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 * Manages the objs in a slab. Placed either at the beginning of mem allocated
170 * for a slab, or allocated from an general cache.
171 * Slabs are chained into three list: fully used, partial, fully free slabs.
175 struct list_head list;
176 void *s_mem; /* including colour offset */
177 unsigned int inuse; /* num of objs active in slab */
186 * - LIFO ordering, to hand out cache-warm objects from _alloc
187 * - reduce the number of linked list operations
188 * - reduce spinlock operations
190 * The limit is stored in the per-cpu structure to reduce the data cache
197 unsigned int batchcount;
198 unsigned int touched;
201 * Must have this definition in here for the proper
202 * alignment of array_cache. Also simplifies accessing
205 * Entries should not be directly dereferenced as
206 * entries belonging to slabs marked pfmemalloc will
207 * have the lower bits set SLAB_OBJ_PFMEMALLOC
211 #define SLAB_OBJ_PFMEMALLOC 1
212 static inline bool is_obj_pfmemalloc(void *objp)
214 return (unsigned long)objp & SLAB_OBJ_PFMEMALLOC;
217 static inline void set_obj_pfmemalloc(void **objp)
219 *objp = (void *)((unsigned long)*objp | SLAB_OBJ_PFMEMALLOC);
223 static inline void clear_obj_pfmemalloc(void **objp)
225 *objp = (void *)((unsigned long)*objp & ~SLAB_OBJ_PFMEMALLOC);
229 * bootstrap: The caches do not work without cpuarrays anymore, but the
230 * cpuarrays are allocated from the generic caches...
232 #define BOOT_CPUCACHE_ENTRIES 1
233 struct arraycache_init {
234 struct array_cache cache;
235 void *entries[BOOT_CPUCACHE_ENTRIES];
239 * Need this for bootstrapping a per node allocator.
241 #define NUM_INIT_LISTS (3 * MAX_NUMNODES)
242 static struct kmem_cache_node __initdata init_kmem_cache_node[NUM_INIT_LISTS];
243 #define CACHE_CACHE 0
244 #define SIZE_AC MAX_NUMNODES
245 #define SIZE_NODE (2 * MAX_NUMNODES)
247 static int drain_freelist(struct kmem_cache *cache,
248 struct kmem_cache_node *n, int tofree);
249 static void free_block(struct kmem_cache *cachep, void **objpp, int len,
251 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
252 static void cache_reap(struct work_struct *unused);
254 static int slab_early_init = 1;
256 #define INDEX_AC kmalloc_index(sizeof(struct arraycache_init))
257 #define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
259 static void kmem_cache_node_init(struct kmem_cache_node *parent)
261 INIT_LIST_HEAD(&parent->slabs_full);
262 INIT_LIST_HEAD(&parent->slabs_partial);
263 INIT_LIST_HEAD(&parent->slabs_free);
264 parent->shared = NULL;
265 parent->alien = NULL;
266 parent->colour_next = 0;
267 spin_lock_init(&parent->list_lock);
268 parent->free_objects = 0;
269 parent->free_touched = 0;
272 #define MAKE_LIST(cachep, listp, slab, nodeid) \
274 INIT_LIST_HEAD(listp); \
275 list_splice(&(cachep->node[nodeid]->slab), listp); \
278 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
280 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
281 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
282 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
285 #define CFLGS_OFF_SLAB (0x80000000UL)
286 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
288 #define BATCHREFILL_LIMIT 16
290 * Optimization question: fewer reaps means less probability for unnessary
291 * cpucache drain/refill cycles.
293 * OTOH the cpuarrays can contain lots of objects,
294 * which could lock up otherwise freeable slabs.
296 #define REAPTIMEOUT_CPUC (2*HZ)
297 #define REAPTIMEOUT_LIST3 (4*HZ)
300 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
301 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
302 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
303 #define STATS_INC_GROWN(x) ((x)->grown++)
304 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
305 #define STATS_SET_HIGH(x) \
307 if ((x)->num_active > (x)->high_mark) \
308 (x)->high_mark = (x)->num_active; \
310 #define STATS_INC_ERR(x) ((x)->errors++)
311 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
312 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
313 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
314 #define STATS_SET_FREEABLE(x, i) \
316 if ((x)->max_freeable < i) \
317 (x)->max_freeable = i; \
319 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
320 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
321 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
322 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
324 #define STATS_INC_ACTIVE(x) do { } while (0)
325 #define STATS_DEC_ACTIVE(x) do { } while (0)
326 #define STATS_INC_ALLOCED(x) do { } while (0)
327 #define STATS_INC_GROWN(x) do { } while (0)
328 #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
329 #define STATS_SET_HIGH(x) do { } while (0)
330 #define STATS_INC_ERR(x) do { } while (0)
331 #define STATS_INC_NODEALLOCS(x) do { } while (0)
332 #define STATS_INC_NODEFREES(x) do { } while (0)
333 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
334 #define STATS_SET_FREEABLE(x, i) do { } while (0)
335 #define STATS_INC_ALLOCHIT(x) do { } while (0)
336 #define STATS_INC_ALLOCMISS(x) do { } while (0)
337 #define STATS_INC_FREEHIT(x) do { } while (0)
338 #define STATS_INC_FREEMISS(x) do { } while (0)
344 * memory layout of objects:
346 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
347 * the end of an object is aligned with the end of the real
348 * allocation. Catches writes behind the end of the allocation.
349 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
351 * cachep->obj_offset: The real object.
352 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
353 * cachep->size - 1* BYTES_PER_WORD: last caller address
354 * [BYTES_PER_WORD long]
356 static int obj_offset(struct kmem_cache *cachep)
358 return cachep->obj_offset;
361 static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
363 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
364 return (unsigned long long*) (objp + obj_offset(cachep) -
365 sizeof(unsigned long long));
368 static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
370 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
371 if (cachep->flags & SLAB_STORE_USER)
372 return (unsigned long long *)(objp + cachep->size -
373 sizeof(unsigned long long) -
375 return (unsigned long long *) (objp + cachep->size -
376 sizeof(unsigned long long));
379 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
381 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
382 return (void **)(objp + cachep->size - BYTES_PER_WORD);
387 #define obj_offset(x) 0
388 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
389 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
390 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
395 * Do not go above this order unless 0 objects fit into the slab or
396 * overridden on the command line.
398 #define SLAB_MAX_ORDER_HI 1
399 #define SLAB_MAX_ORDER_LO 0
400 static int slab_max_order = SLAB_MAX_ORDER_LO;
401 static bool slab_max_order_set __initdata;
403 static inline struct kmem_cache *virt_to_cache(const void *obj)
405 struct page *page = virt_to_head_page(obj);
406 return page->slab_cache;
409 static inline struct slab *virt_to_slab(const void *obj)
411 struct page *page = virt_to_head_page(obj);
413 VM_BUG_ON(!PageSlab(page));
414 return page->slab_page;
417 static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
420 return slab->s_mem + cache->size * idx;
424 * We want to avoid an expensive divide : (offset / cache->size)
425 * Using the fact that size is a constant for a particular cache,
426 * we can replace (offset / cache->size) by
427 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
429 static inline unsigned int obj_to_index(const struct kmem_cache *cache,
430 const struct slab *slab, void *obj)
432 u32 offset = (obj - slab->s_mem);
433 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
436 static struct arraycache_init initarray_generic =
437 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
439 /* internal cache of cache description objs */
440 static struct kmem_cache kmem_cache_boot = {
442 .limit = BOOT_CPUCACHE_ENTRIES,
444 .size = sizeof(struct kmem_cache),
445 .name = "kmem_cache",
448 #define BAD_ALIEN_MAGIC 0x01020304ul
450 #ifdef CONFIG_LOCKDEP
453 * Slab sometimes uses the kmalloc slabs to store the slab headers
454 * for other slabs "off slab".
455 * The locking for this is tricky in that it nests within the locks
456 * of all other slabs in a few places; to deal with this special
457 * locking we put on-slab caches into a separate lock-class.
459 * We set lock class for alien array caches which are up during init.
460 * The lock annotation will be lost if all cpus of a node goes down and
461 * then comes back up during hotplug
463 static struct lock_class_key on_slab_l3_key;
464 static struct lock_class_key on_slab_alc_key;
466 static struct lock_class_key debugobj_l3_key;
467 static struct lock_class_key debugobj_alc_key;
469 static void slab_set_lock_classes(struct kmem_cache *cachep,
470 struct lock_class_key *l3_key, struct lock_class_key *alc_key,
473 struct array_cache **alc;
474 struct kmem_cache_node *n;
481 lockdep_set_class(&n->list_lock, l3_key);
484 * FIXME: This check for BAD_ALIEN_MAGIC
485 * should go away when common slab code is taught to
486 * work even without alien caches.
487 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
488 * for alloc_alien_cache,
490 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
494 lockdep_set_class(&alc[r]->lock, alc_key);
498 static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
500 slab_set_lock_classes(cachep, &debugobj_l3_key, &debugobj_alc_key, node);
503 static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
507 for_each_online_node(node)
508 slab_set_debugobj_lock_classes_node(cachep, node);
511 static void init_node_lock_keys(int q)
518 for (i = 1; i <= KMALLOC_SHIFT_HIGH; i++) {
519 struct kmem_cache_node *n;
520 struct kmem_cache *cache = kmalloc_caches[i];
526 if (!n || OFF_SLAB(cache))
529 slab_set_lock_classes(cache, &on_slab_l3_key,
530 &on_slab_alc_key, q);
534 static void on_slab_lock_classes_node(struct kmem_cache *cachep, int q)
536 if (!cachep->node[q])
539 slab_set_lock_classes(cachep, &on_slab_l3_key,
540 &on_slab_alc_key, q);
543 static inline void on_slab_lock_classes(struct kmem_cache *cachep)
547 VM_BUG_ON(OFF_SLAB(cachep));
549 on_slab_lock_classes_node(cachep, node);
552 static inline void init_lock_keys(void)
557 init_node_lock_keys(node);
560 static void init_node_lock_keys(int q)
564 static inline void init_lock_keys(void)
568 static inline void on_slab_lock_classes(struct kmem_cache *cachep)
572 static inline void on_slab_lock_classes_node(struct kmem_cache *cachep, int node)
576 static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
580 static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
585 static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
587 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
589 return cachep->array[smp_processor_id()];
592 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
594 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(unsigned int), align);
598 * Calculate the number of objects and left-over bytes for a given buffer size.
600 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
601 size_t align, int flags, size_t *left_over,
606 size_t slab_size = PAGE_SIZE << gfporder;
609 * The slab management structure can be either off the slab or
610 * on it. For the latter case, the memory allocated for a
614 * - One unsigned int for each object
615 * - Padding to respect alignment of @align
616 * - @buffer_size bytes for each object
618 * If the slab management structure is off the slab, then the
619 * alignment will already be calculated into the size. Because
620 * the slabs are all pages aligned, the objects will be at the
621 * correct alignment when allocated.
623 if (flags & CFLGS_OFF_SLAB) {
625 nr_objs = slab_size / buffer_size;
629 * Ignore padding for the initial guess. The padding
630 * is at most @align-1 bytes, and @buffer_size is at
631 * least @align. In the worst case, this result will
632 * be one greater than the number of objects that fit
633 * into the memory allocation when taking the padding
636 nr_objs = (slab_size - sizeof(struct slab)) /
637 (buffer_size + sizeof(unsigned int));
640 * This calculated number will be either the right
641 * amount, or one greater than what we want.
643 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
647 mgmt_size = slab_mgmt_size(nr_objs, align);
650 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
654 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
656 static void __slab_error(const char *function, struct kmem_cache *cachep,
659 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
660 function, cachep->name, msg);
662 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
667 * By default on NUMA we use alien caches to stage the freeing of
668 * objects allocated from other nodes. This causes massive memory
669 * inefficiencies when using fake NUMA setup to split memory into a
670 * large number of small nodes, so it can be disabled on the command
674 static int use_alien_caches __read_mostly = 1;
675 static int __init noaliencache_setup(char *s)
677 use_alien_caches = 0;
680 __setup("noaliencache", noaliencache_setup);
682 static int __init slab_max_order_setup(char *str)
684 get_option(&str, &slab_max_order);
685 slab_max_order = slab_max_order < 0 ? 0 :
686 min(slab_max_order, MAX_ORDER - 1);
687 slab_max_order_set = true;
691 __setup("slab_max_order=", slab_max_order_setup);
695 * Special reaping functions for NUMA systems called from cache_reap().
