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
160 #define FREELIST_BYTE_INDEX (((PAGE_SIZE >> BITS_PER_BYTE) \
161 <= SLAB_OBJ_MIN_SIZE) ? 1 : 0)
163 #if FREELIST_BYTE_INDEX
164 typedef unsigned char freelist_idx_t;
166 typedef unsigned short freelist_idx_t;
169 #define SLAB_OBJ_MAX_NUM ((1 << sizeof(freelist_idx_t) * BITS_PER_BYTE) - 1)
172 * true if a page was allocated from pfmemalloc reserves for network-based
175 static bool pfmemalloc_active __read_mostly;
181 * - LIFO ordering, to hand out cache-warm objects from _alloc
182 * - reduce the number of linked list operations
183 * - reduce spinlock operations
185 * The limit is stored in the per-cpu structure to reduce the data cache
192 unsigned int batchcount;
193 unsigned int touched;
195 * Must have this definition in here for the proper
196 * alignment of array_cache. Also simplifies accessing
199 * Entries should not be directly dereferenced as
200 * entries belonging to slabs marked pfmemalloc will
201 * have the lower bits set SLAB_OBJ_PFMEMALLOC
207 struct array_cache ac;
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,
249 int node, struct list_head *list);
250 static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list);
251 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
252 static void cache_reap(struct work_struct *unused);
254 static int slab_early_init = 1;
256 #define INDEX_AC kmalloc_index(sizeof(struct arraycache_init))
257 #define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
259 static void kmem_cache_node_init(struct kmem_cache_node *parent)
261 INIT_LIST_HEAD(&parent->slabs_full);
262 INIT_LIST_HEAD(&parent->slabs_partial);
263 INIT_LIST_HEAD(&parent->slabs_free);
264 parent->shared = NULL;
265 parent->alien = NULL;
266 parent->colour_next = 0;
267 spin_lock_init(&parent->list_lock);
268 parent->free_objects = 0;
269 parent->free_touched = 0;
272 #define MAKE_LIST(cachep, listp, slab, nodeid) \
274 INIT_LIST_HEAD(listp); \
275 list_splice(&get_node(cachep, nodeid)->slab, listp); \
278 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
280 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
281 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
282 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
285 #define CFLGS_OFF_SLAB (0x80000000UL)
286 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
288 #define BATCHREFILL_LIMIT 16
290 * Optimization question: fewer reaps means less probability for unnessary
291 * cpucache drain/refill cycles.
293 * OTOH the cpuarrays can contain lots of objects,
294 * which could lock up otherwise freeable slabs.
296 #define REAPTIMEOUT_AC (2*HZ)
297 #define REAPTIMEOUT_NODE (4*HZ)
300 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
301 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
302 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
303 #define STATS_INC_GROWN(x) ((x)->grown++)
304 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
305 #define STATS_SET_HIGH(x) \
307 if ((x)->num_active > (x)->high_mark) \
308 (x)->high_mark = (x)->num_active; \
310 #define STATS_INC_ERR(x) ((x)->errors++)
311 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
312 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
313 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
314 #define STATS_SET_FREEABLE(x, i) \
316 if ((x)->max_freeable < i) \
317 (x)->max_freeable = i; \
319 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
320 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
321 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
322 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
324 #define STATS_INC_ACTIVE(x) do { } while (0)
325 #define STATS_DEC_ACTIVE(x) do { } while (0)
326 #define STATS_INC_ALLOCED(x) do { } while (0)
327 #define STATS_INC_GROWN(x) do { } while (0)
328 #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
329 #define STATS_SET_HIGH(x) do { } while (0)
330 #define STATS_INC_ERR(x) do { } while (0)
331 #define STATS_INC_NODEALLOCS(x) do { } while (0)
332 #define STATS_INC_NODEFREES(x) do { } while (0)
333 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
334 #define STATS_SET_FREEABLE(x, i) do { } while (0)
335 #define STATS_INC_ALLOCHIT(x) do { } while (0)
336 #define STATS_INC_ALLOCMISS(x) do { } while (0)
337 #define STATS_INC_FREEHIT(x) do { } while (0)
338 #define STATS_INC_FREEMISS(x) do { } while (0)
344 * memory layout of objects:
346 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
347 * the end of an object is aligned with the end of the real
348 * allocation. Catches writes behind the end of the allocation.
349 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
351 * cachep->obj_offset: The real object.
352 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
353 * cachep->size - 1* BYTES_PER_WORD: last caller address
354 * [BYTES_PER_WORD long]
356 static int obj_offset(struct kmem_cache *cachep)
358 return cachep->obj_offset;
361 static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
363 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
364 return (unsigned long long*) (objp + obj_offset(cachep) -
365 sizeof(unsigned long long));
368 static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
370 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
371 if (cachep->flags & SLAB_STORE_USER)
372 return (unsigned long long *)(objp + cachep->size -
373 sizeof(unsigned long long) -
375 return (unsigned long long *) (objp + cachep->size -
376 sizeof(unsigned long long));
379 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
381 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
382 return (void **)(objp + cachep->size - BYTES_PER_WORD);
387 #define obj_offset(x) 0
388 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
389 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
390 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
394 #define OBJECT_FREE (0)
395 #define OBJECT_ACTIVE (1)
397 #ifdef CONFIG_DEBUG_SLAB_LEAK
399 static void set_obj_status(struct page *page, int idx, int val)
403 struct kmem_cache *cachep = page->slab_cache;
405 freelist_size = cachep->num * sizeof(freelist_idx_t);
406 status = (char *)page->freelist + freelist_size;
410 static inline unsigned int get_obj_status(struct page *page, int idx)
414 struct kmem_cache *cachep = page->slab_cache;
416 freelist_size = cachep->num * sizeof(freelist_idx_t);
417 status = (char *)page->freelist + freelist_size;
423 static inline void set_obj_status(struct page *page, int idx, int val) {}
428 * Do not go above this order unless 0 objects fit into the slab or
429 * overridden on the command line.
431 #define SLAB_MAX_ORDER_HI 1
432 #define SLAB_MAX_ORDER_LO 0
433 static int slab_max_order = SLAB_MAX_ORDER_LO;
434 static bool slab_max_order_set __initdata;
436 static inline struct kmem_cache *virt_to_cache(const void *obj)
438 struct page *page = virt_to_head_page(obj);
439 return page->slab_cache;
442 static inline void *index_to_obj(struct kmem_cache *cache, struct page *page,
445 return page->s_mem + cache->size * idx;
449 * We want to avoid an expensive divide : (offset / cache->size)
450 * Using the fact that size is a constant for a particular cache,
451 * we can replace (offset / cache->size) by
452 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
454 static inline unsigned int obj_to_index(const struct kmem_cache *cache,
455 const struct page *page, void *obj)
457 u32 offset = (obj - page->s_mem);
458 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
461 static struct arraycache_init initarray_generic =
462 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
464 /* internal cache of cache description objs */
465 static struct kmem_cache kmem_cache_boot = {
467 .limit = BOOT_CPUCACHE_ENTRIES,
469 .size = sizeof(struct kmem_cache),
470 .name = "kmem_cache",
473 #define BAD_ALIEN_MAGIC 0x01020304ul
475 #ifdef CONFIG_LOCKDEP
478 * Slab sometimes uses the kmalloc slabs to store the slab headers
479 * for other slabs "off slab".
480 * The locking for this is tricky in that it nests within the locks
481 * of all other slabs in a few places; to deal with this special
482 * locking we put on-slab caches into a separate lock-class.
484 * We set lock class for alien array caches which are up during init.
485 * The lock annotation will be lost if all cpus of a node goes down and
486 * then comes back up during hotplug
488 static struct lock_class_key on_slab_l3_key;
489 static struct lock_class_key on_slab_alc_key;
491 static struct lock_class_key debugobj_l3_key;
492 static struct lock_class_key debugobj_alc_key;
494 static void slab_set_lock_classes(struct kmem_cache *cachep,
495 struct lock_class_key *l3_key, struct lock_class_key *alc_key,
496 struct kmem_cache_node *n)
498 struct alien_cache **alc;
501 lockdep_set_class(&n->list_lock, l3_key);
504 * FIXME: This check for BAD_ALIEN_MAGIC
505 * should go away when common slab code is taught to
506 * work even without alien caches.
507 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
508 * for alloc_alien_cache,
510 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
514 lockdep_set_class(&(alc[r]->lock), alc_key);
518 static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep,
519 struct kmem_cache_node *n)
521 slab_set_lock_classes(cachep, &debugobj_l3_key, &debugobj_alc_key, n);
524 static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
527 struct kmem_cache_node *n;
529 for_each_kmem_cache_node(cachep, node, n)
530 slab_set_debugobj_lock_classes_node(cachep, n);
533 static void init_node_lock_keys(int q)
540 for (i = 1; i <= KMALLOC_SHIFT_HIGH; i++) {
541 struct kmem_cache_node *n;
542 struct kmem_cache *cache = kmalloc_caches[i];
547 n = get_node(cache, q);
548 if (!n || OFF_SLAB(cache))
551 slab_set_lock_classes(cache, &on_slab_l3_key,
552 &on_slab_alc_key, n);
556 static void on_slab_lock_classes_node(struct kmem_cache *cachep,
557 struct kmem_cache_node *n)
559 slab_set_lock_classes(cachep, &on_slab_l3_key,
560 &on_slab_alc_key, n);
563 static inline void on_slab_lock_classes(struct kmem_cache *cachep)
566 struct kmem_cache_node *n;
568 VM_BUG_ON(OFF_SLAB(cachep));
569 for_each_kmem_cache_node(cachep, node, n)
570 on_slab_lock_classes_node(cachep, n);
573 static inline void __init init_lock_keys(void)
578 init_node_lock_keys(node);
581 static void __init init_node_lock_keys(int q)
585 static inline void init_lock_keys(void)
589 static inline void on_slab_lock_classes(struct kmem_cache *cachep)
593 static inline void on_slab_lock_classes_node(struct kmem_cache *cachep,
594 struct kmem_cache_node *n)
598 static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep,
599 struct kmem_cache_node *n)
603 static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
608 static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
610 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
612 return cachep->array[smp_processor_id()];
615 static size_t calculate_freelist_size(int nr_objs, size_t align)
617 size_t freelist_size;
619 freelist_size = nr_objs * sizeof(freelist_idx_t);
620 if (IS_ENABLED(CONFIG_DEBUG_SLAB_LEAK))
621 freelist_size += nr_objs * sizeof(char);
624 freelist_size = ALIGN(freelist_size, align);
626 return freelist_size;
629 static int calculate_nr_objs(size_t slab_size, size_t buffer_size,
630 size_t idx_size, size_t align)
633 size_t remained_size;
634 size_t freelist_size;
637 if (IS_ENABLED(CONFIG_DEBUG_SLAB_LEAK))
638 extra_space = sizeof(char);
640 * Ignore padding for the initial guess. The padding
641 * is at most @align-1 bytes, and @buffer_size is at
642 * least @align. In the worst case, this result will
643 * be one greater than the number of objects that fit
644 * into the memory allocation when taking the padding
647 nr_objs = slab_size / (buffer_size + idx_size + extra_space);
650 * This calculated number will be either the right
651 * amount, or one greater than what we want.
653 remained_size = slab_size - nr_objs * buffer_size;
654 freelist_size = calculate_freelist_size(nr_objs, align);
655 if (remained_size < freelist_size)
662 * Calculate the number of objects and left-over bytes for a given buffer size.
