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 active; /* num of objs active in slab */
185 * - LIFO ordering, to hand out cache-warm objects from _alloc
186 * - reduce the number of linked list operations
187 * - reduce spinlock operations
189 * The limit is stored in the per-cpu structure to reduce the data cache
196 unsigned int batchcount;
197 unsigned int touched;
200 * Must have this definition in here for the proper
201 * alignment of array_cache. Also simplifies accessing
204 * Entries should not be directly dereferenced as
205 * entries belonging to slabs marked pfmemalloc will
206 * have the lower bits set SLAB_OBJ_PFMEMALLOC
210 #define SLAB_OBJ_PFMEMALLOC 1
211 static inline bool is_obj_pfmemalloc(void *objp)
213 return (unsigned long)objp & SLAB_OBJ_PFMEMALLOC;
216 static inline void set_obj_pfmemalloc(void **objp)
218 *objp = (void *)((unsigned long)*objp | SLAB_OBJ_PFMEMALLOC);
222 static inline void clear_obj_pfmemalloc(void **objp)
224 *objp = (void *)((unsigned long)*objp & ~SLAB_OBJ_PFMEMALLOC);
228 * bootstrap: The caches do not work without cpuarrays anymore, but the
229 * cpuarrays are allocated from the generic caches...
231 #define BOOT_CPUCACHE_ENTRIES 1
232 struct arraycache_init {
233 struct array_cache cache;
234 void *entries[BOOT_CPUCACHE_ENTRIES];
238 * Need this for bootstrapping a per node allocator.
240 #define NUM_INIT_LISTS (3 * MAX_NUMNODES)
241 static struct kmem_cache_node __initdata init_kmem_cache_node[NUM_INIT_LISTS];
242 #define CACHE_CACHE 0
243 #define SIZE_AC MAX_NUMNODES
244 #define SIZE_NODE (2 * MAX_NUMNODES)
246 static int drain_freelist(struct kmem_cache *cache,
247 struct kmem_cache_node *n, int tofree);
248 static void free_block(struct kmem_cache *cachep, void **objpp, int len,
250 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
251 static void cache_reap(struct work_struct *unused);
253 static int slab_early_init = 1;
255 #define INDEX_AC kmalloc_index(sizeof(struct arraycache_init))
256 #define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
258 static void kmem_cache_node_init(struct kmem_cache_node *parent)
260 INIT_LIST_HEAD(&parent->slabs_full);
261 INIT_LIST_HEAD(&parent->slabs_partial);
262 INIT_LIST_HEAD(&parent->slabs_free);
263 parent->shared = NULL;
264 parent->alien = NULL;
265 parent->colour_next = 0;
266 spin_lock_init(&parent->list_lock);
267 parent->free_objects = 0;
268 parent->free_touched = 0;
271 #define MAKE_LIST(cachep, listp, slab, nodeid) \
273 INIT_LIST_HEAD(listp); \
274 list_splice(&(cachep->node[nodeid]->slab), listp); \
277 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
279 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
280 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
281 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
284 #define CFLGS_OFF_SLAB (0x80000000UL)
285 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
287 #define BATCHREFILL_LIMIT 16
289 * Optimization question: fewer reaps means less probability for unnessary
290 * cpucache drain/refill cycles.
292 * OTOH the cpuarrays can contain lots of objects,
293 * which could lock up otherwise freeable slabs.
295 #define REAPTIMEOUT_CPUC (2*HZ)
296 #define REAPTIMEOUT_LIST3 (4*HZ)
299 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
300 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
301 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
302 #define STATS_INC_GROWN(x) ((x)->grown++)
303 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
304 #define STATS_SET_HIGH(x) \
306 if ((x)->num_active > (x)->high_mark) \
307 (x)->high_mark = (x)->num_active; \
309 #define STATS_INC_ERR(x) ((x)->errors++)
310 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
311 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
312 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
313 #define STATS_SET_FREEABLE(x, i) \
315 if ((x)->max_freeable < i) \
316 (x)->max_freeable = i; \
318 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
319 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
320 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
321 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
323 #define STATS_INC_ACTIVE(x) do { } while (0)
324 #define STATS_DEC_ACTIVE(x) do { } while (0)
325 #define STATS_INC_ALLOCED(x) do { } while (0)
326 #define STATS_INC_GROWN(x) do { } while (0)
327 #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
328 #define STATS_SET_HIGH(x) do { } while (0)
329 #define STATS_INC_ERR(x) do { } while (0)
330 #define STATS_INC_NODEALLOCS(x) do { } while (0)
331 #define STATS_INC_NODEFREES(x) do { } while (0)
332 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
333 #define STATS_SET_FREEABLE(x, i) do { } while (0)
334 #define STATS_INC_ALLOCHIT(x) do { } while (0)
335 #define STATS_INC_ALLOCMISS(x) do { } while (0)
336 #define STATS_INC_FREEHIT(x) do { } while (0)
337 #define STATS_INC_FREEMISS(x) do { } while (0)
343 * memory layout of objects:
345 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
346 * the end of an object is aligned with the end of the real
347 * allocation. Catches writes behind the end of the allocation.
348 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
350 * cachep->obj_offset: The real object.
351 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
352 * cachep->size - 1* BYTES_PER_WORD: last caller address
353 * [BYTES_PER_WORD long]
355 static int obj_offset(struct kmem_cache *cachep)
357 return cachep->obj_offset;
360 static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
362 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
363 return (unsigned long long*) (objp + obj_offset(cachep) -
364 sizeof(unsigned long long));
367 static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
369 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
370 if (cachep->flags & SLAB_STORE_USER)
371 return (unsigned long long *)(objp + cachep->size -
372 sizeof(unsigned long long) -
374 return (unsigned long long *) (objp + cachep->size -
375 sizeof(unsigned long long));
378 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
380 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
381 return (void **)(objp + cachep->size - BYTES_PER_WORD);
386 #define obj_offset(x) 0
387 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
388 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
389 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
394 * Do not go above this order unless 0 objects fit into the slab or
395 * overridden on the command line.
397 #define SLAB_MAX_ORDER_HI 1
398 #define SLAB_MAX_ORDER_LO 0
399 static int slab_max_order = SLAB_MAX_ORDER_LO;
400 static bool slab_max_order_set __initdata;
402 static inline struct kmem_cache *virt_to_cache(const void *obj)
404 struct page *page = virt_to_head_page(obj);
405 return page->slab_cache;
408 static inline struct slab *virt_to_slab(const void *obj)
410 struct page *page = virt_to_head_page(obj);
412 VM_BUG_ON(!PageSlab(page));
413 return page->slab_page;
416 static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
419 return slab->s_mem + cache->size * idx;
423 * We want to avoid an expensive divide : (offset / cache->size)
424 * Using the fact that size is a constant for a particular cache,
425 * we can replace (offset / cache->size) by
426 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
428 static inline unsigned int obj_to_index(const struct kmem_cache *cache,
429 const struct slab *slab, void *obj)
431 u32 offset = (obj - slab->s_mem);
432 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
435 static struct arraycache_init initarray_generic =
436 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
438 /* internal cache of cache description objs */
439 static struct kmem_cache kmem_cache_boot = {
441 .limit = BOOT_CPUCACHE_ENTRIES,
443 .size = sizeof(struct kmem_cache),
444 .name = "kmem_cache",
447 #define BAD_ALIEN_MAGIC 0x01020304ul
449 #ifdef CONFIG_LOCKDEP
452 * Slab sometimes uses the kmalloc slabs to store the slab headers
453 * for other slabs "off slab".
454 * The locking for this is tricky in that it nests within the locks
455 * of all other slabs in a few places; to deal with this special
456 * locking we put on-slab caches into a separate lock-class.
458 * We set lock class for alien array caches which are up during init.
459 * The lock annotation will be lost if all cpus of a node goes down and
460 * then comes back up during hotplug
462 static struct lock_class_key on_slab_l3_key;
463 static struct lock_class_key on_slab_alc_key;
465 static struct lock_class_key debugobj_l3_key;
466 static struct lock_class_key debugobj_alc_key;
468 static void slab_set_lock_classes(struct kmem_cache *cachep,
469 struct lock_class_key *l3_key, struct lock_class_key *alc_key,
472 struct array_cache **alc;
473 struct kmem_cache_node *n;
480 lockdep_set_class(&n->list_lock, l3_key);
483 * FIXME: This check for BAD_ALIEN_MAGIC
484 * should go away when common slab code is taught to
485 * work even without alien caches.
486 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
487 * for alloc_alien_cache,
489 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
493 lockdep_set_class(&alc[r]->lock, alc_key);
497 static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
499 slab_set_lock_classes(cachep, &debugobj_l3_key, &debugobj_alc_key, node);
502 static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
506 for_each_online_node(node)
507 slab_set_debugobj_lock_classes_node(cachep, node);
510 static void init_node_lock_keys(int q)
517 for (i = 1; i <= KMALLOC_SHIFT_HIGH; i++) {
518 struct kmem_cache_node *n;
519 struct kmem_cache *cache = kmalloc_caches[i];
525 if (!n || OFF_SLAB(cache))
528 slab_set_lock_classes(cache, &on_slab_l3_key,
529 &on_slab_alc_key, q);
533 static void on_slab_lock_classes_node(struct kmem_cache *cachep, int q)
535 if (!cachep->node[q])
538 slab_set_lock_classes(cachep, &on_slab_l3_key,
539 &on_slab_alc_key, q);
542 static inline void on_slab_lock_classes(struct kmem_cache *cachep)
546 VM_BUG_ON(OFF_SLAB(cachep));
548 on_slab_lock_classes_node(cachep, node);
551 static inline void init_lock_keys(void)
556 init_node_lock_keys(node);
559 static void init_node_lock_keys(int q)
563 static inline void init_lock_keys(void)
567 static inline void on_slab_lock_classes(struct kmem_cache *cachep)
571 static inline void on_slab_lock_classes_node(struct kmem_cache *cachep, int node)
575 static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
579 static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
584 static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
586 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
588 return cachep->array[smp_processor_id()];
591 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
593 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(unsigned int), align);
597 * Calculate the number of objects and left-over bytes for a given buffer size.
599 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
600 size_t align, int flags, size_t *left_over,
605 size_t slab_size = PAGE_SIZE << gfporder;
608 * The slab management structure can be either off the slab or
609 * on it. For the latter case, the memory allocated for a
613 * - One unsigned int for each object
614 * - Padding to respect alignment of @align
615 * - @buffer_size bytes for each object
617 * If the slab management structure is off the slab, then the
618 * alignment will already be calculated into the size. Because
619 * the slabs are all pages aligned, the objects will be at the
620 * correct alignment when allocated.
622 if (flags & CFLGS_OFF_SLAB) {
624 nr_objs = slab_size / buffer_size;
628 * Ignore padding for the initial guess. The padding
629 * is at most @align-1 bytes, and @buffer_size is at
630 * least @align. In the worst case, this result will
631 * be one greater than the number of objects that fit
632 * into the memory allocation when taking the padding
635 nr_objs = (slab_size - sizeof(struct slab)) /
636 (buffer_size + sizeof(unsigned int));
639 * This calculated number will be either the right
640 * amount, or one greater than what we want.
642 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
646 mgmt_size = slab_mgmt_size(nr_objs, align);
649 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
653 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
655 static void __slab_error(const char *function, struct kmem_cache *cachep,
658 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
659 function, cachep->name, msg);
661 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
666 * By default on NUMA we use alien caches to stage the freeing of
667 * objects allocated from other nodes. This causes massive memory
668 * inefficiencies when using fake NUMA setup to split memory into a
669 * large number of small nodes, so it can be disabled on the command
673 static int use_alien_caches __read_mostly = 1;
674 static int __init noaliencache_setup(char *s)
676 use_alien_caches = 0;
679 __setup("noaliencache", noaliencache_setup);
681 static int __init slab_max_order_setup(char *str)
683 get_option(&str, &slab_max_order);
684 slab_max_order = slab_max_order < 0 ? 0 :
685 min(slab_max_order, MAX_ORDER - 1);
686 slab_max_order_set = true;
690 __setup("slab_max_order=", slab_max_order_setup);
694 * Special reaping functions for NUMA systems called from cache_reap().
695 * These take care of doing round robin flushing of alien caches (containing
696 * objects freed on different nodes from which they were allocated) and the
697 * flushing of remote pcps by calling drain_node_pages.
