3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
28 * This means, that your constructor is used only for newly allocated
29 * slabs and you must pass objects with the same initializations to
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
39 * empty slabs with no allocated objects
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
53 * The c_cpuarray may not be read with enabled local interrupts -
54 * it's changed with a smp_call_function().
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
58 * Several members in struct kmem_cache and struct slab never change, they
59 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
68 * Further notes from the original documentation:
70 * 11 April '97. Started multi-threading - markhe
71 * The global cache-chain is protected by the mutex 'slab_mutex'.
72 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
76 * At present, each engine can be growing a cache. This should be blocked.
78 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
89 #include <linux/slab.h>
91 #include <linux/poison.h>
92 #include <linux/swap.h>
93 #include <linux/cache.h>
94 #include <linux/interrupt.h>
95 #include <linux/init.h>
96 #include <linux/compiler.h>
97 #include <linux/cpuset.h>
98 #include <linux/proc_fs.h>
99 #include <linux/seq_file.h>
100 #include <linux/notifier.h>
101 #include <linux/kallsyms.h>
102 #include <linux/cpu.h>
103 #include <linux/sysctl.h>
104 #include <linux/module.h>
105 #include <linux/rcupdate.h>
106 #include <linux/string.h>
107 #include <linux/uaccess.h>
108 #include <linux/nodemask.h>
109 #include <linux/kmemleak.h>
110 #include <linux/mempolicy.h>
111 #include <linux/mutex.h>
112 #include <linux/fault-inject.h>
113 #include <linux/rtmutex.h>
114 #include <linux/reciprocal_div.h>
115 #include <linux/debugobjects.h>
116 #include <linux/kmemcheck.h>
117 #include <linux/memory.h>
118 #include <linux/prefetch.h>
120 #include <net/sock.h>
122 #include <asm/cacheflush.h>
123 #include <asm/tlbflush.h>
124 #include <asm/page.h>
126 #include <trace/events/kmem.h>
128 #include "internal.h"
133 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
134 * 0 for faster, smaller code (especially in the critical paths).
136 * STATS - 1 to collect stats for /proc/slabinfo.
137 * 0 for faster, smaller code (especially in the critical paths).
139 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
142 #ifdef CONFIG_DEBUG_SLAB
145 #define FORCED_DEBUG 1
149 #define FORCED_DEBUG 0
152 /* Shouldn't this be in a header file somewhere? */
153 #define BYTES_PER_WORD sizeof(void *)
154 #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
156 #ifndef ARCH_KMALLOC_FLAGS
157 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
161 * true if a page was allocated from pfmemalloc reserves for network-based
164 static bool pfmemalloc_active __read_mostly;
169 * Bufctl's are used for linking objs within a slab
172 * This implementation relies on "struct page" for locating the cache &
173 * slab an object belongs to.
174 * This allows the bufctl structure to be small (one int), but limits
175 * the number of objects a slab (not a cache) can contain when off-slab
176 * bufctls are used. The limit is the size of the largest general cache
177 * that does not use off-slab slabs.
178 * For 32bit archs with 4 kB pages, is this 56.
179 * This is not serious, as it is only for large objects, when it is unwise
180 * to have too many per slab.
181 * Note: This limit can be raised by introducing a general cache whose size
182 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
185 typedef unsigned int kmem_bufctl_t;
186 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
187 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
188 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
189 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
194 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
195 * arrange for kmem_freepages to be called via RCU. This is useful if
196 * we need to approach a kernel structure obliquely, from its address
197 * obtained without the usual locking. We can lock the structure to
198 * stabilize it and check it's still at the given address, only if we
199 * can be sure that the memory has not been meanwhile reused for some
200 * other kind of object (which our subsystem's lock might corrupt).
202 * rcu_read_lock before reading the address, then rcu_read_unlock after
203 * taking the spinlock within the structure expected at that address.
206 struct rcu_head head;
207 struct kmem_cache *cachep;
214 * Manages the objs in a slab. Placed either at the beginning of mem allocated
215 * for a slab, or allocated from an general cache.
216 * Slabs are chained into three list: fully used, partial, fully free slabs.
221 struct list_head list;
222 unsigned long colouroff;
223 void *s_mem; /* including colour offset */
224 unsigned int inuse; /* num of objs active in slab */
226 unsigned short nodeid;
228 struct slab_rcu __slab_cover_slab_rcu;
236 * - LIFO ordering, to hand out cache-warm objects from _alloc
237 * - reduce the number of linked list operations
238 * - reduce spinlock operations
240 * The limit is stored in the per-cpu structure to reduce the data cache
247 unsigned int batchcount;
248 unsigned int touched;
251 * Must have this definition in here for the proper
252 * alignment of array_cache. Also simplifies accessing
255 * Entries should not be directly dereferenced as
256 * entries belonging to slabs marked pfmemalloc will
257 * have the lower bits set SLAB_OBJ_PFMEMALLOC
261 #define SLAB_OBJ_PFMEMALLOC 1
262 static inline bool is_obj_pfmemalloc(void *objp)
264 return (unsigned long)objp & SLAB_OBJ_PFMEMALLOC;
267 static inline void set_obj_pfmemalloc(void **objp)
269 *objp = (void *)((unsigned long)*objp | SLAB_OBJ_PFMEMALLOC);
273 static inline void clear_obj_pfmemalloc(void **objp)
275 *objp = (void *)((unsigned long)*objp & ~SLAB_OBJ_PFMEMALLOC);
279 * bootstrap: The caches do not work without cpuarrays anymore, but the
280 * cpuarrays are allocated from the generic caches...
282 #define BOOT_CPUCACHE_ENTRIES 1
283 struct arraycache_init {
284 struct array_cache cache;
285 void *entries[BOOT_CPUCACHE_ENTRIES];
289 * The slab lists for all objects.
291 struct kmem_cache_node {
292 struct list_head slabs_partial; /* partial list first, better asm code */
293 struct list_head slabs_full;
294 struct list_head slabs_free;
295 unsigned long free_objects;
296 unsigned int free_limit;
297 unsigned int colour_next; /* Per-node cache coloring */
298 spinlock_t list_lock;
299 struct array_cache *shared; /* shared per node */
300 struct array_cache **alien; /* on other nodes */
301 unsigned long next_reap; /* updated without locking */
302 int free_touched; /* updated without locking */
306 * Need this for bootstrapping a per node allocator.
308 #define NUM_INIT_LISTS (3 * MAX_NUMNODES)
309 static struct kmem_cache_node __initdata initkmem_list3[NUM_INIT_LISTS];
310 #define CACHE_CACHE 0
311 #define SIZE_AC MAX_NUMNODES
312 #define SIZE_L3 (2 * MAX_NUMNODES)
314 static int drain_freelist(struct kmem_cache *cache,
315 struct kmem_cache_node *l3, int tofree);
316 static void free_block(struct kmem_cache *cachep, void **objpp, int len,
318 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
319 static void cache_reap(struct work_struct *unused);
321 struct kmem_cache *kmalloc_caches[KMALLOC_SHIFT_HIGH + 1];
322 EXPORT_SYMBOL(kmalloc_caches);
324 #ifdef CONFIG_ZONE_DMA
325 struct kmem_cache *kmalloc_dma_caches[KMALLOC_SHIFT_HIGH + 1];
326 EXPORT_SYMBOL(kmalloc_dma_caches);
329 static int slab_early_init = 1;
331 #define INDEX_AC kmalloc_index(sizeof(struct arraycache_init))
332 #define INDEX_L3 kmalloc_index(sizeof(struct kmem_cache_node))
334 static void kmem_list3_init(struct kmem_cache_node *parent)
336 INIT_LIST_HEAD(&parent->slabs_full);
337 INIT_LIST_HEAD(&parent->slabs_partial);
338 INIT_LIST_HEAD(&parent->slabs_free);
339 parent->shared = NULL;
340 parent->alien = NULL;
341 parent->colour_next = 0;
342 spin_lock_init(&parent->list_lock);
343 parent->free_objects = 0;
344 parent->free_touched = 0;
347 #define MAKE_LIST(cachep, listp, slab, nodeid) \
349 INIT_LIST_HEAD(listp); \
350 list_splice(&(cachep->node[nodeid]->slab), listp); \
353 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
355 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
356 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
357 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
360 #define CFLGS_OFF_SLAB (0x80000000UL)
361 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
363 #define BATCHREFILL_LIMIT 16
365 * Optimization question: fewer reaps means less probability for unnessary
366 * cpucache drain/refill cycles.
368 * OTOH the cpuarrays can contain lots of objects,
369 * which could lock up otherwise freeable slabs.
371 #define REAPTIMEOUT_CPUC (2*HZ)
372 #define REAPTIMEOUT_LIST3 (4*HZ)
375 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
376 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
377 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
378 #define STATS_INC_GROWN(x) ((x)->grown++)
379 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
380 #define STATS_SET_HIGH(x) \
382 if ((x)->num_active > (x)->high_mark) \
383 (x)->high_mark = (x)->num_active; \
385 #define STATS_INC_ERR(x) ((x)->errors++)
386 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
387 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
388 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
389 #define STATS_SET_FREEABLE(x, i) \
391 if ((x)->max_freeable < i) \
392 (x)->max_freeable = i; \
394 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
395 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
396 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
397 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
399 #define STATS_INC_ACTIVE(x) do { } while (0)
400 #define STATS_DEC_ACTIVE(x) do { } while (0)
401 #define STATS_INC_ALLOCED(x) do { } while (0)
402 #define STATS_INC_GROWN(x) do { } while (0)
403 #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
404 #define STATS_SET_HIGH(x) do { } while (0)
405 #define STATS_INC_ERR(x) do { } while (0)
406 #define STATS_INC_NODEALLOCS(x) do { } while (0)
407 #define STATS_INC_NODEFREES(x) do { } while (0)
408 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
409 #define STATS_SET_FREEABLE(x, i) do { } while (0)
410 #define STATS_INC_ALLOCHIT(x) do { } while (0)
411 #define STATS_INC_ALLOCMISS(x) do { } while (0)
412 #define STATS_INC_FREEHIT(x) do { } while (0)
413 #define STATS_INC_FREEMISS(x) do { } while (0)
419 * memory layout of objects:
421 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
422 * the end of an object is aligned with the end of the real
423 * allocation. Catches writes behind the end of the allocation.
424 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
426 * cachep->obj_offset: The real object.
427 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
428 * cachep->size - 1* BYTES_PER_WORD: last caller address
429 * [BYTES_PER_WORD long]
431 static int obj_offset(struct kmem_cache *cachep)
433 return cachep->obj_offset;
436 static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
438 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
439 return (unsigned long long*) (objp + obj_offset(cachep) -
440 sizeof(unsigned long long));
443 static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
445 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
446 if (cachep->flags & SLAB_STORE_USER)
447 return (unsigned long long *)(objp + cachep->size -
448 sizeof(unsigned long long) -
450 return (unsigned long long *) (objp + cachep->size -
451 sizeof(unsigned long long));
454 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
456 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
457 return (void **)(objp + cachep->size - BYTES_PER_WORD);
462 #define obj_offset(x) 0
463 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
464 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
465 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
470 * Do not go above this order unless 0 objects fit into the slab or
471 * overridden on the command line.
473 #define SLAB_MAX_ORDER_HI 1
474 #define SLAB_MAX_ORDER_LO 0
475 static int slab_max_order = SLAB_MAX_ORDER_LO;
476 static bool slab_max_order_set __initdata;
478 static inline struct kmem_cache *virt_to_cache(const void *obj)
480 struct page *page = virt_to_head_page(obj);
481 return page->slab_cache;
484 static inline struct slab *virt_to_slab(const void *obj)
486 struct page *page = virt_to_head_page(obj);
488 VM_BUG_ON(!PageSlab(page));
489 return page->slab_page;
492 static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
495 return slab->s_mem + cache->size * idx;
499 * We want to avoid an expensive divide : (offset / cache->size)
500 * Using the fact that size is a constant for a particular cache,
501 * we can replace (offset / cache->size) by
502 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
504 static inline unsigned int obj_to_index(const struct kmem_cache *cache,
505 const struct slab *slab, void *obj)
507 u32 offset = (obj - slab->s_mem);
508 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
511 static struct arraycache_init initarray_generic =
512 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
514 /* internal cache of cache description objs */
515 static struct kmem_cache kmem_cache_boot = {
517 .limit = BOOT_CPUCACHE_ENTRIES,
519 .size = sizeof(struct kmem_cache),
520 .name = "kmem_cache",
523 #define BAD_ALIEN_MAGIC 0x01020304ul
525 #ifdef CONFIG_LOCKDEP
528 * Slab sometimes uses the kmalloc slabs to store the slab headers
529 * for other slabs "off slab".
530 * The locking for this is tricky in that it nests within the locks
531 * of all other slabs in a few places; to deal with this special
532 * locking we put on-slab caches into a separate lock-class.
534 * We set lock class for alien array caches which are up during init.
535 * The lock annotation will be lost if all cpus of a node goes down and
536 * then comes back up during hotplug
538 static struct lock_class_key on_slab_l3_key;
539 static struct lock_class_key on_slab_alc_key;
541 static struct lock_class_key debugobj_l3_key;
542 static struct lock_class_key debugobj_alc_key;
544 static void slab_set_lock_classes(struct kmem_cache *cachep,
545 struct lock_class_key *l3_key, struct lock_class_key *alc_key,
548 struct array_cache **alc;
549 struct kmem_cache_node *l3;
552 l3 = cachep->node[q];
556 lockdep_set_class(&l3->list_lock, l3_key);
559 * FIXME: This check for BAD_ALIEN_MAGIC
560 * should go away when common slab code is taught to
561 * work even without alien caches.
