all.
2491 * But leave a BUG_ON for some lucky dude.
2493 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
2496 err = setup_cpu_cache(cachep, gfp);
2498 __kmem_cache_shutdown(cachep);
2502 if (flags & SLAB_DEBUG_OBJECTS) {
2504 * Would deadlock through slab_destroy()->call_rcu()->
2505 * debug_object_activate()->kmem_cache_alloc().
2507 WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU);
2509 slab_set_debugobj_lock_classes(cachep);
2510 } else if (!OFF_SLAB(cachep) && !(flags & SLAB_DESTROY_BY_RCU))
2511 on_slab_lock_classes(cachep);
2517 static void check_irq_off(void)
2519 BUG_ON(!irqs_disabled());
2522 static void check_irq_on(void)
2524 BUG_ON(irqs_disabled());
2527 static void check_spinlock_acquired(struct kmem_cache *cachep)
2531 assert_spin_locked(&cachep->nodelists[numa_mem_id()]->list_lock);
2535 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2539 assert_spin_locked(&cachep->nodelists[node]->list_lock);
2544 #define check_irq_off() do { } while(0)
2545 #define check_irq_on() do { } while(0)
2546 #define check_spinlock_acquired(x) do { } while(0)
2547 #define check_spinlock_acquired_node(x, y) do { } while(0)
2550 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
2551 struct array_cache *ac,
2552 int force, int node);
2554 static void do_drain(void *arg)
2556 struct kmem_cache *cachep = arg;
2557 struct array_cache *ac;
2558 int node = numa_mem_id();
2561 ac = cpu_cache_get(cachep);
2562 spin_lock(&cachep->nodelists[node]->list_lock);
2563 free_block(cachep, ac->entry, ac->avail, node);
2564 spin_unlock(&cachep->nodelists[node]->list_lock);
2568 static void drain_cpu_caches(struct kmem_cache *cachep)
2570 struct kmem_list3 *l3;
2573 on_each_cpu(do_drain, cachep, 1);
2575 for_each_online_node(node) {
2576 l3 = cachep->nodelists[node];
2577 if (l3 && l3->alien)
2578 drain_alien_cache(cachep, l3->alien);
2581 for_each_online_node(node) {
2582 l3 = cachep->nodelists[node];
2584 drain_array(cachep, l3, l3->shared, 1, node);
2589 * Remove slabs from the list of free slabs.
2590 * Specify the number of slabs to drain in tofree.
2592 * Returns the actual number of slabs released.
2594 static int drain_freelist(struct kmem_cache *cache,
2595 struct kmem_list3 *l3, int tofree)
2597 struct list_head *p;
2602 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
2604 spin_lock_irq(&l3->list_lock);
2605 p = l3->slabs_free.prev;
2606 if (p == &l3->slabs_free) {
2607 spin_unlock_irq(&l3->list_lock);
2611 slabp = list_entry(p, struct slab, list);
2613 BUG_ON(slabp->inuse);
2615 list_del(&slabp->list);
2617 * Safe to drop the lock. The slab is no longer linked
2620 l3->free_objects -= cache->num;
2621 spin_unlock_irq(&l3->list_lock);
2622 slab_destroy(cache, slabp);
2629 /* Called with slab_mutex held to protect against cpu hotplug */
2630 static int __cache_shrink(struct kmem_cache *cachep)
2633 struct kmem_list3 *l3;
2635 drain_cpu_caches(cachep);
2638 for_each_online_node(i) {
2639 l3 = cachep->nodelists[i];
2643 drain_freelist(cachep, l3, l3->free_objects);
2645 ret += !list_empty(&l3->slabs_full) ||
2646 !list_empty(&l3->slabs_partial);
2648 return (ret ? 1 : 0);
2652 * kmem_cache_shrink - Shrink a cache.
2653 * @cachep: The cache to shrink.
2655 * Releases as many slabs as possible for a cache.
2656 * To help debugging, a zero exit status indicates all slabs were released.
2658 int kmem_cache_shrink(struct kmem_cache *cachep)
2661 BUG_ON(!cachep || in_interrupt());
2664 mutex_lock(&slab_mutex);
2665 ret = __cache_shrink(cachep);
2666 mutex_unlock(&slab_mutex);
2670 EXPORT_SYMBOL(kmem_cache_shrink);
2672 int __kmem_cache_shutdown(struct kmem_cache *cachep)
2675 struct kmem_list3 *l3;
2676 int rc = __cache_shrink(cachep);
2681 for_each_online_cpu(i)
2682 kfree(cachep->array[i]);
2684 /* NUMA: free the list3 structures */
2685 for_each_online_node(i) {
2686 l3 = cachep->nodelists[i];
2689 free_alien_cache(l3->alien);
2697 * Get the memory for a slab management obj.
2698 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2699 * always come from malloc_sizes caches. The slab descriptor cannot
2700 * come from the same cache which is getting created because,
2701 * when we are searching for an appropriate cache for these
2702 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2703 * If we are creating a malloc_sizes cache here it would not be visible to
2704 * kmem_find_general_cachep till the initialization is complete.
2705 * Hence we cannot have slabp_cache same as the original cache.
2707 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2708 int colour_off, gfp_t local_flags,
2713 if (OFF_SLAB(cachep)) {
2714 /* Slab management obj is off-slab. */
2715 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2716 local_flags, nodeid);
2718 * If the first object in the slab is leaked (it's allocated
2719 * but no one has a reference to it), we want to make sure
2720 * kmemleak does not treat the ->s_mem pointer as a reference
2721 * to the object. Otherwise we will not report the leak.