696 * These take care of doing round robin flushing of alien caches (containing
697 * objects freed on different nodes from which they were allocated) and the
698 * flushing of remote pcps by calling drain_node_pages.
700 static DEFINE_PER_CPU(unsigned long, slab_reap_node);
702 static void init_reap_node(int cpu)
706 node = next_node(cpu_to_mem(cpu), node_online_map);
707 if (node == MAX_NUMNODES)
708 node = first_node(node_online_map);
710 per_cpu(slab_reap_node, cpu) = node;
713 static void next_reap_node(void)
715 int node = __this_cpu_read(slab_reap_node);
717 node = next_node(node, node_online_map);
718 if (unlikely(node >= MAX_NUMNODES))
719 node = first_node(node_online_map);
720 __this_cpu_write(slab_reap_node, node);
724 #define init_reap_node(cpu) do { } while (0)
725 #define next_reap_node(void) do { } while (0)
729 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
730 * via the workqueue/eventd.
731 * Add the CPU number into the expiration time to minimize the possibility of
732 * the CPUs getting into lockstep and contending for the global cache chain
735 static void start_cpu_timer(int cpu)
737 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
740 * When this gets called from do_initcalls via cpucache_init(),
741 * init_workqueues() has already run, so keventd will be setup
744 if (keventd_up() && reap_work->work.func == NULL) {
746 INIT_DEFERRABLE_WORK(reap_work, cache_reap);
747 schedule_delayed_work_on(cpu, reap_work,
748 __round_jiffies_relative(HZ, cpu));
752 static struct array_cache *alloc_arraycache(int node, int entries,
753 int batchcount, gfp_t gfp)
755 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
756 struct array_cache *nc = NULL;
758 nc = kmalloc_node(memsize, gfp, node);
760 * The array_cache structures contain pointers to free object.
761 * However, when such objects are allocated or transferred to another
762 * cache the pointers are not cleared and they could be counted as
763 * valid references during a kmemleak scan. Therefore, kmemleak must
764 * not scan such objects.
766 kmemleak_no_scan(nc);
770 nc->batchcount = batchcount;
772 spin_lock_init(&nc->lock);
777 static inline bool is_slab_pfmemalloc(struct slab *slabp)
779 struct page *page = virt_to_page(slabp->s_mem);
781 return PageSlabPfmemalloc(page);
784 /* Clears pfmemalloc_active if no slabs have pfmalloc set */
785 static void recheck_pfmemalloc_active(struct kmem_cache *cachep,
786 struct array_cache *ac)
788 struct kmem_cache_node *n = cachep->node[numa_mem_id()];
792 if (!pfmemalloc_active)
795 spin_lock_irqsave(&n->list_lock, flags);
796 list_for_each_entry(slabp, &n->slabs_full, list)
797 if (is_slab_pfmemalloc(slabp))
800 list_for_each_entry(slabp, &n->slabs_partial, list)
801 if (is_slab_pfmemalloc(slabp))
804 list_for_each_entry(slabp, &n->slabs_free, list)
805 if (is_slab_pfmemalloc(slabp))
808 pfmemalloc_active = false;
810 spin_unlock_irqrestore(&n->list_lock, flags);
813 static void *__ac_get_obj(struct kmem_cache *cachep, struct array_cache *ac,
814 gfp_t flags, bool force_refill)
817 void *objp = ac->entry[--ac->avail];
819 /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
820 if (unlikely(is_obj_pfmemalloc(objp))) {
821 struct kmem_cache_node *n;
823 if (gfp_pfmemalloc_allowed(flags)) {
824 clear_obj_pfmemalloc(&objp);
828 /* The caller cannot use PFMEMALLOC objects, find another one */
829 for (i = 0; i < ac->avail; i++) {
830 /* If a !PFMEMALLOC object is found, swap them */
831 if (!is_obj_pfmemalloc(ac->entry[i])) {
833 ac->entry[i] = ac->entry[ac->avail];
834 ac->entry[ac->avail] = objp;
840 * If there are empty slabs on the slabs_free list and we are
841 * being forced to refill the cache, mark this one !pfmemalloc.
843 n = cachep->node[numa_mem_id()];
844 if (!list_empty(&n->slabs_free) && force_refill) {
845 struct slab *slabp = virt_to_slab(objp);
846 ClearPageSlabPfmemalloc(virt_to_head_page(slabp->s_mem));
847 clear_obj_pfmemalloc(&objp);
848 recheck_pfmemalloc_active(cachep, ac);
852 /* No !PFMEMALLOC objects available */
860 static inline void *ac_get_obj(struct kmem_cache *cachep,
861 struct array_cache *ac, gfp_t flags, bool force_refill)
865 if (unlikely(sk_memalloc_socks()))
866 objp = __ac_get_obj(cachep, ac, flags, force_refill);
868 objp = ac->entry[--ac->avail];
873 static void *__ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
876 if (unlikely(pfmemalloc_active)) {
877 /* Some pfmemalloc slabs exist, check if this is one */
878 struct slab *slabp = virt_to_slab(objp);
879 struct page *page = virt_to_head_page(slabp->s_mem);
880 if (PageSlabPfmemalloc(page))
881 set_obj_pfmemalloc(&objp);
887 static inline void ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
890 if (unlikely(sk_memalloc_socks()))
891 objp = __ac_put_obj(cachep, ac, objp);
893 ac->entry[ac->avail++] = objp;
897 * Transfer objects in one arraycache to another.
898 * Locking must be handled by the caller.
900 * Return the number of entries transferred.
902 static int transfer_objects(struct array_cache *to,
903 struct array_cache *from, unsigned int max)
905 /* Figure out how many entries to transfer */
906 int nr = min3(from->avail, max, to->limit - to->avail);
911 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
921 #define drain_alien_cache(cachep, alien) do { } while (0)
922 #define reap_alien(cachep, n) do { } while (0)
924 static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
926 return (struct array_cache **)BAD_ALIEN_MAGIC;
929 static inline void free_alien_cache(struct array_cache **ac_ptr)
933 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
938 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
944 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
945 gfp_t flags, int nodeid)
950 #else /* CONFIG_NUMA */
952 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
953 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
955 static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
957 struct array_cache **ac_ptr;
958 int memsize = sizeof(void *) * nr_node_ids;
963 ac_ptr = kzalloc_node(memsize, gfp, node);
966 if (i == node || !node_online(i))
968 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
970 for (i--; i >= 0; i--)
980 static void free_alien_cache(struct array_cache **ac_ptr)
991 static void __drain_alien_cache(struct kmem_cache *cachep,
992 struct array_cache *ac, int node)
994 struct kmem_cache_node *n = cachep->node[node];
997 spin_lock(&n->list_lock);
999 * Stuff objects into the remote nodes shared array first.
1000 * That way we could avoid the overhead of putting the objects
1001 * into the free lists and getting them back later.
1004 transfer_objects(n->shared, ac, ac->limit);
1006 free_block(cachep, ac->entry, ac->avail, node);
1008 spin_unlock(&n->list_lock);
1013 * Called from cache_reap() to regularly drain alien caches round robin.
1015 static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *n)
1017 int node = __this_cpu_read(slab_reap_node);
1020 struct array_cache *ac = n->alien[node];
1022 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1023 __drain_alien_cache(cachep, ac, node);
1024 spin_unlock_irq(&ac->lock);
1029 static void drain_alien_cache(struct kmem_cache *cachep,
1030 struct array_cache **alien)
1033 struct array_cache *ac;
1034 unsigned long flags;
1036 for_each_online_node(i) {
1039 spin_lock_irqsave(&ac->lock, flags);
1040 __drain_alien_cache(cachep, ac, i);
1041 spin_unlock_irqrestore(&ac->lock, flags);
1046 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1048 int nodeid = page_to_nid(virt_to_page(objp));
1049 struct kmem_cache_node *n;
1050 struct array_cache *alien = NULL;
1053 node = numa_mem_id();
1056 * Make sure we are not freeing a object from another node to the array
1057 * cache on this cpu.
1059 if (likely(nodeid == node))
1062 n = cachep->node[node];
1063 STATS_INC_NODEFREES(cachep);
1064 if (n->alien && n->alien[nodeid]) {
1065 alien = n->alien[nodeid];
1066 spin_lock(&alien->lock);
1067 if (unlikely(alien->avail == alien->limit)) {
1068 STATS_INC_ACOVERFLOW(cachep);
1069 __drain_alien_cache(cachep, alien, nodeid);
1071 ac_put_obj(cachep, alien, objp);
1072 spin_unlock(&alien->lock);
1074 spin_lock(&(cachep->node[nodeid])->list_lock);
1075 free_block(cachep, &objp, 1, nodeid);
1076 spin_unlock(&(cachep->node[nodeid])->list_lock);
1083 * Allocates and initializes node for a node on each slab cache, used for
1084 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
1085 * will be allocated off-node since memory is not yet online for the new node.
1086 * When hotplugging memory or a cpu, existing node are not replaced if
1089 * Must hold slab_mutex.
1091 static int init_cache_node_node(int node)
1093 struct kmem_cache *cachep;
1094 struct kmem_cache_node *n;
1095 const int memsize = sizeof(struct kmem_cache_node);
1097 list_for_each_entry(cachep, &slab_caches, list) {
1099 * Set up the size64 kmemlist for cpu before we can
1100 * begin anything. Make sure some other cpu on this
1101 * node has not already allocated this
1103 if (!cachep->node[node]) {
1104 n = kmalloc_node(memsize, GFP_KERNEL, node);
1107 kmem_cache_node_init(n);
1108 n->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1109 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1112 * The l3s don't come and go as CPUs come and
1113 * go. slab_mutex is sufficient
1116 cachep->node[node] = n;
1119 spin_lock_irq(&cachep->node[node]->list_lock);
1120 cachep->node[node]->free_limit =
1121 (1 + nr_cpus_node(node)) *
1122 cachep->batchcount + cachep->num;
1123 spin_unlock_irq(&cachep->node[node]->list_lock);
1128 static inline int slabs_tofree(struct kmem_cache *cachep,
1129 struct kmem_cache_node *n)
1131 return (n->free_objects + cachep->num - 1) / cachep->num;
1134 static void cpuup_canceled(long cpu)
1136 struct kmem_cache *cachep;
1137 struct kmem_cache_node *n = NULL;
1138 int node = cpu_to_mem(cpu);
1139 const struct cpumask *mask = cpumask_of_node(node);
1141 list_for_each_entry(cachep, &slab_caches, list) {
1142 struct array_cache *nc;
1143 struct array_cache *shared;
1144 struct array_cache **alien;
1146 /* cpu is dead; no one can alloc from it. */
1147 nc = cachep->array[cpu];
1148 cachep->array[cpu] = NULL;
1149 n = cachep->node[node];
1152 goto free_array_cache;
1154 spin_lock_irq(&n->list_lock);
1156 /* Free limit for this kmem_cache_node */
1157 n->free_limit -= cachep->batchcount;
1159 free_block(cachep, nc->entry, nc->avail, node);
1161 if (!cpumask_empty(mask)) {
1162 spin_unlock_irq(&n->list_lock);
1163 goto free_array_cache;
1168 free_block(cachep, shared->entry,
1169 shared->avail, node);
1176 spin_unlock_irq(&n->list_lock);
1180 drain_alien_cache(cachep, alien);
1181 free_alien_cache(alien);
1187 * In the previous loop, all the objects were freed to
1188 * the respective cache's slabs, now we can go ahead and
1189 * shrink each nodelist to its limit.
1191 list_for_each_entry(cachep, &slab_caches, list) {
1192 n = cachep->node[node];
1195 drain_freelist(cachep, n, slabs_tofree(cachep, n));
1199 static int cpuup_prepare(long cpu)
1201 struct kmem_cache *cachep;
1202 struct kmem_cache_node *n = NULL;
1203 int node = cpu_to_mem(cpu);
1207 * We need to do this right in the beginning since
1208 * alloc_arraycache's are going to use this list.