699 static DEFINE_PER_CPU(unsigned long, slab_reap_node);
701 static void init_reap_node(int cpu)
705 node = next_node(cpu_to_mem(cpu), node_online_map);
706 if (node == MAX_NUMNODES)
707 node = first_node(node_online_map);
709 per_cpu(slab_reap_node, cpu) = node;
712 static void next_reap_node(void)
714 int node = __this_cpu_read(slab_reap_node);
716 node = next_node(node, node_online_map);
717 if (unlikely(node >= MAX_NUMNODES))
718 node = first_node(node_online_map);
719 __this_cpu_write(slab_reap_node, node);
723 #define init_reap_node(cpu) do { } while (0)
724 #define next_reap_node(void) do { } while (0)
728 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
729 * via the workqueue/eventd.
730 * Add the CPU number into the expiration time to minimize the possibility of
731 * the CPUs getting into lockstep and contending for the global cache chain
734 static void start_cpu_timer(int cpu)
736 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
739 * When this gets called from do_initcalls via cpucache_init(),
740 * init_workqueues() has already run, so keventd will be setup
743 if (keventd_up() && reap_work->work.func == NULL) {
745 INIT_DEFERRABLE_WORK(reap_work, cache_reap);
746 schedule_delayed_work_on(cpu, reap_work,
747 __round_jiffies_relative(HZ, cpu));
751 static struct array_cache *alloc_arraycache(int node, int entries,
752 int batchcount, gfp_t gfp)
754 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
755 struct array_cache *nc = NULL;
757 nc = kmalloc_node(memsize, gfp, node);
759 * The array_cache structures contain pointers to free object.
760 * However, when such objects are allocated or transferred to another
761 * cache the pointers are not cleared and they could be counted as
762 * valid references during a kmemleak scan. Therefore, kmemleak must
763 * not scan such objects.
765 kmemleak_no_scan(nc);
769 nc->batchcount = batchcount;
771 spin_lock_init(&nc->lock);
776 static inline bool is_slab_pfmemalloc(struct slab *slabp)
778 struct page *page = virt_to_page(slabp->s_mem);
780 return PageSlabPfmemalloc(page);
783 /* Clears pfmemalloc_active if no slabs have pfmalloc set */
784 static void recheck_pfmemalloc_active(struct kmem_cache *cachep,
785 struct array_cache *ac)
787 struct kmem_cache_node *n = cachep->node[numa_mem_id()];
791 if (!pfmemalloc_active)
794 spin_lock_irqsave(&n->list_lock, flags);
795 list_for_each_entry(slabp, &n->slabs_full, list)
796 if (is_slab_pfmemalloc(slabp))
799 list_for_each_entry(slabp, &n->slabs_partial, list)
800 if (is_slab_pfmemalloc(slabp))
803 list_for_each_entry(slabp, &n->slabs_free, list)
804 if (is_slab_pfmemalloc(slabp))
807 pfmemalloc_active = false;
809 spin_unlock_irqrestore(&n->list_lock, flags);
812 static void *__ac_get_obj(struct kmem_cache *cachep, struct array_cache *ac,
813 gfp_t flags, bool force_refill)
816 void *objp = ac->entry[--ac->avail];
818 /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
819 if (unlikely(is_obj_pfmemalloc(objp))) {
820 struct kmem_cache_node *n;
822 if (gfp_pfmemalloc_allowed(flags)) {
823 clear_obj_pfmemalloc(&objp);
827 /* The caller cannot use PFMEMALLOC objects, find another one */
828 for (i = 0; i < ac->avail; i++) {
829 /* If a !PFMEMALLOC object is found, swap them */
830 if (!is_obj_pfmemalloc(ac->entry[i])) {
832 ac->entry[i] = ac->entry[ac->avail];
833 ac->entry[ac->avail] = objp;
839 * If there are empty slabs on the slabs_free list and we are
840 * being forced to refill the cache, mark this one !pfmemalloc.
842 n = cachep->node[numa_mem_id()];
843 if (!list_empty(&n->slabs_free) && force_refill) {
844 struct slab *slabp = virt_to_slab(objp);
845 ClearPageSlabPfmemalloc(virt_to_head_page(slabp->s_mem));
846 clear_obj_pfmemalloc(&objp);
847 recheck_pfmemalloc_active(cachep, ac);
851 /* No !PFMEMALLOC objects available */
859 static inline void *ac_get_obj(struct kmem_cache *cachep,
860 struct array_cache *ac, gfp_t flags, bool force_refill)
864 if (unlikely(sk_memalloc_socks()))
865 objp = __ac_get_obj(cachep, ac, flags, force_refill);
867 objp = ac->entry[--ac->avail];
872 static void *__ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
875 if (unlikely(pfmemalloc_active)) {
876 /* Some pfmemalloc slabs exist, check if this is one */
877 struct slab *slabp = virt_to_slab(objp);
878 struct page *page = virt_to_head_page(slabp->s_mem);
879 if (PageSlabPfmemalloc(page))
880 set_obj_pfmemalloc(&objp);
886 static inline void ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
889 if (unlikely(sk_memalloc_socks()))
890 objp = __ac_put_obj(cachep, ac, objp);
892 ac->entry[ac->avail++] = objp;
896 * Transfer objects in one arraycache to another.
897 * Locking must be handled by the caller.
899 * Return the number of entries transferred.
901 static int transfer_objects(struct array_cache *to,
902 struct array_cache *from, unsigned int max)
904 /* Figure out how many entries to transfer */
905 int nr = min3(from->avail, max, to->limit - to->avail);
910 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
920 #define drain_alien_cache(cachep, alien) do { } while (0)
921 #define reap_alien(cachep, n) do { } while (0)
923 static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
925 return (struct array_cache **)BAD_ALIEN_MAGIC;
928 static inline void free_alien_cache(struct array_cache **ac_ptr)
932 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
937 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
943 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
944 gfp_t flags, int nodeid)
949 #else /* CONFIG_NUMA */
951 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
952 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
954 static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
956 struct array_cache **ac_ptr;
957 int memsize = sizeof(void *) * nr_node_ids;
962 ac_ptr = kzalloc_node(memsize, gfp, node);
965 if (i == node || !node_online(i))
967 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
969 for (i--; i >= 0; i--)
979 static void free_alien_cache(struct array_cache **ac_ptr)
990 static void __drain_alien_cache(struct kmem_cache *cachep,
991 struct array_cache *ac, int node)
993 struct kmem_cache_node *n = cachep->node[node];
996 spin_lock(&n->list_lock);
998 * Stuff objects into the remote nodes shared array first.
999 * That way we could avoid the overhead of putting the objects
1000 * into the free lists and getting them back later.
1003 transfer_objects(n->shared, ac, ac->limit);
1005 free_block(cachep, ac->entry, ac->avail, node);
1007 spin_unlock(&n->list_lock);
1012 * Called from cache_reap() to regularly drain alien caches round robin.
1014 static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *n)
1016 int node = __this_cpu_read(slab_reap_node);
1019 struct array_cache *ac = n->alien[node];
1021 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1022 __drain_alien_cache(cachep, ac, node);
1023 spin_unlock_irq(&ac->lock);
1028 static void drain_alien_cache(struct kmem_cache *cachep,
1029 struct array_cache **alien)
1032 struct array_cache *ac;
1033 unsigned long flags;
1035 for_each_online_node(i) {
1038 spin_lock_irqsave(&ac->lock, flags);
1039 __drain_alien_cache(cachep, ac, i);
1040 spin_unlock_irqrestore(&ac->lock, flags);
1045 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1047 int nodeid = page_to_nid(virt_to_page(objp));
1048 struct kmem_cache_node *n;
1049 struct array_cache *alien = NULL;
1052 node = numa_mem_id();
1055 * Make sure we are not freeing a object from another node to the array
1056 * cache on this cpu.
1058 if (likely(nodeid == node))
1061 n = cachep->node[node];
1062 STATS_INC_NODEFREES(cachep);
1063 if (n->alien && n->alien[nodeid]) {
1064 alien = n->alien[nodeid];
1065 spin_lock(&alien->lock);
1066 if (unlikely(alien->avail == alien->limit)) {
1067 STATS_INC_ACOVERFLOW(cachep);
1068 __drain_alien_cache(cachep, alien, nodeid);
1070 ac_put_obj(cachep, alien, objp);
1071 spin_unlock(&alien->lock);
1073 spin_lock(&(cachep->node[nodeid])->list_lock);
1074 free_block(cachep, &objp, 1, nodeid);
1075 spin_unlock(&(cachep->node[nodeid])->list_lock);
1082 * Allocates and initializes node for a node on each slab cache, used for
1083 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
1084 * will be allocated off-node since memory is not yet online for the new node.
1085 * When hotplugging memory or a cpu, existing node are not replaced if
1088 * Must hold slab_mutex.
1090 static int init_cache_node_node(int node)
1092 struct kmem_cache *cachep;
1093 struct kmem_cache_node *n;
1094 const int memsize = sizeof(struct kmem_cache_node);
1096 list_for_each_entry(cachep, &slab_caches, list) {
1098 * Set up the size64 kmemlist for cpu before we can
1099 * begin anything. Make sure some other cpu on this
1100 * node has not already allocated this
1102 if (!cachep->node[node]) {
1103 n = kmalloc_node(memsize, GFP_KERNEL, node);
1106 kmem_cache_node_init(n);
1107 n->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1108 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1111 * The l3s don't come and go as CPUs come and
1112 * go. slab_mutex is sufficient
1115 cachep->node[node] = n;
1118 spin_lock_irq(&cachep->node[node]->list_lock);
1119 cachep->node[node]->free_limit =
1120 (1 + nr_cpus_node(node)) *
1121 cachep->batchcount + cachep->num;
1122 spin_unlock_irq(&cachep->node[node]->list_lock);
1127 static inline int slabs_tofree(struct kmem_cache *cachep,
1128 struct kmem_cache_node *n)
1130 return (n->free_objects + cachep->num - 1) / cachep->num;
1133 static void cpuup_canceled(long cpu)
1135 struct kmem_cache *cachep;
1136 struct kmem_cache_node *n = NULL;
1137 int node = cpu_to_mem(cpu);
1138 const struct cpumask *mask = cpumask_of_node(node);
1140 list_for_each_entry(cachep, &slab_caches, list) {
1141 struct array_cache *nc;
1142 struct array_cache *shared;
1143 struct array_cache **alien;
1145 /* cpu is dead; no one can alloc from it. */
1146 nc = cachep->array[cpu];
1147 cachep->array[cpu] = NULL;
1148 n = cachep->node[node];
1151 goto free_array_cache;
1153 spin_lock_irq(&n->list_lock);
1155 /* Free limit for this kmem_cache_node */
1156 n->free_limit -= cachep->batchcount;
1158 free_block(cachep, nc->entry, nc->avail, node);
1160 if (!cpumask_empty(mask)) {
1161 spin_unlock_irq(&n->list_lock);
1162 goto free_array_cache;
1167 free_block(cachep, shared->entry,
1168 shared->avail, node);
1175 spin_unlock_irq(&n->list_lock);
1179 drain_alien_cache(cachep, alien);
1180 free_alien_cache(alien);
1186 * In the previous loop, all the objects were freed to
1187 * the respective cache's slabs, now we can go ahead and
1188 * shrink each nodelist to its limit.
1190 list_for_each_entry(cachep, &slab_caches, list) {
1191 n = cachep->node[node];
1194 drain_freelist(cachep, n, slabs_tofree(cachep, n));
1198 static int cpuup_prepare(long cpu)
1200 struct kmem_cache *cachep;
1201 struct kmem_cache_node *n = NULL;
1202 int node = cpu_to_mem(cpu);
1206 * We need to do this right in the beginning since
1207 * alloc_arraycache's are going to use this list.