562 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
563 * for alloc_alien_cache,
565 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
569 lockdep_set_class(&alc[r]->lock, alc_key);
573 static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
575 slab_set_lock_classes(cachep, &debugobj_l3_key, &debugobj_alc_key, node);
578 static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
582 for_each_online_node(node)
583 slab_set_debugobj_lock_classes_node(cachep, node);
586 static void init_node_lock_keys(int q)
593 for (i = 1; i < PAGE_SHIFT + MAX_ORDER; i++) {
594 struct kmem_cache_node *l3;
595 struct kmem_cache *cache = kmalloc_caches[i];
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 if (!cachep->node[q])
614 slab_set_lock_classes(cachep, &on_slab_l3_key,
615 &on_slab_alc_key, q);
618 static inline void on_slab_lock_classes(struct kmem_cache *cachep)
622 VM_BUG_ON(OFF_SLAB(cachep));
624 on_slab_lock_classes_node(cachep, node);
627 static inline void init_lock_keys(void)
632 init_node_lock_keys(node);
635 static void init_node_lock_keys(int q)
639 static inline void init_lock_keys(void)
643 static inline void on_slab_lock_classes(struct kmem_cache *cachep)
647 static inline void on_slab_lock_classes_node(struct kmem_cache *cachep, int node)
651 static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
655 static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
660 static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
662 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
664 return cachep->array[smp_processor_id()];
667 static inline struct kmem_cache *__find_general_cachep(size_t size,
673 /* This happens if someone tries to call
674 * kmem_cache_create(), or __kmalloc(), before
675 * the generic caches are initialized.
677 BUG_ON(kmalloc_caches[INDEX_AC] == NULL);
680 return ZERO_SIZE_PTR;
682 i = kmalloc_index(size);
685 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
686 * has cs_{dma,}cachep==NULL. Thus no special case
687 * for large kmalloc calls required.
689 #ifdef CONFIG_ZONE_DMA
690 if (unlikely(gfpflags & GFP_DMA))
691 return kmalloc_dma_caches[i];
693 return kmalloc_caches[i];
696 static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
698 return __find_general_cachep(size, gfpflags);
701 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
703 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
707 * Calculate the number of objects and left-over bytes for a given buffer size.
709 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
710 size_t align, int flags, size_t *left_over,
715 size_t slab_size = PAGE_SIZE << gfporder;
718 * The slab management structure can be either off the slab or
719 * on it. For the latter case, the memory allocated for a
723 * - One kmem_bufctl_t for each object
724 * - Padding to respect alignment of @align
725 * - @buffer_size bytes for each object
727 * If the slab management structure is off the slab, then the
728 * alignment will already be calculated into the size. Because
729 * the slabs are all pages aligned, the objects will be at the
730 * correct alignment when allocated.
732 if (flags & CFLGS_OFF_SLAB) {
734 nr_objs = slab_size / buffer_size;
736 if (nr_objs > SLAB_LIMIT)
737 nr_objs = SLAB_LIMIT;
740 * Ignore padding for the initial guess. The padding
741 * is at most @align-1 bytes, and @buffer_size is at
742 * least @align. In the worst case, this result will
743 * be one greater than the number of objects that fit
744 * into the memory allocation when taking the padding
747 nr_objs = (slab_size - sizeof(struct slab)) /
748 (buffer_size + sizeof(kmem_bufctl_t));
751 * This calculated number will be either the right
752 * amount, or one greater than what we want.
754 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
758 if (nr_objs > SLAB_LIMIT)
759 nr_objs = SLAB_LIMIT;
761 mgmt_size = slab_mgmt_size(nr_objs, align);
764 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
768 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
770 static void __slab_error(const char *function, struct kmem_cache *cachep,
773 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
774 function, cachep->name, msg);
776 add_taint(TAINT_BAD_PAGE);
781 * By default on NUMA we use alien caches to stage the freeing of
782 * objects allocated from other nodes. This causes massive memory
783 * inefficiencies when using fake NUMA setup to split memory into a
784 * large number of small nodes, so it can be disabled on the command
788 static int use_alien_caches __read_mostly = 1;
789 static int __init noaliencache_setup(char *s)
791 use_alien_caches = 0;
794 __setup("noaliencache", noaliencache_setup);
796 static int __init slab_max_order_setup(char *str)
798 get_option(&str, &slab_max_order);
799 slab_max_order = slab_max_order < 0 ? 0 :
800 min(slab_max_order, MAX_ORDER - 1);
801 slab_max_order_set = true;
805 __setup("slab_max_order=", slab_max_order_setup);
809 * Special reaping functions for NUMA systems called from cache_reap().
810 * These take care of doing round robin flushing of alien caches (containing
811 * objects freed on different nodes from which they were allocated) and the
812 * flushing of remote pcps by calling drain_node_pages.
814 static DEFINE_PER_CPU(unsigned long, slab_reap_node);
816 static void init_reap_node(int cpu)
820 node = next_node(cpu_to_mem(cpu), node_online_map);
821 if (node == MAX_NUMNODES)
822 node = first_node(node_online_map);
824 per_cpu(slab_reap_node, cpu) = node;
827 static void next_reap_node(void)
829 int node = __this_cpu_read(slab_reap_node);
831 node = next_node(node, node_online_map);
832 if (unlikely(node >= MAX_NUMNODES))
833 node = first_node(node_online_map);
834 __this_cpu_write(slab_reap_node, node);
838 #define init_reap_node(cpu) do { } while (0)
839 #define next_reap_node(void) do { } while (0)
843 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
844 * via the workqueue/eventd.
845 * Add the CPU number into the expiration time to minimize the possibility of
846 * the CPUs getting into lockstep and contending for the global cache chain
849 static void __cpuinit start_cpu_timer(int cpu)
851 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
854 * When this gets called from do_initcalls via cpucache_init(),
855 * init_workqueues() has already run, so keventd will be setup
858 if (keventd_up() && reap_work->work.func == NULL) {
860 INIT_DEFERRABLE_WORK(reap_work, cache_reap);
861 schedule_delayed_work_on(cpu, reap_work,
862 __round_jiffies_relative(HZ, cpu));
866 static struct array_cache *alloc_arraycache(int node, int entries,
867 int batchcount, gfp_t gfp)
869 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
870 struct array_cache *nc = NULL;
872 nc = kmalloc_node(memsize, gfp, node);
874 * The array_cache structures contain pointers to free object.
875 * However, when such objects are allocated or transferred to another
876 * cache the pointers are not cleared and they could be counted as
877 * valid references during a kmemleak scan. Therefore, kmemleak must
878 * not scan such objects.
880 kmemleak_no_scan(nc);
884 nc->batchcount = batchcount;
886 spin_lock_init(&nc->lock);
891 static inline bool is_slab_pfmemalloc(struct slab *slabp)
893 struct page *page = virt_to_page(slabp->s_mem);
895 return PageSlabPfmemalloc(page);
898 /* Clears pfmemalloc_active if no slabs have pfmalloc set */
899 static void recheck_pfmemalloc_active(struct kmem_cache *cachep,
900 struct array_cache *ac)
902 struct kmem_cache_node *l3 = cachep->node[numa_mem_id()];
906 if (!pfmemalloc_active)
909 spin_lock_irqsave(&l3->list_lock, flags);
910 list_for_each_entry(slabp, &l3->slabs_full, list)
911 if (is_slab_pfmemalloc(slabp))
914 list_for_each_entry(slabp, &l3->slabs_partial, list)
915 if (is_slab_pfmemalloc(slabp))
918 list_for_each_entry(slabp, &l3->slabs_free, list)
919 if (is_slab_pfmemalloc(slabp))
922 pfmemalloc_active = false;
924 spin_unlock_irqrestore(&l3->list_lock, flags);
927 static void *__ac_get_obj(struct kmem_cache *cachep, struct array_cache *ac,
928 gfp_t flags, bool force_refill)
931 void *objp = ac->entry[--ac->avail];
933 /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
934 if (unlikely(is_obj_pfmemalloc(objp))) {
935 struct kmem_cache_node *l3;
937 if (gfp_pfmemalloc_allowed(flags)) {
938 clear_obj_pfmemalloc(&objp);
942 /* The caller cannot use PFMEMALLOC objects, find another one */
943 for (i = 0; i < ac->avail; i++) {
944 /* If a !PFMEMALLOC object is found, swap them */
945 if (!is_obj_pfmemalloc(ac->entry[i])) {
947 ac->entry[i] = ac->entry[ac->avail];
948 ac->entry[ac->avail] = objp;
954 * If there are empty slabs on the slabs_free list and we are
955 * being forced to refill the cache, mark this one !pfmemalloc.
957 l3 = cachep->node[numa_mem_id()];
958 if (!list_empty(&l3->slabs_free) && force_refill) {
959 struct slab *slabp = virt_to_slab(objp);
960 ClearPageSlabPfmemalloc(virt_to_head_page(slabp->s_mem));
961 clear_obj_pfmemalloc(&objp);
962 recheck_pfmemalloc_active(cachep, ac);
966 /* No !PFMEMALLOC objects available */
974 static inline void *ac_get_obj(struct kmem_cache *cachep,
975 struct array_cache *ac, gfp_t flags, bool force_refill)
979 if (unlikely(sk_memalloc_socks()))
980 objp = __ac_get_obj(cachep, ac, flags, force_refill);
982 objp = ac->entry[--ac->avail];
987 static void *__ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
990 if (unlikely(pfmemalloc_active)) {
991 /* Some pfmemalloc slabs exist, check if this is one */
992 struct page *page = virt_to_head_page(objp);
993 if (PageSlabPfmemalloc(page))
994 set_obj_pfmemalloc(&objp);
1000 static inline void ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
1003 if (unlikely(sk_memalloc_socks()))
1004 objp = __ac_put_obj(cachep, ac, objp);
1006 ac->entry[ac->avail++] = objp;
1010 * Transfer objects in one arraycache to another.
1011 * Locking must be handled by the caller.
1013 * Return the number of entries transferred.
1015 static int transfer_objects(struct array_cache *to,
1016 struct array_cache *from, unsigned int max)
1018 /* Figure out how many entries to transfer */
1019 int nr = min3(from->avail, max, to->limit - to->avail);
1024 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
1025 sizeof(void *) *nr);
1034 #define drain_alien_cache(cachep, alien) do { } while (0)
1035 #define reap_alien(cachep, l3) do { } while (0)
1037 static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
1039 return (struct array_cache **)BAD_ALIEN_MAGIC;
1042 static inline void free_alien_cache(struct array_cache **ac_ptr)
1046 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1051 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
1057 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
1058 gfp_t flags, int nodeid)
1063 #else /* CONFIG_NUMA */
1065 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
1066 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
1068 static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
1070 struct array_cache **ac_ptr;
1071 int memsize = sizeof(void *) * nr_node_ids;
1076 ac_ptr = kzalloc_node(memsize, gfp, node);
1079 if (i == node || !node_online(i))
1081 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
1083 for (i--; i >= 0; i--)
1093 static void free_alien_cache(struct array_cache **ac_ptr)
1104 static void __drain_alien_cache(struct kmem_cache *cachep,
1105 struct array_cache *ac, int node)
1107 struct kmem_cache_node *rl3 = cachep->node[node];
1110 spin_lock(&rl3->list_lock);
1112 * Stuff objects into the remote nodes shared array first.
1113 * That way we could avoid the overhead of putting the objects
1114 * into the free lists and getting them back later.
1117 transfer_objects(rl3->shared, ac, ac->limit);
1119 free_block(cachep, ac->entry, ac->avail, node);
1121 spin_unlock(&rl3->list_lock);
1126 * Called from cache_reap() to regularly drain alien caches round robin.
1128 static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *l3)
1130 int node = __this_cpu_read(slab_reap_node);
1133 struct array_cache *ac = l3->alien[node];
1135 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1136 __drain_alien_cache(cachep, ac, node);
1137 spin_unlock_irq(&ac->lock);
1142 static void drain_alien_cache(struct kmem_cache *cachep,
1143 struct array_cache **alien)
1146 struct array_cache *ac;
1147 unsigned long flags;
1149 for_each_online_node(i) {
1152 spin_lock_irqsave(&ac->lock, flags);
1153 __drain_alien_cache(cachep, ac, i);
1154 spin_unlock_irqrestore(&ac->lock, flags);
1159 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1161 struct slab *slabp = virt_to_slab(objp);
1162 int nodeid = slabp->nodeid;
1163 struct kmem_cache_node *l3;
1164 struct array_cache *alien = NULL;
1167 node = numa_mem_id();
1170 * Make sure we are not freeing a object from another node to the array
1171 * cache on this cpu.
1173 if (likely(slabp->nodeid == node))
1176 l3 = cachep->node[node];
1177 STATS_INC_NODEFREES(cachep);
1178 if (l3->alien && l3->alien[nodeid]) {
1179 alien = l3->alien[nodeid];
1180 spin_lock(&alien->lock);
1181 if (unlikely(alien->avail == alien->limit)) {
1182 STATS_INC_ACOVERFLOW(cachep);
1183 __drain_alien_cache(cachep, alien, nodeid);
1185 ac_put_obj(cachep, alien, objp);
1186 spin_unlock(&alien->lock);
1188 spin_lock(&(cachep->node[nodeid])->list_lock);
1189 free_block(cachep, &objp, 1, nodeid);
1190 spin_unlock(&(cachep->node[nodeid])->list_lock);
1197 * Allocates and initializes node for a node on each slab cache, used for
1198 * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3
1199 * will be allocated off-node since memory is not yet online for the new node.