664 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
665 size_t align, int flags, size_t *left_over,
670 size_t slab_size = PAGE_SIZE << gfporder;
673 * The slab management structure can be either off the slab or
674 * on it. For the latter case, the memory allocated for a
677 * - One unsigned int for each object
678 * - Padding to respect alignment of @align
679 * - @buffer_size bytes for each object
681 * If the slab management structure is off the slab, then the
682 * alignment will already be calculated into the size. Because
683 * the slabs are all pages aligned, the objects will be at the
684 * correct alignment when allocated.
686 if (flags & CFLGS_OFF_SLAB) {
688 nr_objs = slab_size / buffer_size;
691 nr_objs = calculate_nr_objs(slab_size, buffer_size,
692 sizeof(freelist_idx_t), align);
693 mgmt_size = calculate_freelist_size(nr_objs, align);
696 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
700 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
702 static void __slab_error(const char *function, struct kmem_cache *cachep,
705 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
706 function, cachep->name, msg);
708 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
713 * By default on NUMA we use alien caches to stage the freeing of
714 * objects allocated from other nodes. This causes massive memory
715 * inefficiencies when using fake NUMA setup to split memory into a
716 * large number of small nodes, so it can be disabled on the command
720 static int use_alien_caches __read_mostly = 1;
721 static int __init noaliencache_setup(char *s)
723 use_alien_caches = 0;
726 __setup("noaliencache", noaliencache_setup);
728 static int __init slab_max_order_setup(char *str)
730 get_option(&str, &slab_max_order);
731 slab_max_order = slab_max_order < 0 ? 0 :
732 min(slab_max_order, MAX_ORDER - 1);
733 slab_max_order_set = true;
737 __setup("slab_max_order=", slab_max_order_setup);
741 * Special reaping functions for NUMA systems called from cache_reap().
742 * These take care of doing round robin flushing of alien caches (containing
743 * objects freed on different nodes from which they were allocated) and the
744 * flushing of remote pcps by calling drain_node_pages.
746 static DEFINE_PER_CPU(unsigned long, slab_reap_node);
748 static void init_reap_node(int cpu)
752 node = next_node(cpu_to_mem(cpu), node_online_map);
753 if (node == MAX_NUMNODES)
754 node = first_node(node_online_map);
756 per_cpu(slab_reap_node, cpu) = node;
759 static void next_reap_node(void)
761 int node = __this_cpu_read(slab_reap_node);
763 node = next_node(node, node_online_map);
764 if (unlikely(node >= MAX_NUMNODES))
765 node = first_node(node_online_map);
766 __this_cpu_write(slab_reap_node, node);
770 #define init_reap_node(cpu) do { } while (0)
771 #define next_reap_node(void) do { } while (0)
775 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
776 * via the workqueue/eventd.
777 * Add the CPU number into the expiration time to minimize the possibility of
778 * the CPUs getting into lockstep and contending for the global cache chain
781 static void start_cpu_timer(int cpu)
783 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
786 * When this gets called from do_initcalls via cpucache_init(),
787 * init_workqueues() has already run, so keventd will be setup
790 if (keventd_up() && reap_work->work.func == NULL) {
792 INIT_DEFERRABLE_WORK(reap_work, cache_reap);
793 schedule_delayed_work_on(cpu, reap_work,
794 __round_jiffies_relative(HZ, cpu));
798 static void init_arraycache(struct array_cache *ac, int limit, int batch)
801 * The array_cache structures contain pointers to free object.
802 * However, when such objects are allocated or transferred to another
803 * cache the pointers are not cleared and they could be counted as
804 * valid references during a kmemleak scan. Therefore, kmemleak must
805 * not scan such objects.
807 kmemleak_no_scan(ac);
811 ac->batchcount = batch;
816 static struct array_cache *alloc_arraycache(int node, int entries,
817 int batchcount, gfp_t gfp)
819 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
820 struct array_cache *ac = NULL;
822 ac = kmalloc_node(memsize, gfp, node);
823 init_arraycache(ac, entries, batchcount);
827 static inline bool is_slab_pfmemalloc(struct page *page)
829 return PageSlabPfmemalloc(page);
832 /* Clears pfmemalloc_active if no slabs have pfmalloc set */
833 static void recheck_pfmemalloc_active(struct kmem_cache *cachep,
834 struct array_cache *ac)
836 struct kmem_cache_node *n = get_node(cachep, numa_mem_id());
840 if (!pfmemalloc_active)
843 spin_lock_irqsave(&n->list_lock, flags);
844 list_for_each_entry(page, &n->slabs_full, lru)
845 if (is_slab_pfmemalloc(page))
848 list_for_each_entry(page, &n->slabs_partial, lru)
849 if (is_slab_pfmemalloc(page))
852 list_for_each_entry(page, &n->slabs_free, lru)
853 if (is_slab_pfmemalloc(page))
856 pfmemalloc_active = false;
858 spin_unlock_irqrestore(&n->list_lock, flags);
861 static void *__ac_get_obj(struct kmem_cache *cachep, struct array_cache *ac,
862 gfp_t flags, bool force_refill)
865 void *objp = ac->entry[--ac->avail];
867 /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
868 if (unlikely(is_obj_pfmemalloc(objp))) {
869 struct kmem_cache_node *n;
871 if (gfp_pfmemalloc_allowed(flags)) {
872 clear_obj_pfmemalloc(&objp);
876 /* The caller cannot use PFMEMALLOC objects, find another one */
877 for (i = 0; i < ac->avail; i++) {
878 /* If a !PFMEMALLOC object is found, swap them */
879 if (!is_obj_pfmemalloc(ac->entry[i])) {
881 ac->entry[i] = ac->entry[ac->avail];
882 ac->entry[ac->avail] = objp;
888 * If there are empty slabs on the slabs_free list and we are
889 * being forced to refill the cache, mark this one !pfmemalloc.
891 n = get_node(cachep, numa_mem_id());
892 if (!list_empty(&n->slabs_free) && force_refill) {
893 struct page *page = virt_to_head_page(objp);
894 ClearPageSlabPfmemalloc(page);
895 clear_obj_pfmemalloc(&objp);
896 recheck_pfmemalloc_active(cachep, ac);
900 /* No !PFMEMALLOC objects available */
908 static inline void *ac_get_obj(struct kmem_cache *cachep,
909 struct array_cache *ac, gfp_t flags, bool force_refill)
913 if (unlikely(sk_memalloc_socks()))
914 objp = __ac_get_obj(cachep, ac, flags, force_refill);
916 objp = ac->entry[--ac->avail];
921 static void *__ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
924 if (unlikely(pfmemalloc_active)) {
925 /* Some pfmemalloc slabs exist, check if this is one */
926 struct page *page = virt_to_head_page(objp);
927 if (PageSlabPfmemalloc(page))
928 set_obj_pfmemalloc(&objp);
934 static inline void ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
937 if (unlikely(sk_memalloc_socks()))
938 objp = __ac_put_obj(cachep, ac, objp);
940 ac->entry[ac->avail++] = objp;
944 * Transfer objects in one arraycache to another.
945 * Locking must be handled by the caller.
947 * Return the number of entries transferred.
949 static int transfer_objects(struct array_cache *to,
950 struct array_cache *from, unsigned int max)
952 /* Figure out how many entries to transfer */
953 int nr = min3(from->avail, max, to->limit - to->avail);
958 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
968 #define drain_alien_cache(cachep, alien) do { } while (0)
969 #define reap_alien(cachep, n) do { } while (0)
971 static inline struct alien_cache **alloc_alien_cache(int node,
972 int limit, gfp_t gfp)
974 return (struct alien_cache **)BAD_ALIEN_MAGIC;
977 static inline void free_alien_cache(struct alien_cache **ac_ptr)
981 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
986 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
992 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
993 gfp_t flags, int nodeid)
998 #else /* CONFIG_NUMA */
1000 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
1001 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
1003 static struct alien_cache *__alloc_alien_cache(int node, int entries,
1004 int batch, gfp_t gfp)
1006 int memsize = sizeof(void *) * entries + sizeof(struct alien_cache);
1007 struct alien_cache *alc = NULL;
1009 alc = kmalloc_node(memsize, gfp, node);
1010 init_arraycache(&alc->ac, entries, batch);
1011 spin_lock_init(&alc->lock);
1015 static struct alien_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
1017 struct alien_cache **alc_ptr;
1018 int memsize = sizeof(void *) * nr_node_ids;
1023 alc_ptr = kzalloc_node(memsize, gfp, node);
1028 if (i == node || !node_online(i))
1030 alc_ptr[i] = __alloc_alien_cache(node, limit, 0xbaadf00d, gfp);
1032 for (i--; i >= 0; i--)
1041 static void free_alien_cache(struct alien_cache **alc_ptr)
1052 static void __drain_alien_cache(struct kmem_cache *cachep,
1053 struct array_cache *ac, int node)
1055 struct kmem_cache_node *n = get_node(cachep, node);
1059 spin_lock(&n->list_lock);
1061 * Stuff objects into the remote nodes shared array first.
1062 * That way we could avoid the overhead of putting the objects
1063 * into the free lists and getting them back later.
1066 transfer_objects(n->shared, ac, ac->limit);
1068 free_block(cachep, ac->entry, ac->avail, node, &list);
1070 spin_unlock(&n->list_lock);
1071 slabs_destroy(cachep, &list);
1076 * Called from cache_reap() to regularly drain alien caches round robin.
1078 static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *n)
1080 int node = __this_cpu_read(slab_reap_node);
1083 struct alien_cache *alc = n->alien[node];
1084 struct array_cache *ac;
1088 if (ac->avail && spin_trylock_irq(&alc->lock)) {
1089 __drain_alien_cache(cachep, ac, node);
1090 spin_unlock_irq(&alc->lock);
1096 static void drain_alien_cache(struct kmem_cache *cachep,
1097 struct alien_cache **alien)
1100 struct alien_cache *alc;
1101 struct array_cache *ac;
1102 unsigned long flags;
1104 for_each_online_node(i) {
1108 spin_lock_irqsave(&alc->lock, flags);
1109 __drain_alien_cache(cachep, ac, i);
1110 spin_unlock_irqrestore(&alc->lock, flags);
1115 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1117 int nodeid = page_to_nid(virt_to_page(objp));
1118 struct kmem_cache_node *n;
1119 struct alien_cache *alien = NULL;
1120 struct array_cache *ac;
1124 node = numa_mem_id();
1127 * Make sure we are not freeing a object from another node to the array
1128 * cache on this cpu.
1130 if (likely(nodeid == node))
1133 n = get_node(cachep, node);
1134 STATS_INC_NODEFREES(cachep);
1135 if (n->alien && n->alien[nodeid]) {
1136 alien = n->alien[nodeid];
1138 spin_lock(&alien->lock);
1139 if (unlikely(ac->avail == ac->limit)) {
1140 STATS_INC_ACOVERFLOW(cachep);
1141 __drain_alien_cache(cachep, ac, nodeid);
1143 ac_put_obj(cachep, ac, objp);
1144 spin_unlock(&alien->lock);
1146 n = get_node(cachep, nodeid);
1147 spin_lock(&n->list_lock);
1148 free_block(cachep, &objp, 1, nodeid, &list);
1149 spin_unlock(&n->list_lock);
1150 slabs_destroy(cachep, &list);
1157 * Allocates and initializes node for a node on each slab cache, used for
1158 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
1159 * will be allocated off-node since memory is not yet online for the new node.
1160 * When hotplugging memory or a cpu, existing node are not replaced if
1163 * Must hold slab_mutex.