1208 * kmalloc_node allows us to add the slab to the right
1209 * kmem_cache_node and not this cpu's kmem_cache_node
1211 err = init_cache_node_node(node);
1216 * Now we can go ahead with allocating the shared arrays and
1219 list_for_each_entry(cachep, &slab_caches, list) {
1220 struct array_cache *nc;
1221 struct array_cache *shared = NULL;
1222 struct array_cache **alien = NULL;
1224 nc = alloc_arraycache(node, cachep->limit,
1225 cachep->batchcount, GFP_KERNEL);
1228 if (cachep->shared) {
1229 shared = alloc_arraycache(node,
1230 cachep->shared * cachep->batchcount,
1231 0xbaadf00d, GFP_KERNEL);
1237 if (use_alien_caches) {
1238 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
1245 cachep->array[cpu] = nc;
1246 n = cachep->node[node];
1249 spin_lock_irq(&n->list_lock);
1252 * We are serialised from CPU_DEAD or
1253 * CPU_UP_CANCELLED by the cpucontrol lock
1264 spin_unlock_irq(&n->list_lock);
1266 free_alien_cache(alien);
1267 if (cachep->flags & SLAB_DEBUG_OBJECTS)
1268 slab_set_debugobj_lock_classes_node(cachep, node);
1269 else if (!OFF_SLAB(cachep) &&
1270 !(cachep->flags & SLAB_DESTROY_BY_RCU))
1271 on_slab_lock_classes_node(cachep, node);
1273 init_node_lock_keys(node);
1277 cpuup_canceled(cpu);
1281 static int cpuup_callback(struct notifier_block *nfb,
1282 unsigned long action, void *hcpu)
1284 long cpu = (long)hcpu;
1288 case CPU_UP_PREPARE:
1289 case CPU_UP_PREPARE_FROZEN:
1290 mutex_lock(&slab_mutex);
1291 err = cpuup_prepare(cpu);
1292 mutex_unlock(&slab_mutex);
1295 case CPU_ONLINE_FROZEN:
1296 start_cpu_timer(cpu);
1298 #ifdef CONFIG_HOTPLUG_CPU
1299 case CPU_DOWN_PREPARE:
1300 case CPU_DOWN_PREPARE_FROZEN:
1302 * Shutdown cache reaper. Note that the slab_mutex is
1303 * held so that if cache_reap() is invoked it cannot do
1304 * anything expensive but will only modify reap_work
1305 * and reschedule the timer.
1307 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1308 /* Now the cache_reaper is guaranteed to be not running. */
1309 per_cpu(slab_reap_work, cpu).work.func = NULL;
1311 case CPU_DOWN_FAILED:
1312 case CPU_DOWN_FAILED_FROZEN:
1313 start_cpu_timer(cpu);
1316 case CPU_DEAD_FROZEN:
1318 * Even if all the cpus of a node are down, we don't free the
1319 * kmem_cache_node of any cache. This to avoid a race between
1320 * cpu_down, and a kmalloc allocation from another cpu for
1321 * memory from the node of the cpu going down. The node
1322 * structure is usually allocated from kmem_cache_create() and
1323 * gets destroyed at kmem_cache_destroy().
1327 case CPU_UP_CANCELED:
1328 case CPU_UP_CANCELED_FROZEN:
1329 mutex_lock(&slab_mutex);
1330 cpuup_canceled(cpu);
1331 mutex_unlock(&slab_mutex);
1334 return notifier_from_errno(err);
1337 static struct notifier_block cpucache_notifier = {
1338 &cpuup_callback, NULL, 0
1341 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1343 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1344 * Returns -EBUSY if all objects cannot be drained so that the node is not
1347 * Must hold slab_mutex.
1349 static int __meminit drain_cache_node_node(int node)
1351 struct kmem_cache *cachep;
1354 list_for_each_entry(cachep, &slab_caches, list) {
1355 struct kmem_cache_node *n;
1357 n = cachep->node[node];
1361 drain_freelist(cachep, n, slabs_tofree(cachep, n));
1363 if (!list_empty(&n->slabs_full) ||
1364 !list_empty(&n->slabs_partial)) {
1372 static int __meminit slab_memory_callback(struct notifier_block *self,
1373 unsigned long action, void *arg)
1375 struct memory_notify *mnb = arg;
1379 nid = mnb->status_change_nid;
1384 case MEM_GOING_ONLINE:
1385 mutex_lock(&slab_mutex);
1386 ret = init_cache_node_node(nid);
1387 mutex_unlock(&slab_mutex);
1389 case MEM_GOING_OFFLINE:
1390 mutex_lock(&slab_mutex);
1391 ret = drain_cache_node_node(nid);
1392 mutex_unlock(&slab_mutex);
1396 case MEM_CANCEL_ONLINE:
1397 case MEM_CANCEL_OFFLINE:
1401 return notifier_from_errno(ret);
1403 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1406 * swap the static kmem_cache_node with kmalloced memory
1408 static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list,
1411 struct kmem_cache_node *ptr;
1413 ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid);
1416 memcpy(ptr, list, sizeof(struct kmem_cache_node));
1418 * Do not assume that spinlocks can be initialized via memcpy:
1420 spin_lock_init(&ptr->list_lock);
1422 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1423 cachep->node[nodeid] = ptr;
1427 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1428 * size of kmem_cache_node.
1430 static void __init set_up_node(struct kmem_cache *cachep, int index)
1434 for_each_online_node(node) {
1435 cachep->node[node] = &init_kmem_cache_node[index + node];
1436 cachep->node[node]->next_reap = jiffies +
1438 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1443 * The memory after the last cpu cache pointer is used for the
1446 static void setup_node_pointer(struct kmem_cache *cachep)
1448 cachep->node = (struct kmem_cache_node **)&cachep->array[nr_cpu_ids];
1452 * Initialisation. Called after the page allocator have been initialised and
1453 * before smp_init().
1455 void __init kmem_cache_init(void)
1459 BUILD_BUG_ON(sizeof(((struct page *)NULL)->lru) <
1460 sizeof(struct rcu_head));
1461 kmem_cache = &kmem_cache_boot;
1462 setup_node_pointer(kmem_cache);
1464 if (num_possible_nodes() == 1)
1465 use_alien_caches = 0;
1467 for (i = 0; i < NUM_INIT_LISTS; i++)
1468 kmem_cache_node_init(&init_kmem_cache_node[i]);
1470 set_up_node(kmem_cache, CACHE_CACHE);
1473 * Fragmentation resistance on low memory - only use bigger
1474 * page orders on machines with more than 32MB of memory if
1475 * not overridden on the command line.
1477 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
1478 slab_max_order = SLAB_MAX_ORDER_HI;
1480 /* Bootstrap is tricky, because several objects are allocated
1481 * from caches that do not exist yet:
1482 * 1) initialize the kmem_cache cache: it contains the struct
1483 * kmem_cache structures of all caches, except kmem_cache itself:
1484 * kmem_cache is statically allocated.
1485 * Initially an __init data area is used for the head array and the
1486 * kmem_cache_node structures, it's replaced with a kmalloc allocated
1487 * array at the end of the bootstrap.
1488 * 2) Create the first kmalloc cache.
1489 * The struct kmem_cache for the new cache is allocated normally.
1490 * An __init data area is used for the head array.
1491 * 3) Create the remaining kmalloc caches, with minimally sized
1493 * 4) Replace the __init data head arrays for kmem_cache and the first
1494 * kmalloc cache with kmalloc allocated arrays.
1495 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1496 * the other cache's with kmalloc allocated memory.
1497 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1500 /* 1) create the kmem_cache */
1503 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1505 create_boot_cache(kmem_cache, "kmem_cache",
1506 offsetof(struct kmem_cache, array[nr_cpu_ids]) +
1507 nr_node_ids * sizeof(struct kmem_cache_node *),
1508 SLAB_HWCACHE_ALIGN);
1509 list_add(&kmem_cache->list, &slab_caches);
1511 /* 2+3) create the kmalloc caches */
1514 * Initialize the caches that provide memory for the array cache and the
1515 * kmem_cache_node structures first. Without this, further allocations will
1519 kmalloc_caches[INDEX_AC] = create_kmalloc_cache("kmalloc-ac",
1520 kmalloc_size(INDEX_AC), ARCH_KMALLOC_FLAGS);
1522 if (INDEX_AC != INDEX_NODE)
1523 kmalloc_caches[INDEX_NODE] =
1524 create_kmalloc_cache("kmalloc-node",
1525 kmalloc_size(INDEX_NODE), ARCH_KMALLOC_FLAGS);
1527 slab_early_init = 0;
1529 /* 4) Replace the bootstrap head arrays */
1531 struct array_cache *ptr;
1533 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1535 memcpy(ptr, cpu_cache_get(kmem_cache),
1536 sizeof(struct arraycache_init));
1538 * Do not assume that spinlocks can be initialized via memcpy:
1540 spin_lock_init(&ptr->lock);
1542 kmem_cache->array[smp_processor_id()] = ptr;
1544 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1546 BUG_ON(cpu_cache_get(kmalloc_caches[INDEX_AC])
1547 != &initarray_generic.cache);
1548 memcpy(ptr, cpu_cache_get(kmalloc_caches[INDEX_AC]),
1549 sizeof(struct arraycache_init));
1551 * Do not assume that spinlocks can be initialized via memcpy:
1553 spin_lock_init(&ptr->lock);
1555 kmalloc_caches[INDEX_AC]->array[smp_processor_id()] = ptr;
1557 /* 5) Replace the bootstrap kmem_cache_node */
1561 for_each_online_node(nid) {
1562 init_list(kmem_cache, &init_kmem_cache_node[CACHE_CACHE + nid], nid);
1564 init_list(kmalloc_caches[INDEX_AC],
1565 &init_kmem_cache_node[SIZE_AC + nid], nid);
1567 if (INDEX_AC != INDEX_NODE) {
1568 init_list(kmalloc_caches[INDEX_NODE],
1569 &init_kmem_cache_node[SIZE_NODE + nid], nid);
1574 create_kmalloc_caches(ARCH_KMALLOC_FLAGS);
1577 void __init kmem_cache_init_late(void)
1579 struct kmem_cache *cachep;
1583 /* 6) resize the head arrays to their final sizes */
1584 mutex_lock(&slab_mutex);
1585 list_for_each_entry(cachep, &slab_caches, list)
1586 if (enable_cpucache(cachep, GFP_NOWAIT))
1588 mutex_unlock(&slab_mutex);
1590 /* Annotate slab for lockdep -- annotate the malloc caches */
1597 * Register a cpu startup notifier callback that initializes
1598 * cpu_cache_get for all new cpus
1600 register_cpu_notifier(&cpucache_notifier);
1604 * Register a memory hotplug callback that initializes and frees
1607 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1611 * The reap timers are started later, with a module init call: That part
1612 * of the kernel is not yet operational.
1616 static int __init cpucache_init(void)
1621 * Register the timers that return unneeded pages to the page allocator
1623 for_each_online_cpu(cpu)
1624 start_cpu_timer(cpu);
1630 __initcall(cpucache_init);
1632 static noinline void
1633 slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1635 struct kmem_cache_node *n;
1637 unsigned long flags;
1641 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1643 printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n",
1644 cachep->name, cachep->size, cachep->gfporder);
1646 for_each_online_node(node) {
1647 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1648 unsigned long active_slabs = 0, num_slabs = 0;
1650 n = cachep->node[node];
1654 spin_lock_irqsave(&n->list_lock, flags);
1655 list_for_each_entry(slabp, &n->slabs_full, list) {
1656 active_objs += cachep->num;
1659 list_for_each_entry(slabp, &n->slabs_partial, list) {
1660 active_objs += slabp->active;
1663 list_for_each_entry(slabp, &n->slabs_free, list)
1666 free_objects += n->free_objects;
1667 spin_unlock_irqrestore(&n->list_lock, flags);
1669 num_slabs += active_slabs;
1670 num_objs = num_slabs * cachep->num;
1672 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1673 node, active_slabs, num_slabs, active_objs, num_objs,
1679 * Interface to system's page allocator. No need to hold the cache-lock.
1681 * If we requested dmaable memory, we will get it. Even if we
1682 * did not request dmaable memory, we might get it, but that
1683 * would be relatively rare and ignorable.
1685 static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags,
1691 flags |= cachep->allocflags;
1692 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1693 flags |= __GFP_RECLAIMABLE;
1695 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1697 if (!(flags & __GFP_NOWARN) && printk_ratelimit())
1698 slab_out_of_memory(cachep, flags, nodeid);
1702 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1703 if (unlikely(page->pfmemalloc))
1704 pfmemalloc_active = true;
1706 nr_pages = (1 << cachep->gfporder);
1707 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1708 add_zone_page_state(page_zone(page),
1709 NR_SLAB_RECLAIMABLE, nr_pages);
1711 add_zone_page_state(page_zone(page),
1712 NR_SLAB_UNRECLAIMABLE, nr_pages);
1713 __SetPageSlab(page);
1714 if (page->pfmemalloc)
1715 SetPageSlabPfmemalloc(page);
1716 memcg_bind_pages(cachep, cachep->gfporder);
1718 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1719 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1722 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1724 kmemcheck_mark_unallocated_pages(page, nr_pages);
1731 * Interface to system's page release.