1209 * kmalloc_node allows us to add the slab to the right
1210 * kmem_cache_node and not this cpu's kmem_cache_node
1212 err = init_cache_node_node(node);
1217 * Now we can go ahead with allocating the shared arrays and
1220 list_for_each_entry(cachep, &slab_caches, list) {
1221 struct array_cache *nc;
1222 struct array_cache *shared = NULL;
1223 struct array_cache **alien = NULL;
1225 nc = alloc_arraycache(node, cachep->limit,
1226 cachep->batchcount, GFP_KERNEL);
1229 if (cachep->shared) {
1230 shared = alloc_arraycache(node,
1231 cachep->shared * cachep->batchcount,
1232 0xbaadf00d, GFP_KERNEL);
1238 if (use_alien_caches) {
1239 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
1246 cachep->array[cpu] = nc;
1247 n = cachep->node[node];
1250 spin_lock_irq(&n->list_lock);
1253 * We are serialised from CPU_DEAD or
1254 * CPU_UP_CANCELLED by the cpucontrol lock
1265 spin_unlock_irq(&n->list_lock);
1267 free_alien_cache(alien);
1268 if (cachep->flags & SLAB_DEBUG_OBJECTS)
1269 slab_set_debugobj_lock_classes_node(cachep, node);
1270 else if (!OFF_SLAB(cachep) &&
1271 !(cachep->flags & SLAB_DESTROY_BY_RCU))
1272 on_slab_lock_classes_node(cachep, node);
1274 init_node_lock_keys(node);
1278 cpuup_canceled(cpu);
1282 static int cpuup_callback(struct notifier_block *nfb,
1283 unsigned long action, void *hcpu)
1285 long cpu = (long)hcpu;
1289 case CPU_UP_PREPARE:
1290 case CPU_UP_PREPARE_FROZEN:
1291 mutex_lock(&slab_mutex);
1292 err = cpuup_prepare(cpu);
1293 mutex_unlock(&slab_mutex);
1296 case CPU_ONLINE_FROZEN:
1297 start_cpu_timer(cpu);
1299 #ifdef CONFIG_HOTPLUG_CPU
1300 case CPU_DOWN_PREPARE:
1301 case CPU_DOWN_PREPARE_FROZEN:
1303 * Shutdown cache reaper. Note that the slab_mutex is
1304 * held so that if cache_reap() is invoked it cannot do
1305 * anything expensive but will only modify reap_work
1306 * and reschedule the timer.
1308 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1309 /* Now the cache_reaper is guaranteed to be not running. */
1310 per_cpu(slab_reap_work, cpu).work.func = NULL;
1312 case CPU_DOWN_FAILED:
1313 case CPU_DOWN_FAILED_FROZEN:
1314 start_cpu_timer(cpu);
1317 case CPU_DEAD_FROZEN:
1319 * Even if all the cpus of a node are down, we don't free the
1320 * kmem_cache_node of any cache. This to avoid a race between
1321 * cpu_down, and a kmalloc allocation from another cpu for
1322 * memory from the node of the cpu going down. The node
1323 * structure is usually allocated from kmem_cache_create() and
1324 * gets destroyed at kmem_cache_destroy().
1328 case CPU_UP_CANCELED:
1329 case CPU_UP_CANCELED_FROZEN:
1330 mutex_lock(&slab_mutex);
1331 cpuup_canceled(cpu);
1332 mutex_unlock(&slab_mutex);
1335 return notifier_from_errno(err);
1338 static struct notifier_block cpucache_notifier = {
1339 &cpuup_callback, NULL, 0
1342 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1344 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1345 * Returns -EBUSY if all objects cannot be drained so that the node is not
1348 * Must hold slab_mutex.
1350 static int __meminit drain_cache_node_node(int node)
1352 struct kmem_cache *cachep;
1355 list_for_each_entry(cachep, &slab_caches, list) {
1356 struct kmem_cache_node *n;
1358 n = cachep->node[node];
1362 drain_freelist(cachep, n, slabs_tofree(cachep, n));
1364 if (!list_empty(&n->slabs_full) ||
1365 !list_empty(&n->slabs_partial)) {
1373 static int __meminit slab_memory_callback(struct notifier_block *self,
1374 unsigned long action, void *arg)
1376 struct memory_notify *mnb = arg;
1380 nid = mnb->status_change_nid;
1385 case MEM_GOING_ONLINE:
1386 mutex_lock(&slab_mutex);
1387 ret = init_cache_node_node(nid);
1388 mutex_unlock(&slab_mutex);
1390 case MEM_GOING_OFFLINE:
1391 mutex_lock(&slab_mutex);
1392 ret = drain_cache_node_node(nid);
1393 mutex_unlock(&slab_mutex);
1397 case MEM_CANCEL_ONLINE:
1398 case MEM_CANCEL_OFFLINE:
1402 return notifier_from_errno(ret);
1404 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1407 * swap the static kmem_cache_node with kmalloced memory
1409 static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list,
1412 struct kmem_cache_node *ptr;
1414 ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid);
1417 memcpy(ptr, list, sizeof(struct kmem_cache_node));
1419 * Do not assume that spinlocks can be initialized via memcpy:
1421 spin_lock_init(&ptr->list_lock);
1423 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1424 cachep->node[nodeid] = ptr;
1428 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1429 * size of kmem_cache_node.
1431 static void __init set_up_node(struct kmem_cache *cachep, int index)
1435 for_each_online_node(node) {
1436 cachep->node[node] = &init_kmem_cache_node[index + node];
1437 cachep->node[node]->next_reap = jiffies +
1439 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1444 * The memory after the last cpu cache pointer is used for the
1447 static void setup_node_pointer(struct kmem_cache *cachep)
1449 cachep->node = (struct kmem_cache_node **)&cachep->array[nr_cpu_ids];
1453 * Initialisation. Called after the page allocator have been initialised and
1454 * before smp_init().
1456 void __init kmem_cache_init(void)
1460 BUILD_BUG_ON(sizeof(((struct page *)NULL)->lru) <
1461 sizeof(struct rcu_head));
1462 kmem_cache = &kmem_cache_boot;
1463 setup_node_pointer(kmem_cache);
1465 if (num_possible_nodes() == 1)
1466 use_alien_caches = 0;
1468 for (i = 0; i < NUM_INIT_LISTS; i++)
1469 kmem_cache_node_init(&init_kmem_cache_node[i]);
1471 set_up_node(kmem_cache, CACHE_CACHE);
1474 * Fragmentation resistance on low memory - only use bigger
1475 * page orders on machines with more than 32MB of memory if
1476 * not overridden on the command line.
1478 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
1479 slab_max_order = SLAB_MAX_ORDER_HI;
1481 /* Bootstrap is tricky, because several objects are allocated
1482 * from caches that do not exist yet:
1483 * 1) initialize the kmem_cache cache: it contains the struct
1484 * kmem_cache structures of all caches, except kmem_cache itself:
1485 * kmem_cache is statically allocated.
1486 * Initially an __init data area is used for the head array and the
1487 * kmem_cache_node structures, it's replaced with a kmalloc allocated
1488 * array at the end of the bootstrap.
1489 * 2) Create the first kmalloc cache.
1490 * The struct kmem_cache for the new cache is allocated normally.
1491 * An __init data area is used for the head array.
1492 * 3) Create the remaining kmalloc caches, with minimally sized
1494 * 4) Replace the __init data head arrays for kmem_cache and the first
1495 * kmalloc cache with kmalloc allocated arrays.
1496 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1497 * the other cache's with kmalloc allocated memory.
1498 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1501 /* 1) create the kmem_cache */
1504 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1506 create_boot_cache(kmem_cache, "kmem_cache",
1507 offsetof(struct kmem_cache, array[nr_cpu_ids]) +
1508 nr_node_ids * sizeof(struct kmem_cache_node *),
1509 SLAB_HWCACHE_ALIGN);
1510 list_add(&kmem_cache->list, &slab_caches);
1512 /* 2+3) create the kmalloc caches */
1515 * Initialize the caches that provide memory for the array cache and the
1516 * kmem_cache_node structures first. Without this, further allocations will
1520 kmalloc_caches[INDEX_AC] = create_kmalloc_cache("kmalloc-ac",
1521 kmalloc_size(INDEX_AC), ARCH_KMALLOC_FLAGS);
1523 if (INDEX_AC != INDEX_NODE)
1524 kmalloc_caches[INDEX_NODE] =
1525 create_kmalloc_cache("kmalloc-node",
1526 kmalloc_size(INDEX_NODE), ARCH_KMALLOC_FLAGS);
1528 slab_early_init = 0;
1530 /* 4) Replace the bootstrap head arrays */
1532 struct array_cache *ptr;
1534 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1536 memcpy(ptr, cpu_cache_get(kmem_cache),
1537 sizeof(struct arraycache_init));
1539 * Do not assume that spinlocks can be initialized via memcpy:
1541 spin_lock_init(&ptr->lock);
1543 kmem_cache->array[smp_processor_id()] = ptr;
1545 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1547 BUG_ON(cpu_cache_get(kmalloc_caches[INDEX_AC])
1548 != &initarray_generic.cache);
1549 memcpy(ptr, cpu_cache_get(kmalloc_caches[INDEX_AC]),
1550 sizeof(struct arraycache_init));
1552 * Do not assume that spinlocks can be initialized via memcpy:
1554 spin_lock_init(&ptr->lock);
1556 kmalloc_caches[INDEX_AC]->array[smp_processor_id()] = ptr;
1558 /* 5) Replace the bootstrap kmem_cache_node */
1562 for_each_online_node(nid) {
1563 init_list(kmem_cache, &init_kmem_cache_node[CACHE_CACHE + nid], nid);
1565 init_list(kmalloc_caches[INDEX_AC],
1566 &init_kmem_cache_node[SIZE_AC + nid], nid);
1568 if (INDEX_AC != INDEX_NODE) {
1569 init_list(kmalloc_caches[INDEX_NODE],
1570 &init_kmem_cache_node[SIZE_NODE + nid], nid);
1575 create_kmalloc_caches(ARCH_KMALLOC_FLAGS);
1578 void __init kmem_cache_init_late(void)
1580 struct kmem_cache *cachep;
1584 /* 6) resize the head arrays to their final sizes */
1585 mutex_lock(&slab_mutex);
1586 list_for_each_entry(cachep, &slab_caches, list)
1587 if (enable_cpucache(cachep, GFP_NOWAIT))
1589 mutex_unlock(&slab_mutex);
1591 /* Annotate slab for lockdep -- annotate the malloc caches */
1598 * Register a cpu startup notifier callback that initializes
1599 * cpu_cache_get for all new cpus
1601 register_cpu_notifier(&cpucache_notifier);
1605 * Register a memory hotplug callback that initializes and frees
1608 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1612 * The reap timers are started later, with a module init call: That part
1613 * of the kernel is not yet operational.
1617 static int __init cpucache_init(void)
1622 * Register the timers that return unneeded pages to the page allocator
1624 for_each_online_cpu(cpu)
1625 start_cpu_timer(cpu);
1631 __initcall(cpucache_init);
1633 static noinline void
1634 slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1636 struct kmem_cache_node *n;
1638 unsigned long flags;
1642 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1644 printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n",
1645 cachep->name, cachep->size, cachep->gfporder);
1647 for_each_online_node(node) {
1648 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1649 unsigned long active_slabs = 0, num_slabs = 0;
1651 n = cachep->node[node];
1655 spin_lock_irqsave(&n->list_lock, flags);
1656 list_for_each_entry(slabp, &n->slabs_full, list) {
1657 active_objs += cachep->num;
1660 list_for_each_entry(slabp, &n->slabs_partial, list) {
1661 active_objs += slabp->inuse;
1664 list_for_each_entry(slabp, &n->slabs_free, list)
1667 free_objects += n->free_objects;
1668 spin_unlock_irqrestore(&n->list_lock, flags);
1670 num_slabs += active_slabs;
1671 num_objs = num_slabs * cachep->num;
1673 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1674 node, active_slabs, num_slabs, active_objs, num_objs,
1680 * Interface to system's page allocator. No need to hold the cache-lock.
1682 * If we requested dmaable memory, we will get it. Even if we
1683 * did not request dmaable memory, we might get it, but that
1684 * would be relatively rare and ignorable.
1686 static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags,
1692 flags |= cachep->allocflags;
1693 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1694 flags |= __GFP_RECLAIMABLE;
1696 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1698 if (!(flags & __GFP_NOWARN) && printk_ratelimit())
1699 slab_out_of_memory(cachep, flags, nodeid);
1703 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1704 if (unlikely(page->pfmemalloc))
1705 pfmemalloc_active = true;
1707 nr_pages = (1 << cachep->gfporder);
1708 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1709 add_zone_page_state(page_zone(page),
1710 NR_SLAB_RECLAIMABLE, nr_pages);
1712 add_zone_page_state(page_zone(page),
1713 NR_SLAB_UNRECLAIMABLE, nr_pages);
1714 __SetPageSlab(page);
1715 if (page->pfmemalloc)
1716 SetPageSlabPfmemalloc(page);
1717 memcg_bind_pages(cachep, cachep->gfporder);
1719 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1720 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1723 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1725 kmemcheck_mark_unallocated_pages(page, nr_pages);
1732 * Interface to system's page release.