1165 static int init_cache_node_node(int node)
1167 struct kmem_cache *cachep;
1168 struct kmem_cache_node *n;
1169 const int memsize = sizeof(struct kmem_cache_node);
1171 list_for_each_entry(cachep, &slab_caches, list) {
1173 * Set up the kmem_cache_node for cpu before we can
1174 * begin anything. Make sure some other cpu on this
1175 * node has not already allocated this
1177 n = get_node(cachep, node);
1179 n = kmalloc_node(memsize, GFP_KERNEL, node);
1182 kmem_cache_node_init(n);
1183 n->next_reap = jiffies + REAPTIMEOUT_NODE +
1184 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
1187 * The kmem_cache_nodes don't come and go as CPUs
1188 * come and go. slab_mutex is sufficient
1191 cachep->node[node] = n;
1194 spin_lock_irq(&n->list_lock);
1196 (1 + nr_cpus_node(node)) *
1197 cachep->batchcount + cachep->num;
1198 spin_unlock_irq(&n->list_lock);
1203 static inline int slabs_tofree(struct kmem_cache *cachep,
1204 struct kmem_cache_node *n)
1206 return (n->free_objects + cachep->num - 1) / cachep->num;
1209 static void cpuup_canceled(long cpu)
1211 struct kmem_cache *cachep;
1212 struct kmem_cache_node *n = NULL;
1213 int node = cpu_to_mem(cpu);
1214 const struct cpumask *mask = cpumask_of_node(node);
1216 list_for_each_entry(cachep, &slab_caches, list) {
1217 struct array_cache *nc;
1218 struct array_cache *shared;
1219 struct alien_cache **alien;
1222 /* cpu is dead; no one can alloc from it. */
1223 nc = cachep->array[cpu];
1224 cachep->array[cpu] = NULL;
1225 n = get_node(cachep, node);
1228 goto free_array_cache;
1230 spin_lock_irq(&n->list_lock);
1232 /* Free limit for this kmem_cache_node */
1233 n->free_limit -= cachep->batchcount;
1235 free_block(cachep, nc->entry, nc->avail, node, &list);
1237 if (!cpumask_empty(mask)) {
1238 spin_unlock_irq(&n->list_lock);
1239 goto free_array_cache;
1244 free_block(cachep, shared->entry,
1245 shared->avail, node, &list);
1252 spin_unlock_irq(&n->list_lock);
1256 drain_alien_cache(cachep, alien);
1257 free_alien_cache(alien);
1260 slabs_destroy(cachep, &list);
1264 * In the previous loop, all the objects were freed to
1265 * the respective cache's slabs, now we can go ahead and
1266 * shrink each nodelist to its limit.
1268 list_for_each_entry(cachep, &slab_caches, list) {
1269 n = get_node(cachep, node);
1272 drain_freelist(cachep, n, slabs_tofree(cachep, n));
1276 static int cpuup_prepare(long cpu)
1278 struct kmem_cache *cachep;
1279 struct kmem_cache_node *n = NULL;
1280 int node = cpu_to_mem(cpu);
1284 * We need to do this right in the beginning since
1285 * alloc_arraycache's are going to use this list.
1286 * kmalloc_node allows us to add the slab to the right
1287 * kmem_cache_node and not this cpu's kmem_cache_node
1289 err = init_cache_node_node(node);
1294 * Now we can go ahead with allocating the shared arrays and
1297 list_for_each_entry(cachep, &slab_caches, list) {
1298 struct array_cache *nc;
1299 struct array_cache *shared = NULL;
1300 struct alien_cache **alien = NULL;
1302 nc = alloc_arraycache(node, cachep->limit,
1303 cachep->batchcount, GFP_KERNEL);
1306 if (cachep->shared) {
1307 shared = alloc_arraycache(node,
1308 cachep->shared * cachep->batchcount,
1309 0xbaadf00d, GFP_KERNEL);
1315 if (use_alien_caches) {
1316 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
1323 cachep->array[cpu] = nc;
1324 n = get_node(cachep, node);
1327 spin_lock_irq(&n->list_lock);
1330 * We are serialised from CPU_DEAD or
1331 * CPU_UP_CANCELLED by the cpucontrol lock
1342 spin_unlock_irq(&n->list_lock);
1344 free_alien_cache(alien);
1345 if (cachep->flags & SLAB_DEBUG_OBJECTS)
1346 slab_set_debugobj_lock_classes_node(cachep, n);
1347 else if (!OFF_SLAB(cachep) &&
1348 !(cachep->flags & SLAB_DESTROY_BY_RCU))
1349 on_slab_lock_classes_node(cachep, n);
1351 init_node_lock_keys(node);
1355 cpuup_canceled(cpu);
1359 static int cpuup_callback(struct notifier_block *nfb,
1360 unsigned long action, void *hcpu)
1362 long cpu = (long)hcpu;
1366 case CPU_UP_PREPARE:
1367 case CPU_UP_PREPARE_FROZEN:
1368 mutex_lock(&slab_mutex);
1369 err = cpuup_prepare(cpu);
1370 mutex_unlock(&slab_mutex);
1373 case CPU_ONLINE_FROZEN:
1374 start_cpu_timer(cpu);
1376 #ifdef CONFIG_HOTPLUG_CPU
1377 case CPU_DOWN_PREPARE:
1378 case CPU_DOWN_PREPARE_FROZEN:
1380 * Shutdown cache reaper. Note that the slab_mutex is
1381 * held so that if cache_reap() is invoked it cannot do
1382 * anything expensive but will only modify reap_work
1383 * and reschedule the timer.
1385 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1386 /* Now the cache_reaper is guaranteed to be not running. */
1387 per_cpu(slab_reap_work, cpu).work.func = NULL;
1389 case CPU_DOWN_FAILED:
1390 case CPU_DOWN_FAILED_FROZEN:
1391 start_cpu_timer(cpu);
1394 case CPU_DEAD_FROZEN:
1396 * Even if all the cpus of a node are down, we don't free the
1397 * kmem_cache_node of any cache. This to avoid a race between
1398 * cpu_down, and a kmalloc allocation from another cpu for
1399 * memory from the node of the cpu going down. The node
1400 * structure is usually allocated from kmem_cache_create() and
1401 * gets destroyed at kmem_cache_destroy().
1405 case CPU_UP_CANCELED:
1406 case CPU_UP_CANCELED_FROZEN:
1407 mutex_lock(&slab_mutex);
1408 cpuup_canceled(cpu);
1409 mutex_unlock(&slab_mutex);
1412 return notifier_from_errno(err);
1415 static struct notifier_block cpucache_notifier = {
1416 &cpuup_callback, NULL, 0
1419 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1421 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1422 * Returns -EBUSY if all objects cannot be drained so that the node is not
1425 * Must hold slab_mutex.
1427 static int __meminit drain_cache_node_node(int node)
1429 struct kmem_cache *cachep;
1432 list_for_each_entry(cachep, &slab_caches, list) {
1433 struct kmem_cache_node *n;
1435 n = get_node(cachep, node);
1439 drain_freelist(cachep, n, slabs_tofree(cachep, n));
1441 if (!list_empty(&n->slabs_full) ||
1442 !list_empty(&n->slabs_partial)) {
1450 static int __meminit slab_memory_callback(struct notifier_block *self,
1451 unsigned long action, void *arg)
1453 struct memory_notify *mnb = arg;
1457 nid = mnb->status_change_nid;
1462 case MEM_GOING_ONLINE:
1463 mutex_lock(&slab_mutex);
1464 ret = init_cache_node_node(nid);
1465 mutex_unlock(&slab_mutex);
1467 case MEM_GOING_OFFLINE:
1468 mutex_lock(&slab_mutex);
1469 ret = drain_cache_node_node(nid);
1470 mutex_unlock(&slab_mutex);
1474 case MEM_CANCEL_ONLINE:
1475 case MEM_CANCEL_OFFLINE:
1479 return notifier_from_errno(ret);
1481 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1484 * swap the static kmem_cache_node with kmalloced memory
1486 static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list,
1489 struct kmem_cache_node *ptr;
1491 ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid);
1494 memcpy(ptr, list, sizeof(struct kmem_cache_node));
1496 * Do not assume that spinlocks can be initialized via memcpy:
1498 spin_lock_init(&ptr->list_lock);
1500 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1501 cachep->node[nodeid] = ptr;
1505 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1506 * size of kmem_cache_node.
1508 static void __init set_up_node(struct kmem_cache *cachep, int index)
1512 for_each_online_node(node) {
1513 cachep->node[node] = &init_kmem_cache_node[index + node];
1514 cachep->node[node]->next_reap = jiffies +
1516 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
1521 * The memory after the last cpu cache pointer is used for the
1524 static void setup_node_pointer(struct kmem_cache *cachep)
1526 cachep->node = (struct kmem_cache_node **)&cachep->array[nr_cpu_ids];
1530 * Initialisation. Called after the page allocator have been initialised and
1531 * before smp_init().
1533 void __init kmem_cache_init(void)
1537 BUILD_BUG_ON(sizeof(((struct page *)NULL)->lru) <
1538 sizeof(struct rcu_head));
1539 kmem_cache = &kmem_cache_boot;
1540 setup_node_pointer(kmem_cache);
1542 if (num_possible_nodes() == 1)
1543 use_alien_caches = 0;
1545 for (i = 0; i < NUM_INIT_LISTS; i++)
1546 kmem_cache_node_init(&init_kmem_cache_node[i]);
1548 set_up_node(kmem_cache, CACHE_CACHE);
1551 * Fragmentation resistance on low memory - only use bigger
1552 * page orders on machines with more than 32MB of memory if
1553 * not overridden on the command line.
1555 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
1556 slab_max_order = SLAB_MAX_ORDER_HI;
1558 /* Bootstrap is tricky, because several objects are allocated
1559 * from caches that do not exist yet:
1560 * 1) initialize the kmem_cache cache: it contains the struct
1561 * kmem_cache structures of all caches, except kmem_cache itself:
1562 * kmem_cache is statically allocated.
1563 * Initially an __init data area is used for the head array and the
1564 * kmem_cache_node structures, it's replaced with a kmalloc allocated
1565 * array at the end of the bootstrap.
1566 * 2) Create the first kmalloc cache.
1567 * The struct kmem_cache for the new cache is allocated normally.
1568 * An __init data area is used for the head array.
1569 * 3) Create the remaining kmalloc caches, with minimally sized
1571 * 4) Replace the __init data head arrays for kmem_cache and the first
1572 * kmalloc cache with kmalloc allocated arrays.
1573 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1574 * the other cache's with kmalloc allocated memory.