1733 static void kmem_freepages(struct kmem_cache *cachep, struct page *page)
1735 const unsigned long nr_freed = (1 << cachep->gfporder);
1737 kmemcheck_free_shadow(page, cachep->gfporder);
1739 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1740 sub_zone_page_state(page_zone(page),
1741 NR_SLAB_RECLAIMABLE, nr_freed);
1743 sub_zone_page_state(page_zone(page),
1744 NR_SLAB_UNRECLAIMABLE, nr_freed);
1746 BUG_ON(!PageSlab(page));
1747 __ClearPageSlabPfmemalloc(page);
1748 __ClearPageSlab(page);
1750 memcg_release_pages(cachep, cachep->gfporder);
1751 if (current->reclaim_state)
1752 current->reclaim_state->reclaimed_slab += nr_freed;
1753 __free_memcg_kmem_pages(page, cachep->gfporder);
1756 static void kmem_rcu_free(struct rcu_head *head)
1758 struct kmem_cache *cachep;
1761 page = container_of(head, struct page, rcu_head);
1762 cachep = page->slab_cache;
1764 kmem_freepages(cachep, page);
1769 #ifdef CONFIG_DEBUG_PAGEALLOC
1770 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1771 unsigned long caller)
1773 int size = cachep->object_size;
1775 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1777 if (size < 5 * sizeof(unsigned long))
1780 *addr++ = 0x12345678;
1782 *addr++ = smp_processor_id();
1783 size -= 3 * sizeof(unsigned long);
1785 unsigned long *sptr = &caller;
1786 unsigned long svalue;
1788 while (!kstack_end(sptr)) {
1790 if (kernel_text_address(svalue)) {
1792 size -= sizeof(unsigned long);
1793 if (size <= sizeof(unsigned long))
1799 *addr++ = 0x87654321;
1803 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1805 int size = cachep->object_size;
1806 addr = &((char *)addr)[obj_offset(cachep)];
1808 memset(addr, val, size);
1809 *(unsigned char *)(addr + size - 1) = POISON_END;
1812 static void dump_line(char *data, int offset, int limit)
1815 unsigned char error = 0;
1818 printk(KERN_ERR "%03x: ", offset);
1819 for (i = 0; i < limit; i++) {
1820 if (data[offset + i] != POISON_FREE) {
1821 error = data[offset + i];
1825 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1826 &data[offset], limit, 1);
1828 if (bad_count == 1) {
1829 error ^= POISON_FREE;
1830 if (!(error & (error - 1))) {
1831 printk(KERN_ERR "Single bit error detected. Probably "
1834 printk(KERN_ERR "Run memtest86+ or a similar memory "
1837 printk(KERN_ERR "Run a memory test tool.\n");
1846 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1851 if (cachep->flags & SLAB_RED_ZONE) {
1852 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
1853 *dbg_redzone1(cachep, objp),
1854 *dbg_redzone2(cachep, objp));
1857 if (cachep->flags & SLAB_STORE_USER) {
1858 printk(KERN_ERR "Last user: [<%p>](%pSR)\n",
1859 *dbg_userword(cachep, objp),
1860 *dbg_userword(cachep, objp));
1862 realobj = (char *)objp + obj_offset(cachep);
1863 size = cachep->object_size;
1864 for (i = 0; i < size && lines; i += 16, lines--) {
1867 if (i + limit > size)
1869 dump_line(realobj, i, limit);
1873 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1879 realobj = (char *)objp + obj_offset(cachep);
1880 size = cachep->object_size;
1882 for (i = 0; i < size; i++) {
1883 char exp = POISON_FREE;
1886 if (realobj[i] != exp) {
1892 "Slab corruption (%s): %s start=%p, len=%d\n",
1893 print_tainted(), cachep->name, realobj, size);
1894 print_objinfo(cachep, objp, 0);
1896 /* Hexdump the affected line */
1899 if (i + limit > size)
1901 dump_line(realobj, i, limit);
1904 /* Limit to 5 lines */
1910 /* Print some data about the neighboring objects, if they
1913 struct slab *slabp = virt_to_slab(objp);
1916 objnr = obj_to_index(cachep, slabp, objp);
1918 objp = index_to_obj(cachep, slabp, objnr - 1);
1919 realobj = (char *)objp + obj_offset(cachep);
1920 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1922 print_objinfo(cachep, objp, 2);
1924 if (objnr + 1 < cachep->num) {
1925 objp = index_to_obj(cachep, slabp, objnr + 1);
1926 realobj = (char *)objp + obj_offset(cachep);
1927 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1929 print_objinfo(cachep, objp, 2);
1936 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
1939 for (i = 0; i < cachep->num; i++) {
1940 void *objp = index_to_obj(cachep, slabp, i);
1942 if (cachep->flags & SLAB_POISON) {
1943 #ifdef CONFIG_DEBUG_PAGEALLOC
1944 if (cachep->size % PAGE_SIZE == 0 &&
1946 kernel_map_pages(virt_to_page(objp),
1947 cachep->size / PAGE_SIZE, 1);
1949 check_poison_obj(cachep, objp);
1951 check_poison_obj(cachep, objp);
1954 if (cachep->flags & SLAB_RED_ZONE) {
1955 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1956 slab_error(cachep, "start of a freed object "
1958 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1959 slab_error(cachep, "end of a freed object "
1965 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
1971 * slab_destroy - destroy and release all objects in a slab
1972 * @cachep: cache pointer being destroyed
1973 * @slabp: slab pointer being destroyed
1975 * Destroy all the objs in a slab, and release the mem back to the system.
1976 * Before calling the slab must have been unlinked from the cache. The
1977 * cache-lock is not held/needed.
1979 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
1981 struct page *page = virt_to_head_page(slabp->s_mem);
1983 slab_destroy_debugcheck(cachep, slabp);
1984 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1985 struct rcu_head *head;
1988 * RCU free overloads the RCU head over the LRU.
1989 * slab_page has been overloeaded over the LRU,
1990 * however it is not used from now on so that
1991 * we can use it safely.
1993 head = (void *)&page->rcu_head;
1994 call_rcu(head, kmem_rcu_free);
1997 kmem_freepages(cachep, page);
2001 * From now on, we don't use slab management
2002 * although actual page can be freed in rcu context
2004 if (OFF_SLAB(cachep))
2005 kmem_cache_free(cachep->slabp_cache, slabp);
2009 * calculate_slab_order - calculate size (page order) of slabs
2010 * @cachep: pointer to the cache that is being created
2011 * @size: size of objects to be created in this cache.
2012 * @align: required alignment for the objects.
2013 * @flags: slab allocation flags
2015 * Also calculates the number of objects per slab.
2017 * This could be made much more intelligent. For now, try to avoid using
2018 * high order pages for slabs. When the gfp() functions are more friendly
2019 * towards high-order requests, this should be changed.
2021 static size_t calculate_slab_order(struct kmem_cache *cachep,
2022 size_t size, size_t align, unsigned long flags)
2024 unsigned long offslab_limit;
2025 size_t left_over = 0;
2028 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
2032 cache_estimate(gfporder, size, align, flags, &remainder, &num);
2036 if (flags & CFLGS_OFF_SLAB) {
2038 * Max number of objs-per-slab for caches which
2039 * use off-slab slabs. Needed to avoid a possible
2040 * looping condition in cache_grow().
2042 offslab_limit = size - sizeof(struct slab);
2043 offslab_limit /= sizeof(unsigned int);
2045 if (num > offslab_limit)
2049 /* Found something acceptable - save it away */
2051 cachep->gfporder = gfporder;
2052 left_over = remainder;
2055 * A VFS-reclaimable slab tends to have most allocations
2056 * as GFP_NOFS and we really don't want to have to be allocating
2057 * higher-order pages when we are unable to shrink dcache.
2059 if (flags & SLAB_RECLAIM_ACCOUNT)
2063 * Large number of objects is good, but very large slabs are
2064 * currently bad for the gfp()s.
2066 if (gfporder >= slab_max_order)
2070 * Acceptable internal fragmentation?
2072 if (left_over * 8 <= (PAGE_SIZE << gfporder))
2078 static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
2080 if (slab_state >= FULL)
2081 return enable_cpucache(cachep, gfp);
2083 if (slab_state == DOWN) {
2085 * Note: Creation of first cache (kmem_cache).
2086 * The setup_node is taken care
2087 * of by the caller of __kmem_cache_create
2089 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2090 slab_state = PARTIAL;
2091 } else if (slab_state == PARTIAL) {
2093 * Note: the second kmem_cache_create must create the cache
2094 * that's used by kmalloc(24), otherwise the creation of
2095 * further caches will BUG().
2097 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2100 * If the cache that's used by kmalloc(sizeof(kmem_cache_node)) is
2101 * the second cache, then we need to set up all its node/,
2102 * otherwise the creation of further caches will BUG().
2104 set_up_node(cachep, SIZE_AC);
2105 if (INDEX_AC == INDEX_NODE)
2106 slab_state = PARTIAL_NODE;
2108 slab_state = PARTIAL_ARRAYCACHE;
2110 /* Remaining boot caches */
2111 cachep->array[smp_processor_id()] =
2112 kmalloc(sizeof(struct arraycache_init), gfp);
2114 if (slab_state == PARTIAL_ARRAYCACHE) {
2115 set_up_node(cachep, SIZE_NODE);
2116 slab_state = PARTIAL_NODE;
2119 for_each_online_node(node) {
2120 cachep->node[node] =
2121 kmalloc_node(sizeof(struct kmem_cache_node),
2123 BUG_ON(!cachep->node[node]);
2124 kmem_cache_node_init(cachep->node[node]);
2128 cachep->node[numa_mem_id()]->next_reap =
2129 jiffies + REAPTIMEOUT_LIST3 +
2130 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2132 cpu_cache_get(cachep)->avail = 0;
2133 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2134 cpu_cache_get(cachep)->batchcount = 1;
2135 cpu_cache_get(cachep)->touched = 0;
2136 cachep->batchcount = 1;
2137 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2142 * __kmem_cache_create - Create a cache.
2143 * @cachep: cache management descriptor
2144 * @flags: SLAB flags
2146 * Returns a ptr to the cache on success, NULL on failure.
2147 * Cannot be called within a int, but can be interrupted.
2148 * The @ctor is run when new pages are allocated by the cache.
2152 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2153 * to catch references to uninitialised memory.
2155 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2156 * for buffer overruns.
2158 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2159 * cacheline. This can be beneficial if you're counting cycles as closely
2163 __kmem_cache_create (struct kmem_cache *cachep, unsigned long flags)
2165 size_t left_over, slab_size, ralign;
2168 size_t size = cachep->size;
2173 * Enable redzoning and last user accounting, except for caches with
2174 * large objects, if the increased size would increase the object size
2175 * above the next power of two: caches with object sizes just above a
2176 * power of two have a significant amount of internal fragmentation.
2178 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2179 2 * sizeof(unsigned long long)))
2180 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2181 if (!(flags & SLAB_DESTROY_BY_RCU))
2182 flags |= SLAB_POISON;
2184 if (flags & SLAB_DESTROY_BY_RCU)
2185 BUG_ON(flags & SLAB_POISON);
2189 * Check that size is in terms of words. This is needed to avoid
2190 * unaligned accesses for some archs when redzoning is used, and makes
2191 * sure any on-slab bufctl's are also correctly aligned.
2193 if (size & (BYTES_PER_WORD - 1)) {
2194 size += (BYTES_PER_WORD - 1);
2195 size &= ~(BYTES_PER_WORD - 1);
2199 * Redzoning and user store require word alignment or possibly larger.
2200 * Note this will be overridden by architecture or caller mandated
2201 * alignment if either is greater than BYTES_PER_WORD.