1734 static void kmem_freepages(struct kmem_cache *cachep, struct page *page)
1736 const unsigned long nr_freed = (1 << cachep->gfporder);
1738 kmemcheck_free_shadow(page, cachep->gfporder);
1740 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1741 sub_zone_page_state(page_zone(page),
1742 NR_SLAB_RECLAIMABLE, nr_freed);
1744 sub_zone_page_state(page_zone(page),
1745 NR_SLAB_UNRECLAIMABLE, nr_freed);
1747 BUG_ON(!PageSlab(page));
1748 __ClearPageSlabPfmemalloc(page);
1749 __ClearPageSlab(page);
1751 memcg_release_pages(cachep, cachep->gfporder);
1752 if (current->reclaim_state)
1753 current->reclaim_state->reclaimed_slab += nr_freed;
1754 __free_memcg_kmem_pages(page, cachep->gfporder);
1757 static void kmem_rcu_free(struct rcu_head *head)
1759 struct kmem_cache *cachep;
1762 page = container_of(head, struct page, rcu_head);
1763 cachep = page->slab_cache;
1765 kmem_freepages(cachep, page);
1770 #ifdef CONFIG_DEBUG_PAGEALLOC
1771 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1772 unsigned long caller)
1774 int size = cachep->object_size;
1776 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1778 if (size < 5 * sizeof(unsigned long))
1781 *addr++ = 0x12345678;
1783 *addr++ = smp_processor_id();
1784 size -= 3 * sizeof(unsigned long);
1786 unsigned long *sptr = &caller;
1787 unsigned long svalue;
1789 while (!kstack_end(sptr)) {
1791 if (kernel_text_address(svalue)) {
1793 size -= sizeof(unsigned long);
1794 if (size <= sizeof(unsigned long))
1800 *addr++ = 0x87654321;
1804 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1806 int size = cachep->object_size;
1807 addr = &((char *)addr)[obj_offset(cachep)];
1809 memset(addr, val, size);
1810 *(unsigned char *)(addr + size - 1) = POISON_END;
1813 static void dump_line(char *data, int offset, int limit)
1816 unsigned char error = 0;
1819 printk(KERN_ERR "%03x: ", offset);
1820 for (i = 0; i < limit; i++) {
1821 if (data[offset + i] != POISON_FREE) {
1822 error = data[offset + i];
1826 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1827 &data[offset], limit, 1);
1829 if (bad_count == 1) {
1830 error ^= POISON_FREE;
1831 if (!(error & (error - 1))) {
1832 printk(KERN_ERR "Single bit error detected. Probably "
1835 printk(KERN_ERR "Run memtest86+ or a similar memory "
1838 printk(KERN_ERR "Run a memory test tool.\n");
1847 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1852 if (cachep->flags & SLAB_RED_ZONE) {
1853 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
1854 *dbg_redzone1(cachep, objp),
1855 *dbg_redzone2(cachep, objp));
1858 if (cachep->flags & SLAB_STORE_USER) {
1859 printk(KERN_ERR "Last user: [<%p>](%pSR)\n",
1860 *dbg_userword(cachep, objp),
1861 *dbg_userword(cachep, objp));
1863 realobj = (char *)objp + obj_offset(cachep);
1864 size = cachep->object_size;
1865 for (i = 0; i < size && lines; i += 16, lines--) {
1868 if (i + limit > size)
1870 dump_line(realobj, i, limit);
1874 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1880 realobj = (char *)objp + obj_offset(cachep);
1881 size = cachep->object_size;
1883 for (i = 0; i < size; i++) {
1884 char exp = POISON_FREE;
1887 if (realobj[i] != exp) {
1893 "Slab corruption (%s): %s start=%p, len=%d\n",
1894 print_tainted(), cachep->name, realobj, size);
1895 print_objinfo(cachep, objp, 0);
1897 /* Hexdump the affected line */
1900 if (i + limit > size)
1902 dump_line(realobj, i, limit);
1905 /* Limit to 5 lines */
1911 /* Print some data about the neighboring objects, if they
1914 struct slab *slabp = virt_to_slab(objp);
1917 objnr = obj_to_index(cachep, slabp, objp);
1919 objp = index_to_obj(cachep, slabp, objnr - 1);
1920 realobj = (char *)objp + obj_offset(cachep);
1921 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1923 print_objinfo(cachep, objp, 2);
1925 if (objnr + 1 < cachep->num) {
1926 objp = index_to_obj(cachep, slabp, objnr + 1);
1927 realobj = (char *)objp + obj_offset(cachep);
1928 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1930 print_objinfo(cachep, objp, 2);
1937 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
1940 for (i = 0; i < cachep->num; i++) {
1941 void *objp = index_to_obj(cachep, slabp, i);
1943 if (cachep->flags & SLAB_POISON) {
1944 #ifdef CONFIG_DEBUG_PAGEALLOC
1945 if (cachep->size % PAGE_SIZE == 0 &&
1947 kernel_map_pages(virt_to_page(objp),
1948 cachep->size / PAGE_SIZE, 1);
1950 check_poison_obj(cachep, objp);
1952 check_poison_obj(cachep, objp);
1955 if (cachep->flags & SLAB_RED_ZONE) {
1956 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1957 slab_error(cachep, "start of a freed object "
1959 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1960 slab_error(cachep, "end of a freed object "
1966 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
1972 * slab_destroy - destroy and release all objects in a slab
1973 * @cachep: cache pointer being destroyed
1974 * @slabp: slab pointer being destroyed
1976 * Destroy all the objs in a slab, and release the mem back to the system.
1977 * Before calling the slab must have been unlinked from the cache. The
1978 * cache-lock is not held/needed.
1980 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
1982 struct page *page = virt_to_head_page(slabp->s_mem);
1984 slab_destroy_debugcheck(cachep, slabp);
1985 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1986 struct rcu_head *head;
1989 * RCU free overloads the RCU head over the LRU.
1990 * slab_page has been overloeaded over the LRU,
1991 * however it is not used from now on so that
1992 * we can use it safely.
1994 head = (void *)&page->rcu_head;
1995 call_rcu(head, kmem_rcu_free);
1998 kmem_freepages(cachep, page);
2002 * From now on, we don't use slab management
2003 * although actual page can be freed in rcu context
2005 if (OFF_SLAB(cachep))
2006 kmem_cache_free(cachep->slabp_cache, slabp);
2010 * calculate_slab_order - calculate size (page order) of slabs
2011 * @cachep: pointer to the cache that is being created
2012 * @size: size of objects to be created in this cache.
2013 * @align: required alignment for the objects.
2014 * @flags: slab allocation flags
2016 * Also calculates the number of objects per slab.
2018 * This could be made much more intelligent. For now, try to avoid using
2019 * high order pages for slabs. When the gfp() functions are more friendly
2020 * towards high-order requests, this should be changed.
2022 static size_t calculate_slab_order(struct kmem_cache *cachep,
2023 size_t size, size_t align, unsigned long flags)
2025 unsigned long offslab_limit;
2026 size_t left_over = 0;
2029 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
2033 cache_estimate(gfporder, size, align, flags, &remainder, &num);
2037 if (flags & CFLGS_OFF_SLAB) {
2039 * Max number of objs-per-slab for caches which
2040 * use off-slab slabs. Needed to avoid a possible
2041 * looping condition in cache_grow().
2043 offslab_limit = size - sizeof(struct slab);
2044 offslab_limit /= sizeof(unsigned int);
2046 if (num > offslab_limit)
2050 /* Found something acceptable - save it away */
2052 cachep->gfporder = gfporder;
2053 left_over = remainder;
2056 * A VFS-reclaimable slab tends to have most allocations
2057 * as GFP_NOFS and we really don't want to have to be allocating
2058 * higher-order pages when we are unable to shrink dcache.
2060 if (flags & SLAB_RECLAIM_ACCOUNT)
2064 * Large number of objects is good, but very large slabs are
2065 * currently bad for the gfp()s.
2067 if (gfporder >= slab_max_order)
2071 * Acceptable internal fragmentation?
2073 if (left_over * 8 <= (PAGE_SIZE << gfporder))
2079 static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
2081 if (slab_state >= FULL)
2082 return enable_cpucache(cachep, gfp);
2084 if (slab_state == DOWN) {
2086 * Note: Creation of first cache (kmem_cache).
2087 * The setup_node is taken care
2088 * of by the caller of __kmem_cache_create
2090 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2091 slab_state = PARTIAL;
2092 } else if (slab_state == PARTIAL) {
2094 * Note: the second kmem_cache_create must create the cache
2095 * that's used by kmalloc(24), otherwise the creation of
2096 * further caches will BUG().
2098 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2101 * If the cache that's used by kmalloc(sizeof(kmem_cache_node)) is
2102 * the second cache, then we need to set up all its node/,
2103 * otherwise the creation of further caches will BUG().
2105 set_up_node(cachep, SIZE_AC);
2106 if (INDEX_AC == INDEX_NODE)
2107 slab_state = PARTIAL_NODE;
2109 slab_state = PARTIAL_ARRAYCACHE;
2111 /* Remaining boot caches */
2112 cachep->array[smp_processor_id()] =
2113 kmalloc(sizeof(struct arraycache_init), gfp);
2115 if (slab_state == PARTIAL_ARRAYCACHE) {
2116 set_up_node(cachep, SIZE_NODE);
2117 slab_state = PARTIAL_NODE;
2120 for_each_online_node(node) {
2121 cachep->node[node] =
2122 kmalloc_node(sizeof(struct kmem_cache_node),
2124 BUG_ON(!cachep->node[node]);
2125 kmem_cache_node_init(cachep->node[node]);
2129 cachep->node[numa_mem_id()]->next_reap =
2130 jiffies + REAPTIMEOUT_LIST3 +
2131 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2133 cpu_cache_get(cachep)->avail = 0;
2134 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2135 cpu_cache_get(cachep)->batchcount = 1;
2136 cpu_cache_get(cachep)->touched = 0;
2137 cachep->batchcount = 1;
2138 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2143 * __kmem_cache_create - Create a cache.
2144 * @cachep: cache management descriptor
2145 * @flags: SLAB flags
2147 * Returns a ptr to the cache on success, NULL on failure.
2148 * Cannot be called within a int, but can be interrupted.
2149 * The @ctor is run when new pages are allocated by the cache.
2153 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2154 * to catch references to uninitialised memory.
2156 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2157 * for buffer overruns.
2159 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2160 * cacheline. This can be beneficial if you're counting cycles as closely
2164 __kmem_cache_create (struct kmem_cache *cachep, unsigned long flags)
2166 size_t left_over, slab_size, ralign;
2169 size_t size = cachep->size;
2174 * Enable redzoning and last user accounting, except for caches with
2175 * large objects, if the increased size would increase the object size
2176 * above the next power of two: caches with object sizes just above a
2177 * power of two have a significant amount of internal fragmentation.
2179 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2180 2 * sizeof(unsigned long long)))
2181 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2182 if (!(flags & SLAB_DESTROY_BY_RCU))
2183 flags |= SLAB_POISON;
2185 if (flags & SLAB_DESTROY_BY_RCU)
2186 BUG_ON(flags & SLAB_POISON);
2190 * Check that size is in terms of words. This is needed to avoid
2191 * unaligned accesses for some archs when redzoning is used, and makes
2192 * sure any on-slab bufctl's are also correctly aligned.
2194 if (size & (BYTES_PER_WORD - 1)) {
2195 size += (BYTES_PER_WORD - 1);
2196 size &= ~(BYTES_PER_WORD - 1);
2200 * Redzoning and user store require word alignment or possibly larger.
2201 * Note this will be overridden by architecture or caller mandated
2202 * alignment if either is greater than BYTES_PER_WORD.