2723 kmemleak_scan_area(&slabp->list, sizeof(struct list_head),
2728 slabp = objp + colour_off;
2729 colour_off += cachep->slab_size;
2732 slabp->colouroff = colour_off;
2733 slabp->s_mem = objp + colour_off;
2734 slabp->nodeid = nodeid;
2739 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2741 return (kmem_bufctl_t *) (slabp + 1);
2744 static void cache_init_objs(struct kmem_cache *cachep,
2749 for (i = 0; i < cachep->num; i++) {
2750 void *objp = index_to_obj(cachep, slabp, i);
2752 /* need to poison the objs? */
2753 if (cachep->flags & SLAB_POISON)
2754 poison_obj(cachep, objp, POISON_FREE);
2755 if (cachep->flags & SLAB_STORE_USER)
2756 *dbg_userword(cachep, objp) = NULL;
2758 if (cachep->flags & SLAB_RED_ZONE) {
2759 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2760 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2763 * Constructors are not allowed to allocate memory from the same
2764 * cache which they are a constructor for. Otherwise, deadlock.
2765 * They must also be threaded.
2767 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2768 cachep->ctor(objp + obj_offset(cachep));
2770 if (cachep->flags & SLAB_RED_ZONE) {
2771 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2772 slab_error(cachep, "constructor overwrote the"
2773 " end of an object");
2774 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2775 slab_error(cachep, "constructor overwrote the"
2776 " start of an object");
2778 if ((cachep->size % PAGE_SIZE) == 0 &&
2779 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2780 kernel_map_pages(virt_to_page(objp),
2781 cachep->size / PAGE_SIZE, 0);
2786 slab_bufctl(slabp)[i] = i + 1;
2788 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2791 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2793 if (CONFIG_ZONE_DMA_FLAG) {
2794 if (flags & GFP_DMA)
2795 BUG_ON(!(cachep->allocflags & GFP_DMA));
2797 BUG_ON(cachep->allocflags & GFP_DMA);
2801 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2804 void *objp = index_to_obj(cachep, slabp, slabp->free);
2808 next = slab_bufctl(slabp)[slabp->free];
2810 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2811 WARN_ON(slabp->nodeid != nodeid);
2818 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2819 void *objp, int nodeid)
2821 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2824 /* Verify that the slab belongs to the intended node */
2825 WARN_ON(slabp->nodeid != nodeid);
2827 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2828 printk(KERN_ERR "slab: double free detected in cache "
2829 "'%s', objp %p\n", cachep->name, objp);
2833 slab_bufctl(slabp)[objnr] = slabp->free;
2834 slabp->free = objnr;
2839 * Map pages beginning at addr to the given cache and slab. This is required
2840 * for the slab allocator to be able to lookup the cache and slab of a
2841 * virtual address for kfree, ksize, and slab debugging.
2843 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2849 page = virt_to_page(addr);
2852 if (likely(!PageCompound(page)))
2853 nr_pages <<= cache->gfporder;
2856 page->slab_cache = cache;
2857 page->slab_page = slab;
2859 } while (--nr_pages);
2863 * Grow (by 1) the number of slabs within a cache. This is called by
2864 * kmem_cache_alloc() when there are no active objs left in a cache.
2866 static int cache_grow(struct kmem_cache *cachep,
2867 gfp_t flags, int nodeid, void *objp)
2872 struct kmem_list3 *l3;
2875 * Be lazy and only check for valid flags here, keeping it out of the
2876 * critical path in kmem_cache_alloc().
2878 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2879 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2881 /* Take the l3 list lock to change the colour_next on this node */
2883 l3 = cachep->nodelists[nodeid];
2884 spin_lock(&l3->list_lock);
2886 /* Get colour for the slab, and cal the next value. */
2887 offset = l3->colour_next;
2889 if (l3->colour_next >= cachep->colour)
2890 l3->colour_next = 0;
2891 spin_unlock(&l3->list_lock);
2893 offset *= cachep->colour_off;
2895 if (local_flags & __GFP_WAIT)
2899 * The test for missing atomic flag is performed here, rather than
2900 * the more obvious place, simply to reduce the critical path length
2901 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2902 * will eventually be caught here (where it matters).
2904 kmem_flagcheck(cachep, flags);
2907 * Get mem for the objs. Attempt to allocate a physical page from
2911 objp = kmem_getpages(cachep, local_flags, nodeid);
2915 /* Get slab management. */
2916 slabp = alloc_slabmgmt(cachep, objp, offset,
2917 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2921 slab_map_pages(cachep, slabp, objp);
2923 cache_init_objs(cachep, slabp);
2925 if (local_flags & __GFP_WAIT)
2926 local_irq_disable();
2928 spin_lock(&l3->list_lock);
2930 /* Make slab active. */
2931 list_add_tail(&slabp->list, &(l3->slabs_free));
2932 STATS_INC_GROWN(cachep);
2933 l3->free_objects += cachep->num;
2934 spin_unlock(&l3->list_lock);
2937 kmem_freepages(cachep, objp);
2939 if (local_flags & __GFP_WAIT)
2940 local_irq_disable();
2947 * Perform extra freeing checks:
2948 * - detect bad pointers.