2203 if (flags & SLAB_STORE_USER)
2204 ralign = BYTES_PER_WORD;
2206 if (flags & SLAB_RED_ZONE) {
2207 ralign = REDZONE_ALIGN;
2208 /* If redzoning, ensure that the second redzone is suitably
2209 * aligned, by adjusting the object size accordingly. */
2210 size += REDZONE_ALIGN - 1;
2211 size &= ~(REDZONE_ALIGN - 1);
2214 /* 3) caller mandated alignment */
2215 if (ralign < cachep->align) {
2216 ralign = cachep->align;
2218 /* disable debug if necessary */
2219 if (ralign > __alignof__(unsigned long long))
2220 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2224 cachep->align = ralign;
2226 if (slab_is_available())
2231 setup_node_pointer(cachep);
2235 * Both debugging options require word-alignment which is calculated
2238 if (flags & SLAB_RED_ZONE) {
2239 /* add space for red zone words */
2240 cachep->obj_offset += sizeof(unsigned long long);
2241 size += 2 * sizeof(unsigned long long);
2243 if (flags & SLAB_STORE_USER) {
2244 /* user store requires one word storage behind the end of
2245 * the real object. But if the second red zone needs to be
2246 * aligned to 64 bits, we must allow that much space.
2248 if (flags & SLAB_RED_ZONE)
2249 size += REDZONE_ALIGN;
2251 size += BYTES_PER_WORD;
2253 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2254 if (size >= kmalloc_size(INDEX_NODE + 1)
2255 && cachep->object_size > cache_line_size()
2256 && ALIGN(size, cachep->align) < PAGE_SIZE) {
2257 cachep->obj_offset += PAGE_SIZE - ALIGN(size, cachep->align);
2264 * Determine if the slab management is 'on' or 'off' slab.
2265 * (bootstrapping cannot cope with offslab caches so don't do
2266 * it too early on. Always use on-slab management when
2267 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2269 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
2270 !(flags & SLAB_NOLEAKTRACE))
2272 * Size is large, assume best to place the slab management obj
2273 * off-slab (should allow better packing of objs).
2275 flags |= CFLGS_OFF_SLAB;
2277 size = ALIGN(size, cachep->align);
2279 left_over = calculate_slab_order(cachep, size, cachep->align, flags);
2284 slab_size = ALIGN(cachep->num * sizeof(unsigned int)
2285 + sizeof(struct slab), cachep->align);
2288 * If the slab has been placed off-slab, and we have enough space then
2289 * move it on-slab. This is at the expense of any extra colouring.
2291 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2292 flags &= ~CFLGS_OFF_SLAB;
2293 left_over -= slab_size;
2296 if (flags & CFLGS_OFF_SLAB) {
2297 /* really off slab. No need for manual alignment */
2299 cachep->num * sizeof(unsigned int) + sizeof(struct slab);
2301 #ifdef CONFIG_PAGE_POISONING
2302 /* If we're going to use the generic kernel_map_pages()
2303 * poisoning, then it's going to smash the contents of
2304 * the redzone and userword anyhow, so switch them off.
2306 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2307 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2311 cachep->colour_off = cache_line_size();
2312 /* Offset must be a multiple of the alignment. */
2313 if (cachep->colour_off < cachep->align)
2314 cachep->colour_off = cachep->align;
2315 cachep->colour = left_over / cachep->colour_off;
2316 cachep->slab_size = slab_size;
2317 cachep->flags = flags;
2318 cachep->allocflags = __GFP_COMP;
2319 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2320 cachep->allocflags |= GFP_DMA;
2321 cachep->size = size;
2322 cachep->reciprocal_buffer_size = reciprocal_value(size);
2324 if (flags & CFLGS_OFF_SLAB) {
2325 cachep->slabp_cache = kmalloc_slab(slab_size, 0u);
2327 * This is a possibility for one of the malloc_sizes caches.
2328 * But since we go off slab only for object size greater than
2329 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2330 * this should not happen at all.
2331 * But leave a BUG_ON for some lucky dude.
2333 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
2336 err = setup_cpu_cache(cachep, gfp);
2338 __kmem_cache_shutdown(cachep);
2342 if (flags & SLAB_DEBUG_OBJECTS) {
2344 * Would deadlock through slab_destroy()->call_rcu()->
2345 * debug_object_activate()->kmem_cache_alloc().
2347 WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU);
2349 slab_set_debugobj_lock_classes(cachep);
2350 } else if (!OFF_SLAB(cachep) && !(flags & SLAB_DESTROY_BY_RCU))
2351 on_slab_lock_classes(cachep);
2357 static void check_irq_off(void)
2359 BUG_ON(!irqs_disabled());
2362 static void check_irq_on(void)
2364 BUG_ON(irqs_disabled());
2367 static void check_spinlock_acquired(struct kmem_cache *cachep)
2371 assert_spin_locked(&cachep->node[numa_mem_id()]->list_lock);
2375 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2379 assert_spin_locked(&cachep->node[node]->list_lock);
2384 #define check_irq_off() do { } while(0)
2385 #define check_irq_on() do { } while(0)
2386 #define check_spinlock_acquired(x) do { } while(0)
2387 #define check_spinlock_acquired_node(x, y) do { } while(0)
2390 static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
2391 struct array_cache *ac,
2392 int force, int node);
2394 static void do_drain(void *arg)
2396 struct kmem_cache *cachep = arg;
2397 struct array_cache *ac;
2398 int node = numa_mem_id();
2401 ac = cpu_cache_get(cachep);
2402 spin_lock(&cachep->node[node]->list_lock);
2403 free_block(cachep, ac->entry, ac->avail, node);
2404 spin_unlock(&cachep->node[node]->list_lock);
2408 static void drain_cpu_caches(struct kmem_cache *cachep)
2410 struct kmem_cache_node *n;
2413 on_each_cpu(do_drain, cachep, 1);
2415 for_each_online_node(node) {
2416 n = cachep->node[node];
2418 drain_alien_cache(cachep, n->alien);
2421 for_each_online_node(node) {
2422 n = cachep->node[node];
2424 drain_array(cachep, n, n->shared, 1, node);
2429 * Remove slabs from the list of free slabs.
2430 * Specify the number of slabs to drain in tofree.
2432 * Returns the actual number of slabs released.
2434 static int drain_freelist(struct kmem_cache *cache,
2435 struct kmem_cache_node *n, int tofree)
2437 struct list_head *p;
2442 while (nr_freed < tofree && !list_empty(&n->slabs_free)) {
2444 spin_lock_irq(&n->list_lock);
2445 p = n->slabs_free.prev;
2446 if (p == &n->slabs_free) {
2447 spin_unlock_irq(&n->list_lock);
2451 slabp = list_entry(p, struct slab, list);
2453 BUG_ON(slabp->active);
2455 list_del(&slabp->list);
2457 * Safe to drop the lock. The slab is no longer linked
2460 n->free_objects -= cache->num;
2461 spin_unlock_irq(&n->list_lock);
2462 slab_destroy(cache, slabp);
2469 /* Called with slab_mutex held to protect against cpu hotplug */
2470 static int __cache_shrink(struct kmem_cache *cachep)
2473 struct kmem_cache_node *n;
2475 drain_cpu_caches(cachep);
2478 for_each_online_node(i) {
2479 n = cachep->node[i];
2483 drain_freelist(cachep, n, slabs_tofree(cachep, n));
2485 ret += !list_empty(&n->slabs_full) ||
2486 !list_empty(&n->slabs_partial);
2488 return (ret ? 1 : 0);
2492 * kmem_cache_shrink - Shrink a cache.
2493 * @cachep: The cache to shrink.
2495 * Releases as many slabs as possible for a cache.
2496 * To help debugging, a zero exit status indicates all slabs were released.
2498 int kmem_cache_shrink(struct kmem_cache *cachep)
2501 BUG_ON(!cachep || in_interrupt());
2504 mutex_lock(&slab_mutex);
2505 ret = __cache_shrink(cachep);
2506 mutex_unlock(&slab_mutex);
2510 EXPORT_SYMBOL(kmem_cache_shrink);
2512 int __kmem_cache_shutdown(struct kmem_cache *cachep)
2515 struct kmem_cache_node *n;
2516 int rc = __cache_shrink(cachep);
2521 for_each_online_cpu(i)
2522 kfree(cachep->array[i]);
2524 /* NUMA: free the node structures */
2525 for_each_online_node(i) {
2526 n = cachep->node[i];
2529 free_alien_cache(n->alien);
2537 * Get the memory for a slab management obj.
2538 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2539 * always come from malloc_sizes caches. The slab descriptor cannot
2540 * come from the same cache which is getting created because,
2541 * when we are searching for an appropriate cache for these
2542 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2543 * If we are creating a malloc_sizes cache here it would not be visible to
2544 * kmem_find_general_cachep till the initialization is complete.
2545 * Hence we cannot have slabp_cache same as the original cache.
2547 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep,
2548 struct page *page, int colour_off,
2549 gfp_t local_flags, int nodeid)
2552 void *addr = page_address(page);
2554 if (OFF_SLAB(cachep)) {
2555 /* Slab management obj is off-slab. */
2556 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2557 local_flags, nodeid);
2559 * If the first object in the slab is leaked (it's allocated
2560 * but no one has a reference to it), we want to make sure
2561 * kmemleak does not treat the ->s_mem pointer as a reference
2562 * to the object. Otherwise we will not report the leak.
2564 kmemleak_scan_area(&slabp->list, sizeof(struct list_head),
2569 slabp = addr + colour_off;
2570 colour_off += cachep->slab_size;
2573 slabp->s_mem = addr + colour_off;
2577 static inline unsigned int *slab_bufctl(struct slab *slabp)
2579 return (unsigned int *) (slabp + 1);
2582 static void cache_init_objs(struct kmem_cache *cachep,
2587 for (i = 0; i < cachep->num; i++) {
2588 void *objp = index_to_obj(cachep, slabp, i);
2590 /* need to poison the objs? */
2591 if (cachep->flags & SLAB_POISON)
2592 poison_obj(cachep, objp, POISON_FREE);
2593 if (cachep->flags & SLAB_STORE_USER)
2594 *dbg_userword(cachep, objp) = NULL;
2596 if (cachep->flags & SLAB_RED_ZONE) {
2597 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2598 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2601 * Constructors are not allowed to allocate memory from the same
2602 * cache which they are a constructor for. Otherwise, deadlock.
2603 * They must also be threaded.
2605 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2606 cachep->ctor(objp + obj_offset(cachep));
2608 if (cachep->flags & SLAB_RED_ZONE) {
2609 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2610 slab_error(cachep, "constructor overwrote the"
2611 " end of an object");
2612 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2613 slab_error(cachep, "constructor overwrote the"
2614 " start of an object");
2616 if ((cachep->size % PAGE_SIZE) == 0 &&
2617 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2618 kernel_map_pages(virt_to_page(objp),
2619 cachep->size / PAGE_SIZE, 0);
2624 slab_bufctl(slabp)[i] = i;
2628 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2630 if (CONFIG_ZONE_DMA_FLAG) {
2631 if (flags & GFP_DMA)
2632 BUG_ON(!(cachep->allocflags & GFP_DMA));
2634 BUG_ON(cachep->allocflags & GFP_DMA);
2638 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2643 objp = index_to_obj(cachep, slabp, slab_bufctl(slabp)[slabp->active]);
2646 WARN_ON(page_to_nid(virt_to_page(objp)) != nodeid);
2652 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2653 void *objp, int nodeid)
2655 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2659 /* Verify that the slab belongs to the intended node */
2660 WARN_ON(page_to_nid(virt_to_page(objp)) != nodeid);
2662 /* Verify double free bug */
2663 for (i = slabp->active; i < cachep->num; i++) {
2664 if (slab_bufctl(slabp)[i] == objnr) {
2665 printk(KERN_ERR "slab: double free detected in cache "
2666 "'%s', objp %p\n", cachep->name, objp);
2672 slab_bufctl(slabp)[slabp->active] = objnr;
2676 * Map pages beginning at addr to the given cache and slab. This is required
2677 * for the slab allocator to be able to lookup the cache and slab of a
2678 * virtual address for kfree, ksize, and slab debugging.
2680 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2683 page->slab_cache = cache;
2684 page->slab_page = slab;
2688 * Grow (by 1) the number of slabs within a cache. This is called by
2689 * kmem_cache_alloc() when there are no active objs left in a cache.
2691 static int cache_grow(struct kmem_cache *cachep,
2692 gfp_t flags, int nodeid, struct page *page)
2697 struct kmem_cache_node *n;
2700 * Be lazy and only check for valid flags here, keeping it out of the
2701 * critical path in kmem_cache_alloc().