2204 if (flags & SLAB_STORE_USER)
2205 ralign = BYTES_PER_WORD;
2207 if (flags & SLAB_RED_ZONE) {
2208 ralign = REDZONE_ALIGN;
2209 /* If redzoning, ensure that the second redzone is suitably
2210 * aligned, by adjusting the object size accordingly. */
2211 size += REDZONE_ALIGN - 1;
2212 size &= ~(REDZONE_ALIGN - 1);
2215 /* 3) caller mandated alignment */
2216 if (ralign < cachep->align) {
2217 ralign = cachep->align;
2219 /* disable debug if necessary */
2220 if (ralign > __alignof__(unsigned long long))
2221 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2225 cachep->align = ralign;
2227 if (slab_is_available())
2232 setup_node_pointer(cachep);
2236 * Both debugging options require word-alignment which is calculated
2239 if (flags & SLAB_RED_ZONE) {
2240 /* add space for red zone words */
2241 cachep->obj_offset += sizeof(unsigned long long);
2242 size += 2 * sizeof(unsigned long long);
2244 if (flags & SLAB_STORE_USER) {
2245 /* user store requires one word storage behind the end of
2246 * the real object. But if the second red zone needs to be
2247 * aligned to 64 bits, we must allow that much space.
2249 if (flags & SLAB_RED_ZONE)
2250 size += REDZONE_ALIGN;
2252 size += BYTES_PER_WORD;
2254 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2255 if (size >= kmalloc_size(INDEX_NODE + 1)
2256 && cachep->object_size > cache_line_size()
2257 && ALIGN(size, cachep->align) < PAGE_SIZE) {
2258 cachep->obj_offset += PAGE_SIZE - ALIGN(size, cachep->align);
2265 * Determine if the slab management is 'on' or 'off' slab.
2266 * (bootstrapping cannot cope with offslab caches so don't do
2267 * it too early on. Always use on-slab management when
2268 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2270 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
2271 !(flags & SLAB_NOLEAKTRACE))
2273 * Size is large, assume best to place the slab management obj
2274 * off-slab (should allow better packing of objs).
2276 flags |= CFLGS_OFF_SLAB;
2278 size = ALIGN(size, cachep->align);
2280 left_over = calculate_slab_order(cachep, size, cachep->align, flags);
2285 slab_size = ALIGN(cachep->num * sizeof(unsigned int)
2286 + sizeof(struct slab), cachep->align);
2289 * If the slab has been placed off-slab, and we have enough space then
2290 * move it on-slab. This is at the expense of any extra colouring.
2292 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2293 flags &= ~CFLGS_OFF_SLAB;
2294 left_over -= slab_size;
2297 if (flags & CFLGS_OFF_SLAB) {
2298 /* really off slab. No need for manual alignment */
2300 cachep->num * sizeof(unsigned int) + sizeof(struct slab);
2302 #ifdef CONFIG_PAGE_POISONING
2303 /* If we're going to use the generic kernel_map_pages()
2304 * poisoning, then it's going to smash the contents of
2305 * the redzone and userword anyhow, so switch them off.
2307 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2308 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2312 cachep->colour_off = cache_line_size();
2313 /* Offset must be a multiple of the alignment. */
2314 if (cachep->colour_off < cachep->align)
2315 cachep->colour_off = cachep->align;
2316 cachep->colour = left_over / cachep->colour_off;
2317 cachep->slab_size = slab_size;
2318 cachep->flags = flags;
2319 cachep->allocflags = __GFP_COMP;
2320 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2321 cachep->allocflags |= GFP_DMA;
2322 cachep->size = size;
2323 cachep->reciprocal_buffer_size = reciprocal_value(size);
2325 if (flags & CFLGS_OFF_SLAB) {
2326 cachep->slabp_cache = kmalloc_slab(slab_size, 0u);
2328 * This is a possibility for one of the malloc_sizes caches.
2329 * But since we go off slab only for object size greater than
2330 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2331 * this should not happen at all.
2332 * But leave a BUG_ON for some lucky dude.
2334 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
2337 err = setup_cpu_cache(cachep, gfp);
2339 __kmem_cache_shutdown(cachep);
2343 if (flags & SLAB_DEBUG_OBJECTS) {
2345 * Would deadlock through slab_destroy()->call_rcu()->
2346 * debug_object_activate()->kmem_cache_alloc().
2348 WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU);
2350 slab_set_debugobj_lock_classes(cachep);
2351 } else if (!OFF_SLAB(cachep) && !(flags & SLAB_DESTROY_BY_RCU))
2352 on_slab_lock_classes(cachep);
2358 static void check_irq_off(void)
2360 BUG_ON(!irqs_disabled());
2363 static void check_irq_on(void)
2365 BUG_ON(irqs_disabled());
2368 static void check_spinlock_acquired(struct kmem_cache *cachep)
2372 assert_spin_locked(&cachep->node[numa_mem_id()]->list_lock);
2376 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2380 assert_spin_locked(&cachep->node[node]->list_lock);
2385 #define check_irq_off() do { } while(0)
2386 #define check_irq_on() do { } while(0)
2387 #define check_spinlock_acquired(x) do { } while(0)
2388 #define check_spinlock_acquired_node(x, y) do { } while(0)
2391 static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
2392 struct array_cache *ac,
2393 int force, int node);
2395 static void do_drain(void *arg)
2397 struct kmem_cache *cachep = arg;
2398 struct array_cache *ac;
2399 int node = numa_mem_id();
2402 ac = cpu_cache_get(cachep);
2403 spin_lock(&cachep->node[node]->list_lock);
2404 free_block(cachep, ac->entry, ac->avail, node);
2405 spin_unlock(&cachep->node[node]->list_lock);
2409 static void drain_cpu_caches(struct kmem_cache *cachep)
2411 struct kmem_cache_node *n;
2414 on_each_cpu(do_drain, cachep, 1);
2416 for_each_online_node(node) {
2417 n = cachep->node[node];
2419 drain_alien_cache(cachep, n->alien);
2422 for_each_online_node(node) {
2423 n = cachep->node[node];
2425 drain_array(cachep, n, n->shared, 1, node);
2430 * Remove slabs from the list of free slabs.
2431 * Specify the number of slabs to drain in tofree.
2433 * Returns the actual number of slabs released.
2435 static int drain_freelist(struct kmem_cache *cache,
2436 struct kmem_cache_node *n, int tofree)
2438 struct list_head *p;
2443 while (nr_freed < tofree && !list_empty(&n->slabs_free)) {
2445 spin_lock_irq(&n->list_lock);
2446 p = n->slabs_free.prev;
2447 if (p == &n->slabs_free) {
2448 spin_unlock_irq(&n->list_lock);
2452 slabp = list_entry(p, struct slab, list);
2454 BUG_ON(slabp->inuse);
2456 list_del(&slabp->list);
2458 * Safe to drop the lock. The slab is no longer linked
2461 n->free_objects -= cache->num;
2462 spin_unlock_irq(&n->list_lock);
2463 slab_destroy(cache, slabp);
2470 /* Called with slab_mutex held to protect against cpu hotplug */
2471 static int __cache_shrink(struct kmem_cache *cachep)
2474 struct kmem_cache_node *n;
2476 drain_cpu_caches(cachep);
2479 for_each_online_node(i) {
2480 n = cachep->node[i];
2484 drain_freelist(cachep, n, slabs_tofree(cachep, n));
2486 ret += !list_empty(&n->slabs_full) ||
2487 !list_empty(&n->slabs_partial);
2489 return (ret ? 1 : 0);
2493 * kmem_cache_shrink - Shrink a cache.
2494 * @cachep: The cache to shrink.
2496 * Releases as many slabs as possible for a cache.
2497 * To help debugging, a zero exit status indicates all slabs were released.
2499 int kmem_cache_shrink(struct kmem_cache *cachep)
2502 BUG_ON(!cachep || in_interrupt());
2505 mutex_lock(&slab_mutex);
2506 ret = __cache_shrink(cachep);
2507 mutex_unlock(&slab_mutex);
2511 EXPORT_SYMBOL(kmem_cache_shrink);
2513 int __kmem_cache_shutdown(struct kmem_cache *cachep)
2516 struct kmem_cache_node *n;
2517 int rc = __cache_shrink(cachep);
2522 for_each_online_cpu(i)
2523 kfree(cachep->array[i]);
2525 /* NUMA: free the node structures */
2526 for_each_online_node(i) {
2527 n = cachep->node[i];
2530 free_alien_cache(n->alien);
2538 * Get the memory for a slab management obj.
2539 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2540 * always come from malloc_sizes caches. The slab descriptor cannot
2541 * come from the same cache which is getting created because,
2542 * when we are searching for an appropriate cache for these
2543 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2544 * If we are creating a malloc_sizes cache here it would not be visible to
2545 * kmem_find_general_cachep till the initialization is complete.
2546 * Hence we cannot have slabp_cache same as the original cache.
2548 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep,
2549 struct page *page, int colour_off,
2550 gfp_t local_flags, int nodeid)
2553 void *addr = page_address(page);
2555 if (OFF_SLAB(cachep)) {
2556 /* Slab management obj is off-slab. */
2557 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2558 local_flags, nodeid);
2560 * If the first object in the slab is leaked (it's allocated
2561 * but no one has a reference to it), we want to make sure
2562 * kmemleak does not treat the ->s_mem pointer as a reference
2563 * to the object. Otherwise we will not report the leak.
2565 kmemleak_scan_area(&slabp->list, sizeof(struct list_head),
2570 slabp = addr + colour_off;
2571 colour_off += cachep->slab_size;
2574 slabp->s_mem = addr + colour_off;
2579 static inline unsigned int *slab_bufctl(struct slab *slabp)
2581 return (unsigned int *) (slabp + 1);
2584 static void cache_init_objs(struct kmem_cache *cachep,
2589 for (i = 0; i < cachep->num; i++) {
2590 void *objp = index_to_obj(cachep, slabp, i);
2592 /* need to poison the objs? */
2593 if (cachep->flags & SLAB_POISON)
2594 poison_obj(cachep, objp, POISON_FREE);
2595 if (cachep->flags & SLAB_STORE_USER)
2596 *dbg_userword(cachep, objp) = NULL;
2598 if (cachep->flags & SLAB_RED_ZONE) {
2599 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2600 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2603 * Constructors are not allowed to allocate memory from the same
2604 * cache which they are a constructor for. Otherwise, deadlock.
2605 * They must also be threaded.
2607 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2608 cachep->ctor(objp + obj_offset(cachep));
2610 if (cachep->flags & SLAB_RED_ZONE) {
2611 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2612 slab_error(cachep, "constructor overwrote the"
2613 " end of an object");
2614 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2615 slab_error(cachep, "constructor overwrote the"
2616 " start of an object");
2618 if ((cachep->size % PAGE_SIZE) == 0 &&
2619 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2620 kernel_map_pages(virt_to_page(objp),
2621 cachep->size / PAGE_SIZE, 0);
2626 slab_bufctl(slabp)[i] = i;
2630 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2632 if (CONFIG_ZONE_DMA_FLAG) {
2633 if (flags & GFP_DMA)
2634 BUG_ON(!(cachep->allocflags & GFP_DMA));
2636 BUG_ON(cachep->allocflags & GFP_DMA);
2640 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2646 objp = index_to_obj(cachep, slabp, slab_bufctl(slabp)[slabp->free]);
2648 WARN_ON(page_to_nid(virt_to_page(objp)) != nodeid);
2655 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2656 void *objp, int nodeid)
2658 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2662 /* Verify that the slab belongs to the intended node */
2663 WARN_ON(page_to_nid(virt_to_page(objp)) != nodeid);
2665 /* Verify double free bug */
2666 for (i = slabp->free; i < cachep->num; i++) {
2667 if (slab_bufctl(slabp)[i] == objnr) {
2668 printk(KERN_ERR "slab: double free detected in cache "
2669 "'%s', objp %p\n", cachep->name, objp);
2675 slab_bufctl(slabp)[slabp->free] = objnr;
2680 * Map pages beginning at addr to the given cache and slab. This is required
2681 * for the slab allocator to be able to lookup the cache and slab of a
2682 * virtual address for kfree, ksize, and slab debugging.
2684 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2687 page->slab_cache = cache;
2688 page->slab_page = slab;
2692 * Grow (by 1) the number of slabs within a cache. This is called by
2693 * kmem_cache_alloc() when there are no active objs left in a cache.
2695 static int cache_grow(struct kmem_cache *cachep,
2696 gfp_t flags, int nodeid, struct page *page)
2701 struct kmem_cache_node *n;
2704 * Be lazy and only check for valid flags here, keeping it out of the
2705 * critical path in kmem_cache_alloc().