1575 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1578 /* 1) create the kmem_cache */
1581 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1583 create_boot_cache(kmem_cache, "kmem_cache",
1584 offsetof(struct kmem_cache, array[nr_cpu_ids]) +
1585 nr_node_ids * sizeof(struct kmem_cache_node *),
1586 SLAB_HWCACHE_ALIGN);
1587 list_add(&kmem_cache->list, &slab_caches);
1589 /* 2+3) create the kmalloc caches */
1592 * Initialize the caches that provide memory for the array cache and the
1593 * kmem_cache_node structures first. Without this, further allocations will
1597 kmalloc_caches[INDEX_AC] = create_kmalloc_cache("kmalloc-ac",
1598 kmalloc_size(INDEX_AC), ARCH_KMALLOC_FLAGS);
1600 if (INDEX_AC != INDEX_NODE)
1601 kmalloc_caches[INDEX_NODE] =
1602 create_kmalloc_cache("kmalloc-node",
1603 kmalloc_size(INDEX_NODE), ARCH_KMALLOC_FLAGS);
1605 slab_early_init = 0;
1607 /* 4) Replace the bootstrap head arrays */
1609 struct array_cache *ptr;
1611 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1613 memcpy(ptr, cpu_cache_get(kmem_cache),
1614 sizeof(struct arraycache_init));
1616 kmem_cache->array[smp_processor_id()] = ptr;
1618 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1620 BUG_ON(cpu_cache_get(kmalloc_caches[INDEX_AC])
1621 != &initarray_generic.cache);
1622 memcpy(ptr, cpu_cache_get(kmalloc_caches[INDEX_AC]),
1623 sizeof(struct arraycache_init));
1625 kmalloc_caches[INDEX_AC]->array[smp_processor_id()] = ptr;
1627 /* 5) Replace the bootstrap kmem_cache_node */
1631 for_each_online_node(nid) {
1632 init_list(kmem_cache, &init_kmem_cache_node[CACHE_CACHE + nid], nid);
1634 init_list(kmalloc_caches[INDEX_AC],
1635 &init_kmem_cache_node[SIZE_AC + nid], nid);
1637 if (INDEX_AC != INDEX_NODE) {
1638 init_list(kmalloc_caches[INDEX_NODE],
1639 &init_kmem_cache_node[SIZE_NODE + nid], nid);
1644 create_kmalloc_caches(ARCH_KMALLOC_FLAGS);
1647 void __init kmem_cache_init_late(void)
1649 struct kmem_cache *cachep;
1653 /* 6) resize the head arrays to their final sizes */
1654 mutex_lock(&slab_mutex);
1655 list_for_each_entry(cachep, &slab_caches, list)
1656 if (enable_cpucache(cachep, GFP_NOWAIT))
1658 mutex_unlock(&slab_mutex);
1660 /* Annotate slab for lockdep -- annotate the malloc caches */
1667 * Register a cpu startup notifier callback that initializes
1668 * cpu_cache_get for all new cpus
1670 register_cpu_notifier(&cpucache_notifier);
1674 * Register a memory hotplug callback that initializes and frees
1677 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1681 * The reap timers are started later, with a module init call: That part
1682 * of the kernel is not yet operational.
1686 static int __init cpucache_init(void)
1691 * Register the timers that return unneeded pages to the page allocator
1693 for_each_online_cpu(cpu)
1694 start_cpu_timer(cpu);
1700 __initcall(cpucache_init);
1702 static noinline void
1703 slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1706 struct kmem_cache_node *n;
1708 unsigned long flags;
1710 static DEFINE_RATELIMIT_STATE(slab_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
1711 DEFAULT_RATELIMIT_BURST);
1713 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slab_oom_rs))
1717 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1719 printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n",
1720 cachep->name, cachep->size, cachep->gfporder);
1722 for_each_kmem_cache_node(cachep, node, n) {
1723 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1724 unsigned long active_slabs = 0, num_slabs = 0;
1726 spin_lock_irqsave(&n->list_lock, flags);
1727 list_for_each_entry(page, &n->slabs_full, lru) {
1728 active_objs += cachep->num;
1731 list_for_each_entry(page, &n->slabs_partial, lru) {
1732 active_objs += page->active;
1735 list_for_each_entry(page, &n->slabs_free, lru)
1738 free_objects += n->free_objects;
1739 spin_unlock_irqrestore(&n->list_lock, flags);
1741 num_slabs += active_slabs;
1742 num_objs = num_slabs * cachep->num;
1744 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1745 node, active_slabs, num_slabs, active_objs, num_objs,
1752 * Interface to system's page allocator. No need to hold the cache-lock.
1754 * If we requested dmaable memory, we will get it. Even if we
1755 * did not request dmaable memory, we might get it, but that
1756 * would be relatively rare and ignorable.
1758 static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags,
1764 flags |= cachep->allocflags;
1765 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1766 flags |= __GFP_RECLAIMABLE;
1768 if (memcg_charge_slab(cachep, flags, cachep->gfporder))
1771 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1773 memcg_uncharge_slab(cachep, cachep->gfporder);
1774 slab_out_of_memory(cachep, flags, nodeid);
1778 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1779 if (unlikely(page->pfmemalloc))
1780 pfmemalloc_active = true;
1782 nr_pages = (1 << cachep->gfporder);
1783 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1784 add_zone_page_state(page_zone(page),
1785 NR_SLAB_RECLAIMABLE, nr_pages);
1787 add_zone_page_state(page_zone(page),
1788 NR_SLAB_UNRECLAIMABLE, nr_pages);
1789 __SetPageSlab(page);
1790 if (page->pfmemalloc)
1791 SetPageSlabPfmemalloc(page);
1793 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1794 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1797 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1799 kmemcheck_mark_unallocated_pages(page, nr_pages);
1806 * Interface to system's page release.
1808 static void kmem_freepages(struct kmem_cache *cachep, struct page *page)
1810 const unsigned long nr_freed = (1 << cachep->gfporder);
1812 kmemcheck_free_shadow(page, cachep->gfporder);
1814 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1815 sub_zone_page_state(page_zone(page),
1816 NR_SLAB_RECLAIMABLE, nr_freed);
1818 sub_zone_page_state(page_zone(page),
1819 NR_SLAB_UNRECLAIMABLE, nr_freed);
1821 BUG_ON(!PageSlab(page));
1822 __ClearPageSlabPfmemalloc(page);
1823 __ClearPageSlab(page);
1824 page_mapcount_reset(page);
1825 page->mapping = NULL;
1827 if (current->reclaim_state)
1828 current->reclaim_state->reclaimed_slab += nr_freed;
1829 __free_pages(page, cachep->gfporder);
1830 memcg_uncharge_slab(cachep, cachep->gfporder);
1833 static void kmem_rcu_free(struct rcu_head *head)
1835 struct kmem_cache *cachep;
1838 page = container_of(head, struct page, rcu_head);
1839 cachep = page->slab_cache;
1841 kmem_freepages(cachep, page);
1846 #ifdef CONFIG_DEBUG_PAGEALLOC
1847 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1848 unsigned long caller)
1850 int size = cachep->object_size;
1852 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1854 if (size < 5 * sizeof(unsigned long))
1857 *addr++ = 0x12345678;
1859 *addr++ = smp_processor_id();
1860 size -= 3 * sizeof(unsigned long);
1862 unsigned long *sptr = &caller;
1863 unsigned long svalue;
1865 while (!kstack_end(sptr)) {
1867 if (kernel_text_address(svalue)) {
1869 size -= sizeof(unsigned long);
1870 if (size <= sizeof(unsigned long))
1876 *addr++ = 0x87654321;
1880 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1882 int size = cachep->object_size;
1883 addr = &((char *)addr)[obj_offset(cachep)];
1885 memset(addr, val, size);
1886 *(unsigned char *)(addr + size - 1) = POISON_END;
1889 static void dump_line(char *data, int offset, int limit)
1892 unsigned char error = 0;
1895 printk(KERN_ERR "%03x: ", offset);
1896 for (i = 0; i < limit; i++) {
1897 if (data[offset + i] != POISON_FREE) {
1898 error = data[offset + i];
1902 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1903 &data[offset], limit, 1);
1905 if (bad_count == 1) {
1906 error ^= POISON_FREE;
1907 if (!(error & (error - 1))) {
1908 printk(KERN_ERR "Single bit error detected. Probably "
1911 printk(KERN_ERR "Run memtest86+ or a similar memory "
1914 printk(KERN_ERR "Run a memory test tool.\n");
1923 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1928 if (cachep->flags & SLAB_RED_ZONE) {
1929 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
1930 *dbg_redzone1(cachep, objp),
1931 *dbg_redzone2(cachep, objp));
1934 if (cachep->flags & SLAB_STORE_USER) {
1935 printk(KERN_ERR "Last user: [<%p>](%pSR)\n",
1936 *dbg_userword(cachep, objp),
1937 *dbg_userword(cachep, objp));
1939 realobj = (char *)objp + obj_offset(cachep);
1940 size = cachep->object_size;
1941 for (i = 0; i < size && lines; i += 16, lines--) {
1944 if (i + limit > size)
1946 dump_line(realobj, i, limit);
1950 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1956 realobj = (char *)objp + obj_offset(cachep);
1957 size = cachep->object_size;
1959 for (i = 0; i < size; i++) {
1960 char exp = POISON_FREE;
1963 if (realobj[i] != exp) {
1969 "Slab corruption (%s): %s start=%p, len=%d\n",
1970 print_tainted(), cachep->name, realobj, size);
1971 print_objinfo(cachep, objp, 0);
1973 /* Hexdump the affected line */
1976 if (i + limit > size)
1978 dump_line(realobj, i, limit);
1981 /* Limit to 5 lines */
1987 /* Print some data about the neighboring objects, if they
1990 struct page *page = virt_to_head_page(objp);
1993 objnr = obj_to_index(cachep, page, objp);
1995 objp = index_to_obj(cachep, page, objnr - 1);
1996 realobj = (char *)objp + obj_offset(cachep);
1997 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1999 print_objinfo(cachep, objp, 2);
2001 if (objnr + 1 < cachep->num) {
2002 objp = index_to_obj(cachep, page, objnr + 1);
2003 realobj = (char *)objp + obj_offset(cachep);
2004 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
2006 print_objinfo(cachep, objp, 2);
2013 static void slab_destroy_debugcheck(struct kmem_cache *cachep,
2017 for (i = 0; i < cachep->num; i++) {
2018 void *objp = index_to_obj(cachep, page, i);
2020 if (cachep->flags & SLAB_POISON) {
2021 #ifdef CONFIG_DEBUG_PAGEALLOC
2022 if (cachep->size % PAGE_SIZE == 0 &&
2024 kernel_map_pages(virt_to_page(objp),
2025 cachep->size / PAGE_SIZE, 1);
2027 check_poison_obj(cachep, objp);
2029 check_poison_obj(cachep, objp);
2032 if (cachep->flags & SLAB_RED_ZONE) {
2033 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2034 slab_error(cachep, "start of a freed object "
2036 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2037 slab_error(cachep, "end of a freed object "
2043 static void slab_destroy_debugcheck(struct kmem_cache *cachep,
2050 * slab_destroy - destroy and release all objects in a slab
2051 * @cachep: cache pointer being destroyed
2052 * @page: page pointer being destroyed
2054 * Destroy all the objs in a slab, and release the mem back to the system.
2055 * Before calling the slab must have been unlinked from the cache. The
2056 * cache-lock is not held/needed.
2058 static void slab_destroy(struct kmem_cache *cachep, struct page *page)
2062 freelist = page->freelist;
2063 slab_destroy_debugcheck(cachep, page);
2064 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
2065 struct rcu_head *head;
2068 * RCU free overloads the RCU head over the LRU.
2069 * slab_page has been overloeaded over the LRU,
2070 * however it is not used from now on so that
2071 * we can use it safely.
2073 head = (void *)&page->rcu_head;
2074 call_rcu(head, kmem_rcu_free);
2077 kmem_freepages(cachep, page);
2081 * From now on, we don't use freelist
2082 * although actual page can be freed in rcu context
2084 if (OFF_SLAB(cachep))
2085 kmem_cache_free(cachep->freelist_cache, freelist);
2088 static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list)
2090 struct page *page, *n;
2092 list_for_each_entry_safe(page, n, list, lru) {
2093 list_del(&page->lru);
2094 slab_destroy(cachep, page);
2099 * calculate_slab_order - calculate size (page order) of slabs
2100 * @cachep: pointer to the cache that is being created
2101 * @size: size of objects to be created in this cache.