2703 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2704 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2706 /* Take the node list lock to change the colour_next on this node */
2708 n = cachep->node[nodeid];
2709 spin_lock(&n->list_lock);
2711 /* Get colour for the slab, and cal the next value. */
2712 offset = n->colour_next;
2714 if (n->colour_next >= cachep->colour)
2716 spin_unlock(&n->list_lock);
2718 offset *= cachep->colour_off;
2720 if (local_flags & __GFP_WAIT)
2724 * The test for missing atomic flag is performed here, rather than
2725 * the more obvious place, simply to reduce the critical path length
2726 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2727 * will eventually be caught here (where it matters).
2729 kmem_flagcheck(cachep, flags);
2732 * Get mem for the objs. Attempt to allocate a physical page from
2736 page = kmem_getpages(cachep, local_flags, nodeid);
2740 /* Get slab management. */
2741 slabp = alloc_slabmgmt(cachep, page, offset,
2742 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2746 slab_map_pages(cachep, slabp, page);
2748 cache_init_objs(cachep, slabp);
2750 if (local_flags & __GFP_WAIT)
2751 local_irq_disable();
2753 spin_lock(&n->list_lock);
2755 /* Make slab active. */
2756 list_add_tail(&slabp->list, &(n->slabs_free));
2757 STATS_INC_GROWN(cachep);
2758 n->free_objects += cachep->num;
2759 spin_unlock(&n->list_lock);
2762 kmem_freepages(cachep, page);
2764 if (local_flags & __GFP_WAIT)
2765 local_irq_disable();
2772 * Perform extra freeing checks:
2773 * - detect bad pointers.
2774 * - POISON/RED_ZONE checking
2776 static void kfree_debugcheck(const void *objp)
2778 if (!virt_addr_valid(objp)) {
2779 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2780 (unsigned long)objp);
2785 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2787 unsigned long long redzone1, redzone2;
2789 redzone1 = *dbg_redzone1(cache, obj);
2790 redzone2 = *dbg_redzone2(cache, obj);
2795 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2798 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2799 slab_error(cache, "double free detected");
2801 slab_error(cache, "memory outside object was overwritten");
2803 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2804 obj, redzone1, redzone2);
2807 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2808 unsigned long caller)
2813 BUG_ON(virt_to_cache(objp) != cachep);
2815 objp -= obj_offset(cachep);
2816 kfree_debugcheck(objp);
2817 slabp = virt_to_slab(objp);
2819 if (cachep->flags & SLAB_RED_ZONE) {
2820 verify_redzone_free(cachep, objp);
2821 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2822 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2824 if (cachep->flags & SLAB_STORE_USER)
2825 *dbg_userword(cachep, objp) = (void *)caller;
2827 objnr = obj_to_index(cachep, slabp, objp);
2829 BUG_ON(objnr >= cachep->num);
2830 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
2832 if (cachep->flags & SLAB_POISON) {
2833 #ifdef CONFIG_DEBUG_PAGEALLOC
2834 if ((cachep->size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
2835 store_stackinfo(cachep, objp, caller);
2836 kernel_map_pages(virt_to_page(objp),
2837 cachep->size / PAGE_SIZE, 0);
2839 poison_obj(cachep, objp, POISON_FREE);
2842 poison_obj(cachep, objp, POISON_FREE);
2849 #define kfree_debugcheck(x) do { } while(0)
2850 #define cache_free_debugcheck(x,objp,z) (objp)
2853 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags,
2857 struct kmem_cache_node *n;
2858 struct array_cache *ac;
2862 node = numa_mem_id();
2863 if (unlikely(force_refill))
2866 ac = cpu_cache_get(cachep);
2867 batchcount = ac->batchcount;
2868 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2870 * If there was little recent activity on this cache, then
2871 * perform only a partial refill. Otherwise we could generate
2874 batchcount = BATCHREFILL_LIMIT;
2876 n = cachep->node[node];
2878 BUG_ON(ac->avail > 0 || !n);
2879 spin_lock(&n->list_lock);
2881 /* See if we can refill from the shared array */
2882 if (n->shared && transfer_objects(ac, n->shared, batchcount)) {
2883 n->shared->touched = 1;
2887 while (batchcount > 0) {
2888 struct list_head *entry;
2890 /* Get slab alloc is to come from. */
2891 entry = n->slabs_partial.next;
2892 if (entry == &n->slabs_partial) {
2893 n->free_touched = 1;
2894 entry = n->slabs_free.next;
2895 if (entry == &n->slabs_free)
2899 slabp = list_entry(entry, struct slab, list);
2900 check_spinlock_acquired(cachep);
2903 * The slab was either on partial or free list so
2904 * there must be at least one object available for
2907 BUG_ON(slabp->active >= cachep->num);
2909 while (slabp->active < cachep->num && batchcount--) {
2910 STATS_INC_ALLOCED(cachep);
2911 STATS_INC_ACTIVE(cachep);
2912 STATS_SET_HIGH(cachep);
2914 ac_put_obj(cachep, ac, slab_get_obj(cachep, slabp,
2918 /* move slabp to correct slabp list: */
2919 list_del(&slabp->list);
2920 if (slabp->active == cachep->num)
2921 list_add(&slabp->list, &n->slabs_full);
2923 list_add(&slabp->list, &n->slabs_partial);
2927 n->free_objects -= ac->avail;
2929 spin_unlock(&n->list_lock);
2931 if (unlikely(!ac->avail)) {
2934 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
2936 /* cache_grow can reenable interrupts, then ac could change. */
2937 ac = cpu_cache_get(cachep);
2938 node = numa_mem_id();
2940 /* no objects in sight? abort */
2941 if (!x && (ac->avail == 0 || force_refill))
2944 if (!ac->avail) /* objects refilled by interrupt? */
2949 return ac_get_obj(cachep, ac, flags, force_refill);
2952 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
2955 might_sleep_if(flags & __GFP_WAIT);
2957 kmem_flagcheck(cachep, flags);
2962 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
2963 gfp_t flags, void *objp, unsigned long caller)
2967 if (cachep->flags & SLAB_POISON) {
2968 #ifdef CONFIG_DEBUG_PAGEALLOC
2969 if ((cachep->size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
2970 kernel_map_pages(virt_to_page(objp),
2971 cachep->size / PAGE_SIZE, 1);
2973 check_poison_obj(cachep, objp);
2975 check_poison_obj(cachep, objp);
2977 poison_obj(cachep, objp, POISON_INUSE);
2979 if (cachep->flags & SLAB_STORE_USER)
2980 *dbg_userword(cachep, objp) = (void *)caller;
2982 if (cachep->flags & SLAB_RED_ZONE) {
2983 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
2984 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2985 slab_error(cachep, "double free, or memory outside"
2986 " object was overwritten");
2988 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
2989 objp, *dbg_redzone1(cachep, objp),
2990 *dbg_redzone2(cachep, objp));
2992 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2993 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2995 objp += obj_offset(cachep);
2996 if (cachep->ctor && cachep->flags & SLAB_POISON)
2998 if (ARCH_SLAB_MINALIGN &&
2999 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
3000 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3001 objp, (int)ARCH_SLAB_MINALIGN);
3006 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3009 static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
3011 if (cachep == kmem_cache)
3014 return should_failslab(cachep->object_size, flags, cachep->flags);
3017 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3020 struct array_cache *ac;
3021 bool force_refill = false;
3025 ac = cpu_cache_get(cachep);
3026 if (likely(ac->avail)) {
3028 objp = ac_get_obj(cachep, ac, flags, false);
3031 * Allow for the possibility all avail objects are not allowed
3032 * by the current flags
3035 STATS_INC_ALLOCHIT(cachep);
3038 force_refill = true;
3041 STATS_INC_ALLOCMISS(cachep);
3042 objp = cache_alloc_refill(cachep, flags, force_refill);
3044 * the 'ac' may be updated by cache_alloc_refill(),
3045 * and kmemleak_erase() requires its correct value.
3047 ac = cpu_cache_get(cachep);
3051 * To avoid a false negative, if an object that is in one of the
3052 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3053 * treat the array pointers as a reference to the object.
3056 kmemleak_erase(&ac->entry[ac->avail]);
3062 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3064 * If we are in_interrupt, then process context, including cpusets and
3065 * mempolicy, may not apply and should not be used for allocation policy.
3067 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3069 int nid_alloc, nid_here;
3071 if (in_interrupt() || (flags & __GFP_THISNODE))
3073 nid_alloc = nid_here = numa_mem_id();
3074 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3075 nid_alloc = cpuset_slab_spread_node();
3076 else if (current->mempolicy)
3077 nid_alloc = slab_node();
3078 if (nid_alloc != nid_here)
3079 return ____cache_alloc_node(cachep, flags, nid_alloc);
3084 * Fallback function if there was no memory available and no objects on a
3085 * certain node and fall back is permitted. First we scan all the
3086 * available node for available objects. If that fails then we
3087 * perform an allocation without specifying a node. This allows the page
3088 * allocator to do its reclaim / fallback magic. We then insert the
3089 * slab into the proper nodelist and then allocate from it.
3091 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3093 struct zonelist *zonelist;
3097 enum zone_type high_zoneidx = gfp_zone(flags);
3100 unsigned int cpuset_mems_cookie;
3102 if (flags & __GFP_THISNODE)
3105 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3108 cpuset_mems_cookie = get_mems_allowed();
3109 zonelist = node_zonelist(slab_node(), flags);
3113 * Look through allowed nodes for objects available
3114 * from existing per node queues.
3116 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3117 nid = zone_to_nid(zone);
3119 if (cpuset_zone_allowed_hardwall(zone, flags) &&
3121 cache->node[nid]->free_objects) {
3122 obj = ____cache_alloc_node(cache,
3123 flags | GFP_THISNODE, nid);
3131 * This allocation will be performed within the constraints
3132 * of the current cpuset / memory policy requirements.
3133 * We may trigger various forms of reclaim on the allowed
3134 * set and go into memory reserves if necessary.
3138 if (local_flags & __GFP_WAIT)
3140 kmem_flagcheck(cache, flags);
3141 page = kmem_getpages(cache, local_flags, numa_mem_id());
3142 if (local_flags & __GFP_WAIT)
3143 local_irq_disable();
3146 * Insert into the appropriate per node queues
3148 nid = page_to_nid(page);
3149 if (cache_grow(cache, flags, nid, page)) {
3150 obj = ____cache_alloc_node(cache,
3151 flags | GFP_THISNODE, nid);
3154 * Another processor may allocate the
3155 * objects in the slab since we are
3156 * not holding any locks.
3160 /* cache_grow already freed obj */
3166 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !obj))
3172 * A interface to enable slab creation on nodeid
3174 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3177 struct list_head *entry;
3179 struct kmem_cache_node *n;
3183 VM_BUG_ON(nodeid > num_online_nodes());
3184 n = cachep->node[nodeid];
3189 spin_lock(&n->list_lock);
3190 entry = n->slabs_partial.next;
3191 if (entry == &n->slabs_partial) {
3192 n->free_touched = 1;
3193 entry = n->slabs_free.next;
3194 if (entry == &n->slabs_free)
3198 slabp = list_entry(entry, struct slab, list);
3199 check_spinlock_acquired_node(cachep, nodeid);
3201 STATS_INC_NODEALLOCS(cachep);
3202 STATS_INC_ACTIVE(cachep);
3203 STATS_SET_HIGH(cachep);
3205 BUG_ON(slabp->active == cachep->num);
3207 obj = slab_get_obj(cachep, slabp, nodeid);
3209 /* move slabp to correct slabp list: */
3210 list_del(&slabp->list);
3212 if (slabp->active == cachep->num)
3213 list_add(&slabp->list, &n->slabs_full);
3215 list_add(&slabp->list, &n->slabs_partial);
3217 spin_unlock(&n->list_lock);
3221 spin_unlock(&n->list_lock);
3222 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
3226 return fallback_alloc(cachep, flags);
3232 static __always_inline void *
3233 slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3234 unsigned long caller)
3236 unsigned long save_flags;
3238 int slab_node = numa_mem_id();
3240 flags &= gfp_allowed_mask;
3242 lockdep_trace_alloc(flags);
3244 if (slab_should_failslab(cachep, flags))
3247 cachep = memcg_kmem_get_cache(cachep, flags);
3249 cache_alloc_debugcheck_before(cachep, flags);
3250 local_irq_save(save_flags);
3252 if (nodeid == NUMA_NO_NODE)
3255 if (unlikely(!cachep->node[nodeid])) {
3256 /* Node not bootstrapped yet */
3257 ptr = fallback_alloc(cachep, flags);
3261 if (nodeid == slab_node) {
3263 * Use the locally cached objects if possible.