2707 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2708 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2710 /* Take the node list lock to change the colour_next on this node */
2712 n = cachep->node[nodeid];
2713 spin_lock(&n->list_lock);
2715 /* Get colour for the slab, and cal the next value. */
2716 offset = n->colour_next;
2718 if (n->colour_next >= cachep->colour)
2720 spin_unlock(&n->list_lock);
2722 offset *= cachep->colour_off;
2724 if (local_flags & __GFP_WAIT)
2728 * The test for missing atomic flag is performed here, rather than
2729 * the more obvious place, simply to reduce the critical path length
2730 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2731 * will eventually be caught here (where it matters).
2733 kmem_flagcheck(cachep, flags);
2736 * Get mem for the objs. Attempt to allocate a physical page from
2740 page = kmem_getpages(cachep, local_flags, nodeid);
2744 /* Get slab management. */
2745 slabp = alloc_slabmgmt(cachep, page, offset,
2746 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2750 slab_map_pages(cachep, slabp, page);
2752 cache_init_objs(cachep, slabp);
2754 if (local_flags & __GFP_WAIT)
2755 local_irq_disable();
2757 spin_lock(&n->list_lock);
2759 /* Make slab active. */
2760 list_add_tail(&slabp->list, &(n->slabs_free));
2761 STATS_INC_GROWN(cachep);
2762 n->free_objects += cachep->num;
2763 spin_unlock(&n->list_lock);
2766 kmem_freepages(cachep, page);
2768 if (local_flags & __GFP_WAIT)
2769 local_irq_disable();
2776 * Perform extra freeing checks:
2777 * - detect bad pointers.
2778 * - POISON/RED_ZONE checking
2780 static void kfree_debugcheck(const void *objp)
2782 if (!virt_addr_valid(objp)) {
2783 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2784 (unsigned long)objp);
2789 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2791 unsigned long long redzone1, redzone2;
2793 redzone1 = *dbg_redzone1(cache, obj);
2794 redzone2 = *dbg_redzone2(cache, obj);
2799 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2802 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2803 slab_error(cache, "double free detected");
2805 slab_error(cache, "memory outside object was overwritten");
2807 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2808 obj, redzone1, redzone2);
2811 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2812 unsigned long caller)
2817 BUG_ON(virt_to_cache(objp) != cachep);
2819 objp -= obj_offset(cachep);
2820 kfree_debugcheck(objp);
2821 slabp = virt_to_slab(objp);
2823 if (cachep->flags & SLAB_RED_ZONE) {
2824 verify_redzone_free(cachep, objp);
2825 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2826 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2828 if (cachep->flags & SLAB_STORE_USER)
2829 *dbg_userword(cachep, objp) = (void *)caller;
2831 objnr = obj_to_index(cachep, slabp, objp);
2833 BUG_ON(objnr >= cachep->num);
2834 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
2836 if (cachep->flags & SLAB_POISON) {
2837 #ifdef CONFIG_DEBUG_PAGEALLOC
2838 if ((cachep->size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
2839 store_stackinfo(cachep, objp, caller);
2840 kernel_map_pages(virt_to_page(objp),
2841 cachep->size / PAGE_SIZE, 0);
2843 poison_obj(cachep, objp, POISON_FREE);
2846 poison_obj(cachep, objp, POISON_FREE);
2853 #define kfree_debugcheck(x) do { } while(0)
2854 #define cache_free_debugcheck(x,objp,z) (objp)
2857 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags,
2861 struct kmem_cache_node *n;
2862 struct array_cache *ac;
2866 node = numa_mem_id();
2867 if (unlikely(force_refill))
2870 ac = cpu_cache_get(cachep);
2871 batchcount = ac->batchcount;
2872 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2874 * If there was little recent activity on this cache, then
2875 * perform only a partial refill. Otherwise we could generate
2878 batchcount = BATCHREFILL_LIMIT;
2880 n = cachep->node[node];
2882 BUG_ON(ac->avail > 0 || !n);
2883 spin_lock(&n->list_lock);
2885 /* See if we can refill from the shared array */
2886 if (n->shared && transfer_objects(ac, n->shared, batchcount)) {
2887 n->shared->touched = 1;
2891 while (batchcount > 0) {
2892 struct list_head *entry;
2894 /* Get slab alloc is to come from. */
2895 entry = n->slabs_partial.next;
2896 if (entry == &n->slabs_partial) {
2897 n->free_touched = 1;
2898 entry = n->slabs_free.next;
2899 if (entry == &n->slabs_free)
2903 slabp = list_entry(entry, struct slab, list);
2904 check_spinlock_acquired(cachep);
2907 * The slab was either on partial or free list so
2908 * there must be at least one object available for
2911 BUG_ON(slabp->inuse >= cachep->num);
2913 while (slabp->inuse < cachep->num && batchcount--) {
2914 STATS_INC_ALLOCED(cachep);
2915 STATS_INC_ACTIVE(cachep);
2916 STATS_SET_HIGH(cachep);
2918 ac_put_obj(cachep, ac, slab_get_obj(cachep, slabp,
2922 /* move slabp to correct slabp list: */
2923 list_del(&slabp->list);
2924 if (slabp->free == cachep->num)
2925 list_add(&slabp->list, &n->slabs_full);
2927 list_add(&slabp->list, &n->slabs_partial);
2931 n->free_objects -= ac->avail;
2933 spin_unlock(&n->list_lock);
2935 if (unlikely(!ac->avail)) {
2938 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
2940 /* cache_grow can reenable interrupts, then ac could change. */
2941 ac = cpu_cache_get(cachep);
2942 node = numa_mem_id();
2944 /* no objects in sight? abort */
2945 if (!x && (ac->avail == 0 || force_refill))
2948 if (!ac->avail) /* objects refilled by interrupt? */
2953 return ac_get_obj(cachep, ac, flags, force_refill);
2956 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
2959 might_sleep_if(flags & __GFP_WAIT);
2961 kmem_flagcheck(cachep, flags);
2966 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
2967 gfp_t flags, void *objp, unsigned long caller)
2971 if (cachep->flags & SLAB_POISON) {
2972 #ifdef CONFIG_DEBUG_PAGEALLOC
2973 if ((cachep->size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
2974 kernel_map_pages(virt_to_page(objp),
2975 cachep->size / PAGE_SIZE, 1);
2977 check_poison_obj(cachep, objp);
2979 check_poison_obj(cachep, objp);
2981 poison_obj(cachep, objp, POISON_INUSE);
2983 if (cachep->flags & SLAB_STORE_USER)
2984 *dbg_userword(cachep, objp) = (void *)caller;
2986 if (cachep->flags & SLAB_RED_ZONE) {
2987 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
2988 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2989 slab_error(cachep, "double free, or memory outside"
2990 " object was overwritten");
2992 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
2993 objp, *dbg_redzone1(cachep, objp),
2994 *dbg_redzone2(cachep, objp));
2996 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2997 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2999 objp += obj_offset(cachep);
3000 if (cachep->ctor && cachep->flags & SLAB_POISON)
3002 if (ARCH_SLAB_MINALIGN &&
3003 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
3004 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3005 objp, (int)ARCH_SLAB_MINALIGN);
3010 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3013 static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
3015 if (cachep == kmem_cache)
3018 return should_failslab(cachep->object_size, flags, cachep->flags);
3021 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3024 struct array_cache *ac;
3025 bool force_refill = false;
3029 ac = cpu_cache_get(cachep);
3030 if (likely(ac->avail)) {
3032 objp = ac_get_obj(cachep, ac, flags, false);
3035 * Allow for the possibility all avail objects are not allowed
3036 * by the current flags
3039 STATS_INC_ALLOCHIT(cachep);
3042 force_refill = true;
3045 STATS_INC_ALLOCMISS(cachep);
3046 objp = cache_alloc_refill(cachep, flags, force_refill);
3048 * the 'ac' may be updated by cache_alloc_refill(),
3049 * and kmemleak_erase() requires its correct value.
3051 ac = cpu_cache_get(cachep);
3055 * To avoid a false negative, if an object that is in one of the
3056 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3057 * treat the array pointers as a reference to the object.
3060 kmemleak_erase(&ac->entry[ac->avail]);
3066 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3068 * If we are in_interrupt, then process context, including cpusets and
3069 * mempolicy, may not apply and should not be used for allocation policy.
3071 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3073 int nid_alloc, nid_here;
3075 if (in_interrupt() || (flags & __GFP_THISNODE))
3077 nid_alloc = nid_here = numa_mem_id();
3078 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3079 nid_alloc = cpuset_slab_spread_node();
3080 else if (current->mempolicy)
3081 nid_alloc = slab_node();
3082 if (nid_alloc != nid_here)
3083 return ____cache_alloc_node(cachep, flags, nid_alloc);
3088 * Fallback function if there was no memory available and no objects on a
3089 * certain node and fall back is permitted. First we scan all the
3090 * available node for available objects. If that fails then we
3091 * perform an allocation without specifying a node. This allows the page
3092 * allocator to do its reclaim / fallback magic. We then insert the
3093 * slab into the proper nodelist and then allocate from it.
3095 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3097 struct zonelist *zonelist;
3101 enum zone_type high_zoneidx = gfp_zone(flags);
3104 unsigned int cpuset_mems_cookie;
3106 if (flags & __GFP_THISNODE)
3109 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3112 cpuset_mems_cookie = get_mems_allowed();
3113 zonelist = node_zonelist(slab_node(), flags);
3117 * Look through allowed nodes for objects available
3118 * from existing per node queues.
3120 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3121 nid = zone_to_nid(zone);
3123 if (cpuset_zone_allowed_hardwall(zone, flags) &&
3125 cache->node[nid]->free_objects) {
3126 obj = ____cache_alloc_node(cache,
3127 flags | GFP_THISNODE, nid);
3135 * This allocation will be performed within the constraints
3136 * of the current cpuset / memory policy requirements.
3137 * We may trigger various forms of reclaim on the allowed
3138 * set and go into memory reserves if necessary.
3142 if (local_flags & __GFP_WAIT)
3144 kmem_flagcheck(cache, flags);
3145 page = kmem_getpages(cache, local_flags, numa_mem_id());
3146 if (local_flags & __GFP_WAIT)
3147 local_irq_disable();
3150 * Insert into the appropriate per node queues
3152 nid = page_to_nid(page);
3153 if (cache_grow(cache, flags, nid, page)) {
3154 obj = ____cache_alloc_node(cache,
3155 flags | GFP_THISNODE, nid);
3158 * Another processor may allocate the
3159 * objects in the slab since we are
3160 * not holding any locks.
3164 /* cache_grow already freed obj */
3170 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !obj))
3176 * A interface to enable slab creation on nodeid
3178 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3181 struct list_head *entry;
3183 struct kmem_cache_node *n;
3187 VM_BUG_ON(nodeid > num_online_nodes());
3188 n = cachep->node[nodeid];
3193 spin_lock(&n->list_lock);
3194 entry = n->slabs_partial.next;
3195 if (entry == &n->slabs_partial) {
3196 n->free_touched = 1;
3197 entry = n->slabs_free.next;
3198 if (entry == &n->slabs_free)
3202 slabp = list_entry(entry, struct slab, list);
3203 check_spinlock_acquired_node(cachep, nodeid);
3205 STATS_INC_NODEALLOCS(cachep);
3206 STATS_INC_ACTIVE(cachep);
3207 STATS_SET_HIGH(cachep);
3209 BUG_ON(slabp->inuse == cachep->num);
3211 obj = slab_get_obj(cachep, slabp, nodeid);
3213 /* move slabp to correct slabp list: */
3214 list_del(&slabp->list);
3216 if (slabp->free == cachep->num)
3217 list_add(&slabp->list, &n->slabs_full);
3219 list_add(&slabp->list, &n->slabs_partial);
3221 spin_unlock(&n->list_lock);
3225 spin_unlock(&n->list_lock);
3226 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
3230 return fallback_alloc(cachep, flags);
3236 static __always_inline void *
3237 slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3238 unsigned long caller)
3240 unsigned long save_flags;
3242 int slab_node = numa_mem_id();
3244 flags &= gfp_allowed_mask;
3246 lockdep_trace_alloc(flags);
3248 if (slab_should_failslab(cachep, flags))
3251 cachep = memcg_kmem_get_cache(cachep, flags);
3253 cache_alloc_debugcheck_before(cachep, flags);
3254 local_irq_save(save_flags);
3256 if (nodeid == NUMA_NO_NODE)
3259 if (unlikely(!cachep->node[nodeid])) {
3260 /* Node not bootstrapped yet */
3261 ptr = fallback_alloc(cachep, flags);
3265 if (nodeid == slab_node) {
3267 * Use the locally cached objects if possible.