2949 * - POISON/RED_ZONE checking
2951 static void kfree_debugcheck(const void *objp)
2953 if (!virt_addr_valid(objp)) {
2954 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2955 (unsigned long)objp);
2960 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2962 unsigned long long redzone1, redzone2;
2964 redzone1 = *dbg_redzone1(cache, obj);
2965 redzone2 = *dbg_redzone2(cache, obj);
2970 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2973 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2974 slab_error(cache, "double free detected");
2976 slab_error(cache, "memory outside object was overwritten");
2978 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2979 obj, redzone1, redzone2);
2982 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2983 unsigned long caller)
2989 BUG_ON(virt_to_cache(objp) != cachep);
2991 objp -= obj_offset(cachep);
2992 kfree_debugcheck(objp);
2993 page = virt_to_head_page(objp);
2995 slabp = page->slab_page;
2997 if (cachep->flags & SLAB_RED_ZONE) {
2998 verify_redzone_free(cachep, objp);
2999 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
3000 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
3002 if (cachep->flags & SLAB_STORE_USER)
3003 *dbg_userword(cachep, objp) = (void *)caller;
3005 objnr = obj_to_index(cachep, slabp, objp);
3007 BUG_ON(objnr >= cachep->num);
3008 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
3010 #ifdef CONFIG_DEBUG_SLAB_LEAK
3011 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
3013 if (cachep->flags & SLAB_POISON) {
3014 #ifdef CONFIG_DEBUG_PAGEALLOC
3015 if ((cachep->size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
3016 store_stackinfo(cachep, objp, caller);
3017 kernel_map_pages(virt_to_page(objp),
3018 cachep->size / PAGE_SIZE, 0);
3020 poison_obj(cachep, objp, POISON_FREE);
3023 poison_obj(cachep, objp, POISON_FREE);
3029 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
3034 /* Check slab's freelist to see if this obj is there. */
3035 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
3037 if (entries > cachep->num || i >= cachep->num)
3040 if (entries != cachep->num - slabp->inuse) {
3042 printk(KERN_ERR "slab: Internal list corruption detected in "
3043 "cache '%s'(%d), slabp %p(%d). Tainted(%s). Hexdump:\n",
3044 cachep->name, cachep->num, slabp, slabp->inuse,
3046 print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 16, 1, slabp,
3047 sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t),
3053 #define kfree_debugcheck(x) do { } while(0)
3054 #define cache_free_debugcheck(x,objp,z) (objp)
3055 #define check_slabp(x,y) do { } while(0)
3058 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags,
3062 struct kmem_list3 *l3;
3063 struct array_cache *ac;
3067 node = numa_mem_id();
3068 if (unlikely(force_refill))
3071 ac = cpu_cache_get(cachep);
3072 batchcount = ac->batchcount;
3073 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
3075 * If there was little recent activity on this cache, then
3076 * perform only a partial refill. Otherwise we could generate
3079 batchcount = BATCHREFILL_LIMIT;
3081 l3 = cachep->nodelists[node];
3083 BUG_ON(ac->avail > 0 || !l3);
3084 spin_lock(&l3->list_lock);
3086 /* See if we can refill from the shared array */
3087 if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) {
3088 l3->shared->touched = 1;
3092 while (batchcount > 0) {
3093 struct list_head *entry;
3095 /* Get slab alloc is to come from. */
3096 entry = l3->slabs_partial.next;
3097 if (entry == &l3->slabs_partial) {
3098 l3->free_touched = 1;
3099 entry = l3->slabs_free.next;
3100 if (entry == &l3->slabs_free)
3104 slabp = list_entry(entry, struct slab, list);
3105 check_slabp(cachep, slabp);
3106 check_spinlock_acquired(cachep);
3109 * The slab was either on partial or free list so
3110 * there must be at least one object available for
3113 BUG_ON(slabp->inuse >= cachep->num);
3115 while (slabp->inuse < cachep->num && batchcount--) {
3116 STATS_INC_ALLOCED(cachep);
3117 STATS_INC_ACTIVE(cachep);
3118 STATS_SET_HIGH(cachep);
3120 ac_put_obj(cachep, ac, slab_get_obj(cachep, slabp,
3123 check_slabp(cachep, slabp);
3125 /* move slabp to correct slabp list: */
3126 list_del(&slabp->list);
3127 if (slabp->free == BUFCTL_END)
3128 list_add(&slabp->list, &l3->slabs_full);
3130 list_add(&slabp->list, &l3->slabs_partial);
3134 l3->free_objects -= ac->avail;
3136 spin_unlock(&l3->list_lock);
3138 if (unlikely(!ac->avail)) {
3141 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
3143 /* cache_grow can reenable interrupts, then ac could change. */
3144 ac = cpu_cache_get(cachep);
3145 node = numa_mem_id();
3147 /* no objects in sight? abort */
3148 if (!x && (ac->avail == 0 || force_refill))
3151 if (!ac->avail) /* objects refilled by interrupt? */
3156 return ac_get_obj(cachep, ac, flags, force_refill);
3159 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3162 might_sleep_if(flags & __GFP_WAIT);
3164 kmem_flagcheck(cachep, flags);
3169 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3170 gfp_t flags, void *objp, unsigned long caller)
3174 if (cachep->flags & SLAB_POISON) {
3175 #ifdef CONFIG_DEBUG_PAGEALLOC
3176 if ((cachep->size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
3177 kernel_map_pages(virt_to_page(objp),
3178 cachep->size / PAGE_SIZE, 1);
3180 check_poison_obj(cachep, objp);
3182 check_poison_obj(cachep, objp);
3184 poison_obj(cachep, objp, POISON_INUSE);
3186 if (cachep->flags & SLAB_STORE_USER)
3187 *dbg_userword(cachep, objp) = (void *)caller;
3189 if (cachep->flags & SLAB_RED_ZONE) {
3190 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3191 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3192 slab_error(cachep, "double free, or memory outside"
3193 " object was overwritten");
3195 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3196 objp, *dbg_redzone1(cachep, objp),
3197 *dbg_redzone2(cachep, objp));
3199 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3200 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3202 #ifdef CONFIG_DEBUG_SLAB_LEAK
3207 slabp = virt_to_head_page(objp)->slab_page;
3208 objnr = (unsigned)(objp - slabp->s_mem) / cachep->size;
3209 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3212 objp += obj_offset(cachep);
3213 if (cachep->ctor && cachep->flags & SLAB_POISON)
3215 if (ARCH_SLAB_MINALIGN &&
3216 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
3217 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3218 objp, (int)ARCH_SLAB_MINALIGN);
3223 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3226 static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
3228 if (cachep == kmem_cache)
3231 return should_failslab(cachep->object_size, flags, cachep->flags);
3234 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3237 struct array_cache *ac;
3238 bool force_refill = false;
3242 ac = cpu_cache_get(cachep);
3243 if (likely(ac->avail)) {
3245 objp = ac_get_obj(cachep, ac, flags, false);
3248 * Allow for the possibility all avail objects are not allowed
3249 * by the current flags
3252 STATS_INC_ALLOCHIT(cachep);
3255 force_refill = true;
3258 STATS_INC_ALLOCMISS(cachep);
3259 objp = cache_alloc_refill(cachep, flags, force_refill);
3261 * the 'ac' may be updated by cache_alloc_refill(),
3262 * and kmemleak_erase() requires its correct value.