1200 * When hotplugging memory or a cpu, existing node are not replaced if
1203 * Must hold slab_mutex.
1205 static int init_cache_node_node(int node)
1207 struct kmem_cache *cachep;
1208 struct kmem_cache_node *l3;
1209 const int memsize = sizeof(struct kmem_cache_node);
1211 list_for_each_entry(cachep, &slab_caches, list) {
1213 * Set up the size64 kmemlist for cpu before we can
1214 * begin anything. Make sure some other cpu on this
1215 * node has not already allocated this
1217 if (!cachep->node[node]) {
1218 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1221 kmem_list3_init(l3);
1222 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1223 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1226 * The l3s don't come and go as CPUs come and
1227 * go. slab_mutex is sufficient
1230 cachep->node[node] = l3;
1233 spin_lock_irq(&cachep->node[node]->list_lock);
1234 cachep->node[node]->free_limit =
1235 (1 + nr_cpus_node(node)) *
1236 cachep->batchcount + cachep->num;
1237 spin_unlock_irq(&cachep->node[node]->list_lock);
1242 static void __cpuinit cpuup_canceled(long cpu)
1244 struct kmem_cache *cachep;
1245 struct kmem_cache_node *l3 = NULL;
1246 int node = cpu_to_mem(cpu);
1247 const struct cpumask *mask = cpumask_of_node(node);
1249 list_for_each_entry(cachep, &slab_caches, list) {
1250 struct array_cache *nc;
1251 struct array_cache *shared;
1252 struct array_cache **alien;
1254 /* cpu is dead; no one can alloc from it. */
1255 nc = cachep->array[cpu];
1256 cachep->array[cpu] = NULL;
1257 l3 = cachep->node[node];
1260 goto free_array_cache;
1262 spin_lock_irq(&l3->list_lock);
1264 /* Free limit for this kmem_list3 */
1265 l3->free_limit -= cachep->batchcount;
1267 free_block(cachep, nc->entry, nc->avail, node);
1269 if (!cpumask_empty(mask)) {
1270 spin_unlock_irq(&l3->list_lock);
1271 goto free_array_cache;
1274 shared = l3->shared;
1276 free_block(cachep, shared->entry,
1277 shared->avail, node);
1284 spin_unlock_irq(&l3->list_lock);
1288 drain_alien_cache(cachep, alien);
1289 free_alien_cache(alien);
1295 * In the previous loop, all the objects were freed to
1296 * the respective cache's slabs, now we can go ahead and
1297 * shrink each nodelist to its limit.
1299 list_for_each_entry(cachep, &slab_caches, list) {
1300 l3 = cachep->node[node];
1303 drain_freelist(cachep, l3, l3->free_objects);
1307 static int __cpuinit cpuup_prepare(long cpu)
1309 struct kmem_cache *cachep;
1310 struct kmem_cache_node *l3 = NULL;
1311 int node = cpu_to_mem(cpu);
1315 * We need to do this right in the beginning since
1316 * alloc_arraycache's are going to use this list.
1317 * kmalloc_node allows us to add the slab to the right
1318 * kmem_list3 and not this cpu's kmem_list3
1320 err = init_cache_node_node(node);
1325 * Now we can go ahead with allocating the shared arrays and
1328 list_for_each_entry(cachep, &slab_caches, list) {
1329 struct array_cache *nc;
1330 struct array_cache *shared = NULL;
1331 struct array_cache **alien = NULL;
1333 nc = alloc_arraycache(node, cachep->limit,
1334 cachep->batchcount, GFP_KERNEL);
1337 if (cachep->shared) {
1338 shared = alloc_arraycache(node,
1339 cachep->shared * cachep->batchcount,
1340 0xbaadf00d, GFP_KERNEL);
1346 if (use_alien_caches) {
1347 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
1354 cachep->array[cpu] = nc;
1355 l3 = cachep->node[node];
1358 spin_lock_irq(&l3->list_lock);
1361 * We are serialised from CPU_DEAD or
1362 * CPU_UP_CANCELLED by the cpucontrol lock
1364 l3->shared = shared;
1373 spin_unlock_irq(&l3->list_lock);
1375 free_alien_cache(alien);
1376 if (cachep->flags & SLAB_DEBUG_OBJECTS)
1377 slab_set_debugobj_lock_classes_node(cachep, node);
1378 else if (!OFF_SLAB(cachep) &&
1379 !(cachep->flags & SLAB_DESTROY_BY_RCU))
1380 on_slab_lock_classes_node(cachep, node);
1382 init_node_lock_keys(node);
1386 cpuup_canceled(cpu);
1390 static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1391 unsigned long action, void *hcpu)
1393 long cpu = (long)hcpu;
1397 case CPU_UP_PREPARE:
1398 case CPU_UP_PREPARE_FROZEN:
1399 mutex_lock(&slab_mutex);
1400 err = cpuup_prepare(cpu);
1401 mutex_unlock(&slab_mutex);
1404 case CPU_ONLINE_FROZEN:
1405 start_cpu_timer(cpu);
1407 #ifdef CONFIG_HOTPLUG_CPU
1408 case CPU_DOWN_PREPARE:
1409 case CPU_DOWN_PREPARE_FROZEN:
1411 * Shutdown cache reaper. Note that the slab_mutex is
1412 * held so that if cache_reap() is invoked it cannot do
1413 * anything expensive but will only modify reap_work
1414 * and reschedule the timer.
1416 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1417 /* Now the cache_reaper is guaranteed to be not running. */
1418 per_cpu(slab_reap_work, cpu).work.func = NULL;
1420 case CPU_DOWN_FAILED:
1421 case CPU_DOWN_FAILED_FROZEN:
1422 start_cpu_timer(cpu);
1425 case CPU_DEAD_FROZEN:
1427 * Even if all the cpus of a node are down, we don't free the
1428 * kmem_list3 of any cache. This to avoid a race between
1429 * cpu_down, and a kmalloc allocation from another cpu for
1430 * memory from the node of the cpu going down. The list3
1431 * structure is usually allocated from kmem_cache_create() and
1432 * gets destroyed at kmem_cache_destroy().
1436 case CPU_UP_CANCELED:
1437 case CPU_UP_CANCELED_FROZEN:
1438 mutex_lock(&slab_mutex);
1439 cpuup_canceled(cpu);
1440 mutex_unlock(&slab_mutex);
1443 return notifier_from_errno(err);
1446 static struct notifier_block __cpuinitdata cpucache_notifier = {
1447 &cpuup_callback, NULL, 0
1450 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1452 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1453 * Returns -EBUSY if all objects cannot be drained so that the node is not
1456 * Must hold slab_mutex.
1458 static int __meminit drain_cache_node_node(int node)
1460 struct kmem_cache *cachep;
1463 list_for_each_entry(cachep, &slab_caches, list) {
1464 struct kmem_cache_node *l3;
1466 l3 = cachep->node[node];
1470 drain_freelist(cachep, l3, l3->free_objects);
1472 if (!list_empty(&l3->slabs_full) ||
1473 !list_empty(&l3->slabs_partial)) {
1481 static int __meminit slab_memory_callback(struct notifier_block *self,
1482 unsigned long action, void *arg)
1484 struct memory_notify *mnb = arg;
1488 nid = mnb->status_change_nid;
1493 case MEM_GOING_ONLINE:
1494 mutex_lock(&slab_mutex);
1495 ret = init_cache_node_node(nid);
1496 mutex_unlock(&slab_mutex);
1498 case MEM_GOING_OFFLINE:
1499 mutex_lock(&slab_mutex);
1500 ret = drain_cache_node_node(nid);
1501 mutex_unlock(&slab_mutex);
1505 case MEM_CANCEL_ONLINE:
1506 case MEM_CANCEL_OFFLINE:
1510 return notifier_from_errno(ret);
1512 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1515 * swap the static kmem_list3 with kmalloced memory
1517 static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list,
1520 struct kmem_cache_node *ptr;
1522 ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid);
1525 memcpy(ptr, list, sizeof(struct kmem_cache_node));
1527 * Do not assume that spinlocks can be initialized via memcpy:
1529 spin_lock_init(&ptr->list_lock);
1531 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1532 cachep->node[nodeid] = ptr;
1536 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1537 * size of kmem_list3.
1539 static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1543 for_each_online_node(node) {
1544 cachep->node[node] = &initkmem_list3[index + node];
1545 cachep->node[node]->next_reap = jiffies +
1547 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1552 * The memory after the last cpu cache pointer is used for the
1555 static void setup_node_pointer(struct kmem_cache *cachep)
1557 cachep->node = (struct kmem_cache_node **)&cachep->array[nr_cpu_ids];
1561 * Initialisation. Called after the page allocator have been initialised and
1562 * before smp_init().
1564 void __init kmem_cache_init(void)
1568 kmem_cache = &kmem_cache_boot;
1569 setup_node_pointer(kmem_cache);
1571 if (num_possible_nodes() == 1)
1572 use_alien_caches = 0;
1574 for (i = 0; i < NUM_INIT_LISTS; i++)
1575 kmem_list3_init(&initkmem_list3[i]);
1577 set_up_list3s(kmem_cache, CACHE_CACHE);
1580 * Fragmentation resistance on low memory - only use bigger
1581 * page orders on machines with more than 32MB of memory if
1582 * not overridden on the command line.
1584 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
1585 slab_max_order = SLAB_MAX_ORDER_HI;
1587 /* Bootstrap is tricky, because several objects are allocated
1588 * from caches that do not exist yet:
1589 * 1) initialize the kmem_cache cache: it contains the struct
1590 * kmem_cache structures of all caches, except kmem_cache itself:
1591 * kmem_cache is statically allocated.
1592 * Initially an __init data area is used for the head array and the
1593 * kmem_list3 structures, it's replaced with a kmalloc allocated
1594 * array at the end of the bootstrap.
1595 * 2) Create the first kmalloc cache.
1596 * The struct kmem_cache for the new cache is allocated normally.
1597 * An __init data area is used for the head array.
1598 * 3) Create the remaining kmalloc caches, with minimally sized
1600 * 4) Replace the __init data head arrays for kmem_cache and the first
1601 * kmalloc cache with kmalloc allocated arrays.
1602 * 5) Replace the __init data for kmem_list3 for kmem_cache and
1603 * the other cache's with kmalloc allocated memory.
1604 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1607 /* 1) create the kmem_cache */
1610 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1612 create_boot_cache(kmem_cache, "kmem_cache",
1613 offsetof(struct kmem_cache, array[nr_cpu_ids]) +
1614 nr_node_ids * sizeof(struct kmem_cache_node *),
1615 SLAB_HWCACHE_ALIGN);
1616 list_add(&kmem_cache->list, &slab_caches);
1618 /* 2+3) create the kmalloc caches */
1621 * Initialize the caches that provide memory for the array cache and the
1622 * kmem_list3 structures first. Without this, further allocations will
1626 kmalloc_caches[INDEX_AC] = create_kmalloc_cache("kmalloc-ac",
1627 kmalloc_size(INDEX_AC), ARCH_KMALLOC_FLAGS);
1629 if (INDEX_AC != INDEX_L3)
1630 kmalloc_caches[INDEX_L3] =
1631 create_kmalloc_cache("kmalloc-l3",
1632 kmalloc_size(INDEX_L3), ARCH_KMALLOC_FLAGS);
1634 slab_early_init = 0;
1636 for (i = 1; i < PAGE_SHIFT + MAX_ORDER; i++) {
1637 size_t cs_size = kmalloc_size(i);
1639 if (cs_size < KMALLOC_MIN_SIZE)
1642 if (!kmalloc_caches[i]) {
1644 * For performance, all the general caches are L1 aligned.
1645 * This should be particularly beneficial on SMP boxes, as it
1646 * eliminates "false sharing".
1647 * Note for systems short on memory removing the alignment will
1648 * allow tighter packing of the smaller caches.