2102 * @align: required alignment for the objects.
2103 * @flags: slab allocation flags
2105 * Also calculates the number of objects per slab.
2107 * This could be made much more intelligent. For now, try to avoid using
2108 * high order pages for slabs. When the gfp() functions are more friendly
2109 * towards high-order requests, this should be changed.
2111 static size_t calculate_slab_order(struct kmem_cache *cachep,
2112 size_t size, size_t align, unsigned long flags)
2114 unsigned long offslab_limit;
2115 size_t left_over = 0;
2118 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
2122 cache_estimate(gfporder, size, align, flags, &remainder, &num);
2126 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
2127 if (num > SLAB_OBJ_MAX_NUM)
2130 if (flags & CFLGS_OFF_SLAB) {
2131 size_t freelist_size_per_obj = sizeof(freelist_idx_t);
2133 * Max number of objs-per-slab for caches which
2134 * use off-slab slabs. Needed to avoid a possible
2135 * looping condition in cache_grow().
2137 if (IS_ENABLED(CONFIG_DEBUG_SLAB_LEAK))
2138 freelist_size_per_obj += sizeof(char);
2139 offslab_limit = size;
2140 offslab_limit /= freelist_size_per_obj;
2142 if (num > offslab_limit)
2146 /* Found something acceptable - save it away */
2148 cachep->gfporder = gfporder;
2149 left_over = remainder;
2152 * A VFS-reclaimable slab tends to have most allocations
2153 * as GFP_NOFS and we really don't want to have to be allocating
2154 * higher-order pages when we are unable to shrink dcache.
2156 if (flags & SLAB_RECLAIM_ACCOUNT)
2160 * Large number of objects is good, but very large slabs are
2161 * currently bad for the gfp()s.
2163 if (gfporder >= slab_max_order)
2167 * Acceptable internal fragmentation?
2169 if (left_over * 8 <= (PAGE_SIZE << gfporder))
2175 static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
2177 if (slab_state >= FULL)
2178 return enable_cpucache(cachep, gfp);
2180 if (slab_state == DOWN) {
2182 * Note: Creation of first cache (kmem_cache).
2183 * The setup_node is taken care
2184 * of by the caller of __kmem_cache_create
2186 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2187 slab_state = PARTIAL;
2188 } else if (slab_state == PARTIAL) {
2190 * Note: the second kmem_cache_create must create the cache
2191 * that's used by kmalloc(24), otherwise the creation of
2192 * further caches will BUG().
2194 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2197 * If the cache that's used by kmalloc(sizeof(kmem_cache_node)) is
2198 * the second cache, then we need to set up all its node/,
2199 * otherwise the creation of further caches will BUG().
2201 set_up_node(cachep, SIZE_AC);
2202 if (INDEX_AC == INDEX_NODE)
2203 slab_state = PARTIAL_NODE;
2205 slab_state = PARTIAL_ARRAYCACHE;
2207 /* Remaining boot caches */
2208 cachep->array[smp_processor_id()] =
2209 kmalloc(sizeof(struct arraycache_init), gfp);
2211 if (slab_state == PARTIAL_ARRAYCACHE) {
2212 set_up_node(cachep, SIZE_NODE);
2213 slab_state = PARTIAL_NODE;
2216 for_each_online_node(node) {
2217 cachep->node[node] =
2218 kmalloc_node(sizeof(struct kmem_cache_node),
2220 BUG_ON(!cachep->node[node]);
2221 kmem_cache_node_init(cachep->node[node]);
2225 cachep->node[numa_mem_id()]->next_reap =
2226 jiffies + REAPTIMEOUT_NODE +
2227 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
2229 cpu_cache_get(cachep)->avail = 0;
2230 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2231 cpu_cache_get(cachep)->batchcount = 1;
2232 cpu_cache_get(cachep)->touched = 0;
2233 cachep->batchcount = 1;
2234 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2239 * __kmem_cache_create - Create a cache.
2240 * @cachep: cache management descriptor
2241 * @flags: SLAB flags
2243 * Returns a ptr to the cache on success, NULL on failure.
2244 * Cannot be called within a int, but can be interrupted.
2245 * The @ctor is run when new pages are allocated by the cache.
2249 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2250 * to catch references to uninitialised memory.
2252 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2253 * for buffer overruns.
2255 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2256 * cacheline. This can be beneficial if you're counting cycles as closely
2260 __kmem_cache_create (struct kmem_cache *cachep, unsigned long flags)
2262 size_t left_over, freelist_size, ralign;
2265 size_t size = cachep->size;
2270 * Enable redzoning and last user accounting, except for caches with
2271 * large objects, if the increased size would increase the object size
2272 * above the next power of two: caches with object sizes just above a
2273 * power of two have a significant amount of internal fragmentation.
2275 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2276 2 * sizeof(unsigned long long)))
2277 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2278 if (!(flags & SLAB_DESTROY_BY_RCU))
2279 flags |= SLAB_POISON;
2281 if (flags & SLAB_DESTROY_BY_RCU)
2282 BUG_ON(flags & SLAB_POISON);
2286 * Check that size is in terms of words. This is needed to avoid
2287 * unaligned accesses for some archs when redzoning is used, and makes
2288 * sure any on-slab bufctl's are also correctly aligned.
2290 if (size & (BYTES_PER_WORD - 1)) {
2291 size += (BYTES_PER_WORD - 1);
2292 size &= ~(BYTES_PER_WORD - 1);
2296 * Redzoning and user store require word alignment or possibly larger.
2297 * Note this will be overridden by architecture or caller mandated
2298 * alignment if either is greater than BYTES_PER_WORD.
2300 if (flags & SLAB_STORE_USER)
2301 ralign = BYTES_PER_WORD;
2303 if (flags & SLAB_RED_ZONE) {
2304 ralign = REDZONE_ALIGN;
2305 /* If redzoning, ensure that the second redzone is suitably
2306 * aligned, by adjusting the object size accordingly. */
2307 size += REDZONE_ALIGN - 1;
2308 size &= ~(REDZONE_ALIGN - 1);
2311 /* 3) caller mandated alignment */
2312 if (ralign < cachep->align) {
2313 ralign = cachep->align;
2315 /* disable debug if necessary */
2316 if (ralign > __alignof__(unsigned long long))
2317 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2321 cachep->align = ralign;
2323 if (slab_is_available())
2328 setup_node_pointer(cachep);
2332 * Both debugging options require word-alignment which is calculated
2335 if (flags & SLAB_RED_ZONE) {
2336 /* add space for red zone words */
2337 cachep->obj_offset += sizeof(unsigned long long);
2338 size += 2 * sizeof(unsigned long long);
2340 if (flags & SLAB_STORE_USER) {
2341 /* user store requires one word storage behind the end of
2342 * the real object. But if the second red zone needs to be
2343 * aligned to 64 bits, we must allow that much space.
2345 if (flags & SLAB_RED_ZONE)
2346 size += REDZONE_ALIGN;
2348 size += BYTES_PER_WORD;
2350 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2351 if (size >= kmalloc_size(INDEX_NODE + 1)
2352 && cachep->object_size > cache_line_size()
2353 && ALIGN(size, cachep->align) < PAGE_SIZE) {
2354 cachep->obj_offset += PAGE_SIZE - ALIGN(size, cachep->align);
2361 * Determine if the slab management is 'on' or 'off' slab.
2362 * (bootstrapping cannot cope with offslab caches so don't do
2363 * it too early on. Always use on-slab management when
2364 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2366 if ((size >= (PAGE_SIZE >> 5)) && !slab_early_init &&
2367 !(flags & SLAB_NOLEAKTRACE))
2369 * Size is large, assume best to place the slab management obj
2370 * off-slab (should allow better packing of objs).
2372 flags |= CFLGS_OFF_SLAB;
2374 size = ALIGN(size, cachep->align);
2376 * We should restrict the number of objects in a slab to implement
2377 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2379 if (FREELIST_BYTE_INDEX && size < SLAB_OBJ_MIN_SIZE)
2380 size = ALIGN(SLAB_OBJ_MIN_SIZE, cachep->align);
2382 left_over = calculate_slab_order(cachep, size, cachep->align, flags);
2387 freelist_size = calculate_freelist_size(cachep->num, cachep->align);
2390 * If the slab has been placed off-slab, and we have enough space then
2391 * move it on-slab. This is at the expense of any extra colouring.
2393 if (flags & CFLGS_OFF_SLAB && left_over >= freelist_size) {
2394 flags &= ~CFLGS_OFF_SLAB;
2395 left_over -= freelist_size;
2398 if (flags & CFLGS_OFF_SLAB) {
2399 /* really off slab. No need for manual alignment */
2400 freelist_size = calculate_freelist_size(cachep->num, 0);
2402 #ifdef CONFIG_PAGE_POISONING
2403 /* If we're going to use the generic kernel_map_pages()
2404 * poisoning, then it's going to smash the contents of
2405 * the redzone and userword anyhow, so switch them off.
2407 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2408 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2412 cachep->colour_off = cache_line_size();
2413 /* Offset must be a multiple of the alignment. */
2414 if (cachep->colour_off < cachep->align)
2415 cachep->colour_off = cachep->align;
2416 cachep->colour = left_over / cachep->colour_off;
2417 cachep->freelist_size = freelist_size;
2418 cachep->flags = flags;
2419 cachep->allocflags = __GFP_COMP;
2420 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2421 cachep->allocflags |= GFP_DMA;
2422 cachep->size = size;
2423 cachep->reciprocal_buffer_size = reciprocal_value(size);
2425 if (flags & CFLGS_OFF_SLAB) {
2426 cachep->freelist_cache = kmalloc_slab(freelist_size, 0u);
2428 * This is a possibility for one of the kmalloc_{dma,}_caches.
2429 * But since we go off slab only for object size greater than
2430 * PAGE_SIZE/8, and kmalloc_{dma,}_caches get created
2431 * in ascending order,this should not happen at all.
2432 * But leave a BUG_ON for some lucky dude.
2434 BUG_ON(ZERO_OR_NULL_PTR(cachep->freelist_cache));
2437 err = setup_cpu_cache(cachep, gfp);
2439 __kmem_cache_shutdown(cachep);
2443 if (flags & SLAB_DEBUG_OBJECTS) {
2445 * Would deadlock through slab_destroy()->call_rcu()->
2446 * debug_object_activate()->kmem_cache_alloc().