3268 * However ____cache_alloc does not allow fallback
3269 * to other nodes. It may fail while we still have
3270 * objects on other nodes available.
3272 ptr = ____cache_alloc(cachep, flags);
3276 /* ___cache_alloc_node can fall back to other nodes */
3277 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3279 local_irq_restore(save_flags);
3280 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3281 kmemleak_alloc_recursive(ptr, cachep->object_size, 1, cachep->flags,
3285 kmemcheck_slab_alloc(cachep, flags, ptr, cachep->object_size);
3287 if (unlikely((flags & __GFP_ZERO) && ptr))
3288 memset(ptr, 0, cachep->object_size);
3293 static __always_inline void *
3294 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3298 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3299 objp = alternate_node_alloc(cache, flags);
3303 objp = ____cache_alloc(cache, flags);
3306 * We may just have run out of memory on the local node.
3307 * ____cache_alloc_node() knows how to locate memory on other nodes
3310 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3317 static __always_inline void *
3318 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3320 return ____cache_alloc(cachep, flags);
3323 #endif /* CONFIG_NUMA */
3325 static __always_inline void *
3326 slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
3328 unsigned long save_flags;
3331 flags &= gfp_allowed_mask;
3333 lockdep_trace_alloc(flags);
3335 if (slab_should_failslab(cachep, flags))
3338 cachep = memcg_kmem_get_cache(cachep, flags);
3340 cache_alloc_debugcheck_before(cachep, flags);
3341 local_irq_save(save_flags);
3342 objp = __do_cache_alloc(cachep, flags);
3343 local_irq_restore(save_flags);
3344 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3345 kmemleak_alloc_recursive(objp, cachep->object_size, 1, cachep->flags,
3350 kmemcheck_slab_alloc(cachep, flags, objp, cachep->object_size);
3352 if (unlikely((flags & __GFP_ZERO) && objp))
3353 memset(objp, 0, cachep->object_size);
3359 * Caller needs to acquire correct kmem_list's list_lock
3361 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3365 struct kmem_cache_node *n;
3367 for (i = 0; i < nr_objects; i++) {
3371 clear_obj_pfmemalloc(&objpp[i]);
3374 slabp = virt_to_slab(objp);
3375 n = cachep->node[node];
3376 list_del(&slabp->list);
3377 check_spinlock_acquired_node(cachep, node);
3378 slab_put_obj(cachep, slabp, objp, node);
3379 STATS_DEC_ACTIVE(cachep);
3382 /* fixup slab chains */
3383 if (slabp->inuse == 0) {
3384 if (n->free_objects > n->free_limit) {
3385 n->free_objects -= cachep->num;
3386 /* No need to drop any previously held
3387 * lock here, even if we have a off-slab slab
3388 * descriptor it is guaranteed to come from
3389 * a different cache, refer to comments before
3392 slab_destroy(cachep, slabp);
3394 list_add(&slabp->list, &n->slabs_free);
3397 /* Unconditionally move a slab to the end of the
3398 * partial list on free - maximum time for the
3399 * other objects to be freed, too.
3401 list_add_tail(&slabp->list, &n->slabs_partial);
3406 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3409 struct kmem_cache_node *n;
3410 int node = numa_mem_id();
3412 batchcount = ac->batchcount;
3414 BUG_ON(!batchcount || batchcount > ac->avail);
3417 n = cachep->node[node];
3418 spin_lock(&n->list_lock);
3420 struct array_cache *shared_array = n->shared;
3421 int max = shared_array->limit - shared_array->avail;
3423 if (batchcount > max)
3425 memcpy(&(shared_array->entry[shared_array->avail]),
3426 ac->entry, sizeof(void *) * batchcount);
3427 shared_array->avail += batchcount;
3432 free_block(cachep, ac->entry, batchcount, node);
3437 struct list_head *p;
3439 p = n->slabs_free.next;
3440 while (p != &(n->slabs_free)) {
3443 slabp = list_entry(p, struct slab, list);
3444 BUG_ON(slabp->inuse);
3449 STATS_SET_FREEABLE(cachep, i);
3452 spin_unlock(&n->list_lock);
3453 ac->avail -= batchcount;
3454 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3458 * Release an obj back to its cache. If the obj has a constructed state, it must
3459 * be in this state _before_ it is released. Called with disabled ints.
3461 static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3462 unsigned long caller)
3464 struct array_cache *ac = cpu_cache_get(cachep);
3467 kmemleak_free_recursive(objp, cachep->flags);
3468 objp = cache_free_debugcheck(cachep, objp, caller);
3470 kmemcheck_slab_free(cachep, objp, cachep->object_size);
3473 * Skip calling cache_free_alien() when the platform is not numa.
3474 * This will avoid cache misses that happen while accessing slabp (which
3475 * is per page memory reference) to get nodeid. Instead use a global
3476 * variable to skip the call, which is mostly likely to be present in
3479 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3482 if (likely(ac->avail < ac->limit)) {
3483 STATS_INC_FREEHIT(cachep);
3485 STATS_INC_FREEMISS(cachep);
3486 cache_flusharray(cachep, ac);
3489 ac_put_obj(cachep, ac, objp);
3493 * kmem_cache_alloc - Allocate an object
3494 * @cachep: The cache to allocate from.
3495 * @flags: See kmalloc().
3497 * Allocate an object from this cache. The flags are only relevant
3498 * if the cache has no available objects.
3500 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3502 void *ret = slab_alloc(cachep, flags, _RET_IP_);
3504 trace_kmem_cache_alloc(_RET_IP_, ret,
3505 cachep->object_size, cachep->size, flags);
3509 EXPORT_SYMBOL(kmem_cache_alloc);
3511 #ifdef CONFIG_TRACING
3513 kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
3517 ret = slab_alloc(cachep, flags, _RET_IP_);
3519 trace_kmalloc(_RET_IP_, ret,
3520 size, cachep->size, flags);
3523 EXPORT_SYMBOL(kmem_cache_alloc_trace);
3528 * kmem_cache_alloc_node - Allocate an object on the specified node
3529 * @cachep: The cache to allocate from.
3530 * @flags: See kmalloc().
3531 * @nodeid: node number of the target node.
3533 * Identical to kmem_cache_alloc but it will allocate memory on the given
3534 * node, which can improve the performance for cpu bound structures.
3536 * Fallback to other node is possible if __GFP_THISNODE is not set.
3538 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3540 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3542 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3543 cachep->object_size, cachep->size,
3548 EXPORT_SYMBOL(kmem_cache_alloc_node);
3550 #ifdef CONFIG_TRACING
3551 void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
3558 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3560 trace_kmalloc_node(_RET_IP_, ret,
3565 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3568 static __always_inline void *
3569 __do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
3571 struct kmem_cache *cachep;
3573 cachep = kmalloc_slab(size, flags);
3574 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3576 return kmem_cache_alloc_node_trace(cachep, flags, node, size);
3579 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3580 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3582 return __do_kmalloc_node(size, flags, node, _RET_IP_);
3584 EXPORT_SYMBOL(__kmalloc_node);
3586 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3587 int node, unsigned long caller)
3589 return __do_kmalloc_node(size, flags, node, caller);
3591 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3593 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3595 return __do_kmalloc_node(size, flags, node, 0);
3597 EXPORT_SYMBOL(__kmalloc_node);
3598 #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3599 #endif /* CONFIG_NUMA */
3602 * __do_kmalloc - allocate memory
3603 * @size: how many bytes of memory are required.
3604 * @flags: the type of memory to allocate (see kmalloc).
3605 * @caller: function caller for debug tracking of the caller
3607 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3608 unsigned long caller)
3610 struct kmem_cache *cachep;
3613 /* If you want to save a few bytes .text space: replace
3615 * Then kmalloc uses the uninlined functions instead of the inline
3618 cachep = kmalloc_slab(size, flags);
3619 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3621 ret = slab_alloc(cachep, flags, caller);
3623 trace_kmalloc(caller, ret,
3624 size, cachep->size, flags);
3630 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3631 void *__kmalloc(size_t size, gfp_t flags)
3633 return __do_kmalloc(size, flags, _RET_IP_);
3635 EXPORT_SYMBOL(__kmalloc);
3637 void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3639 return __do_kmalloc(size, flags, caller);
3641 EXPORT_SYMBOL(__kmalloc_track_caller);
3644 void *__kmalloc(size_t size, gfp_t flags)
3646 return __do_kmalloc(size, flags, 0);
3648 EXPORT_SYMBOL(__kmalloc);
3652 * kmem_cache_free - Deallocate an object
3653 * @cachep: The cache the allocation was from.
3654 * @objp: The previously allocated object.
3656 * Free an object which was previously allocated from this
3659 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3661 unsigned long flags;
3662 cachep = cache_from_obj(cachep, objp);
3666 local_irq_save(flags);
3667 debug_check_no_locks_freed(objp, cachep->object_size);
3668 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3669 debug_check_no_obj_freed(objp, cachep->object_size);
3670 __cache_free(cachep, objp, _RET_IP_);
3671 local_irq_restore(flags);
3673 trace_kmem_cache_free(_RET_IP_, objp);
3675 EXPORT_SYMBOL(kmem_cache_free);
3678 * kfree - free previously allocated memory
3679 * @objp: pointer returned by kmalloc.
3681 * If @objp is NULL, no operation is performed.
3683 * Don't free memory not originally allocated by kmalloc()
3684 * or you will run into trouble.
3686 void kfree(const void *objp)
3688 struct kmem_cache *c;
3689 unsigned long flags;
3691 trace_kfree(_RET_IP_, objp);
3693 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3695 local_irq_save(flags);
3696 kfree_debugcheck(objp);
3697 c = virt_to_cache(objp);
3698 debug_check_no_locks_freed(objp, c->object_size);
3700 debug_check_no_obj_freed(objp, c->object_size);
3701 __cache_free(c, (void *)objp, _RET_IP_);
3702 local_irq_restore(flags);
3704 EXPORT_SYMBOL(kfree);
3707 * This initializes kmem_cache_node or resizes various caches for all nodes.
3709 static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
3712 struct kmem_cache_node *n;
3713 struct array_cache *new_shared;
3714 struct array_cache **new_alien = NULL;
3716 for_each_online_node(node) {
3718 if (use_alien_caches) {
3719 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
3725 if (cachep->shared) {
3726 new_shared = alloc_arraycache(node,
3727 cachep->shared*cachep->batchcount,
3730 free_alien_cache(new_alien);
3735 n = cachep->node[node];
3737 struct array_cache *shared = n->shared;
3739 spin_lock_irq(&n->list_lock);
3742 free_block(cachep, shared->entry,
3743 shared->avail, node);
3745 n->shared = new_shared;
3747 n->alien = new_alien;
3750 n->free_limit = (1 + nr_cpus_node(node)) *
3751 cachep->batchcount + cachep->num;
3752 spin_unlock_irq(&n->list_lock);
3754 free_alien_cache(new_alien);
3757 n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
3759 free_alien_cache(new_alien);
3764 kmem_cache_node_init(n);
3765 n->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3766 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3767 n->shared = new_shared;
3768 n->alien = new_alien;
3769 n->free_limit = (1 + nr_cpus_node(node)) *
3770 cachep->batchcount + cachep->num;
3771 cachep->node[node] = n;
3776 if (!cachep->list.next) {
3777 /* Cache is not active yet. Roll back what we did */
3780 if (cachep->node[node]) {
3781 n = cachep->node[node];
3784 free_alien_cache(n->alien);
3786 cachep->node[node] = NULL;
3794 struct ccupdate_struct {
3795 struct kmem_cache *cachep;
3796 struct array_cache *new[0];
3799 static void do_ccupdate_local(void *info)
3801 struct ccupdate_struct *new = info;
3802 struct array_cache *old;
3805 old = cpu_cache_get(new->cachep);
3807 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3808 new->new[smp_processor_id()] = old;
3811 /* Always called with the slab_mutex held */
3812 static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
3813 int batchcount, int shared, gfp_t gfp)
3815 struct ccupdate_struct *new;
3818 new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *),
3823 for_each_online_cpu(i) {
3824 new->new[i] = alloc_arraycache(cpu_to_mem(i), limit,
3827 for (i--; i >= 0; i--)
3833 new->cachep = cachep;
3835 on_each_cpu(do_ccupdate_local, (void *)new, 1);
3838 cachep->batchcount = batchcount;
3839 cachep->limit = limit;
3840 cachep->shared = shared;
3842 for_each_online_cpu(i) {
3843 struct array_cache *ccold = new->new[i];
3846 spin_lock_irq(&cachep->node[cpu_to_mem(i)]->list_lock);
3847 free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i));
3848 spin_unlock_irq(&cachep->node[cpu_to_mem(i)]->list_lock);
3852 return alloc_kmemlist(cachep, gfp);
3855 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3856 int batchcount, int shared, gfp_t gfp)
3859 struct kmem_cache *c = NULL;
3862 ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3864 if (slab_state < FULL)
3867 if ((ret < 0) || !is_root_cache(cachep))
3870 VM_BUG_ON(!mutex_is_locked(&slab_mutex));
3871 for_each_memcg_cache_index(i) {
3872 c = cache_from_memcg(cachep, i);
3874 /* return value determined by the parent cache only */
3875 __do_tune_cpucache(c, limit, batchcount, shared, gfp);
3881 /* Called with slab_mutex held always */
3882 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
3889 if (!is_root_cache(cachep)) {
3890 struct kmem_cache *root = memcg_root_cache(cachep);
3891 limit = root->limit;
3892 shared = root->shared;
3893 batchcount = root->batchcount;
3896 if (limit && shared && batchcount)
3899 * The head array serves three purposes:
3900 * - create a LIFO ordering, i.e. return objects that are cache-warm
3901 * - reduce the number of spinlock operations.