1650 kmalloc_caches[i] = create_kmalloc_cache("kmalloc",
1651 cs_size, ARCH_KMALLOC_FLAGS);
1654 #ifdef CONFIG_ZONE_DMA
1655 kmalloc_dma_caches[i] = create_kmalloc_cache(
1656 "kmalloc-dma", cs_size,
1657 SLAB_CACHE_DMA|ARCH_KMALLOC_FLAGS);
1660 /* 4) Replace the bootstrap head arrays */
1662 struct array_cache *ptr;
1664 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1666 memcpy(ptr, cpu_cache_get(kmem_cache),
1667 sizeof(struct arraycache_init));
1669 * Do not assume that spinlocks can be initialized via memcpy:
1671 spin_lock_init(&ptr->lock);
1673 kmem_cache->array[smp_processor_id()] = ptr;
1675 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1677 BUG_ON(cpu_cache_get(kmalloc_caches[INDEX_AC])
1678 != &initarray_generic.cache);
1679 memcpy(ptr, cpu_cache_get(kmalloc_caches[INDEX_AC]),
1680 sizeof(struct arraycache_init));
1682 * Do not assume that spinlocks can be initialized via memcpy:
1684 spin_lock_init(&ptr->lock);
1686 kmalloc_caches[INDEX_AC]->array[smp_processor_id()] = ptr;
1688 /* 5) Replace the bootstrap kmem_list3's */
1692 for_each_online_node(nid) {
1693 init_list(kmem_cache, &initkmem_list3[CACHE_CACHE + nid], nid);
1695 init_list(kmalloc_caches[INDEX_AC],
1696 &initkmem_list3[SIZE_AC + nid], nid);
1698 if (INDEX_AC != INDEX_L3) {
1699 init_list(kmalloc_caches[INDEX_L3],
1700 &initkmem_list3[SIZE_L3 + nid], nid);
1707 /* Create the proper names */
1708 for (i = 1; i < PAGE_SHIFT + MAX_ORDER; i++) {
1710 struct kmem_cache *c = kmalloc_caches[i];
1715 s = kasprintf(GFP_NOWAIT, "kmalloc-%d", kmalloc_size(i));
1720 #ifdef CONFIG_ZONE_DMA
1721 c = kmalloc_dma_caches[i];
1723 s = kasprintf(GFP_NOWAIT, "dma-kmalloc-%d", kmalloc_size(i));
1730 void __init kmem_cache_init_late(void)
1732 struct kmem_cache *cachep;
1736 /* 6) resize the head arrays to their final sizes */
1737 mutex_lock(&slab_mutex);
1738 list_for_each_entry(cachep, &slab_caches, list)
1739 if (enable_cpucache(cachep, GFP_NOWAIT))
1741 mutex_unlock(&slab_mutex);
1743 /* Annotate slab for lockdep -- annotate the malloc caches */
1750 * Register a cpu startup notifier callback that initializes
1751 * cpu_cache_get for all new cpus
1753 register_cpu_notifier(&cpucache_notifier);
1757 * Register a memory hotplug callback that initializes and frees
1760 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1764 * The reap timers are started later, with a module init call: That part
1765 * of the kernel is not yet operational.
1769 static int __init cpucache_init(void)
1774 * Register the timers that return unneeded pages to the page allocator
1776 for_each_online_cpu(cpu)
1777 start_cpu_timer(cpu);
1783 __initcall(cpucache_init);
1785 static noinline void
1786 slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1788 struct kmem_cache_node *l3;
1790 unsigned long flags;
1794 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1796 printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n",
1797 cachep->name, cachep->size, cachep->gfporder);
1799 for_each_online_node(node) {
1800 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1801 unsigned long active_slabs = 0, num_slabs = 0;
1803 l3 = cachep->node[node];
1807 spin_lock_irqsave(&l3->list_lock, flags);
1808 list_for_each_entry(slabp, &l3->slabs_full, list) {
1809 active_objs += cachep->num;
1812 list_for_each_entry(slabp, &l3->slabs_partial, list) {
1813 active_objs += slabp->inuse;
1816 list_for_each_entry(slabp, &l3->slabs_free, list)
1819 free_objects += l3->free_objects;
1820 spin_unlock_irqrestore(&l3->list_lock, flags);
1822 num_slabs += active_slabs;
1823 num_objs = num_slabs * cachep->num;
1825 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1826 node, active_slabs, num_slabs, active_objs, num_objs,
1832 * Interface to system's page allocator. No need to hold the cache-lock.
1834 * If we requested dmaable memory, we will get it. Even if we
1835 * did not request dmaable memory, we might get it, but that
1836 * would be relatively rare and ignorable.
1838 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1846 * Nommu uses slab's for process anonymous memory allocations, and thus
1847 * requires __GFP_COMP to properly refcount higher order allocations
1849 flags |= __GFP_COMP;
1852 flags |= cachep->allocflags;
1853 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1854 flags |= __GFP_RECLAIMABLE;
1856 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1858 if (!(flags & __GFP_NOWARN) && printk_ratelimit())
1859 slab_out_of_memory(cachep, flags, nodeid);
1863 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1864 if (unlikely(page->pfmemalloc))
1865 pfmemalloc_active = true;
1867 nr_pages = (1 << cachep->gfporder);
1868 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1869 add_zone_page_state(page_zone(page),
1870 NR_SLAB_RECLAIMABLE, nr_pages);
1872 add_zone_page_state(page_zone(page),
1873 NR_SLAB_UNRECLAIMABLE, nr_pages);
1874 for (i = 0; i < nr_pages; i++) {
1875 __SetPageSlab(page + i);
1877 if (page->pfmemalloc)
1878 SetPageSlabPfmemalloc(page + i);
1880 memcg_bind_pages(cachep, cachep->gfporder);
1882 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1883 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1886 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1888 kmemcheck_mark_unallocated_pages(page, nr_pages);
1891 return page_address(page);
1895 * Interface to system's page release.
1897 static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1899 unsigned long i = (1 << cachep->gfporder);
1900 struct page *page = virt_to_page(addr);
1901 const unsigned long nr_freed = i;
1903 kmemcheck_free_shadow(page, cachep->gfporder);
1905 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1906 sub_zone_page_state(page_zone(page),
1907 NR_SLAB_RECLAIMABLE, nr_freed);
1909 sub_zone_page_state(page_zone(page),
1910 NR_SLAB_UNRECLAIMABLE, nr_freed);
1912 BUG_ON(!PageSlab(page));
1913 __ClearPageSlabPfmemalloc(page);
1914 __ClearPageSlab(page);
1918 memcg_release_pages(cachep, cachep->gfporder);
1919 if (current->reclaim_state)
1920 current->reclaim_state->reclaimed_slab += nr_freed;
1921 free_memcg_kmem_pages((unsigned long)addr, cachep->gfporder);
1924 static void kmem_rcu_free(struct rcu_head *head)
1926 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1927 struct kmem_cache *cachep = slab_rcu->cachep;
1929 kmem_freepages(cachep, slab_rcu->addr);
1930 if (OFF_SLAB(cachep))
1931 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1936 #ifdef CONFIG_DEBUG_PAGEALLOC
1937 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1938 unsigned long caller)
1940 int size = cachep->object_size;
1942 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1944 if (size < 5 * sizeof(unsigned long))
1947 *addr++ = 0x12345678;
1949 *addr++ = smp_processor_id();
1950 size -= 3 * sizeof(unsigned long);
1952 unsigned long *sptr = &caller;
1953 unsigned long svalue;
1955 while (!kstack_end(sptr)) {
1957 if (kernel_text_address(svalue)) {
1959 size -= sizeof(unsigned long);
1960 if (size <= sizeof(unsigned long))
1966 *addr++ = 0x87654321;
1970 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1972 int size = cachep->object_size;
1973 addr = &((char *)addr)[obj_offset(cachep)];
1975 memset(addr, val, size);
1976 *(unsigned char *)(addr + size - 1) = POISON_END;
1979 static void dump_line(char *data, int offset, int limit)
1982 unsigned char error = 0;
1985 printk(KERN_ERR "%03x: ", offset);
1986 for (i = 0; i < limit; i++) {
1987 if (data[offset + i] != POISON_FREE) {
1988 error = data[offset + i];
1992 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1993 &data[offset], limit, 1);
1995 if (bad_count == 1) {
1996 error ^= POISON_FREE;
1997 if (!(error & (error - 1))) {
1998 printk(KERN_ERR "Single bit error detected. Probably "
2001 printk(KERN_ERR "Run memtest86+ or a similar memory "
2004 printk(KERN_ERR "Run a memory test tool.\n");
2013 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
2018 if (cachep->flags & SLAB_RED_ZONE) {
2019 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
2020 *dbg_redzone1(cachep, objp),
2021 *dbg_redzone2(cachep, objp));
2024 if (cachep->flags & SLAB_STORE_USER) {
2025 printk(KERN_ERR "Last user: [<%p>]",
2026 *dbg_userword(cachep, objp));
2027 print_symbol("(%s)",
2028 (unsigned long)*dbg_userword(cachep, objp));
2031 realobj = (char *)objp + obj_offset(cachep);
2032 size = cachep->object_size;
2033 for (i = 0; i < size && lines; i += 16, lines--) {
2036 if (i + limit > size)
2038 dump_line(realobj, i, limit);
2042 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
2048 realobj = (char *)objp + obj_offset(cachep);
2049 size = cachep->object_size;
2051 for (i = 0; i < size; i++) {
2052 char exp = POISON_FREE;
2055 if (realobj[i] != exp) {
2061 "Slab corruption (%s): %s start=%p, len=%d\n",
2062 print_tainted(), cachep->name, realobj, size);
2063 print_objinfo(cachep, objp, 0);
2065 /* Hexdump the affected line */
2068 if (i + limit > size)
2070 dump_line(realobj, i, limit);
2073 /* Limit to 5 lines */
2079 /* Print some data about the neighboring objects, if they
2082 struct slab *slabp = virt_to_slab(objp);
2085 objnr = obj_to_index(cachep, slabp, objp);
2087 objp = index_to_obj(cachep, slabp, objnr - 1);
2088 realobj = (char *)objp + obj_offset(cachep);
2089 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
2091 print_objinfo(cachep, objp, 2);
2093 if (objnr + 1 < cachep->num) {
2094 objp = index_to_obj(cachep, slabp, objnr + 1);
2095 realobj = (char *)objp + obj_offset(cachep);
2096 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
2098 print_objinfo(cachep, objp, 2);
2105 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
2108 for (i = 0; i < cachep->num; i++) {
2109 void *objp = index_to_obj(cachep, slabp, i);
2111 if (cachep->flags & SLAB_POISON) {
2112 #ifdef CONFIG_DEBUG_PAGEALLOC
2113 if (cachep->size % PAGE_SIZE == 0 &&
2115 kernel_map_pages(virt_to_page(objp),
2116 cachep->size / PAGE_SIZE, 1);
2118 check_poison_obj(cachep, objp);
2120 check_poison_obj(cachep, objp);
2123 if (cachep->flags & SLAB_RED_ZONE) {
2124 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2125 slab_error(cachep, "start of a freed object "
2127 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2128 slab_error(cachep, "end of a freed object "
2134 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
2140 * slab_destroy - destroy and release all objects in a slab
2141 * @cachep: cache pointer being destroyed
2142 * @slabp: slab pointer being destroyed
2144 * Destroy all the objs in a slab, and release the mem back to the system.
2145 * Before calling the slab must have been unlinked from the cache. The
2146 * cache-lock is not held/needed.
2148 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
2150 void *addr = slabp->s_mem - slabp->colouroff;
2152 slab_destroy_debugcheck(cachep, slabp);
2153 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
2154 struct slab_rcu *slab_rcu;
2156 slab_rcu = (struct slab_rcu *)slabp;
2157 slab_rcu->cachep = cachep;
2158 slab_rcu->addr = addr;
2159 call_rcu(&slab_rcu->head, kmem_rcu_free);
2161 kmem_freepages(cachep, addr);
2162 if (OFF_SLAB(cachep))
2163 kmem_cache_free(cachep->slabp_cache, slabp);
2168 * calculate_slab_order - calculate size (page order) of slabs
2169 * @cachep: pointer to the cache that is being created
2170 * @size: size of objects to be created in this cache.
2171 * @align: required alignment for the objects.
2172 * @flags: slab allocation flags
2174 * Also calculates the number of objects per slab.
2176 * This could be made much more intelligent. For now, try to avoid using
2177 * high order pages for slabs. When the gfp() functions are more friendly
2178 * towards high-order requests, this should be changed.
2180 static size_t calculate_slab_order(struct kmem_cache *cachep,
2181 size_t size, size_t align, unsigned long flags)
2183 unsigned long offslab_limit;
2184 size_t left_over = 0;
2187 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
2191 cache_estimate(gfporder, size, align, flags, &remainder, &num);
2195 if (flags & CFLGS_OFF_SLAB) {
2197 * Max number of objs-per-slab for caches which
2198 * use off-slab slabs. Needed to avoid a possible
2199 * looping condition in cache_grow().
2201 offslab_limit = size - sizeof(struct slab);
2202 offslab_limit /= sizeof(kmem_bufctl_t);
2204 if (num > offslab_limit)
2208 /* Found something acceptable - save it away */
2210 cachep->gfporder = gfporder;
2211 left_over = remainder;
2214 * A VFS-reclaimable slab tends to have most allocations
2215 * as GFP_NOFS and we really don't want to have to be allocating
2216 * higher-order pages when we are unable to shrink dcache.
2218 if (flags & SLAB_RECLAIM_ACCOUNT)
2222 * Large number of objects is good, but very large slabs are
2223 * currently bad for the gfp()s.
2225 if (gfporder >= slab_max_order)
2229 * Acceptable internal fragmentation?
2231 if (left_over * 8 <= (PAGE_SIZE << gfporder))
2237 static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
2239 if (slab_state >= FULL)
2240 return enable_cpucache(cachep, gfp);
2242 if (slab_state == DOWN) {
2244 * Note: Creation of first cache (kmem_cache).
2245 * The setup_list3s is taken care
2246 * of by the caller of __kmem_cache_create
2248 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2249 slab_state = PARTIAL;
2250 } else if (slab_state == PARTIAL) {
2252 * Note: the second kmem_cache_create must create the cache
2253 * that's used by kmalloc(24), otherwise the creation of
2254 * further caches will BUG().
2256 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2259 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2260 * the second cache, then we need to set up all its list3s,
2261 * otherwise the creation of further caches will BUG().