3264 ac = cpu_cache_get(cachep);
3268 * To avoid a false negative, if an object that is in one of the
3269 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3270 * treat the array pointers as a reference to the object.
3273 kmemleak_erase(&ac->entry[ac->avail]);
3279 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3281 * If we are in_interrupt, then process context, including cpusets and
3282 * mempolicy, may not apply and should not be used for allocation policy.
3284 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3286 int nid_alloc, nid_here;
3288 if (in_interrupt() || (flags & __GFP_THISNODE))
3290 nid_alloc = nid_here = numa_mem_id();
3291 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3292 nid_alloc = cpuset_slab_spread_node();
3293 else if (current->mempolicy)
3294 nid_alloc = slab_node();
3295 if (nid_alloc != nid_here)
3296 return ____cache_alloc_node(cachep, flags, nid_alloc);
3301 * Fallback function if there was no memory available and no objects on a
3302 * certain node and fall back is permitted. First we scan all the
3303 * available nodelists for available objects. If that fails then we
3304 * perform an allocation without specifying a node. This allows the page
3305 * allocator to do its reclaim / fallback magic. We then insert the
3306 * slab into the proper nodelist and then allocate from it.
3308 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3310 struct zonelist *zonelist;
3314 enum zone_type high_zoneidx = gfp_zone(flags);
3317 unsigned int cpuset_mems_cookie;
3319 if (flags & __GFP_THISNODE)
3322 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3325 cpuset_mems_cookie = get_mems_allowed();
3326 zonelist = node_zonelist(slab_node(), flags);
3330 * Look through allowed nodes for objects available
3331 * from existing per node queues.
3333 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3334 nid = zone_to_nid(zone);
3336 if (cpuset_zone_allowed_hardwall(zone, flags) &&
3337 cache->nodelists[nid] &&
3338 cache->nodelists[nid]->free_objects) {
3339 obj = ____cache_alloc_node(cache,
3340 flags | GFP_THISNODE, nid);
3348 * This allocation will be performed within the constraints
3349 * of the current cpuset / memory policy requirements.
3350 * We may trigger various forms of reclaim on the allowed
3351 * set and go into memory reserves if necessary.
3353 if (local_flags & __GFP_WAIT)
3355 kmem_flagcheck(cache, flags);
3356 obj = kmem_getpages(cache, local_flags, numa_mem_id());
3357 if (local_flags & __GFP_WAIT)
3358 local_irq_disable();
3361 * Insert into the appropriate per node queues
3363 nid = page_to_nid(virt_to_page(obj));
3364 if (cache_grow(cache, flags, nid, obj)) {
3365 obj = ____cache_alloc_node(cache,
3366 flags | GFP_THISNODE, nid);
3369 * Another processor may allocate the
3370 * objects in the slab since we are
3371 * not holding any locks.
3375 /* cache_grow already freed obj */
3381 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !obj))
3387 * A interface to enable slab creation on nodeid
3389 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3392 struct list_head *entry;
3394 struct kmem_list3 *l3;
3398 l3 = cachep->nodelists[nodeid];
3403 spin_lock(&l3->list_lock);
3404 entry = l3->slabs_partial.next;
3405 if (entry == &l3->slabs_partial) {
3406 l3->free_touched = 1;
3407 entry = l3->slabs_free.next;
3408 if (entry == &l3->slabs_free)
3412 slabp = list_entry(entry, struct slab, list);
3413 check_spinlock_acquired_node(cachep, nodeid);
3414 check_slabp(cachep, slabp);
3416 STATS_INC_NODEALLOCS(cachep);
3417 STATS_INC_ACTIVE(cachep);
3418 STATS_SET_HIGH(cachep);
3420 BUG_ON(slabp->inuse == cachep->num);
3422 obj = slab_get_obj(cachep, slabp, nodeid);
3423 check_slabp(cachep, slabp);
3425 /* move slabp to correct slabp list: */
3426 list_del(&slabp->list);
3428 if (slabp->free == BUFCTL_END)
3429 list_add(&slabp->list, &l3->slabs_full);
3431 list_add(&slabp->list, &l3->slabs_partial);
3433 spin_unlock(&l3->list_lock);
3437 spin_unlock(&l3->list_lock);
3438 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
3442 return fallback_alloc(cachep, flags);
3449 * kmem_cache_alloc_node - Allocate an object on the specified node
3450 * @cachep: The cache to allocate from.
3451 * @flags: See kmalloc().
3452 * @nodeid: node number of the target node.
3453 * @caller: return address of caller, used for debug information
3455 * Identical to kmem_cache_alloc but it will allocate memory on the given
3456 * node, which can improve the performance for cpu bound structures.
3458 * Fallback to other node is possible if __GFP_THISNODE is not set.
3460 static __always_inline void *
3461 slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3462 unsigned long caller)
3464 unsigned long save_flags;
3466 int slab_node = numa_mem_id();
3468 flags &= gfp_allowed_mask;
3470 lockdep_trace_alloc(flags);
3472 if (slab_should_failslab(cachep, flags))
3475 cachep = memcg_kmem_get_cache(cachep, flags);
3477 cache_alloc_debugcheck_before(cachep, flags);
3478 local_irq_save(save_flags);
3480 if (nodeid == NUMA_NO_NODE)
3483 if (unlikely(!cachep->nodelists[nodeid])) {
3484 /* Node not bootstrapped yet */
3485 ptr = fallback_alloc(cachep, flags);
3489 if (nodeid == slab_node) {
3491 * Use the locally cached objects if possible.
3492 * However ____cache_alloc does not allow fallback
3493 * to other nodes. It may fail while we still have
3494 * objects on other nodes available.