2448 WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU);
2450 slab_set_debugobj_lock_classes(cachep);
2451 } else if (!OFF_SLAB(cachep) && !(flags & SLAB_DESTROY_BY_RCU))
2452 on_slab_lock_classes(cachep);
2458 static void check_irq_off(void)
2460 BUG_ON(!irqs_disabled());
2463 static void check_irq_on(void)
2465 BUG_ON(irqs_disabled());
2468 static void check_spinlock_acquired(struct kmem_cache *cachep)
2472 assert_spin_locked(&get_node(cachep, numa_mem_id())->list_lock);
2476 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2480 assert_spin_locked(&get_node(cachep, node)->list_lock);
2485 #define check_irq_off() do { } while(0)
2486 #define check_irq_on() do { } while(0)
2487 #define check_spinlock_acquired(x) do { } while(0)
2488 #define check_spinlock_acquired_node(x, y) do { } while(0)
2491 static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
2492 struct array_cache *ac,
2493 int force, int node);
2495 static void do_drain(void *arg)
2497 struct kmem_cache *cachep = arg;
2498 struct array_cache *ac;
2499 int node = numa_mem_id();
2500 struct kmem_cache_node *n;
2504 ac = cpu_cache_get(cachep);
2505 n = get_node(cachep, node);
2506 spin_lock(&n->list_lock);
2507 free_block(cachep, ac->entry, ac->avail, node, &list);
2508 spin_unlock(&n->list_lock);
2509 slabs_destroy(cachep, &list);
2513 static void drain_cpu_caches(struct kmem_cache *cachep)
2515 struct kmem_cache_node *n;
2518 on_each_cpu(do_drain, cachep, 1);
2520 for_each_kmem_cache_node(cachep, node, n)
2522 drain_alien_cache(cachep, n->alien);
2524 for_each_kmem_cache_node(cachep, node, n)
2525 drain_array(cachep, n, n->shared, 1, node);
2529 * Remove slabs from the list of free slabs.
2530 * Specify the number of slabs to drain in tofree.
2532 * Returns the actual number of slabs released.
2534 static int drain_freelist(struct kmem_cache *cache,
2535 struct kmem_cache_node *n, int tofree)
2537 struct list_head *p;
2542 while (nr_freed < tofree && !list_empty(&n->slabs_free)) {
2544 spin_lock_irq(&n->list_lock);
2545 p = n->slabs_free.prev;
2546 if (p == &n->slabs_free) {
2547 spin_unlock_irq(&n->list_lock);
2551 page = list_entry(p, struct page, lru);
2553 BUG_ON(page->active);
2555 list_del(&page->lru);
2557 * Safe to drop the lock. The slab is no longer linked
2560 n->free_objects -= cache->num;
2561 spin_unlock_irq(&n->list_lock);
2562 slab_destroy(cache, page);
2569 int __kmem_cache_shrink(struct kmem_cache *cachep)
2573 struct kmem_cache_node *n;
2575 drain_cpu_caches(cachep);
2578 for_each_kmem_cache_node(cachep, node, n) {
2579 drain_freelist(cachep, n, slabs_tofree(cachep, n));
2581 ret += !list_empty(&n->slabs_full) ||
2582 !list_empty(&n->slabs_partial);
2584 return (ret ? 1 : 0);
2587 int __kmem_cache_shutdown(struct kmem_cache *cachep)
2590 struct kmem_cache_node *n;
2591 int rc = __kmem_cache_shrink(cachep);
2596 for_each_online_cpu(i)
2597 kfree(cachep->array[i]);
2599 /* NUMA: free the node structures */
2600 for_each_kmem_cache_node(cachep, i, n) {
2602 free_alien_cache(n->alien);
2604 cachep->node[i] = NULL;
2610 * Get the memory for a slab management obj.
2612 * For a slab cache when the slab descriptor is off-slab, the
2613 * slab descriptor can't come from the same cache which is being created,
2614 * Because if it is the case, that means we defer the creation of
2615 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2616 * And we eventually call down to __kmem_cache_create(), which
2617 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2618 * This is a "chicken-and-egg" problem.
2620 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2621 * which are all initialized during kmem_cache_init().
2623 static void *alloc_slabmgmt(struct kmem_cache *cachep,
2624 struct page *page, int colour_off,
2625 gfp_t local_flags, int nodeid)
2628 void *addr = page_address(page);
2630 if (OFF_SLAB(cachep)) {
2631 /* Slab management obj is off-slab. */
2632 freelist = kmem_cache_alloc_node(cachep->freelist_cache,
2633 local_flags, nodeid);
2637 freelist = addr + colour_off;
2638 colour_off += cachep->freelist_size;
2641 page->s_mem = addr + colour_off;
2645 static inline freelist_idx_t get_free_obj(struct page *page, unsigned int idx)
2647 return ((freelist_idx_t *)page->freelist)[idx];
2650 static inline void set_free_obj(struct page *page,
2651 unsigned int idx, freelist_idx_t val)
2653 ((freelist_idx_t *)(page->freelist))[idx] = val;
2656 static void cache_init_objs(struct kmem_cache *cachep,
2661 for (i = 0; i < cachep->num; i++) {
2662 void *objp = index_to_obj(cachep, page, i);
2664 /* need to poison the objs? */
2665 if (cachep->flags & SLAB_POISON)
2666 poison_obj(cachep, objp, POISON_FREE);
2667 if (cachep->flags & SLAB_STORE_USER)
2668 *dbg_userword(cachep, objp) = NULL;
2670 if (cachep->flags & SLAB_RED_ZONE) {
2671 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2672 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2675 * Constructors are not allowed to allocate memory from the same
2676 * cache which they are a constructor for. Otherwise, deadlock.
2677 * They must also be threaded.
2679 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2680 cachep->ctor(objp + obj_offset(cachep));
2682 if (cachep->flags & SLAB_RED_ZONE) {
2683 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2684 slab_error(cachep, "constructor overwrote the"
2685 " end of an object");
2686 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2687 slab_error(cachep, "constructor overwrote the"
2688 " start of an object");
2690 if ((cachep->size % PAGE_SIZE) == 0 &&
2691 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2692 kernel_map_pages(virt_to_page(objp),
2693 cachep->size / PAGE_SIZE, 0);
2698 set_obj_status(page, i, OBJECT_FREE);
2699 set_free_obj(page, i, i);
2703 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2705 if (CONFIG_ZONE_DMA_FLAG) {
2706 if (flags & GFP_DMA)
2707 BUG_ON(!(cachep->allocflags & GFP_DMA));
2709 BUG_ON(cachep->allocflags & GFP_DMA);
2713 static void *slab_get_obj(struct kmem_cache *cachep, struct page *page,
2718 objp = index_to_obj(cachep, page, get_free_obj(page, page->active));
2721 WARN_ON(page_to_nid(virt_to_page(objp)) != nodeid);
2727 static void slab_put_obj(struct kmem_cache *cachep, struct page *page,
2728 void *objp, int nodeid)
2730 unsigned int objnr = obj_to_index(cachep, page, objp);
2734 /* Verify that the slab belongs to the intended node */
2735 WARN_ON(page_to_nid(virt_to_page(objp)) != nodeid);
2737 /* Verify double free bug */
2738 for (i = page->active; i < cachep->num; i++) {
2739 if (get_free_obj(page, i) == objnr) {
2740 printk(KERN_ERR "slab: double free detected in cache "
2741 "'%s', objp %p\n", cachep->name, objp);
2747 set_free_obj(page, page->active, objnr);
2751 * Map pages beginning at addr to the given cache and slab. This is required
2752 * for the slab allocator to be able to lookup the cache and slab of a
2753 * virtual address for kfree, ksize, and slab debugging.
2755 static void slab_map_pages(struct kmem_cache *cache, struct page *page,
2758 page->slab_cache = cache;
2759 page->freelist = freelist;
2763 * Grow (by 1) the number of slabs within a cache. This is called by
2764 * kmem_cache_alloc() when there are no active objs left in a cache.
2766 static int cache_grow(struct kmem_cache *cachep,
2767 gfp_t flags, int nodeid, struct page *page)
2772 struct kmem_cache_node *n;
2775 * Be lazy and only check for valid flags here, keeping it out of the
2776 * critical path in kmem_cache_alloc().
2778 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2779 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2781 /* Take the node list lock to change the colour_next on this node */
2783 n = get_node(cachep, nodeid);
2784 spin_lock(&n->list_lock);
2786 /* Get colour for the slab, and cal the next value. */
2787 offset = n->colour_next;
2789 if (n->colour_next >= cachep->colour)
2791 spin_unlock(&n->list_lock);
2793 offset *= cachep->colour_off;
2795 if (local_flags & __GFP_WAIT)
2799 * The test for missing atomic flag is performed here, rather than
2800 * the more obvious place, simply to reduce the critical path length
2801 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2802 * will eventually be caught here (where it matters).
2804 kmem_flagcheck(cachep, flags);
2807 * Get mem for the objs. Attempt to allocate a physical page from
2811 page = kmem_getpages(cachep, local_flags, nodeid);
2815 /* Get slab management. */
2816 freelist = alloc_slabmgmt(cachep, page, offset,
2817 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2821 slab_map_pages(cachep, page, freelist);
2823 cache_init_objs(cachep, page);
2825 if (local_flags & __GFP_WAIT)
2826 local_irq_disable();
2828 spin_lock(&n->list_lock);
2830 /* Make slab active. */
2831 list_add_tail(&page->lru, &(n->slabs_free));
2832 STATS_INC_GROWN(cachep);
2833 n->free_objects += cachep->num;
2834 spin_unlock(&n->list_lock);
2837 kmem_freepages(cachep, page);
2839 if (local_flags & __GFP_WAIT)
2840 local_irq_disable();
2847 * Perform extra freeing checks:
2848 * - detect bad pointers.