2263 set_up_list3s(cachep, SIZE_AC);
2264 if (INDEX_AC == INDEX_L3)
2265 slab_state = PARTIAL_L3;
2267 slab_state = PARTIAL_ARRAYCACHE;
2269 /* Remaining boot caches */
2270 cachep->array[smp_processor_id()] =
2271 kmalloc(sizeof(struct arraycache_init), gfp);
2273 if (slab_state == PARTIAL_ARRAYCACHE) {
2274 set_up_list3s(cachep, SIZE_L3);
2275 slab_state = PARTIAL_L3;
2278 for_each_online_node(node) {
2279 cachep->node[node] =
2280 kmalloc_node(sizeof(struct kmem_cache_node),
2282 BUG_ON(!cachep->node[node]);
2283 kmem_list3_init(cachep->node[node]);
2287 cachep->node[numa_mem_id()]->next_reap =
2288 jiffies + REAPTIMEOUT_LIST3 +
2289 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2291 cpu_cache_get(cachep)->avail = 0;
2292 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2293 cpu_cache_get(cachep)->batchcount = 1;
2294 cpu_cache_get(cachep)->touched = 0;
2295 cachep->batchcount = 1;
2296 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2301 * __kmem_cache_create - Create a cache.
2302 * @cachep: cache management descriptor
2303 * @flags: SLAB flags
2305 * Returns a ptr to the cache on success, NULL on failure.
2306 * Cannot be called within a int, but can be interrupted.
2307 * The @ctor is run when new pages are allocated by the cache.
2311 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2312 * to catch references to uninitialised memory.
2314 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2315 * for buffer overruns.
2317 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2318 * cacheline. This can be beneficial if you're counting cycles as closely
2322 __kmem_cache_create (struct kmem_cache *cachep, unsigned long flags)
2324 size_t left_over, slab_size, ralign;
2327 size_t size = cachep->size;
2332 * Enable redzoning and last user accounting, except for caches with
2333 * large objects, if the increased size would increase the object size
2334 * above the next power of two: caches with object sizes just above a
2335 * power of two have a significant amount of internal fragmentation.
2337 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2338 2 * sizeof(unsigned long long)))
2339 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2340 if (!(flags & SLAB_DESTROY_BY_RCU))
2341 flags |= SLAB_POISON;
2343 if (flags & SLAB_DESTROY_BY_RCU)
2344 BUG_ON(flags & SLAB_POISON);
2348 * Check that size is in terms of words. This is needed to avoid
2349 * unaligned accesses for some archs when redzoning is used, and makes
2350 * sure any on-slab bufctl's are also correctly aligned.
2352 if (size & (BYTES_PER_WORD - 1)) {
2353 size += (BYTES_PER_WORD - 1);
2354 size &= ~(BYTES_PER_WORD - 1);
2358 * Redzoning and user store require word alignment or possibly larger.
2359 * Note this will be overridden by architecture or caller mandated
2360 * alignment if either is greater than BYTES_PER_WORD.
2362 if (flags & SLAB_STORE_USER)
2363 ralign = BYTES_PER_WORD;
2365 if (flags & SLAB_RED_ZONE) {
2366 ralign = REDZONE_ALIGN;
2367 /* If redzoning, ensure that the second redzone is suitably
2368 * aligned, by adjusting the object size accordingly. */
2369 size += REDZONE_ALIGN - 1;
2370 size &= ~(REDZONE_ALIGN - 1);
2373 /* 3) caller mandated alignment */
2374 if (ralign < cachep->align) {
2375 ralign = cachep->align;
2377 /* disable debug if necessary */
2378 if (ralign > __alignof__(unsigned long long))
2379 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2383 cachep->align = ralign;
2385 if (slab_is_available())
2390 setup_node_pointer(cachep);
2394 * Both debugging options require word-alignment which is calculated
2397 if (flags & SLAB_RED_ZONE) {
2398 /* add space for red zone words */
2399 cachep->obj_offset += sizeof(unsigned long long);
2400 size += 2 * sizeof(unsigned long long);
2402 if (flags & SLAB_STORE_USER) {
2403 /* user store requires one word storage behind the end of
2404 * the real object. But if the second red zone needs to be
2405 * aligned to 64 bits, we must allow that much space.
2407 if (flags & SLAB_RED_ZONE)
2408 size += REDZONE_ALIGN;
2410 size += BYTES_PER_WORD;
2412 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2413 if (size >= kmalloc_size(INDEX_L3 + 1)
2414 && cachep->object_size > cache_line_size() && ALIGN(size, align) < PAGE_SIZE) {
2415 cachep->obj_offset += PAGE_SIZE - ALIGN(size, align);
2422 * Determine if the slab management is 'on' or 'off' slab.
2423 * (bootstrapping cannot cope with offslab caches so don't do
2424 * it too early on. Always use on-slab management when
2425 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2427 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
2428 !(flags & SLAB_NOLEAKTRACE))
2430 * Size is large, assume best to place the slab management obj
2431 * off-slab (should allow better packing of objs).
2433 flags |= CFLGS_OFF_SLAB;
2435 size = ALIGN(size, cachep->align);
2437 left_over = calculate_slab_order(cachep, size, cachep->align, flags);
2442 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2443 + sizeof(struct slab), cachep->align);
2446 * If the slab has been placed off-slab, and we have enough space then
2447 * move it on-slab. This is at the expense of any extra colouring.
2449 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2450 flags &= ~CFLGS_OFF_SLAB;
2451 left_over -= slab_size;
2454 if (flags & CFLGS_OFF_SLAB) {
2455 /* really off slab. No need for manual alignment */
2457 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2459 #ifdef CONFIG_PAGE_POISONING
2460 /* If we're going to use the generic kernel_map_pages()
2461 * poisoning, then it's going to smash the contents of
2462 * the redzone and userword anyhow, so switch them off.
2464 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2465 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2469 cachep->colour_off = cache_line_size();
2470 /* Offset must be a multiple of the alignment. */
2471 if (cachep->colour_off < cachep->align)
2472 cachep->colour_off = cachep->align;
2473 cachep->colour = left_over / cachep->colour_off;
2474 cachep->slab_size = slab_size;
2475 cachep->flags = flags;
2476 cachep->allocflags = 0;
2477 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2478 cachep->allocflags |= GFP_DMA;
2479 cachep->size = size;
2480 cachep->reciprocal_buffer_size = reciprocal_value(size);
2482 if (flags & CFLGS_OFF_SLAB) {
2483 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2485 * This is a possibility for one of the malloc_sizes caches.
2486 * But since we go off slab only for object size greater than
2487 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2488 * this should not happen at all.
2489 * But leave a BUG_ON for some lucky dude.
2491 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
2494 err = setup_cpu_cache(cachep, gfp);
2496 __kmem_cache_shutdown(cachep);
2500 if (flags & SLAB_DEBUG_OBJECTS) {
2502 * Would deadlock through slab_destroy()->call_rcu()->
2503 * debug_object_activate()->kmem_cache_alloc().
2505 WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU);
2507 slab_set_debugobj_lock_classes(cachep);
2508 } else if (!OFF_SLAB(cachep) && !(flags & SLAB_DESTROY_BY_RCU))
2509 on_slab_lock_classes(cachep);
2515 static void check_irq_off(void)
2517 BUG_ON(!irqs_disabled());
2520 static void check_irq_on(void)
2522 BUG_ON(irqs_disabled());
2525 static void check_spinlock_acquired(struct kmem_cache *cachep)
2529 assert_spin_locked(&cachep->node[numa_mem_id()]->list_lock);
2533 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2537 assert_spin_locked(&cachep->node[node]->list_lock);
2542 #define check_irq_off() do { } while(0)
2543 #define check_irq_on() do { } while(0)
2544 #define check_spinlock_acquired(x) do { } while(0)
2545 #define check_spinlock_acquired_node(x, y) do { } while(0)
2548 static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *l3,
2549 struct array_cache *ac,
2550 int force, int node);
2552 static void do_drain(void *arg)
2554 struct kmem_cache *cachep = arg;
2555 struct array_cache *ac;
2556 int node = numa_mem_id();
2559 ac = cpu_cache_get(cachep);
2560 spin_lock(&cachep->node[node]->list_lock);
2561 free_block(cachep, ac->entry, ac->avail, node);
2562 spin_unlock(&cachep->node[node]->list_lock);
2566 static void drain_cpu_caches(struct kmem_cache *cachep)
2568 struct kmem_cache_node *l3;
2571 on_each_cpu(do_drain, cachep, 1);
2573 for_each_online_node(node) {
2574 l3 = cachep->node[node];
2575 if (l3 && l3->alien)
2576 drain_alien_cache(cachep, l3->alien);
2579 for_each_online_node(node) {
2580 l3 = cachep->node[node];
2582 drain_array(cachep, l3, l3->shared, 1, node);
2587 * Remove slabs from the list of free slabs.
2588 * Specify the number of slabs to drain in tofree.
2590 * Returns the actual number of slabs released.
2592 static int drain_freelist(struct kmem_cache *cache,
2593 struct kmem_cache_node *l3, int tofree)
2595 struct list_head *p;
2600 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
2602 spin_lock_irq(&l3->list_lock);
2603 p = l3->slabs_free.prev;
2604 if (p == &l3->slabs_free) {
2605 spin_unlock_irq(&l3->list_lock);
2609 slabp = list_entry(p, struct slab, list);
2611 BUG_ON(slabp->inuse);
2613 list_del(&slabp->list);
2615 * Safe to drop the lock. The slab is no longer linked
2618 l3->free_objects -= cache->num;
2619 spin_unlock_irq(&l3->list_lock);
2620 slab_destroy(cache, slabp);
2627 /* Called with slab_mutex held to protect against cpu hotplug */
2628 static int __cache_shrink(struct kmem_cache *cachep)
2631 struct kmem_cache_node *l3;
2633 drain_cpu_caches(cachep);
2636 for_each_online_node(i) {
2637 l3 = cachep->node[i];
2641 drain_freelist(cachep, l3, l3->free_objects);
2643 ret += !list_empty(&l3->slabs_full) ||
2644 !list_empty(&l3->slabs_partial);
2646 return (ret ? 1 : 0);
2650 * kmem_cache_shrink - Shrink a cache.
2651 * @cachep: The cache to shrink.
2653 * Releases as many slabs as possible for a cache.
2654 * To help debugging, a zero exit status indicates all slabs were released.
2656 int kmem_cache_shrink(struct kmem_cache *cachep)
2659 BUG_ON(!cachep || in_interrupt());
2662 mutex_lock(&slab_mutex);
2663 ret = __cache_shrink(cachep);
2664 mutex_unlock(&slab_mutex);
2668 EXPORT_SYMBOL(kmem_cache_shrink);
2670 int __kmem_cache_shutdown(struct kmem_cache *cachep)
2673 struct kmem_cache_node *l3;
2674 int rc = __cache_shrink(cachep);
2679 for_each_online_cpu(i)
2680 kfree(cachep->array[i]);
2682 /* NUMA: free the list3 structures */
2683 for_each_online_node(i) {
2684 l3 = cachep->node[i];
2687 free_alien_cache(l3->alien);
2695 * Get the memory for a slab management obj.
2696 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2697 * always come from malloc_sizes caches. The slab descriptor cannot
2698 * come from the same cache which is getting created because,
2699 * when we are searching for an appropriate cache for these
2700 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2701 * If we are creating a malloc_sizes cache here it would not be visible to
2702 * kmem_find_general_cachep till the initialization is complete.
2703 * Hence we cannot have slabp_cache same as the original cache.
2705 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2706 int colour_off, gfp_t local_flags,
2711 if (OFF_SLAB(cachep)) {
2712 /* Slab management obj is off-slab. */
2713 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2714 local_flags, nodeid);
2716 * If the first object in the slab is leaked (it's allocated
2717 * but no one has a reference to it), we want to make sure
2718 * kmemleak does not treat the ->s_mem pointer as a reference
2719 * to the object. Otherwise we will not report the leak.
2721 kmemleak_scan_area(&slabp->list, sizeof(struct list_head),
2726 slabp = objp + colour_off;
2727 colour_off += cachep->slab_size;
2730 slabp->colouroff = colour_off;
2731 slabp->s_mem = objp + colour_off;
2732 slabp->nodeid = nodeid;
2737 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2739 return (kmem_bufctl_t *) (slabp + 1);
2742 static void cache_init_objs(struct kmem_cache *cachep,
2747 for (i = 0; i < cachep->num; i++) {
2748 void *objp = index_to_obj(cachep, slabp, i);
2750 /* need to poison the objs? */
2751 if (cachep->flags & SLAB_POISON)
2752 poison_obj(cachep, objp, POISON_FREE);
2753 if (cachep->flags & SLAB_STORE_USER)
2754 *dbg_userword(cachep, objp) = NULL;
2756 if (cachep->flags & SLAB_RED_ZONE) {
2757 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2758 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2761 * Constructors are not allowed to allocate memory from the same
2762 * cache which they are a constructor for. Otherwise, deadlock.
2763 * They must also be threaded.