3496 ptr = ____cache_alloc(cachep, flags);
3500 /* ___cache_alloc_node can fall back to other nodes */
3501 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3503 local_irq_restore(save_flags);
3504 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3505 kmemleak_alloc_recursive(ptr, cachep->object_size, 1, cachep->flags,
3509 kmemcheck_slab_alloc(cachep, flags, ptr, cachep->object_size);
3511 if (unlikely((flags & __GFP_ZERO) && ptr))
3512 memset(ptr, 0, cachep->object_size);
3517 static __always_inline void *
3518 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3522 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3523 objp = alternate_node_alloc(cache, flags);
3527 objp = ____cache_alloc(cache, flags);
3530 * We may just have run out of memory on the local node.
3531 * ____cache_alloc_node() knows how to locate memory on other nodes
3534 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3541 static __always_inline void *
3542 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3544 return ____cache_alloc(cachep, flags);
3547 #endif /* CONFIG_NUMA */
3549 static __always_inline void *
3550 slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
3552 unsigned long save_flags;
3555 flags &= gfp_allowed_mask;
3557 lockdep_trace_alloc(flags);
3559 if (slab_should_failslab(cachep, flags))
3562 cachep = memcg_kmem_get_cache(cachep, flags);
3564 cache_alloc_debugcheck_before(cachep, flags);
3565 local_irq_save(save_flags);
3566 objp = __do_cache_alloc(cachep, flags);
3567 local_irq_restore(save_flags);
3568 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3569 kmemleak_alloc_recursive(objp, cachep->object_size, 1, cachep->flags,
3574 kmemcheck_slab_alloc(cachep, flags, objp, cachep->object_size);
3576 if (unlikely((flags & __GFP_ZERO) && objp))
3577 memset(objp, 0, cachep->object_size);
3583 * Caller needs to acquire correct kmem_list's list_lock
3585 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3589 struct kmem_list3 *l3;
3591 for (i = 0; i < nr_objects; i++) {
3595 clear_obj_pfmemalloc(&objpp[i]);
3598 slabp = virt_to_slab(objp);
3599 l3 = cachep->nodelists[node];
3600 list_del(&slabp->list);
3601 check_spinlock_acquired_node(cachep, node);
3602 check_slabp(cachep, slabp);
3603 slab_put_obj(cachep, slabp, objp, node);
3604 STATS_DEC_ACTIVE(cachep);
3606 check_slabp(cachep, slabp);
3608 /* fixup slab chains */
3609 if (slabp->inuse == 0) {
3610 if (l3->free_objects > l3->free_limit) {
3611 l3->free_objects -= cachep->num;
3612 /* No need to drop any previously held
3613 * lock here, even if we have a off-slab slab
3614 * descriptor it is guaranteed to come from
3615 * a different cache, refer to comments before
3618 slab_destroy(cachep, slabp);
3620 list_add(&slabp->list, &l3->slabs_free);
3623 /* Unconditionally move a slab to the end of the
3624 * partial list on free - maximum time for the
3625 * other objects to be freed, too.
3627 list_add_tail(&slabp->list, &l3->slabs_partial);
3632 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3635 struct kmem_list3 *l3;
3636 int node = numa_mem_id();
3638 batchcount = ac->batchcount;
3640 BUG_ON(!batchcount || batchcount > ac->avail);
3643 l3 = cachep->nodelists[node];
3644 spin_lock(&l3->list_lock);
3646 struct array_cache *shared_array = l3->shared;
3647 int max = shared_array->limit - shared_array->avail;
3649 if (batchcount > max)
3651 memcpy(&(shared_array->entry[shared_array->avail]),
3652 ac->entry, sizeof(void *) * batchcount);
3653 shared_array->avail += batchcount;
3658 free_block(cachep, ac->entry, batchcount, node);
3663 struct list_head *p;
3665 p = l3->slabs_free.next;
3666 while (p != &(l3->slabs_free)) {
3669 slabp = list_entry(p, struct slab, list);
3670 BUG_ON(slabp->inuse);
3675 STATS_SET_FREEABLE(cachep, i);
3678 spin_unlock(&l3->list_lock);
3679 ac->avail -= batchcount;
3680 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3684 * Release an obj back to its cache. If the obj has a constructed state, it must
3685 * be in this state _before_ it is released. Called with disabled ints.
3687 static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3688 unsigned long caller)
3690 struct array_cache *ac = cpu_cache_get(cachep);
3693 kmemleak_free_recursive(objp, cachep->flags);
3694 objp = cache_free_debugcheck(cachep, objp, caller);
3696 kmemcheck_slab_free(cachep, objp, cachep->object_size);
3699 * Skip calling cache_free_alien() when the platform is not numa.
3700 * This will avoid cache misses that happen while accessing slabp (which
3701 * is per page memory reference) to get nodeid. Instead use a global
3702 * variable to skip the call, which is mostly likely to be present in
3705 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3708 if (likely(ac->avail < ac->limit)) {
3709 STATS_INC_FREEHIT(cachep);
3711 STATS_INC_FREEMISS(cachep);
3712 cache_flusharray(cachep, ac);
3715 ac_put_obj(cachep, ac, objp);
3719 * kmem_cache_alloc - Allocate an object
3720 * @cachep: The cache to allocate from.
3721 * @flags: See kmalloc().
3723 * Allocate an object from this cache. The flags are only relevant
3724 * if the cache has no available objects.
3726 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3728 void *ret = slab_alloc(cachep, flags, _RET_IP_);
3730 trace_kmem_cache_alloc(_RET_IP_, ret,
3731 cachep->object_size, cachep->size, flags);
3735 EXPORT_SYMBOL(kmem_cache_alloc);
3737 #ifdef CONFIG_TRACING
3739 kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
3743 ret = slab_alloc(cachep, flags, _RET_IP_);
3745 trace_kmalloc(_RET_IP_, ret,
3746 size, cachep->size, flags);
3749 EXPORT_SYMBOL(kmem_cache_alloc_trace);
3753 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3755 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3757 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3758 cachep->object_size, cachep->size,
3763 EXPORT_SYMBOL(kmem_cache_alloc_node);
3765 #ifdef CONFIG_TRACING
3766 void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
3773 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3775 trace_kmalloc_node(_RET_IP_, ret,
3780 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3783 static __always_inline void *
3784 __do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
3786 struct kmem_cache *cachep;
3788 cachep = kmem_find_general_cachep(size, flags);
3789 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3791 return kmem_cache_alloc_node_trace(cachep, flags, node, size);
3794 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3795 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3797 return __do_kmalloc_node(size, flags, node, _RET_IP_);
3799 EXPORT_SYMBOL(__kmalloc_node);
3801 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3802 int node, unsigned long caller)
3804 return __do_kmalloc_node(size, flags, node, caller);
3806 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3808 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3810 return __do_kmalloc_node(size, flags, node, 0);
3812 EXPORT_SYMBOL(__kmalloc_node);
3813 #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3814 #endif /* CONFIG_NUMA */
3817 * __do_kmalloc - allocate memory
3818 * @size: how many bytes of memory are required.