2849 * - POISON/RED_ZONE checking
2851 static void kfree_debugcheck(const void *objp)
2853 if (!virt_addr_valid(objp)) {
2854 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2855 (unsigned long)objp);
2860 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2862 unsigned long long redzone1, redzone2;
2864 redzone1 = *dbg_redzone1(cache, obj);
2865 redzone2 = *dbg_redzone2(cache, obj);
2870 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2873 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2874 slab_error(cache, "double free detected");
2876 slab_error(cache, "memory outside object was overwritten");
2878 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2879 obj, redzone1, redzone2);
2882 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2883 unsigned long caller)
2888 BUG_ON(virt_to_cache(objp) != cachep);
2890 objp -= obj_offset(cachep);
2891 kfree_debugcheck(objp);
2892 page = virt_to_head_page(objp);
2894 if (cachep->flags & SLAB_RED_ZONE) {
2895 verify_redzone_free(cachep, objp);
2896 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2897 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2899 if (cachep->flags & SLAB_STORE_USER)
2900 *dbg_userword(cachep, objp) = (void *)caller;
2902 objnr = obj_to_index(cachep, page, objp);
2904 BUG_ON(objnr >= cachep->num);
2905 BUG_ON(objp != index_to_obj(cachep, page, objnr));
2907 set_obj_status(page, objnr, OBJECT_FREE);
2908 if (cachep->flags & SLAB_POISON) {
2909 #ifdef CONFIG_DEBUG_PAGEALLOC
2910 if ((cachep->size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
2911 store_stackinfo(cachep, objp, caller);
2912 kernel_map_pages(virt_to_page(objp),
2913 cachep->size / PAGE_SIZE, 0);
2915 poison_obj(cachep, objp, POISON_FREE);
2918 poison_obj(cachep, objp, POISON_FREE);
2925 #define kfree_debugcheck(x) do { } while(0)
2926 #define cache_free_debugcheck(x,objp,z) (objp)
2929 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags,
2933 struct kmem_cache_node *n;
2934 struct array_cache *ac;
2938 node = numa_mem_id();
2939 if (unlikely(force_refill))
2942 ac = cpu_cache_get(cachep);
2943 batchcount = ac->batchcount;
2944 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2946 * If there was little recent activity on this cache, then
2947 * perform only a partial refill. Otherwise we could generate
2950 batchcount = BATCHREFILL_LIMIT;
2952 n = get_node(cachep, node);
2954 BUG_ON(ac->avail > 0 || !n);
2955 spin_lock(&n->list_lock);
2957 /* See if we can refill from the shared array */
2958 if (n->shared && transfer_objects(ac, n->shared, batchcount)) {
2959 n->shared->touched = 1;
2963 while (batchcount > 0) {
2964 struct list_head *entry;
2966 /* Get slab alloc is to come from. */
2967 entry = n->slabs_partial.next;
2968 if (entry == &n->slabs_partial) {
2969 n->free_touched = 1;
2970 entry = n->slabs_free.next;
2971 if (entry == &n->slabs_free)
2975 page = list_entry(entry, struct page, lru);
2976 check_spinlock_acquired(cachep);
2979 * The slab was either on partial or free list so
2980 * there must be at least one object available for
2983 BUG_ON(page->active >= cachep->num);
2985 while (page->active < cachep->num && batchcount--) {
2986 STATS_INC_ALLOCED(cachep);
2987 STATS_INC_ACTIVE(cachep);
2988 STATS_SET_HIGH(cachep);
2990 ac_put_obj(cachep, ac, slab_get_obj(cachep, page,
2994 /* move slabp to correct slabp list: */
2995 list_del(&page->lru);
2996 if (page->active == cachep->num)
2997 list_add(&page->lru, &n->slabs_full);
2999 list_add(&page->lru, &n->slabs_partial);
3003 n->free_objects -= ac->avail;
3005 spin_unlock(&n->list_lock);
3007 if (unlikely(!ac->avail)) {
3010 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
3012 /* cache_grow can reenable interrupts, then ac could change. */
3013 ac = cpu_cache_get(cachep);
3014 node = numa_mem_id();
3016 /* no objects in sight? abort */
3017 if (!x && (ac->avail == 0 || force_refill))
3020 if (!ac->avail) /* objects refilled by interrupt? */
3025 return ac_get_obj(cachep, ac, flags, force_refill);
3028 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3031 might_sleep_if(flags & __GFP_WAIT);
3033 kmem_flagcheck(cachep, flags);
3038 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3039 gfp_t flags, void *objp, unsigned long caller)
3045 if (cachep->flags & SLAB_POISON) {
3046 #ifdef CONFIG_DEBUG_PAGEALLOC
3047 if ((cachep->size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
3048 kernel_map_pages(virt_to_page(objp),
3049 cachep->size / PAGE_SIZE, 1);
3051 check_poison_obj(cachep, objp);
3053 check_poison_obj(cachep, objp);
3055 poison_obj(cachep, objp, POISON_INUSE);
3057 if (cachep->flags & SLAB_STORE_USER)
3058 *dbg_userword(cachep, objp) = (void *)caller;
3060 if (cachep->flags & SLAB_RED_ZONE) {
3061 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3062 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3063 slab_error(cachep, "double free, or memory outside"
3064 " object was overwritten");
3066 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3067 objp, *dbg_redzone1(cachep, objp),
3068 *dbg_redzone2(cachep, objp));
3070 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3071 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3074 page = virt_to_head_page(objp);
3075 set_obj_status(page, obj_to_index(cachep, page, objp), OBJECT_ACTIVE);
3076 objp += obj_offset(cachep);
3077 if (cachep->ctor && cachep->flags & SLAB_POISON)
3079 if (ARCH_SLAB_MINALIGN &&
3080 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
3081 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3082 objp, (int)ARCH_SLAB_MINALIGN);
3087 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3090 static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
3092 if (unlikely(cachep == kmem_cache))
3095 return should_failslab(cachep->object_size, flags, cachep->flags);
3098 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3101 struct array_cache *ac;
3102 bool force_refill = false;
3106 ac = cpu_cache_get(cachep);
3107 if (likely(ac->avail)) {
3109 objp = ac_get_obj(cachep, ac, flags, false);
3112 * Allow for the possibility all avail objects are not allowed
3113 * by the current flags
3116 STATS_INC_ALLOCHIT(cachep);
3119 force_refill = true;
3122 STATS_INC_ALLOCMISS(cachep);
3123 objp = cache_alloc_refill(cachep, flags, force_refill);
3125 * the 'ac' may be updated by cache_alloc_refill(),
3126 * and kmemleak_erase() requires its correct value.
3128 ac = cpu_cache_get(cachep);
3132 * To avoid a false negative, if an object that is in one of the
3133 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3134 * treat the array pointers as a reference to the object.
3137 kmemleak_erase(&ac->entry[ac->avail]);
3143 * Try allocating on another node if PF_SPREAD_SLAB is a mempolicy is set.
3145 * If we are in_interrupt, then process context, including cpusets and
3146 * mempolicy, may not apply and should not be used for allocation policy.
3148 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3150 int nid_alloc, nid_here;
3152 if (in_interrupt() || (flags & __GFP_THISNODE))
3154 nid_alloc = nid_here = numa_mem_id();
3155 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3156 nid_alloc = cpuset_slab_spread_node();
3157 else if (current->mempolicy)
3158 nid_alloc = mempolicy_slab_node();
3159 if (nid_alloc != nid_here)
3160 return ____cache_alloc_node(cachep, flags, nid_alloc);
3165 * Fallback function if there was no memory available and no objects on a
3166 * certain node and fall back is permitted. First we scan all the
3167 * available node for available objects. If that fails then we
3168 * perform an allocation without specifying a node. This allows the page
3169 * allocator to do its reclaim / fallback magic. We then insert the
3170 * slab into the proper nodelist and then allocate from it.
3172 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3174 struct zonelist *zonelist;
3178 enum zone_type high_zoneidx = gfp_zone(flags);
3181 unsigned int cpuset_mems_cookie;
3183 if (flags & __GFP_THISNODE)
3186 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3189 cpuset_mems_cookie = read_mems_allowed_begin();
3190 zonelist = node_zonelist(mempolicy_slab_node(), flags);
3194 * Look through allowed nodes for objects available
3195 * from existing per node queues.
3197 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3198 nid = zone_to_nid(zone);
3200 if (cpuset_zone_allowed_hardwall(zone, flags) &&
3201 get_node(cache, nid) &&
3202 get_node(cache, nid)->free_objects) {
3203 obj = ____cache_alloc_node(cache,
3204 flags | GFP_THISNODE, nid);
3212 * This allocation will be performed within the constraints
3213 * of the current cpuset / memory policy requirements.
3214 * We may trigger various forms of reclaim on the allowed
3215 * set and go into memory reserves if necessary.
3219 if (local_flags & __GFP_WAIT)
3221 kmem_flagcheck(cache, flags);
3222 page = kmem_getpages(cache, local_flags, numa_mem_id());
3223 if (local_flags & __GFP_WAIT)
3224 local_irq_disable();
3227 * Insert into the appropriate per node queues
3229 nid = page_to_nid(page);
3230 if (cache_grow(cache, flags, nid, page)) {
3231 obj = ____cache_alloc_node(cache,
3232 flags | GFP_THISNODE, nid);
3235 * Another processor may allocate the
3236 * objects in the slab since we are
3237 * not holding any locks.
3241 /* cache_grow already freed obj */
3247 if (unlikely(!obj && read_mems_allowed_retry(cpuset_mems_cookie)))
3253 * A interface to enable slab creation on nodeid
3255 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3258 struct list_head *entry;
3260 struct kmem_cache_node *n;
3264 VM_BUG_ON(nodeid > num_online_nodes());
3265 n = get_node(cachep, nodeid);
3270 spin_lock(&n->list_lock);
3271 entry = n->slabs_partial.next;
3272 if (entry == &n->slabs_partial) {
3273 n->free_touched = 1;
3274 entry = n->slabs_free.next;
3275 if (entry == &n->slabs_free)
3279 page = list_entry(entry, struct page, lru);
3280 check_spinlock_acquired_node(cachep, nodeid);
3282 STATS_INC_NODEALLOCS(cachep);
3283 STATS_INC_ACTIVE(cachep);
3284 STATS_SET_HIGH(cachep);
3286 BUG_ON(page->active == cachep->num);
3288 obj = slab_get_obj(cachep, page, nodeid);
3290 /* move slabp to correct slabp list: */
3291 list_del(&page->lru);
3293 if (page->active == cachep->num)
3294 list_add(&page->lru, &n->slabs_full);
3296 list_add(&page->lru, &n->slabs_partial);
3298 spin_unlock(&n->list_lock);
3302 spin_unlock(&n->list_lock);
3303 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
3307 return fallback_alloc(cachep, flags);
3313 static __always_inline void *
3314 slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3315 unsigned long caller)
3317 unsigned long save_flags;
3319 int slab_node = numa_mem_id();
3321 flags &= gfp_allowed_mask;
3323 lockdep_trace_alloc(flags);
3325 if (slab_should_failslab(cachep, flags))
3328 cachep = memcg_kmem_get_cache(cachep, flags);
3330 cache_alloc_debugcheck_before(cachep, flags);
3331 local_irq_save(save_flags);
3333 if (nodeid == NUMA_NO_NODE)
3336 if (unlikely(!get_node(cachep, nodeid))) {
3337 /* Node not bootstrapped yet */
3338 ptr = fallback_alloc(cachep, flags);
3342 if (nodeid == slab_node) {
3344 * Use the locally cached objects if possible.
3345 * However ____cache_alloc does not allow fallback
3346 * to other nodes. It may fail while we still have
3347 * objects on other nodes available.
3349 ptr = ____cache_alloc(cachep, flags);
3353 /* ___cache_alloc_node can fall back to other nodes */
3354 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3356 local_irq_restore(save_flags);
3357 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3358 kmemleak_alloc_recursive(ptr, cachep->object_size, 1, cachep->flags,
3362 kmemcheck_slab_alloc(cachep, flags, ptr, cachep->object_size);
3363 if (unlikely(flags & __GFP_ZERO))
3364 memset(ptr, 0, cachep->object_size);
3370 static __always_inline void *
3371 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3375 if (current->mempolicy || unlikely(current->flags & PF_SPREAD_SLAB)) {
3376 objp = alternate_node_alloc(cache, flags);
3380 objp = ____cache_alloc(cache, flags);
3383 * We may just have run out of memory on the local node.
3384 * ____cache_alloc_node() knows how to locate memory on other nodes
3387 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3394 static __always_inline void *
3395 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3397 return ____cache_alloc(cachep, flags);
3400 #endif /* CONFIG_NUMA */
3402 static __always_inline void *
3403 slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
3405 unsigned long save_flags;
3408 flags &= gfp_allowed_mask;
3410 lockdep_trace_alloc(flags);
3412 if (slab_should_failslab(cachep, flags))
3415 cachep = memcg_kmem_get_cache(cachep, flags);
3417 cache_alloc_debugcheck_before(cachep, flags);
3418 local_irq_save(save_flags);
3419 objp = __do_cache_alloc(cachep, flags);
3420 local_irq_restore(save_flags);
3421 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3422 kmemleak_alloc_recursive(objp, cachep->object_size, 1, cachep->flags,
3427 kmemcheck_slab_alloc(cachep, flags, objp, cachep->object_size);
3428 if (unlikely(flags & __GFP_ZERO))
3429 memset(objp, 0, cachep->object_size);
3436 * Caller needs to acquire correct kmem_cache_node's list_lock
3437 * @list: List of detached free slabs should be freed by caller
3439 static void free_block(struct kmem_cache *cachep, void **objpp,
3440 int nr_objects, int node, struct list_head *list)
3443 struct kmem_cache_node *n = get_node(cachep, node);
3445 for (i = 0; i < nr_objects; i++) {
3449 clear_obj_pfmemalloc(&objpp[i]);
3452 page = virt_to_head_page(objp);
3453 list_del(&page->lru);
3454 check_spinlock_acquired_node(cachep, node);
3455 slab_put_obj(cachep, page, objp, node);
3456 STATS_DEC_ACTIVE(cachep);
3459 /* fixup slab chains */
3460 if (page->active == 0) {
3461 if (n->free_objects > n->free_limit) {
3462 n->free_objects -= cachep->num;
3463 list_add_tail(&page->lru, list);
3465 list_add(&page->lru, &n->slabs_free);
3468 /* Unconditionally move a slab to the end of the
3469 * partial list on free - maximum time for the
3470 * other objects to be freed, too.