3264 * However ____cache_alloc does not allow fallback
3265 * to other nodes. It may fail while we still have
3266 * objects on other nodes available.
3268 ptr = ____cache_alloc(cachep, flags);
3272 /* ___cache_alloc_node can fall back to other nodes */
3273 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3275 local_irq_restore(save_flags);
3276 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3277 kmemleak_alloc_recursive(ptr, cachep->object_size, 1, cachep->flags,
3281 kmemcheck_slab_alloc(cachep, flags, ptr, cachep->object_size);
3283 if (unlikely((flags & __GFP_ZERO) && ptr))
3284 memset(ptr, 0, cachep->object_size);
3289 static __always_inline void *
3290 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3294 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3295 objp = alternate_node_alloc(cache, flags);
3299 objp = ____cache_alloc(cache, flags);
3302 * We may just have run out of memory on the local node.
3303 * ____cache_alloc_node() knows how to locate memory on other nodes
3306 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3313 static __always_inline void *
3314 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3316 return ____cache_alloc(cachep, flags);
3319 #endif /* CONFIG_NUMA */
3321 static __always_inline void *
3322 slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
3324 unsigned long save_flags;
3327 flags &= gfp_allowed_mask;
3329 lockdep_trace_alloc(flags);
3331 if (slab_should_failslab(cachep, flags))
3334 cachep = memcg_kmem_get_cache(cachep, flags);
3336 cache_alloc_debugcheck_before(cachep, flags);
3337 local_irq_save(save_flags);
3338 objp = __do_cache_alloc(cachep, flags);
3339 local_irq_restore(save_flags);
3340 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3341 kmemleak_alloc_recursive(objp, cachep->object_size, 1, cachep->flags,
3346 kmemcheck_slab_alloc(cachep, flags, objp, cachep->object_size);
3348 if (unlikely((flags & __GFP_ZERO) && objp))
3349 memset(objp, 0, cachep->object_size);
3355 * Caller needs to acquire correct kmem_list's list_lock
3357 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3361 struct kmem_cache_node *n;
3363 for (i = 0; i < nr_objects; i++) {
3367 clear_obj_pfmemalloc(&objpp[i]);
3370 slabp = virt_to_slab(objp);
3371 n = cachep->node[node];
3372 list_del(&slabp->list);
3373 check_spinlock_acquired_node(cachep, node);
3374 slab_put_obj(cachep, slabp, objp, node);
3375 STATS_DEC_ACTIVE(cachep);
3378 /* fixup slab chains */
3379 if (slabp->active == 0) {
3380 if (n->free_objects > n->free_limit) {
3381 n->free_objects -= cachep->num;
3382 /* No need to drop any previously held
3383 * lock here, even if we have a off-slab slab
3384 * descriptor it is guaranteed to come from
3385 * a different cache, refer to comments before
3388 slab_destroy(cachep, slabp);
3390 list_add(&slabp->list, &n->slabs_free);
3393 /* Unconditionally move a slab to the end of the
3394 * partial list on free - maximum time for the
3395 * other objects to be freed, too.
3397 list_add_tail(&slabp->list, &n->slabs_partial);
3402 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3405 struct kmem_cache_node *n;
3406 int node = numa_mem_id();
3408 batchcount = ac->batchcount;
3410 BUG_ON(!batchcount || batchcount > ac->avail);
3413 n = cachep->node[node];
3414 spin_lock(&n->list_lock);
3416 struct array_cache *shared_array = n->shared;
3417 int max = shared_array->limit - shared_array->avail;
3419 if (batchcount > max)
3421 memcpy(&(shared_array->entry[shared_array->avail]),
3422 ac->entry, sizeof(void *) * batchcount);
3423 shared_array->avail += batchcount;
3428 free_block(cachep, ac->entry, batchcount, node);
3433 struct list_head *p;
3435 p = n->slabs_free.next;
3436 while (p != &(n->slabs_free)) {
3439 slabp = list_entry(p, struct slab, list);
3440 BUG_ON(slabp->active);
3445 STATS_SET_FREEABLE(cachep, i);
3448 spin_unlock(&n->list_lock);
3449 ac->avail -= batchcount;
3450 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3454 * Release an obj back to its cache. If the obj has a constructed state, it must
3455 * be in this state _before_ it is released. Called with disabled ints.
3457 static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3458 unsigned long caller)
3460 struct array_cache *ac = cpu_cache_get(cachep);
3463 kmemleak_free_recursive(objp, cachep->flags);
3464 objp = cache_free_debugcheck(cachep, objp, caller);
3466 kmemcheck_slab_free(cachep, objp, cachep->object_size);
3469 * Skip calling cache_free_alien() when the platform is not numa.
3470 * This will avoid cache misses that happen while accessing slabp (which
3471 * is per page memory reference) to get nodeid. Instead use a global
3472 * variable to skip the call, which is mostly likely to be present in
3475 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3478 if (likely(ac->avail < ac->limit)) {
3479 STATS_INC_FREEHIT(cachep);
3481 STATS_INC_FREEMISS(cachep);
3482 cache_flusharray(cachep, ac);
3485 ac_put_obj(cachep, ac, objp);
3489 * kmem_cache_alloc - Allocate an object
3490 * @cachep: The cache to allocate from.
3491 * @flags: See kmalloc().
3493 * Allocate an object from this cache. The flags are only relevant
3494 * if the cache has no available objects.
3496 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3498 void *ret = slab_alloc(cachep, flags, _RET_IP_);
3500 trace_kmem_cache_alloc(_RET_IP_, ret,
3501 cachep->object_size, cachep->size, flags);
3505 EXPORT_SYMBOL(kmem_cache_alloc);
3507 #ifdef CONFIG_TRACING
3509 kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
3513 ret = slab_alloc(cachep, flags, _RET_IP_);
3515 trace_kmalloc(_RET_IP_, ret,
3516 size, cachep->size, flags);
3519 EXPORT_SYMBOL(kmem_cache_alloc_trace);
3524 * kmem_cache_alloc_node - Allocate an object on the specified node
3525 * @cachep: The cache to allocate from.
3526 * @flags: See kmalloc().
3527 * @nodeid: node number of the target node.
3529 * Identical to kmem_cache_alloc but it will allocate memory on the given
3530 * node, which can improve the performance for cpu bound structures.
3532 * Fallback to other node is possible if __GFP_THISNODE is not set.
3534 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3536 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3538 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3539 cachep->object_size, cachep->size,
3544 EXPORT_SYMBOL(kmem_cache_alloc_node);
3546 #ifdef CONFIG_TRACING
3547 void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
3554 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3556 trace_kmalloc_node(_RET_IP_, ret,
3561 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3564 static __always_inline void *
3565 __do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
3567 struct kmem_cache *cachep;
3569 cachep = kmalloc_slab(size, flags);
3570 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3572 return kmem_cache_alloc_node_trace(cachep, flags, node, size);
3575 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3576 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3578 return __do_kmalloc_node(size, flags, node, _RET_IP_);
3580 EXPORT_SYMBOL(__kmalloc_node);
3582 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3583 int node, unsigned long caller)
3585 return __do_kmalloc_node(size, flags, node, caller);
3587 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3589 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3591 return __do_kmalloc_node(size, flags, node, 0);
3593 EXPORT_SYMBOL(__kmalloc_node);
3594 #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3595 #endif /* CONFIG_NUMA */
3598 * __do_kmalloc - allocate memory
3599 * @size: how many bytes of memory are required.
3600 * @flags: the type of memory to allocate (see kmalloc).
3601 * @caller: function caller for debug tracking of the caller
3603 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3604 unsigned long caller)
3606 struct kmem_cache *cachep;
3609 /* If you want to save a few bytes .text space: replace
3611 * Then kmalloc uses the uninlined functions instead of the inline
3614 cachep = kmalloc_slab(size, flags);
3615 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3617 ret = slab_alloc(cachep, flags, caller);
3619 trace_kmalloc(caller, ret,
3620 size, cachep->size, flags);
3626 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3627 void *__kmalloc(size_t size, gfp_t flags)
3629 return __do_kmalloc(size, flags, _RET_IP_);
3631 EXPORT_SYMBOL(__kmalloc);
3633 void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3635 return __do_kmalloc(size, flags, caller);
3637 EXPORT_SYMBOL(__kmalloc_track_caller);
3640 void *__kmalloc(size_t size, gfp_t flags)
3642 return __do_kmalloc(size, flags, 0);
3644 EXPORT_SYMBOL(__kmalloc);
3648 * kmem_cache_free - Deallocate an object
3649 * @cachep: The cache the allocation was from.
3650 * @objp: The previously allocated object.
3652 * Free an object which was previously allocated from this
3655 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3657 unsigned long flags;
3658 cachep = cache_from_obj(cachep, objp);
3662 local_irq_save(flags);
3663 debug_check_no_locks_freed(objp, cachep->object_size);
3664 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3665 debug_check_no_obj_freed(objp, cachep->object_size);
3666 __cache_free(cachep, objp, _RET_IP_);
3667 local_irq_restore(flags);
3669 trace_kmem_cache_free(_RET_IP_, objp);
3671 EXPORT_SYMBOL(kmem_cache_free);
3674 * kfree - free previously allocated memory
3675 * @objp: pointer returned by kmalloc.
3677 * If @objp is NULL, no operation is performed.
3679 * Don't free memory not originally allocated by kmalloc()
3680 * or you will run into trouble.
3682 void kfree(const void *objp)
3684 struct kmem_cache *c;
3685 unsigned long flags;
3687 trace_kfree(_RET_IP_, objp);
3689 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3691 local_irq_save(flags);
3692 kfree_debugcheck(objp);
3693 c = virt_to_cache(objp);
3694 debug_check_no_locks_freed(objp, c->object_size);
3696 debug_check_no_obj_freed(objp, c->object_size);
3697 __cache_free(c, (void *)objp, _RET_IP_);
3698 local_irq_restore(flags);
3700 EXPORT_SYMBOL(kfree);
3703 * This initializes kmem_cache_node or resizes various caches for all nodes.
3705 static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
3708 struct kmem_cache_node *n;
3709 struct array_cache *new_shared;
3710 struct array_cache **new_alien = NULL;
3712 for_each_online_node(node) {
3714 if (use_alien_caches) {
3715 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
3721 if (cachep->shared) {
3722 new_shared = alloc_arraycache(node,
3723 cachep->shared*cachep->batchcount,
3726 free_alien_cache(new_alien);
3731 n = cachep->node[node];
3733 struct array_cache *shared = n->shared;
3735 spin_lock_irq(&n->list_lock);
3738 free_block(cachep, shared->entry,
3739 shared->avail, node);
3741 n->shared = new_shared;
3743 n->alien = new_alien;
3746 n->free_limit = (1 + nr_cpus_node(node)) *
3747 cachep->batchcount + cachep->num;
3748 spin_unlock_irq(&n->list_lock);
3750 free_alien_cache(new_alien);
3753 n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
3755 free_alien_cache(new_alien);
3760 kmem_cache_node_init(n);
3761 n->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3762 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3763 n->shared = new_shared;
3764 n->alien = new_alien;
3765 n->free_limit = (1 + nr_cpus_node(node)) *
3766 cachep->batchcount + cachep->num;
3767 cachep->node[node] = n;
3772 if (!cachep->list.next) {
3773 /* Cache is not active yet. Roll back what we did */
3776 if (cachep->node[node]) {
3777 n = cachep->node[node];
3780 free_alien_cache(n->alien);
3782 cachep->node[node] = NULL;
3790 struct ccupdate_struct {
3791 struct kmem_cache *cachep;
3792 struct array_cache *new[0];
3795 static void do_ccupdate_local(void *info)
3797 struct ccupdate_struct *new = info;
3798 struct array_cache *old;
3801 old = cpu_cache_get(new->cachep);
3803 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3804 new->new[smp_processor_id()] = old;
3807 /* Always called with the slab_mutex held */
3808 static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
3809 int batchcount, int shared, gfp_t gfp)
3811 struct ccupdate_struct *new;
3814 new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *),
3819 for_each_online_cpu(i) {
3820 new->new[i] = alloc_arraycache(cpu_to_mem(i), limit,
3823 for (i--; i >= 0; i--)
3829 new->cachep = cachep;
3831 on_each_cpu(do_ccupdate_local, (void *)new, 1);
3834 cachep->batchcount = batchcount;
3835 cachep->limit = limit;
3836 cachep->shared = shared;
3838 for_each_online_cpu(i) {
3839 struct array_cache *ccold = new->new[i];
3842 spin_lock_irq(&cachep->node[cpu_to_mem(i)]->list_lock);
3843 free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i));
3844 spin_unlock_irq(&cachep->node[cpu_to_mem(i)]->list_lock);
3848 return alloc_kmemlist(cachep, gfp);
3851 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3852 int batchcount, int shared, gfp_t gfp)
3855 struct kmem_cache *c = NULL;
3858 ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3860 if (slab_state < FULL)
3863 if ((ret < 0) || !is_root_cache(cachep))
3866 VM_BUG_ON(!mutex_is_locked(&slab_mutex));
3867 for_each_memcg_cache_index(i) {
3868 c = cache_from_memcg(cachep, i);
3870 /* return value determined by the parent cache only */
3871 __do_tune_cpucache(c, limit, batchcount, shared, gfp);
3877 /* Called with slab_mutex held always */
3878 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
3885 if (!is_root_cache(cachep)) {
3886 struct kmem_cache *root = memcg_root_cache(cachep);
3887 limit = root->limit;
3888 shared = root->shared;
3889 batchcount = root->batchcount;
3892 if (limit && shared && batchcount)
3895 * The head array serves three purposes:
3896 * - create a LIFO ordering, i.e. return objects that are cache-warm
3897 * - reduce the number of spinlock operations.