3902 * - reduce the number of linked list operations on the slab and
3903 * bufctl chains: array operations are cheaper.
3904 * The numbers are guessed, we should auto-tune as described by
3907 if (cachep->size > 131072)
3909 else if (cachep->size > PAGE_SIZE)
3911 else if (cachep->size > 1024)
3913 else if (cachep->size > 256)
3919 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3920 * allocation behaviour: Most allocs on one cpu, most free operations
3921 * on another cpu. For these cases, an efficient object passing between
3922 * cpus is necessary. This is provided by a shared array. The array
3923 * replaces Bonwick's magazine layer.
3924 * On uniprocessor, it's functionally equivalent (but less efficient)
3925 * to a larger limit. Thus disabled by default.
3928 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
3933 * With debugging enabled, large batchcount lead to excessively long
3934 * periods with disabled local interrupts. Limit the batchcount
3939 batchcount = (limit + 1) / 2;
3941 err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3943 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
3944 cachep->name, -err);
3949 * Drain an array if it contains any elements taking the node lock only if
3950 * necessary. Note that the node listlock also protects the array_cache
3951 * if drain_array() is used on the shared array.
3953 static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
3954 struct array_cache *ac, int force, int node)
3958 if (!ac || !ac->avail)
3960 if (ac->touched && !force) {
3963 spin_lock_irq(&n->list_lock);
3965 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3966 if (tofree > ac->avail)
3967 tofree = (ac->avail + 1) / 2;
3968 free_block(cachep, ac->entry, tofree, node);
3969 ac->avail -= tofree;
3970 memmove(ac->entry, &(ac->entry[tofree]),
3971 sizeof(void *) * ac->avail);
3973 spin_unlock_irq(&n->list_lock);
3978 * cache_reap - Reclaim memory from caches.
3979 * @w: work descriptor
3981 * Called from workqueue/eventd every few seconds.
3983 * - clear the per-cpu caches for this CPU.
3984 * - return freeable pages to the main free memory pool.
3986 * If we cannot acquire the cache chain mutex then just give up - we'll try
3987 * again on the next iteration.
3989 static void cache_reap(struct work_struct *w)
3991 struct kmem_cache *searchp;
3992 struct kmem_cache_node *n;
3993 int node = numa_mem_id();
3994 struct delayed_work *work = to_delayed_work(w);
3996 if (!mutex_trylock(&slab_mutex))
3997 /* Give up. Setup the next iteration. */
4000 list_for_each_entry(searchp, &slab_caches, list) {
4004 * We only take the node lock if absolutely necessary and we
4005 * have established with reasonable certainty that
4006 * we can do some work if the lock was obtained.
4008 n = searchp->node[node];
4010 reap_alien(searchp, n);
4012 drain_array(searchp, n, cpu_cache_get(searchp), 0, node);
4015 * These are racy checks but it does not matter
4016 * if we skip one check or scan twice.
4018 if (time_after(n->next_reap, jiffies))
4021 n->next_reap = jiffies + REAPTIMEOUT_LIST3;
4023 drain_array(searchp, n, n->shared, 0, node);
4025 if (n->free_touched)
4026 n->free_touched = 0;
4030 freed = drain_freelist(searchp, n, (n->free_limit +
4031 5 * searchp->num - 1) / (5 * searchp->num));
4032 STATS_ADD_REAPED(searchp, freed);
4038 mutex_unlock(&slab_mutex);
4041 /* Set up the next iteration */
4042 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
4045 #ifdef CONFIG_SLABINFO
4046 void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
4049 unsigned long active_objs;
4050 unsigned long num_objs;
4051 unsigned long active_slabs = 0;
4052 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4056 struct kmem_cache_node *n;
4060 for_each_online_node(node) {
4061 n = cachep->node[node];
4066 spin_lock_irq(&n->list_lock);
4068 list_for_each_entry(slabp, &n->slabs_full, list) {
4069 if (slabp->inuse != cachep->num && !error)
4070 error = "slabs_full accounting error";
4071 active_objs += cachep->num;
4074 list_for_each_entry(slabp, &n->slabs_partial, list) {
4075 if (slabp->inuse == cachep->num && !error)
4076 error = "slabs_partial inuse accounting error";
4077 if (!slabp->inuse && !error)
4078 error = "slabs_partial/inuse accounting error";
4079 active_objs += slabp->inuse;
4082 list_for_each_entry(slabp, &n->slabs_free, list) {
4083 if (slabp->inuse && !error)
4084 error = "slabs_free/inuse accounting error";
4087 free_objects += n->free_objects;
4089 shared_avail += n->shared->avail;
4091 spin_unlock_irq(&n->list_lock);
4093 num_slabs += active_slabs;
4094 num_objs = num_slabs * cachep->num;
4095 if (num_objs - active_objs != free_objects && !error)
4096 error = "free_objects accounting error";
4098 name = cachep->name;
4100 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4102 sinfo->active_objs = active_objs;
4103 sinfo->num_objs = num_objs;
4104 sinfo->active_slabs = active_slabs;
4105 sinfo->num_slabs = num_slabs;
4106 sinfo->shared_avail = shared_avail;
4107 sinfo->limit = cachep->limit;
4108 sinfo->batchcount = cachep->batchcount;
4109 sinfo->shared = cachep->shared;
4110 sinfo->objects_per_slab = cachep->num;
4111 sinfo->cache_order = cachep->gfporder;
4114 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
4118 unsigned long high = cachep->high_mark;
4119 unsigned long allocs = cachep->num_allocations;
4120 unsigned long grown = cachep->grown;
4121 unsigned long reaped = cachep->reaped;
4122 unsigned long errors = cachep->errors;
4123 unsigned long max_freeable = cachep->max_freeable;
4124 unsigned long node_allocs = cachep->node_allocs;
4125 unsigned long node_frees = cachep->node_frees;
4126 unsigned long overflows = cachep->node_overflow;
4128 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4129 "%4lu %4lu %4lu %4lu %4lu",
4130 allocs, high, grown,
4131 reaped, errors, max_freeable, node_allocs,
4132 node_frees, overflows);
4136 unsigned long allochit = atomic_read(&cachep->allochit);
4137 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4138 unsigned long freehit = atomic_read(&cachep->freehit);
4139 unsigned long freemiss = atomic_read(&cachep->freemiss);
4141 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4142 allochit, allocmiss, freehit, freemiss);
4147 #define MAX_SLABINFO_WRITE 128
4149 * slabinfo_write - Tuning for the slab allocator
4151 * @buffer: user buffer
4152 * @count: data length
4155 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4156 size_t count, loff_t *ppos)
4158 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4159 int limit, batchcount, shared, res;
4160 struct kmem_cache *cachep;
4162 if (count > MAX_SLABINFO_WRITE)
4164 if (copy_from_user(&kbuf, buffer, count))
4166 kbuf[MAX_SLABINFO_WRITE] = '\0';
4168 tmp = strchr(kbuf, ' ');
4173 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4176 /* Find the cache in the chain of caches. */
4177 mutex_lock(&slab_mutex);
4179 list_for_each_entry(cachep, &slab_caches, list) {
4180 if (!strcmp(cachep->name, kbuf)) {
4181 if (limit < 1 || batchcount < 1 ||
4182 batchcount > limit || shared < 0) {
4185 res = do_tune_cpucache(cachep, limit,
4192 mutex_unlock(&slab_mutex);
4198 #ifdef CONFIG_DEBUG_SLAB_LEAK
4200 static void *leaks_start(struct seq_file *m, loff_t *pos)
4202 mutex_lock(&slab_mutex);
4203 return seq_list_start(&slab_caches, *pos);
4206 static inline int add_caller(unsigned long *n, unsigned long v)
4216 unsigned long *q = p + 2 * i;
4230 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4236 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4243 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->size) {
4246 for (j = s->free; j < c->num; j++) {
4247 /* Skip freed item */
4248 if (slab_bufctl(s)[j] == i) {
4256 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4261 static void show_symbol(struct seq_file *m, unsigned long address)
4263 #ifdef CONFIG_KALLSYMS
4264 unsigned long offset, size;
4265 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4267 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4268 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4270 seq_printf(m, " [%s]", modname);
4274 seq_printf(m, "%p", (void *)address);
4277 static int leaks_show(struct seq_file *m, void *p)
4279 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
4281 struct kmem_cache_node *n;
4283 unsigned long *x = m->private;
4287 if (!(cachep->flags & SLAB_STORE_USER))
4289 if (!(cachep->flags & SLAB_RED_ZONE))
4292 /* OK, we can do it */
4296 for_each_online_node(node) {
4297 n = cachep->node[node];
4302 spin_lock_irq(&n->list_lock);
4304 list_for_each_entry(slabp, &n->slabs_full, list)
4305 handle_slab(x, cachep, slabp);
4306 list_for_each_entry(slabp, &n->slabs_partial, list)
4307 handle_slab(x, cachep, slabp);
4308 spin_unlock_irq(&n->list_lock);
4310 name = cachep->name;
4312 /* Increase the buffer size */
4313 mutex_unlock(&slab_mutex);
4314 m->private = kzalloc(x[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4316 /* Too bad, we are really out */
4318 mutex_lock(&slab_mutex);
4321 *(unsigned long *)m->private = x[0] * 2;
4323 mutex_lock(&slab_mutex);
4324 /* Now make sure this entry will be retried */
4328 for (i = 0; i < x[1]; i++) {
4329 seq_printf(m, "%s: %lu ", name, x[2*i+3]);
4330 show_symbol(m, x[2*i+2]);
4337 static const struct seq_operations slabstats_op = {
4338 .start = leaks_start,
4344 static int slabstats_open(struct inode *inode, struct file *file)
4346 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4349 ret = seq_open(file, &slabstats_op);
4351 struct seq_file *m = file->private_data;
4352 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4361 static const struct file_operations proc_slabstats_operations = {
4362 .open = slabstats_open,
4364 .llseek = seq_lseek,
4365 .release = seq_release_private,
4369 static int __init slab_proc_init(void)
4371 #ifdef CONFIG_DEBUG_SLAB_LEAK
4372 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4376 module_init(slab_proc_init);
4380 * ksize - get the actual amount of memory allocated for a given object
4381 * @objp: Pointer to the object
4383 * kmalloc may internally round up allocations and return more memory
4384 * than requested. ksize() can be used to determine the actual amount of
4385 * memory allocated. The caller may use this additional memory, even though
4386 * a smaller amount of memory was initially specified with the kmalloc call.
4387 * The caller must guarantee that objp points to a valid object previously
4388 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4389 * must not be freed during the duration of the call.
4391 size_t ksize(const void *objp)
4394 if (unlikely(objp == ZERO_SIZE_PTR))
4397 return virt_to_cache(objp)->object_size;
4399 EXPORT_SYMBOL(ksize);