2765 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2766 cachep->ctor(objp + obj_offset(cachep));
2768 if (cachep->flags & SLAB_RED_ZONE) {
2769 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2770 slab_error(cachep, "constructor overwrote the"
2771 " end of an object");
2772 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2773 slab_error(cachep, "constructor overwrote the"
2774 " start of an object");
2776 if ((cachep->size % PAGE_SIZE) == 0 &&
2777 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2778 kernel_map_pages(virt_to_page(objp),
2779 cachep->size / PAGE_SIZE, 0);
2784 slab_bufctl(slabp)[i] = i + 1;
2786 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2789 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2791 if (CONFIG_ZONE_DMA_FLAG) {
2792 if (flags & GFP_DMA)
2793 BUG_ON(!(cachep->allocflags & GFP_DMA));
2795 BUG_ON(cachep->allocflags & GFP_DMA);
2799 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2802 void *objp = index_to_obj(cachep, slabp, slabp->free);
2806 next = slab_bufctl(slabp)[slabp->free];
2808 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2809 WARN_ON(slabp->nodeid != nodeid);
2816 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2817 void *objp, int nodeid)
2819 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2822 /* Verify that the slab belongs to the intended node */
2823 WARN_ON(slabp->nodeid != nodeid);
2825 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2826 printk(KERN_ERR "slab: double free detected in cache "
2827 "'%s', objp %p\n", cachep->name, objp);
2831 slab_bufctl(slabp)[objnr] = slabp->free;
2832 slabp->free = objnr;
2837 * Map pages beginning at addr to the given cache and slab. This is required
2838 * for the slab allocator to be able to lookup the cache and slab of a
2839 * virtual address for kfree, ksize, and slab debugging.
2841 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2847 page = virt_to_page(addr);
2850 if (likely(!PageCompound(page)))
2851 nr_pages <<= cache->gfporder;
2854 page->slab_cache = cache;
2855 page->slab_page = slab;
2857 } while (--nr_pages);
2861 * Grow (by 1) the number of slabs within a cache. This is called by
2862 * kmem_cache_alloc() when there are no active objs left in a cache.
2864 static int cache_grow(struct kmem_cache *cachep,
2865 gfp_t flags, int nodeid, void *objp)
2870 struct kmem_cache_node *l3;
2873 * Be lazy and only check for valid flags here, keeping it out of the
2874 * critical path in kmem_cache_alloc().
2876 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2877 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2879 /* Take the l3 list lock to change the colour_next on this node */
2881 l3 = cachep->node[nodeid];
2882 spin_lock(&l3->list_lock);
2884 /* Get colour for the slab, and cal the next value. */
2885 offset = l3->colour_next;
2887 if (l3->colour_next >= cachep->colour)
2888 l3->colour_next = 0;
2889 spin_unlock(&l3->list_lock);
2891 offset *= cachep->colour_off;
2893 if (local_flags & __GFP_WAIT)
2897 * The test for missing atomic flag is performed here, rather than
2898 * the more obvious place, simply to reduce the critical path length
2899 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2900 * will eventually be caught here (where it matters).
2902 kmem_flagcheck(cachep, flags);
2905 * Get mem for the objs. Attempt to allocate a physical page from
2909 objp = kmem_getpages(cachep, local_flags, nodeid);
2913 /* Get slab management. */
2914 slabp = alloc_slabmgmt(cachep, objp, offset,
2915 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2919 slab_map_pages(cachep, slabp, objp);
2921 cache_init_objs(cachep, slabp);
2923 if (local_flags & __GFP_WAIT)
2924 local_irq_disable();
2926 spin_lock(&l3->list_lock);
2928 /* Make slab active. */
2929 list_add_tail(&slabp->list, &(l3->slabs_free));
2930 STATS_INC_GROWN(cachep);
2931 l3->free_objects += cachep->num;
2932 spin_unlock(&l3->list_lock);
2935 kmem_freepages(cachep, objp);
2937 if (local_flags & __GFP_WAIT)
2938 local_irq_disable();
2945 * Perform extra freeing checks:
2946 * - detect bad pointers.
2947 * - POISON/RED_ZONE checking
2949 static void kfree_debugcheck(const void *objp)
2951 if (!virt_addr_valid(objp)) {
2952 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2953 (unsigned long)objp);
2958 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2960 unsigned long long redzone1, redzone2;
2962 redzone1 = *dbg_redzone1(cache, obj);
2963 redzone2 = *dbg_redzone2(cache, obj);
2968 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2971 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2972 slab_error(cache, "double free detected");
2974 slab_error(cache, "memory outside object was overwritten");
2976 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2977 obj, redzone1, redzone2);
2980 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2981 unsigned long caller)
2987 BUG_ON(virt_to_cache(objp) != cachep);
2989 objp -= obj_offset(cachep);
2990 kfree_debugcheck(objp);
2991 page = virt_to_head_page(objp);
2993 slabp = page->slab_page;
2995 if (cachep->flags & SLAB_RED_ZONE) {
2996 verify_redzone_free(cachep, objp);
2997 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2998 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
3000 if (cachep->flags & SLAB_STORE_USER)
3001 *dbg_userword(cachep, objp) = (void *)caller;
3003 objnr = obj_to_index(cachep, slabp, objp);
3005 BUG_ON(objnr >= cachep->num);
3006 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
3008 #ifdef CONFIG_DEBUG_SLAB_LEAK
3009 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
3011 if (cachep->flags & SLAB_POISON) {
3012 #ifdef CONFIG_DEBUG_PAGEALLOC
3013 if ((cachep->size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
3014 store_stackinfo(cachep, objp, caller);
3015 kernel_map_pages(virt_to_page(objp),
3016 cachep->size / PAGE_SIZE, 0);
3018 poison_obj(cachep, objp, POISON_FREE);
3021 poison_obj(cachep, objp, POISON_FREE);
3027 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
3032 /* Check slab's freelist to see if this obj is there. */
3033 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
3035 if (entries > cachep->num || i >= cachep->num)
3038 if (entries != cachep->num - slabp->inuse) {
3040 printk(KERN_ERR "slab: Internal list corruption detected in "
3041 "cache '%s'(%d), slabp %p(%d). Tainted(%s). Hexdump:\n",
3042 cachep->name, cachep->num, slabp, slabp->inuse,
3044 print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 16, 1, slabp,
3045 sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t),
3051 #define kfree_debugcheck(x) do { } while(0)
3052 #define cache_free_debugcheck(x,objp,z) (objp)
3053 #define check_slabp(x,y) do { } while(0)
3056 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags,
3060 struct kmem_cache_node *l3;
3061 struct array_cache *ac;
3065 node = numa_mem_id();
3066 if (unlikely(force_refill))
3069 ac = cpu_cache_get(cachep);
3070 batchcount = ac->batchcount;
3071 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
3073 * If there was little recent activity on this cache, then
3074 * perform only a partial refill. Otherwise we could generate
3077 batchcount = BATCHREFILL_LIMIT;
3079 l3 = cachep->node[node];
3081 BUG_ON(ac->avail > 0 || !l3);
3082 spin_lock(&l3->list_lock);
3084 /* See if we can refill from the shared array */
3085 if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) {
3086 l3->shared->touched = 1;
3090 while (batchcount > 0) {
3091 struct list_head *entry;
3093 /* Get slab alloc is to come from. */
3094 entry = l3->slabs_partial.next;
3095 if (entry == &l3->slabs_partial) {
3096 l3->free_touched = 1;
3097 entry = l3->slabs_free.next;
3098 if (entry == &l3->slabs_free)
3102 slabp = list_entry(entry, struct slab, list);
3103 check_slabp(cachep, slabp);
3104 check_spinlock_acquired(cachep);
3107 * The slab was either on partial or free list so
3108 * there must be at least one object available for
3111 BUG_ON(slabp->inuse >= cachep->num);
3113 while (slabp->inuse < cachep->num && batchcount--) {
3114 STATS_INC_ALLOCED(cachep);
3115 STATS_INC_ACTIVE(cachep);
3116 STATS_SET_HIGH(cachep);
3118 ac_put_obj(cachep, ac, slab_get_obj(cachep, slabp,
3121 check_slabp(cachep, slabp);
3123 /* move slabp to correct slabp list: */
3124 list_del(&slabp->list);
3125 if (slabp->free == BUFCTL_END)
3126 list_add(&slabp->list, &l3->slabs_full);
3128 list_add(&slabp->list, &l3->slabs_partial);
3132 l3->free_objects -= ac->avail;
3134 spin_unlock(&l3->list_lock);
3136 if (unlikely(!ac->avail)) {
3139 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
3141 /* cache_grow can reenable interrupts, then ac could change. */
3142 ac = cpu_cache_get(cachep);
3143 node = numa_mem_id();
3145 /* no objects in sight? abort */
3146 if (!x && (ac->avail == 0 || force_refill))
3149 if (!ac->avail) /* objects refilled by interrupt? */
3154 return ac_get_obj(cachep, ac, flags, force_refill);
3157 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3160 might_sleep_if(flags & __GFP_WAIT);
3162 kmem_flagcheck(cachep, flags);
3167 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3168 gfp_t flags, void *objp, unsigned long caller)
3172 if (cachep->flags & SLAB_POISON) {
3173 #ifdef CONFIG_DEBUG_PAGEALLOC
3174 if ((cachep->size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
3175 kernel_map_pages(virt_to_page(objp),
3176 cachep->size / PAGE_SIZE, 1);
3178 check_poison_obj(cachep, objp);
3180 check_poison_obj(cachep, objp);
3182 poison_obj(cachep, objp, POISON_INUSE);
3184 if (cachep->flags & SLAB_STORE_USER)
3185 *dbg_userword(cachep, objp) = (void *)caller;
3187 if (cachep->flags & SLAB_RED_ZONE) {
3188 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3189 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3190 slab_error(cachep, "double free, or memory outside"
3191 " object was overwritten");
3193 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3194 objp, *dbg_redzone1(cachep, objp),
3195 *dbg_redzone2(cachep, objp));
3197 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3198 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3200 #ifdef CONFIG_DEBUG_SLAB_LEAK
3205 slabp = virt_to_head_page(objp)->slab_page;
3206 objnr = (unsigned)(objp - slabp->s_mem) / cachep->size;
3207 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3210 objp += obj_offset(cachep);
3211 if (cachep->ctor && cachep->flags & SLAB_POISON)
3213 if (ARCH_SLAB_MINALIGN &&
3214 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
3215 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3216 objp, (int)ARCH_SLAB_MINALIGN);
3221 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3224 static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
3226 if (cachep == kmem_cache)
3229 return should_failslab(cachep->object_size, flags, cachep->flags);
3232 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3235 struct array_cache *ac;
3236 bool force_refill = false;
3240 ac = cpu_cache_get(cachep);
3241 if (likely(ac->avail)) {
3243 objp = ac_get_obj(cachep, ac, flags, false);
3246 * Allow for the possibility all avail objects are not allowed
3247 * by the current flags
3250 STATS_INC_ALLOCHIT(cachep);
3253 force_refill = true;
3256 STATS_INC_ALLOCMISS(cachep);
3257 objp = cache_alloc_refill(cachep, flags, force_refill);
3259 * the 'ac' may be updated by cache_alloc_refill(),
3260 * and kmemleak_erase() requires its correct value.
3262 ac = cpu_cache_get(cachep);
3266 * To avoid a false negative, if an object that is in one of the
3267 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3268 * treat the array pointers as a reference to the object.
3271 kmemleak_erase(&ac->entry[ac->avail]);
3277 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3279 * If we are in_interrupt, then process context, including cpusets and
3280 * mempolicy, may not apply and should not be used for allocation policy.
3282 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3284 int nid_alloc, nid_here;
3286 if (in_interrupt() || (flags & __GFP_THISNODE))
3288 nid_alloc = nid_here = numa_mem_id();
3289 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3290 nid_alloc = cpuset_slab_spread_node();
3291 else if (current->mempolicy)
3292 nid_alloc = slab_node();
3293 if (nid_alloc != nid_here)
3294 return ____cache_alloc_node(cachep, flags, nid_alloc);
3299 * Fallback function if there was no memory available and no objects on a
3300 * certain node and fall back is permitted. First we scan all the
3301 * available node for available objects. If that fails then we
3302 * perform an allocation without specifying a node. This allows the page
3303 * allocator to do its reclaim / fallback magic. We then insert the
3304 * slab into the proper nodelist and then allocate from it.
3306 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3308 struct zonelist *zonelist;
3312 enum zone_type high_zoneidx = gfp_zone(flags);
3315 unsigned int cpuset_mems_cookie;
3317 if (flags & __GFP_THISNODE)
3320 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3323 cpuset_mems_cookie = get_mems_allowed();
3324 zonelist = node_zonelist(slab_node(), flags);
3328 * Look through allowed nodes for objects available
3329 * from existing per node queues.
3331 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3332 nid = zone_to_nid(zone);
3334 if (cpuset_zone_allowed_hardwall(zone, flags) &&
3336 cache->node[nid]->free_objects) {
3337 obj = ____cache_alloc_node(cache,
3338 flags | GFP_THISNODE, nid);
3346 * This allocation will be performed within the constraints
3347 * of the current cpuset / memory policy requirements.
3348 * We may trigger various forms of reclaim on the allowed
3349 * set and go into memory reserves if necessary.
3351 if (local_flags & __GFP_WAIT)
3353 kmem_flagcheck(cache, flags);
3354 obj = kmem_getpages(cache, local_flags, numa_mem_id());
3355 if (local_flags & __GFP_WAIT)
3356 local_irq_disable();
3359 * Insert into the appropriate per node queues
3361 nid = page_to_nid(virt_to_page(obj));
3362 if (cache_grow(cache, flags, nid, obj)) {
3363 obj = ____cache_alloc_node(cache,
3364 flags | GFP_THISNODE, nid);
3367 * Another processor may allocate the
3368 * objects in the slab since we are
3369 * not holding any locks.