3898 * - reduce the number of linked list operations on the slab and
3899 * bufctl chains: array operations are cheaper.
3900 * The numbers are guessed, we should auto-tune as described by
3903 if (cachep->size > 131072)
3905 else if (cachep->size > PAGE_SIZE)
3907 else if (cachep->size > 1024)
3909 else if (cachep->size > 256)
3915 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3916 * allocation behaviour: Most allocs on one cpu, most free operations
3917 * on another cpu. For these cases, an efficient object passing between
3918 * cpus is necessary. This is provided by a shared array. The array
3919 * replaces Bonwick's magazine layer.
3920 * On uniprocessor, it's functionally equivalent (but less efficient)
3921 * to a larger limit. Thus disabled by default.
3924 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
3929 * With debugging enabled, large batchcount lead to excessively long
3930 * periods with disabled local interrupts. Limit the batchcount
3935 batchcount = (limit + 1) / 2;
3937 err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3939 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
3940 cachep->name, -err);
3945 * Drain an array if it contains any elements taking the node lock only if
3946 * necessary. Note that the node listlock also protects the array_cache
3947 * if drain_array() is used on the shared array.
3949 static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
3950 struct array_cache *ac, int force, int node)
3954 if (!ac || !ac->avail)
3956 if (ac->touched && !force) {
3959 spin_lock_irq(&n->list_lock);
3961 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3962 if (tofree > ac->avail)
3963 tofree = (ac->avail + 1) / 2;
3964 free_block(cachep, ac->entry, tofree, node);
3965 ac->avail -= tofree;
3966 memmove(ac->entry, &(ac->entry[tofree]),
3967 sizeof(void *) * ac->avail);
3969 spin_unlock_irq(&n->list_lock);
3974 * cache_reap - Reclaim memory from caches.
3975 * @w: work descriptor
3977 * Called from workqueue/eventd every few seconds.
3979 * - clear the per-cpu caches for this CPU.
3980 * - return freeable pages to the main free memory pool.
3982 * If we cannot acquire the cache chain mutex then just give up - we'll try
3983 * again on the next iteration.
3985 static void cache_reap(struct work_struct *w)
3987 struct kmem_cache *searchp;
3988 struct kmem_cache_node *n;
3989 int node = numa_mem_id();
3990 struct delayed_work *work = to_delayed_work(w);
3992 if (!mutex_trylock(&slab_mutex))
3993 /* Give up. Setup the next iteration. */
3996 list_for_each_entry(searchp, &slab_caches, list) {
4000 * We only take the node lock if absolutely necessary and we
4001 * have established with reasonable certainty that
4002 * we can do some work if the lock was obtained.
4004 n = searchp->node[node];
4006 reap_alien(searchp, n);
4008 drain_array(searchp, n, cpu_cache_get(searchp), 0, node);
4011 * These are racy checks but it does not matter
4012 * if we skip one check or scan twice.
4014 if (time_after(n->next_reap, jiffies))
4017 n->next_reap = jiffies + REAPTIMEOUT_LIST3;
4019 drain_array(searchp, n, n->shared, 0, node);
4021 if (n->free_touched)
4022 n->free_touched = 0;
4026 freed = drain_freelist(searchp, n, (n->free_limit +
4027 5 * searchp->num - 1) / (5 * searchp->num));
4028 STATS_ADD_REAPED(searchp, freed);
4034 mutex_unlock(&slab_mutex);
4037 /* Set up the next iteration */
4038 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
4041 #ifdef CONFIG_SLABINFO
4042 void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
4045 unsigned long active_objs;
4046 unsigned long num_objs;
4047 unsigned long active_slabs = 0;
4048 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4052 struct kmem_cache_node *n;
4056 for_each_online_node(node) {
4057 n = cachep->node[node];
4062 spin_lock_irq(&n->list_lock);
4064 list_for_each_entry(slabp, &n->slabs_full, list) {
4065 if (slabp->active != cachep->num && !error)
4066 error = "slabs_full accounting error";
4067 active_objs += cachep->num;
4070 list_for_each_entry(slabp, &n->slabs_partial, list) {
4071 if (slabp->active == cachep->num && !error)
4072 error = "slabs_partial accounting error";
4073 if (!slabp->active && !error)
4074 error = "slabs_partial accounting error";
4075 active_objs += slabp->active;
4078 list_for_each_entry(slabp, &n->slabs_free, list) {
4079 if (slabp->active && !error)
4080 error = "slabs_free accounting error";
4083 free_objects += n->free_objects;
4085 shared_avail += n->shared->avail;
4087 spin_unlock_irq(&n->list_lock);
4089 num_slabs += active_slabs;
4090 num_objs = num_slabs * cachep->num;
4091 if (num_objs - active_objs != free_objects && !error)
4092 error = "free_objects accounting error";
4094 name = cachep->name;
4096 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4098 sinfo->active_objs = active_objs;
4099 sinfo->num_objs = num_objs;
4100 sinfo->active_slabs = active_slabs;
4101 sinfo->num_slabs = num_slabs;
4102 sinfo->shared_avail = shared_avail;
4103 sinfo->limit = cachep->limit;
4104 sinfo->batchcount = cachep->batchcount;
4105 sinfo->shared = cachep->shared;
4106 sinfo->objects_per_slab = cachep->num;
4107 sinfo->cache_order = cachep->gfporder;
4110 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
4114 unsigned long high = cachep->high_mark;
4115 unsigned long allocs = cachep->num_allocations;
4116 unsigned long grown = cachep->grown;
4117 unsigned long reaped = cachep->reaped;
4118 unsigned long errors = cachep->errors;
4119 unsigned long max_freeable = cachep->max_freeable;
4120 unsigned long node_allocs = cachep->node_allocs;
4121 unsigned long node_frees = cachep->node_frees;
4122 unsigned long overflows = cachep->node_overflow;
4124 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4125 "%4lu %4lu %4lu %4lu %4lu",
4126 allocs, high, grown,
4127 reaped, errors, max_freeable, node_allocs,
4128 node_frees, overflows);
4132 unsigned long allochit = atomic_read(&cachep->allochit);
4133 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4134 unsigned long freehit = atomic_read(&cachep->freehit);
4135 unsigned long freemiss = atomic_read(&cachep->freemiss);
4137 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4138 allochit, allocmiss, freehit, freemiss);
4143 #define MAX_SLABINFO_WRITE 128
4145 * slabinfo_write - Tuning for the slab allocator
4147 * @buffer: user buffer
4148 * @count: data length
4151 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4152 size_t count, loff_t *ppos)
4154 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4155 int limit, batchcount, shared, res;
4156 struct kmem_cache *cachep;
4158 if (count > MAX_SLABINFO_WRITE)
4160 if (copy_from_user(&kbuf, buffer, count))
4162 kbuf[MAX_SLABINFO_WRITE] = '\0';
4164 tmp = strchr(kbuf, ' ');
4169 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4172 /* Find the cache in the chain of caches. */
4173 mutex_lock(&slab_mutex);
4175 list_for_each_entry(cachep, &slab_caches, list) {
4176 if (!strcmp(cachep->name, kbuf)) {
4177 if (limit < 1 || batchcount < 1 ||
4178 batchcount > limit || shared < 0) {
4181 res = do_tune_cpucache(cachep, limit,
4188 mutex_unlock(&slab_mutex);
4194 #ifdef CONFIG_DEBUG_SLAB_LEAK
4196 static void *leaks_start(struct seq_file *m, loff_t *pos)
4198 mutex_lock(&slab_mutex);
4199 return seq_list_start(&slab_caches, *pos);
4202 static inline int add_caller(unsigned long *n, unsigned long v)
4212 unsigned long *q = p + 2 * i;
4226 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4232 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4239 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->size) {
4242 for (j = s->active; j < c->num; j++) {
4243 /* Skip freed item */
4244 if (slab_bufctl(s)[j] == i) {
4252 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4257 static void show_symbol(struct seq_file *m, unsigned long address)
4259 #ifdef CONFIG_KALLSYMS
4260 unsigned long offset, size;
4261 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4263 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4264 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4266 seq_printf(m, " [%s]", modname);
4270 seq_printf(m, "%p", (void *)address);
4273 static int leaks_show(struct seq_file *m, void *p)
4275 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
4277 struct kmem_cache_node *n;
4279 unsigned long *x = m->private;
4283 if (!(cachep->flags & SLAB_STORE_USER))
4285 if (!(cachep->flags & SLAB_RED_ZONE))
4288 /* OK, we can do it */
4292 for_each_online_node(node) {
4293 n = cachep->node[node];
4298 spin_lock_irq(&n->list_lock);
4300 list_for_each_entry(slabp, &n->slabs_full, list)
4301 handle_slab(x, cachep, slabp);
4302 list_for_each_entry(slabp, &n->slabs_partial, list)
4303 handle_slab(x, cachep, slabp);
4304 spin_unlock_irq(&n->list_lock);
4306 name = cachep->name;
4308 /* Increase the buffer size */
4309 mutex_unlock(&slab_mutex);
4310 m->private = kzalloc(x[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4312 /* Too bad, we are really out */
4314 mutex_lock(&slab_mutex);
4317 *(unsigned long *)m->private = x[0] * 2;
4319 mutex_lock(&slab_mutex);
4320 /* Now make sure this entry will be retried */
4324 for (i = 0; i < x[1]; i++) {
4325 seq_printf(m, "%s: %lu ", name, x[2*i+3]);
4326 show_symbol(m, x[2*i+2]);
4333 static const struct seq_operations slabstats_op = {
4334 .start = leaks_start,
4340 static int slabstats_open(struct inode *inode, struct file *file)
4342 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4345 ret = seq_open(file, &slabstats_op);
4347 struct seq_file *m = file->private_data;
4348 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4357 static const struct file_operations proc_slabstats_operations = {
4358 .open = slabstats_open,
4360 .llseek = seq_lseek,
4361 .release = seq_release_private,
4365 static int __init slab_proc_init(void)
4367 #ifdef CONFIG_DEBUG_SLAB_LEAK
4368 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4372 module_init(slab_proc_init);
4376 * ksize - get the actual amount of memory allocated for a given object
4377 * @objp: Pointer to the object
4379 * kmalloc may internally round up allocations and return more memory
4380 * than requested. ksize() can be used to determine the actual amount of
4381 * memory allocated. The caller may use this additional memory, even though
4382 * a smaller amount of memory was initially specified with the kmalloc call.
4383 * The caller must guarantee that objp points to a valid object previously
4384 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4385 * must not be freed during the duration of the call.
4387 size_t ksize(const void *objp)
4390 if (unlikely(objp == ZERO_SIZE_PTR))
4393 return virt_to_cache(objp)->object_size;
4395 EXPORT_SYMBOL(ksize);