3819 * @flags: the type of memory to allocate (see kmalloc).
3820 * @caller: function caller for debug tracking of the caller
3822 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3823 unsigned long caller)
3825 struct kmem_cache *cachep;
3828 /* If you want to save a few bytes .text space: replace
3830 * Then kmalloc uses the uninlined functions instead of the inline
3833 cachep = __find_general_cachep(size, flags);
3834 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3836 ret = slab_alloc(cachep, flags, caller);
3838 trace_kmalloc(caller, ret,
3839 size, cachep->size, flags);
3845 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3846 void *__kmalloc(size_t size, gfp_t flags)
3848 return __do_kmalloc(size, flags, _RET_IP_);
3850 EXPORT_SYMBOL(__kmalloc);
3852 void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3854 return __do_kmalloc(size, flags, caller);
3856 EXPORT_SYMBOL(__kmalloc_track_caller);
3859 void *__kmalloc(size_t size, gfp_t flags)
3861 return __do_kmalloc(size, flags, 0);
3863 EXPORT_SYMBOL(__kmalloc);
3867 * kmem_cache_free - Deallocate an object
3868 * @cachep: The cache the allocation was from.
3869 * @objp: The previously allocated object.
3871 * Free an object which was previously allocated from this
3874 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3876 unsigned long flags;
3877 cachep = cache_from_obj(cachep, objp);
3881 local_irq_save(flags);
3882 debug_check_no_locks_freed(objp, cachep->object_size);
3883 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3884 debug_check_no_obj_freed(objp, cachep->object_size);
3885 __cache_free(cachep, objp, _RET_IP_);
3886 local_irq_restore(flags);
3888 trace_kmem_cache_free(_RET_IP_, objp);
3890 EXPORT_SYMBOL(kmem_cache_free);
3893 * kfree - free previously allocated memory
3894 * @objp: pointer returned by kmalloc.
3896 * If @objp is NULL, no operation is performed.
3898 * Don't free memory not originally allocated by kmalloc()
3899 * or you will run into trouble.
3901 void kfree(const void *objp)
3903 struct kmem_cache *c;
3904 unsigned long flags;
3906 trace_kfree(_RET_IP_, objp);
3908 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3910 local_irq_save(flags);
3911 kfree_debugcheck(objp);
3912 c = virt_to_cache(objp);
3913 debug_check_no_locks_freed(objp, c->object_size);
3915 debug_check_no_obj_freed(objp, c->object_size);
3916 __cache_free(c, (void *)objp, _RET_IP_);
3917 local_irq_restore(flags);
3919 EXPORT_SYMBOL(kfree);
3922 * This initializes kmem_list3 or resizes various caches for all nodes.
3924 static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
3927 struct kmem_list3 *l3;
3928 struct array_cache *new_shared;
3929 struct array_cache **new_alien = NULL;
3931 for_each_online_node(node) {
3933 if (use_alien_caches) {
3934 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
3940 if (cachep->shared) {
3941 new_shared = alloc_arraycache(node,
3942 cachep->shared*cachep->batchcount,
3945 free_alien_cache(new_alien);
3950 l3 = cachep->nodelists[node];
3952 struct array_cache *shared = l3->shared;
3954 spin_lock_irq(&l3->list_lock);
3957 free_block(cachep, shared->entry,
3958 shared->avail, node);
3960 l3->shared = new_shared;
3962 l3->alien = new_alien;
3965 l3->free_limit = (1 + nr_cpus_node(node)) *
3966 cachep->batchcount + cachep->num;
3967 spin_unlock_irq(&l3->list_lock);
3969 free_alien_cache(new_alien);
3972 l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node);
3974 free_alien_cache(new_alien);
3979 kmem_list3_init(l3);
3980 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3981 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3982 l3->shared = new_shared;
3983 l3->alien = new_alien;
3984 l3->free_limit = (1 + nr_cpus_node(node)) *
3985 cachep->batchcount + cachep->num;
3986 cachep->nodelists[node] = l3;
3991 if (!cachep->list.next) {
3992 /* Cache is not active yet. Roll back what we did */
3995 if (cachep->nodelists[node]) {
3996 l3 = cachep->nodelists[node];
3999 free_alien_cache(l3->alien);
4001 cachep->nodelists[node] = NULL;
4009 struct ccupdate_struct {
4010 struct kmem_cache *cachep;
4011 struct array_cache *new[0];
4014 static void do_ccupdate_local(void *info)
4016 struct ccupdate_struct *new = info;
4017 struct array_cache *old;
4020 old = cpu_cache_get(new->cachep);
4022 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
4023 new->new[smp_processor_id()] = old;
4026 /* Always called with the slab_mutex held */
4027 static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
4028 int batchcount, int shared, gfp_t gfp)
4030 struct ccupdate_struct *new;
4033 new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *),
4038 for_each_online_cpu(i) {
4039 new->new[i] = alloc_arraycache(cpu_to_mem(i), limit,
4042 for (i--; i >= 0; i--)
4048 new->cachep = cachep;
4050 on_each_cpu(do_ccupdate_local, (void *)new, 1);
4053 cachep->batchcount = batchcount;
4054 cachep->limit = limit;
4055 cachep->shared = shared;
4057 for_each_online_cpu(i) {
4058 struct array_cache *ccold = new->new[i];
4061 spin_lock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
4062 free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i));
4063 spin_unlock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock);
4067 return alloc_kmemlist(cachep, gfp);
4070 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
4071 int batchcount, int shared, gfp_t gfp)
4074 struct kmem_cache *c = NULL;
4077 ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
4079 if (slab_state < FULL)
4082 if ((ret < 0) || !is_root_cache(cachep))
4085 VM_BUG_ON(!mutex_is_locked(&slab_mutex));
4086 for_each_memcg_cache_index(i) {
4087 c = cache_from_memcg(cachep, i);
4089 /* return value determined by the parent cache only */
4090 __do_tune_cpucache(c, limit, batchcount, shared, gfp);
4096 /* Called with slab_mutex held always */
4097 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
4104 if (!is_root_cache(cachep)) {
4105 struct kmem_cache *root = memcg_root_cache(cachep);
4106 limit = root->limit;
4107 shared = root->shared;
4108 batchcount = root->batchcount;
4111 if (limit && shared && batchcount)
4114 * The head array serves three purposes:
4115 * - create a LIFO ordering, i.e. return objects that are cache-warm
4116 * - reduce the number of spinlock operations.