3373 /* cache_grow already freed obj */
3379 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !obj))
3385 * A interface to enable slab creation on nodeid
3387 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3390 struct list_head *entry;
3392 struct kmem_cache_node *l3;
3396 l3 = cachep->node[nodeid];
3401 spin_lock(&l3->list_lock);
3402 entry = l3->slabs_partial.next;
3403 if (entry == &l3->slabs_partial) {
3404 l3->free_touched = 1;
3405 entry = l3->slabs_free.next;
3406 if (entry == &l3->slabs_free)
3410 slabp = list_entry(entry, struct slab, list);
3411 check_spinlock_acquired_node(cachep, nodeid);
3412 check_slabp(cachep, slabp);
3414 STATS_INC_NODEALLOCS(cachep);
3415 STATS_INC_ACTIVE(cachep);
3416 STATS_SET_HIGH(cachep);
3418 BUG_ON(slabp->inuse == cachep->num);
3420 obj = slab_get_obj(cachep, slabp, nodeid);
3421 check_slabp(cachep, slabp);
3423 /* move slabp to correct slabp list: */
3424 list_del(&slabp->list);
3426 if (slabp->free == BUFCTL_END)
3427 list_add(&slabp->list, &l3->slabs_full);
3429 list_add(&slabp->list, &l3->slabs_partial);
3431 spin_unlock(&l3->list_lock);
3435 spin_unlock(&l3->list_lock);
3436 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
3440 return fallback_alloc(cachep, flags);
3447 * kmem_cache_alloc_node - Allocate an object on the specified node
3448 * @cachep: The cache to allocate from.
3449 * @flags: See kmalloc().
3450 * @nodeid: node number of the target node.
3451 * @caller: return address of caller, used for debug information
3453 * Identical to kmem_cache_alloc but it will allocate memory on the given
3454 * node, which can improve the performance for cpu bound structures.
3456 * Fallback to other node is possible if __GFP_THISNODE is not set.
3458 static __always_inline void *
3459 slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3460 unsigned long caller)
3462 unsigned long save_flags;
3464 int slab_node = numa_mem_id();
3466 flags &= gfp_allowed_mask;
3468 lockdep_trace_alloc(flags);
3470 if (slab_should_failslab(cachep, flags))
3473 cachep = memcg_kmem_get_cache(cachep, flags);
3475 cache_alloc_debugcheck_before(cachep, flags);
3476 local_irq_save(save_flags);
3478 if (nodeid == NUMA_NO_NODE)
3481 if (unlikely(!cachep->node[nodeid])) {
3482 /* Node not bootstrapped yet */
3483 ptr = fallback_alloc(cachep, flags);
3487 if (nodeid == slab_node) {
3489 * Use the locally cached objects if possible.
3490 * However ____cache_alloc does not allow fallback
3491 * to other nodes. It may fail while we still have
3492 * objects on other nodes available.
3494 ptr = ____cache_alloc(cachep, flags);
3498 /* ___cache_alloc_node can fall back to other nodes */
3499 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3501 local_irq_restore(save_flags);
3502 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3503 kmemleak_alloc_recursive(ptr, cachep->object_size, 1, cachep->flags,
3507 kmemcheck_slab_alloc(cachep, flags, ptr, cachep->object_size);
3509 if (unlikely((flags & __GFP_ZERO) && ptr))
3510 memset(ptr, 0, cachep->object_size);
3515 static __always_inline void *
3516 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3520 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3521 objp = alternate_node_alloc(cache, flags);
3525 objp = ____cache_alloc(cache, flags);
3528 * We may just have run out of memory on the local node.
3529 * ____cache_alloc_node() knows how to locate memory on other nodes
3532 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3539 static __always_inline void *
3540 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3542 return ____cache_alloc(cachep, flags);
3545 #endif /* CONFIG_NUMA */
3547 static __always_inline void *
3548 slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
3550 unsigned long save_flags;
3553 flags &= gfp_allowed_mask;
3555 lockdep_trace_alloc(flags);
3557 if (slab_should_failslab(cachep, flags))
3560 cachep = memcg_kmem_get_cache(cachep, flags);
3562 cache_alloc_debugcheck_before(cachep, flags);
3563 local_irq_save(save_flags);
3564 objp = __do_cache_alloc(cachep, flags);
3565 local_irq_restore(save_flags);
3566 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3567 kmemleak_alloc_recursive(objp, cachep->object_size, 1, cachep->flags,
3572 kmemcheck_slab_alloc(cachep, flags, objp, cachep->object_size);
3574 if (unlikely((flags & __GFP_ZERO) && objp))
3575 memset(objp, 0, cachep->object_size);
3581 * Caller needs to acquire correct kmem_list's list_lock
3583 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3587 struct kmem_cache_node *l3;
3589 for (i = 0; i < nr_objects; i++) {
3593 clear_obj_pfmemalloc(&objpp[i]);
3596 slabp = virt_to_slab(objp);
3597 l3 = cachep->node[node];
3598 list_del(&slabp->list);
3599 check_spinlock_acquired_node(cachep, node);
3600 check_slabp(cachep, slabp);
3601 slab_put_obj(cachep, slabp, objp, node);
3602 STATS_DEC_ACTIVE(cachep);
3604 check_slabp(cachep, slabp);
3606 /* fixup slab chains */
3607 if (slabp->inuse == 0) {
3608 if (l3->free_objects > l3->free_limit) {
3609 l3->free_objects -= cachep->num;
3610 /* No need to drop any previously held
3611 * lock here, even if we have a off-slab slab
3612 * descriptor it is guaranteed to come from
3613 * a different cache, refer to comments before
3616 slab_destroy(cachep, slabp);
3618 list_add(&slabp->list, &l3->slabs_free);
3621 /* Unconditionally move a slab to the end of the
3622 * partial list on free - maximum time for the
3623 * other objects to be freed, too.
3625 list_add_tail(&slabp->list, &l3->slabs_partial);
3630 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3633 struct kmem_cache_node *l3;
3634 int node = numa_mem_id();
3636 batchcount = ac->batchcount;
3638 BUG_ON(!batchcount || batchcount > ac->avail);
3641 l3 = cachep->node[node];
3642 spin_lock(&l3->list_lock);
3644 struct array_cache *shared_array = l3->shared;
3645 int max = shared_array->limit - shared_array->avail;
3647 if (batchcount > max)
3649 memcpy(&(shared_array->entry[shared_array->avail]),
3650 ac->entry, sizeof(void *) * batchcount);
3651 shared_array->avail += batchcount;
3656 free_block(cachep, ac->entry, batchcount, node);
3661 struct list_head *p;
3663 p = l3->slabs_free.next;
3664 while (p != &(l3->slabs_free)) {
3667 slabp = list_entry(p, struct slab, list);
3668 BUG_ON(slabp->inuse);
3673 STATS_SET_FREEABLE(cachep, i);
3676 spin_unlock(&l3->list_lock);
3677 ac->avail -= batchcount;
3678 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3682 * Release an obj back to its cache. If the obj has a constructed state, it must
3683 * be in this state _before_ it is released. Called with disabled ints.
3685 static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3686 unsigned long caller)
3688 struct array_cache *ac = cpu_cache_get(cachep);
3691 kmemleak_free_recursive(objp, cachep->flags);
3692 objp = cache_free_debugcheck(cachep, objp, caller);
3694 kmemcheck_slab_free(cachep, objp, cachep->object_size);
3697 * Skip calling cache_free_alien() when the platform is not numa.
3698 * This will avoid cache misses that happen while accessing slabp (which
3699 * is per page memory reference) to get nodeid. Instead use a global
3700 * variable to skip the call, which is mostly likely to be present in
3703 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3706 if (likely(ac->avail < ac->limit)) {
3707 STATS_INC_FREEHIT(cachep);
3709 STATS_INC_FREEMISS(cachep);
3710 cache_flusharray(cachep, ac);
3713 ac_put_obj(cachep, ac, objp);
3717 * kmem_cache_alloc - Allocate an object
3718 * @cachep: The cache to allocate from.
3719 * @flags: See kmalloc().
3721 * Allocate an object from this cache. The flags are only relevant
3722 * if the cache has no available objects.
3724 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3726 void *ret = slab_alloc(cachep, flags, _RET_IP_);
3728 trace_kmem_cache_alloc(_RET_IP_, ret,
3729 cachep->object_size, cachep->size, flags);
3733 EXPORT_SYMBOL(kmem_cache_alloc);
3735 #ifdef CONFIG_TRACING
3737 kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
3741 ret = slab_alloc(cachep, flags, _RET_IP_);
3743 trace_kmalloc(_RET_IP_, ret,
3744 size, cachep->size, flags);
3747 EXPORT_SYMBOL(kmem_cache_alloc_trace);
3751 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3753 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3755 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3756 cachep->object_size, cachep->size,
3761 EXPORT_SYMBOL(kmem_cache_alloc_node);
3763 #ifdef CONFIG_TRACING
3764 void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
3771 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3773 trace_kmalloc_node(_RET_IP_, ret,
3778 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3781 static __always_inline void *
3782 __do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
3784 struct kmem_cache *cachep;
3786 cachep = kmem_find_general_cachep(size, flags);
3787 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3789 return kmem_cache_alloc_node_trace(cachep, flags, node, size);
3792 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3793 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3795 return __do_kmalloc_node(size, flags, node, _RET_IP_);
3797 EXPORT_SYMBOL(__kmalloc_node);
3799 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3800 int node, unsigned long caller)
3802 return __do_kmalloc_node(size, flags, node, caller);
3804 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3806 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3808 return __do_kmalloc_node(size, flags, node, 0);
3810 EXPORT_SYMBOL(__kmalloc_node);
3811 #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3812 #endif /* CONFIG_NUMA */
3815 * __do_kmalloc - allocate memory
3816 * @size: how many bytes of memory are required.
3817 * @flags: the type of memory to allocate (see kmalloc).
3818 * @caller: function caller for debug tracking of the caller
3820 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3821 unsigned long caller)
3823 struct kmem_cache *cachep;
3826 /* If you want to save a few bytes .text space: replace
3828 * Then kmalloc uses the uninlined functions instead of the inline
3831 cachep = __find_general_cachep(size, flags);
3832 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3834 ret = slab_alloc(cachep, flags, caller);
3836 trace_kmalloc(caller, ret,
3837 size, cachep->size, flags);
3843 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3844 void *__kmalloc(size_t size, gfp_t flags)
3846 return __do_kmalloc(size, flags, _RET_IP_);
3848 EXPORT_SYMBOL(__kmalloc);
3850 void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3852 return __do_kmalloc(size, flags, caller);
3854 EXPORT_SYMBOL(__kmalloc_track_caller);
3857 void *__kmalloc(size_t size, gfp_t flags)
3859 return __do_kmalloc(size, flags, 0);
3861 EXPORT_SYMBOL(__kmalloc);
3865 * kmem_cache_free - Deallocate an object
3866 * @cachep: The cache the allocation was from.
3867 * @objp: The previously allocated object.
3869 * Free an object which was previously allocated from this
3872 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3874 unsigned long flags;
3875 cachep = cache_from_obj(cachep, objp);
3879 local_irq_save(flags);
3880 debug_check_no_locks_freed(objp, cachep->object_size);
3881 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3882 debug_check_no_obj_freed(objp, cachep->object_size);
3883 __cache_free(cachep, objp, _RET_IP_);
3884 local_irq_restore(flags);
3886 trace_kmem_cache_free(_RET_IP_, objp);
3888 EXPORT_SYMBOL(kmem_cache_free);
3891 * kfree - free previously allocated memory
3892 * @objp: pointer returned by kmalloc.
3894 * If @objp is NULL, no operation is performed.
3896 * Don't free memory not originally allocated by kmalloc()
3897 * or you will run into trouble.
3899 void kfree(const void *objp)
3901 struct kmem_cache *c;
3902 unsigned long flags;
3904 trace_kfree(_RET_IP_, objp);
3906 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3908 local_irq_save(flags);
3909 kfree_debugcheck(objp);
3910 c = virt_to_cache(objp);
3911 debug_check_no_locks_freed(objp, c->object_size);
3913 debug_check_no_obj_freed(objp, c->object_size);
3914 __cache_free(c, (void *)objp, _RET_IP_);
3915 local_irq_restore(flags);
3917 EXPORT_SYMBOL(kfree);
3920 * This initializes kmem_list3 or resizes various caches for all nodes.