4117 * - reduce the number of linked list operations on the slab and
4118 * bufctl chains: array operations are cheaper.
4119 * The numbers are guessed, we should auto-tune as described by
4122 if (cachep->size > 131072)
4124 else if (cachep->size > PAGE_SIZE)
4126 else if (cachep->size > 1024)
4128 else if (cachep->size > 256)
4134 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
4135 * allocation behaviour: Most allocs on one cpu, most free operations
4136 * on another cpu. For these cases, an efficient object passing between
4137 * cpus is necessary. This is provided by a shared array. The array
4138 * replaces Bonwick's magazine layer.
4139 * On uniprocessor, it's functionally equivalent (but less efficient)
4140 * to a larger limit. Thus disabled by default.
4143 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
4148 * With debugging enabled, large batchcount lead to excessively long
4149 * periods with disabled local interrupts. Limit the batchcount
4154 batchcount = (limit + 1) / 2;
4156 err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
4158 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
4159 cachep->name, -err);
4164 * Drain an array if it contains any elements taking the l3 lock only if
4165 * necessary. Note that the l3 listlock also protects the array_cache
4166 * if drain_array() is used on the shared array.
4168 static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3,
4169 struct array_cache *ac, int force, int node)
4173 if (!ac || !ac->avail)
4175 if (ac->touched && !force) {
4178 spin_lock_irq(&l3->list_lock);
4180 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4181 if (tofree > ac->avail)
4182 tofree = (ac->avail + 1) / 2;
4183 free_block(cachep, ac->entry, tofree, node);
4184 ac->avail -= tofree;
4185 memmove(ac->entry, &(ac->entry[tofree]),
4186 sizeof(void *) * ac->avail);
4188 spin_unlock_irq(&l3->list_lock);
4193 * cache_reap - Reclaim memory from caches.
4194 * @w: work descriptor
4196 * Called from workqueue/eventd every few seconds.
4198 * - clear the per-cpu caches for this CPU.
4199 * - return freeable pages to the main free memory pool.
4201 * If we cannot acquire the cache chain mutex then just give up - we'll try
4202 * again on the next iteration.
4204 static void cache_reap(struct work_struct *w)
4206 struct kmem_cache *searchp;
4207 struct kmem_list3 *l3;
4208 int node = numa_mem_id();
4209 struct delayed_work *work = to_delayed_work(w);
4211 if (!mutex_trylock(&slab_mutex))
4212 /* Give up. Setup the next iteration. */
4215 list_for_each_entry(searchp, &slab_caches, list) {
4219 * We only take the l3 lock if absolutely necessary and we
4220 * have established with reasonable certainty that
4221 * we can do some work if the lock was obtained.
4223 l3 = searchp->nodelists[node];
4225 reap_alien(searchp, l3);
4227 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
4230 * These are racy checks but it does not matter
4231 * if we skip one check or scan twice.
4233 if (time_after(l3->next_reap, jiffies))
4236 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
4238 drain_array(searchp, l3, l3->shared, 0, node);
4240 if (l3->free_touched)
4241 l3->free_touched = 0;
4245 freed = drain_freelist(searchp, l3, (l3->free_limit +
4246 5 * searchp->num - 1) / (5 * searchp->num));
4247 STATS_ADD_REAPED(searchp, freed);
4253 mutex_unlock(&slab_mutex);
4256 /* Set up the next iteration */
4257 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
4260 #ifdef CONFIG_SLABINFO
4261 void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
4264 unsigned long active_objs;
4265 unsigned long num_objs;
4266 unsigned long active_slabs = 0;
4267 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4271 struct kmem_list3 *l3;
4275 for_each_online_node(node) {
4276 l3 = cachep->nodelists[node];
4281 spin_lock_irq(&l3->list_lock);
4283 list_for_each_entry(slabp, &l3->slabs_full, list) {
4284 if (slabp->inuse != cachep->num && !error)
4285 error = "slabs_full accounting error";
4286 active_objs += cachep->num;
4289 list_for_each_entry(slabp, &l3->slabs_partial, list) {
4290 if (slabp->inuse == cachep->num && !error)
4291 error = "slabs_partial inuse accounting error";
4292 if (!slabp->inuse && !error)
4293 error = "slabs_partial/inuse accounting error";
4294 active_objs += slabp->inuse;
4297 list_for_each_entry(slabp, &l3->slabs_free, list) {
4298 if (slabp->inuse && !error)
4299 error = "slabs_free/inuse accounting error";
4302 free_objects += l3->free_objects;
4304 shared_avail += l3->shared->avail;
4306 spin_unlock_irq(&l3->list_lock);
4308 num_slabs += active_slabs;
4309 num_objs = num_slabs * cachep->num;
4310 if (num_objs - active_objs != free_objects && !error)
4311 error = "free_objects accounting error";
4313 name = cachep->name;
4315 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4317 sinfo->active_objs = active_objs;
4318 sinfo->num_objs = num_objs;
4319 sinfo->active_slabs = active_slabs;
4320 sinfo->num_slabs = num_slabs;
4321 sinfo->shared_avail = shared_avail;
4322 sinfo->limit = cachep->limit;
4323 sinfo->batchcount = cachep->batchcount;
4324 sinfo->shared = cachep->shared;
4325 sinfo->objects_per_slab = cachep->num;
4326 sinfo->cache_order = cachep->gfporder;
4329 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
4333 unsigned long high = cachep->high_mark;
4334 unsigned long allocs = cachep->num_allocations;
4335 unsigned long grown = cachep->grown;
4336 unsigned long reaped = cachep->reaped;
4337 unsigned long errors = cachep->errors;
4338 unsigned long max_freeable = cachep->max_freeable;
4339 unsigned long node_allocs = cachep->node_allocs;
4340 unsigned long node_frees = cachep->node_frees;
4341 unsigned long overflows = cachep->node_overflow;
4343 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4344 "%4lu %4lu %4lu %4lu %4lu",
4345 allocs, high, grown,
4346 reaped, errors, max_freeable, node_allocs,
4347 node_frees, overflows);
4351 unsigned long allochit = atomic_read(&cachep->allochit);
4352 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4353 unsigned long freehit = atomic_read(&cachep->freehit);
4354 unsigned long freemiss = atomic_read(&cachep->freemiss);
4356 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4357 allochit, allocmiss, freehit, freemiss);
4362 #define MAX_SLABINFO_WRITE 128
4364 * slabinfo_write - Tuning for the slab allocator
4366 * @buffer: user buffer
4367 * @count: data length
4370 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4371 size_t count, loff_t *ppos)
4373 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4374 int limit, batchcount, shared, res;
4375 struct kmem_cache *cachep;
4377 if (count > MAX_SLABINFO_WRITE)
4379 if (copy_from_user(&kbuf, buffer, count))
4381 kbuf[MAX_SLABINFO_WRITE] = '\0';
4383 tmp = strchr(kbuf, ' ');
4388 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4391 /* Find the cache in the chain of caches. */
4392 mutex_lock(&slab_mutex);
4394 list_for_each_entry(cachep, &slab_caches, list) {
4395 if (!strcmp(cachep->name, kbuf)) {
4396 if (limit < 1 || batchcount < 1 ||
4397 batchcount > limit || shared < 0) {
4400 res = do_tune_cpucache(cachep, limit,
4407 mutex_unlock(&slab_mutex);
4413 #ifdef CONFIG_DEBUG_SLAB_LEAK
4415 static void *leaks_start(struct seq_file *m, loff_t *pos)
4417 mutex_lock(&slab_mutex);
4418 return seq_list_start(&slab_caches, *pos);
4421 static inline int add_caller(unsigned long *n, unsigned long v)
4431 unsigned long *q = p + 2 * i;
4445 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4451 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4457 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->size) {
4458 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4460 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4465 static void show_symbol(struct seq_file *m, unsigned long address)
4467 #ifdef CONFIG_KALLSYMS
4468 unsigned long offset, size;
4469 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4471 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4472 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4474 seq_printf(m, " [%s]", modname);
4478 seq_printf(m, "%p", (void *)address);
4481 static int leaks_show(struct seq_file *m, void *p)
4483 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
4485 struct kmem_list3 *l3;
4487 unsigned long *n = m->private;
4491 if (!(cachep->flags & SLAB_STORE_USER))
4493 if (!(cachep->flags & SLAB_RED_ZONE))
4496 /* OK, we can do it */
4500 for_each_online_node(node) {
4501 l3 = cachep->nodelists[node];
4506 spin_lock_irq(&l3->list_lock);
4508 list_for_each_entry(slabp, &l3->slabs_full, list)
4509 handle_slab(n, cachep, slabp);
4510 list_for_each_entry(slabp, &l3->slabs_partial, list)
4511 handle_slab(n, cachep, slabp);
4512 spin_unlock_irq(&l3->list_lock);
4514 name = cachep->name;
4516 /* Increase the buffer size */
4517 mutex_unlock(&slab_mutex);
4518 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4520 /* Too bad, we are really out */
4522 mutex_lock(&slab_mutex);
4525 *(unsigned long *)m->private = n[0] * 2;
4527 mutex_lock(&slab_mutex);
4528 /* Now make sure this entry will be retried */
4532 for (i = 0; i < n[1]; i++) {
4533 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4534 show_symbol(m, n[2*i+2]);
4541 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4543 return seq_list_next(p, &slab_caches, pos);
4546 static void s_stop(struct seq_file *m, void *p)
4548 mutex_unlock(&slab_mutex);
4551 static const struct seq_operations slabstats_op = {
4552 .start = leaks_start,
4558 static int slabstats_open(struct inode *inode, struct file *file)
4560 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4563 ret = seq_open(file, &slabstats_op);
4565 struct seq_file *m = file->private_data;
4566 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4575 static const struct file_operations proc_slabstats_operations = {
4576 .open = slabstats_open,
4578 .llseek = seq_lseek,
4579 .release = seq_release_private,
4583 static int __init slab_proc_init(void)
4585 #ifdef CONFIG_DEBUG_SLAB_LEAK
4586 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4590 module_init(slab_proc_init);
4594 * ksize - get the actual amount of memory allocated for a given object
4595 * @objp: Pointer to the object
4597 * kmalloc may internally round up allocations and return more memory
4598 * than requested. ksize() can be used to determine the actual amount of
4599 * memory allocated. The caller may use this additional memory, even though
4600 * a smaller amount of memory was initially specified with the kmalloc call.
4601 * The caller must guarantee that objp points to a valid object previously
4602 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4603 * must not be freed during the duration of the call.
4605 size_t ksize(const void *objp)
4608 if (unlikely(objp == ZERO_SIZE_PTR))
4611 return virt_to_cache(objp)->object_size;
4613 EXPORT_SYMBOL(ksize);