3922 static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
3925 struct kmem_cache_node *l3;
3926 struct array_cache *new_shared;
3927 struct array_cache **new_alien = NULL;
3929 for_each_online_node(node) {
3931 if (use_alien_caches) {
3932 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
3938 if (cachep->shared) {
3939 new_shared = alloc_arraycache(node,
3940 cachep->shared*cachep->batchcount,
3943 free_alien_cache(new_alien);
3948 l3 = cachep->node[node];
3950 struct array_cache *shared = l3->shared;
3952 spin_lock_irq(&l3->list_lock);
3955 free_block(cachep, shared->entry,
3956 shared->avail, node);
3958 l3->shared = new_shared;
3960 l3->alien = new_alien;
3963 l3->free_limit = (1 + nr_cpus_node(node)) *
3964 cachep->batchcount + cachep->num;
3965 spin_unlock_irq(&l3->list_lock);
3967 free_alien_cache(new_alien);
3970 l3 = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
3972 free_alien_cache(new_alien);
3977 kmem_list3_init(l3);
3978 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3979 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3980 l3->shared = new_shared;
3981 l3->alien = new_alien;
3982 l3->free_limit = (1 + nr_cpus_node(node)) *
3983 cachep->batchcount + cachep->num;
3984 cachep->node[node] = l3;
3989 if (!cachep->list.next) {
3990 /* Cache is not active yet. Roll back what we did */
3993 if (cachep->node[node]) {
3994 l3 = cachep->node[node];
3997 free_alien_cache(l3->alien);
3999 cachep->node[node] = NULL;
4007 struct ccupdate_struct {
4008 struct kmem_cache *cachep;
4009 struct array_cache *new[0];
4012 static void do_ccupdate_local(void *info)
4014 struct ccupdate_struct *new = info;
4015 struct array_cache *old;
4018 old = cpu_cache_get(new->cachep);
4020 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
4021 new->new[smp_processor_id()] = old;
4024 /* Always called with the slab_mutex held */
4025 static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
4026 int batchcount, int shared, gfp_t gfp)
4028 struct ccupdate_struct *new;
4031 new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *),
4036 for_each_online_cpu(i) {
4037 new->new[i] = alloc_arraycache(cpu_to_mem(i), limit,
4040 for (i--; i >= 0; i--)
4046 new->cachep = cachep;
4048 on_each_cpu(do_ccupdate_local, (void *)new, 1);
4051 cachep->batchcount = batchcount;
4052 cachep->limit = limit;
4053 cachep->shared = shared;
4055 for_each_online_cpu(i) {
4056 struct array_cache *ccold = new->new[i];
4059 spin_lock_irq(&cachep->node[cpu_to_mem(i)]->list_lock);
4060 free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i));
4061 spin_unlock_irq(&cachep->node[cpu_to_mem(i)]->list_lock);
4065 return alloc_kmemlist(cachep, gfp);
4068 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
4069 int batchcount, int shared, gfp_t gfp)
4072 struct kmem_cache *c = NULL;
4075 ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
4077 if (slab_state < FULL)
4080 if ((ret < 0) || !is_root_cache(cachep))
4083 VM_BUG_ON(!mutex_is_locked(&slab_mutex));
4084 for_each_memcg_cache_index(i) {
4085 c = cache_from_memcg(cachep, i);
4087 /* return value determined by the parent cache only */
4088 __do_tune_cpucache(c, limit, batchcount, shared, gfp);
4094 /* Called with slab_mutex held always */
4095 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
4102 if (!is_root_cache(cachep)) {
4103 struct kmem_cache *root = memcg_root_cache(cachep);
4104 limit = root->limit;
4105 shared = root->shared;
4106 batchcount = root->batchcount;
4109 if (limit && shared && batchcount)
4112 * The head array serves three purposes:
4113 * - create a LIFO ordering, i.e. return objects that are cache-warm
4114 * - reduce the number of spinlock operations.
4115 * - reduce the number of linked list operations on the slab and
4116 * bufctl chains: array operations are cheaper.
4117 * The numbers are guessed, we should auto-tune as described by
4120 if (cachep->size > 131072)
4122 else if (cachep->size > PAGE_SIZE)
4124 else if (cachep->size > 1024)
4126 else if (cachep->size > 256)
4132 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
4133 * allocation behaviour: Most allocs on one cpu, most free operations
4134 * on another cpu. For these cases, an efficient object passing between
4135 * cpus is necessary. This is provided by a shared array. The array
4136 * replaces Bonwick's magazine layer.
4137 * On uniprocessor, it's functionally equivalent (but less efficient)
4138 * to a larger limit. Thus disabled by default.
4141 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
4146 * With debugging enabled, large batchcount lead to excessively long
4147 * periods with disabled local interrupts. Limit the batchcount
4152 batchcount = (limit + 1) / 2;
4154 err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
4156 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
4157 cachep->name, -err);
4162 * Drain an array if it contains any elements taking the l3 lock only if
4163 * necessary. Note that the l3 listlock also protects the array_cache
4164 * if drain_array() is used on the shared array.
4166 static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *l3,
4167 struct array_cache *ac, int force, int node)
4171 if (!ac || !ac->avail)
4173 if (ac->touched && !force) {
4176 spin_lock_irq(&l3->list_lock);
4178 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4179 if (tofree > ac->avail)
4180 tofree = (ac->avail + 1) / 2;
4181 free_block(cachep, ac->entry, tofree, node);
4182 ac->avail -= tofree;
4183 memmove(ac->entry, &(ac->entry[tofree]),
4184 sizeof(void *) * ac->avail);
4186 spin_unlock_irq(&l3->list_lock);
4191 * cache_reap - Reclaim memory from caches.
4192 * @w: work descriptor
4194 * Called from workqueue/eventd every few seconds.
4196 * - clear the per-cpu caches for this CPU.
4197 * - return freeable pages to the main free memory pool.
4199 * If we cannot acquire the cache chain mutex then just give up - we'll try
4200 * again on the next iteration.
4202 static void cache_reap(struct work_struct *w)
4204 struct kmem_cache *searchp;
4205 struct kmem_cache_node *l3;
4206 int node = numa_mem_id();
4207 struct delayed_work *work = to_delayed_work(w);
4209 if (!mutex_trylock(&slab_mutex))
4210 /* Give up. Setup the next iteration. */
4213 list_for_each_entry(searchp, &slab_caches, list) {
4217 * We only take the l3 lock if absolutely necessary and we
4218 * have established with reasonable certainty that
4219 * we can do some work if the lock was obtained.
4221 l3 = searchp->node[node];
4223 reap_alien(searchp, l3);
4225 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
4228 * These are racy checks but it does not matter
4229 * if we skip one check or scan twice.
4231 if (time_after(l3->next_reap, jiffies))
4234 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
4236 drain_array(searchp, l3, l3->shared, 0, node);
4238 if (l3->free_touched)
4239 l3->free_touched = 0;
4243 freed = drain_freelist(searchp, l3, (l3->free_limit +
4244 5 * searchp->num - 1) / (5 * searchp->num));
4245 STATS_ADD_REAPED(searchp, freed);
4251 mutex_unlock(&slab_mutex);
4254 /* Set up the next iteration */
4255 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
4258 #ifdef CONFIG_SLABINFO
4259 void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
4262 unsigned long active_objs;
4263 unsigned long num_objs;
4264 unsigned long active_slabs = 0;
4265 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4269 struct kmem_cache_node *l3;
4273 for_each_online_node(node) {
4274 l3 = cachep->node[node];
4279 spin_lock_irq(&l3->list_lock);
4281 list_for_each_entry(slabp, &l3->slabs_full, list) {
4282 if (slabp->inuse != cachep->num && !error)
4283 error = "slabs_full accounting error";
4284 active_objs += cachep->num;
4287 list_for_each_entry(slabp, &l3->slabs_partial, list) {
4288 if (slabp->inuse == cachep->num && !error)
4289 error = "slabs_partial inuse accounting error";
4290 if (!slabp->inuse && !error)
4291 error = "slabs_partial/inuse accounting error";
4292 active_objs += slabp->inuse;
4295 list_for_each_entry(slabp, &l3->slabs_free, list) {
4296 if (slabp->inuse && !error)
4297 error = "slabs_free/inuse accounting error";
4300 free_objects += l3->free_objects;
4302 shared_avail += l3->shared->avail;
4304 spin_unlock_irq(&l3->list_lock);
4306 num_slabs += active_slabs;
4307 num_objs = num_slabs * cachep->num;
4308 if (num_objs - active_objs != free_objects && !error)
4309 error = "free_objects accounting error";
4311 name = cachep->name;
4313 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4315 sinfo->active_objs = active_objs;
4316 sinfo->num_objs = num_objs;
4317 sinfo->active_slabs = active_slabs;
4318 sinfo->num_slabs = num_slabs;
4319 sinfo->shared_avail = shared_avail;
4320 sinfo->limit = cachep->limit;
4321 sinfo->batchcount = cachep->batchcount;
4322 sinfo->shared = cachep->shared;
4323 sinfo->objects_per_slab = cachep->num;
4324 sinfo->cache_order = cachep->gfporder;
4327 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
4331 unsigned long high = cachep->high_mark;
4332 unsigned long allocs = cachep->num_allocations;
4333 unsigned long grown = cachep->grown;
4334 unsigned long reaped = cachep->reaped;
4335 unsigned long errors = cachep->errors;
4336 unsigned long max_freeable = cachep->max_freeable;
4337 unsigned long node_allocs = cachep->node_allocs;
4338 unsigned long node_frees = cachep->node_frees;
4339 unsigned long overflows = cachep->node_overflow;
4341 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4342 "%4lu %4lu %4lu %4lu %4lu",
4343 allocs, high, grown,
4344 reaped, errors, max_freeable, node_allocs,
4345 node_frees, overflows);
4349 unsigned long allochit = atomic_read(&cachep->allochit);
4350 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4351 unsigned long freehit = atomic_read(&cachep->freehit);
4352 unsigned long freemiss = atomic_read(&cachep->freemiss);
4354 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4355 allochit, allocmiss, freehit, freemiss);
4360 #define MAX_SLABINFO_WRITE 128
4362 * slabinfo_write - Tuning for the slab allocator
4364 * @buffer: user buffer
4365 * @count: data length
4368 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4369 size_t count, loff_t *ppos)
4371 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4372 int limit, batchcount, shared, res;
4373 struct kmem_cache *cachep;
4375 if (count > MAX_SLABINFO_WRITE)
4377 if (copy_from_user(&kbuf, buffer, count))
4379 kbuf[MAX_SLABINFO_WRITE] = '\0';
4381 tmp = strchr(kbuf, ' ');
4386 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4389 /* Find the cache in the chain of caches. */
4390 mutex_lock(&slab_mutex);
4392 list_for_each_entry(cachep, &slab_caches, list) {
4393 if (!strcmp(cachep->name, kbuf)) {
4394 if (limit < 1 || batchcount < 1 ||
4395 batchcount > limit || shared < 0) {
4398 res = do_tune_cpucache(cachep, limit,
4405 mutex_unlock(&slab_mutex);
4411 #ifdef CONFIG_DEBUG_SLAB_LEAK
4413 static void *leaks_start(struct seq_file *m, loff_t *pos)
4415 mutex_lock(&slab_mutex);
4416 return seq_list_start(&slab_caches, *pos);
4419 static inline int add_caller(unsigned long *n, unsigned long v)
4429 unsigned long *q = p + 2 * i;
4443 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4449 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4455 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->size) {
4456 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4458 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4463 static void show_symbol(struct seq_file *m, unsigned long address)
4465 #ifdef CONFIG_KALLSYMS
4466 unsigned long offset, size;
4467 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4469 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4470 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4472 seq_printf(m, " [%s]", modname);
4476 seq_printf(m, "%p", (void *)address);
4479 static int leaks_show(struct seq_file *m, void *p)
4481 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
4483 struct kmem_cache_node *l3;
4485 unsigned long *n = m->private;
4489 if (!(cachep->flags & SLAB_STORE_USER))
4491 if (!(cachep->flags & SLAB_RED_ZONE))
4494 /* OK, we can do it */
4498 for_each_online_node(node) {
4499 l3 = cachep->node[node];
4504 spin_lock_irq(&l3->list_lock);
4506 list_for_each_entry(slabp, &l3->slabs_full, list)
4507 handle_slab(n, cachep, slabp);
4508 list_for_each_entry(slabp, &l3->slabs_partial, list)
4509 handle_slab(n, cachep, slabp);
4510 spin_unlock_irq(&l3->list_lock);
4512 name = cachep->name;
4514 /* Increase the buffer size */
4515 mutex_unlock(&slab_mutex);
4516 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4518 /* Too bad, we are really out */
4520 mutex_lock(&slab_mutex);
4523 *(unsigned long *)m->private = n[0] * 2;
4525 mutex_lock(&slab_mutex);
4526 /* Now make sure this entry will be retried */
4530 for (i = 0; i < n[1]; i++) {
4531 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4532 show_symbol(m, n[2*i+2]);
4539 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4541 return seq_list_next(p, &slab_caches, pos);
4544 static void s_stop(struct seq_file *m, void *p)
4546 mutex_unlock(&slab_mutex);
4549 static const struct seq_operations slabstats_op = {
4550 .start = leaks_start,
4556 static int slabstats_open(struct inode *inode, struct file *file)
4558 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4561 ret = seq_open(file, &slabstats_op);
4563 struct seq_file *m = file->private_data;
4564 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4573 static const struct file_operations proc_slabstats_operations = {
4574 .open = slabstats_open,
4576 .llseek = seq_lseek,
4577 .release = seq_release_private,
4581 static int __init slab_proc_init(void)
4583 #ifdef CONFIG_DEBUG_SLAB_LEAK
4584 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4588 module_init(slab_proc_init);
4592 * ksize - get the actual amount of memory allocated for a given object
4593 * @objp: Pointer to the object
4595 * kmalloc may internally round up allocations and return more memory
4596 * than requested. ksize() can be used to determine the actual amount of
4597 * memory allocated. The caller may use this additional memory, even though
4598 * a smaller amount of memory was initially specified with the kmalloc call.
4599 * The caller must guarantee that objp points to a valid object previously
4600 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4601 * must not be freed during the duration of the call.
4603 size_t ksize(const void *objp)
4606 if (unlikely(objp == ZERO_SIZE_PTR))
4609 return virt_to_cache(objp)->object_size;
4611 EXPORT_SYMBOL(ksize);