3472 list_add_tail(&page->lru, &n->slabs_partial);
3477 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3480 struct kmem_cache_node *n;
3481 int node = numa_mem_id();
3484 batchcount = ac->batchcount;
3486 BUG_ON(!batchcount || batchcount > ac->avail);
3489 n = get_node(cachep, node);
3490 spin_lock(&n->list_lock);
3492 struct array_cache *shared_array = n->shared;
3493 int max = shared_array->limit - shared_array->avail;
3495 if (batchcount > max)
3497 memcpy(&(shared_array->entry[shared_array->avail]),
3498 ac->entry, sizeof(void *) * batchcount);
3499 shared_array->avail += batchcount;
3504 free_block(cachep, ac->entry, batchcount, node, &list);
3509 struct list_head *p;
3511 p = n->slabs_free.next;
3512 while (p != &(n->slabs_free)) {
3515 page = list_entry(p, struct page, lru);
3516 BUG_ON(page->active);
3521 STATS_SET_FREEABLE(cachep, i);
3524 spin_unlock(&n->list_lock);
3525 slabs_destroy(cachep, &list);
3526 ac->avail -= batchcount;
3527 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3531 * Release an obj back to its cache. If the obj has a constructed state, it must
3532 * be in this state _before_ it is released. Called with disabled ints.
3534 static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3535 unsigned long caller)
3537 struct array_cache *ac = cpu_cache_get(cachep);
3540 kmemleak_free_recursive(objp, cachep->flags);
3541 objp = cache_free_debugcheck(cachep, objp, caller);
3543 kmemcheck_slab_free(cachep, objp, cachep->object_size);
3546 * Skip calling cache_free_alien() when the platform is not numa.
3547 * This will avoid cache misses that happen while accessing slabp (which
3548 * is per page memory reference) to get nodeid. Instead use a global
3549 * variable to skip the call, which is mostly likely to be present in
3552 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3555 if (likely(ac->avail < ac->limit)) {
3556 STATS_INC_FREEHIT(cachep);
3558 STATS_INC_FREEMISS(cachep);
3559 cache_flusharray(cachep, ac);
3562 ac_put_obj(cachep, ac, objp);
3566 * kmem_cache_alloc - Allocate an object
3567 * @cachep: The cache to allocate from.
3568 * @flags: See kmalloc().
3570 * Allocate an object from this cache. The flags are only relevant
3571 * if the cache has no available objects.
3573 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3575 void *ret = slab_alloc(cachep, flags, _RET_IP_);
3577 trace_kmem_cache_alloc(_RET_IP_, ret,
3578 cachep->object_size, cachep->size, flags);
3582 EXPORT_SYMBOL(kmem_cache_alloc);
3584 #ifdef CONFIG_TRACING
3586 kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
3590 ret = slab_alloc(cachep, flags, _RET_IP_);
3592 trace_kmalloc(_RET_IP_, ret,
3593 size, cachep->size, flags);
3596 EXPORT_SYMBOL(kmem_cache_alloc_trace);
3601 * kmem_cache_alloc_node - Allocate an object on the specified node
3602 * @cachep: The cache to allocate from.
3603 * @flags: See kmalloc().
3604 * @nodeid: node number of the target node.
3606 * Identical to kmem_cache_alloc but it will allocate memory on the given
3607 * node, which can improve the performance for cpu bound structures.
3609 * Fallback to other node is possible if __GFP_THISNODE is not set.
3611 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3613 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3615 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3616 cachep->object_size, cachep->size,
3621 EXPORT_SYMBOL(kmem_cache_alloc_node);
3623 #ifdef CONFIG_TRACING
3624 void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
3631 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3633 trace_kmalloc_node(_RET_IP_, ret,
3638 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3641 static __always_inline void *
3642 __do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
3644 struct kmem_cache *cachep;
3646 cachep = kmalloc_slab(size, flags);
3647 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3649 return kmem_cache_alloc_node_trace(cachep, flags, node, size);
3652 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3653 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3655 return __do_kmalloc_node(size, flags, node, _RET_IP_);
3657 EXPORT_SYMBOL(__kmalloc_node);
3659 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3660 int node, unsigned long caller)
3662 return __do_kmalloc_node(size, flags, node, caller);
3664 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3666 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3668 return __do_kmalloc_node(size, flags, node, 0);
3670 EXPORT_SYMBOL(__kmalloc_node);
3671 #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3672 #endif /* CONFIG_NUMA */
3675 * __do_kmalloc - allocate memory
3676 * @size: how many bytes of memory are required.
3677 * @flags: the type of memory to allocate (see kmalloc).
3678 * @caller: function caller for debug tracking of the caller
3680 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3681 unsigned long caller)
3683 struct kmem_cache *cachep;
3686 cachep = kmalloc_slab(size, flags);
3687 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3689 ret = slab_alloc(cachep, flags, caller);
3691 trace_kmalloc(caller, ret,
3692 size, cachep->size, flags);
3698 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3699 void *__kmalloc(size_t size, gfp_t flags)
3701 return __do_kmalloc(size, flags, _RET_IP_);
3703 EXPORT_SYMBOL(__kmalloc);
3705 void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3707 return __do_kmalloc(size, flags, caller);
3709 EXPORT_SYMBOL(__kmalloc_track_caller);
3712 void *__kmalloc(size_t size, gfp_t flags)
3714 return __do_kmalloc(size, flags, 0);
3716 EXPORT_SYMBOL(__kmalloc);
3720 * kmem_cache_free - Deallocate an object
3721 * @cachep: The cache the allocation was from.
3722 * @objp: The previously allocated object.
3724 * Free an object which was previously allocated from this
3727 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3729 unsigned long flags;
3730 cachep = cache_from_obj(cachep, objp);
3734 local_irq_save(flags);
3735 debug_check_no_locks_freed(objp, cachep->object_size);
3736 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3737 debug_check_no_obj_freed(objp, cachep->object_size);
3738 __cache_free(cachep, objp, _RET_IP_);
3739 local_irq_restore(flags);
3741 trace_kmem_cache_free(_RET_IP_, objp);
3743 EXPORT_SYMBOL(kmem_cache_free);
3746 * kfree - free previously allocated memory
3747 * @objp: pointer returned by kmalloc.
3749 * If @objp is NULL, no operation is performed.
3751 * Don't free memory not originally allocated by kmalloc()
3752 * or you will run into trouble.
3754 void kfree(const void *objp)
3756 struct kmem_cache *c;
3757 unsigned long flags;
3759 trace_kfree(_RET_IP_, objp);
3761 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3763 local_irq_save(flags);
3764 kfree_debugcheck(objp);
3765 c = virt_to_cache(objp);
3766 debug_check_no_locks_freed(objp, c->object_size);
3768 debug_check_no_obj_freed(objp, c->object_size);
3769 __cache_free(c, (void *)objp, _RET_IP_);
3770 local_irq_restore(flags);
3772 EXPORT_SYMBOL(kfree);
3775 * This initializes kmem_cache_node or resizes various caches for all nodes.
3777 static int alloc_kmem_cache_node(struct kmem_cache *cachep, gfp_t gfp)
3780 struct kmem_cache_node *n;
3781 struct array_cache *new_shared;
3782 struct alien_cache **new_alien = NULL;
3784 for_each_online_node(node) {
3786 if (use_alien_caches) {
3787 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
3793 if (cachep->shared) {
3794 new_shared = alloc_arraycache(node,
3795 cachep->shared*cachep->batchcount,
3798 free_alien_cache(new_alien);
3803 n = get_node(cachep, node);
3805 struct array_cache *shared = n->shared;
3808 spin_lock_irq(&n->list_lock);
3811 free_block(cachep, shared->entry,
3812 shared->avail, node, &list);
3814 n->shared = new_shared;
3816 n->alien = new_alien;
3819 n->free_limit = (1 + nr_cpus_node(node)) *
3820 cachep->batchcount + cachep->num;
3821 spin_unlock_irq(&n->list_lock);
3822 slabs_destroy(cachep, &list);
3824 free_alien_cache(new_alien);
3827 n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
3829 free_alien_cache(new_alien);
3834 kmem_cache_node_init(n);
3835 n->next_reap = jiffies + REAPTIMEOUT_NODE +
3836 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
3837 n->shared = new_shared;
3838 n->alien = new_alien;
3839 n->free_limit = (1 + nr_cpus_node(node)) *
3840 cachep->batchcount + cachep->num;
3841 cachep->node[node] = n;
3846 if (!cachep->list.next) {
3847 /* Cache is not active yet. Roll back what we did */
3850 n = get_node(cachep, node);
3853 free_alien_cache(n->alien);
3855 cachep->node[node] = NULL;
3863 struct ccupdate_struct {
3864 struct kmem_cache *cachep;
3865 struct array_cache *new[0];
3868 static void do_ccupdate_local(void *info)
3870 struct ccupdate_struct *new = info;
3871 struct array_cache *old;
3874 old = cpu_cache_get(new->cachep);
3876 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3877 new->new[smp_processor_id()] = old;
3880 /* Always called with the slab_mutex held */
3881 static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
3882 int batchcount, int shared, gfp_t gfp)
3884 struct ccupdate_struct *new;
3887 new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *),
3892 for_each_online_cpu(i) {
3893 new->new[i] = alloc_arraycache(cpu_to_mem(i), limit,
3896 for (i--; i >= 0; i--)
3902 new->cachep = cachep;
3904 on_each_cpu(do_ccupdate_local, (void *)new, 1);
3907 cachep->batchcount = batchcount;
3908 cachep->limit = limit;
3909 cachep->shared = shared;
3911 for_each_online_cpu(i) {
3913 struct array_cache *ccold = new->new[i];
3915 struct kmem_cache_node *n;
3920 node = cpu_to_mem(i);
3921 n = get_node(cachep, node);
3922 spin_lock_irq(&n->list_lock);
3923 free_block(cachep, ccold->entry, ccold->avail, node, &list);
3924 spin_unlock_irq(&n->list_lock);
3925 slabs_destroy(cachep, &list);
3929 return alloc_kmem_cache_node(cachep, gfp);
3932 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3933 int batchcount, int shared, gfp_t gfp)
3936 struct kmem_cache *c = NULL;
3939 ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3941 if (slab_state < FULL)
3944 if ((ret < 0) || !is_root_cache(cachep))
3947 VM_BUG_ON(!mutex_is_locked(&slab_mutex));
3948 for_each_memcg_cache_index(i) {
3949 c = cache_from_memcg_idx(cachep, i);
3951 /* return value determined by the parent cache only */
3952 __do_tune_cpucache(c, limit, batchcount, shared, gfp);
3958 /* Called with slab_mutex held always */
3959 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
3966 if (!is_root_cache(cachep)) {
3967 struct kmem_cache *root = memcg_root_cache(cachep);
3968 limit = root->limit;
3969 shared = root->shared;
3970 batchcount = root->batchcount;
3973 if (limit && shared && batchcount)
3976 * The head array serves three purposes:
3977 * - create a LIFO ordering, i.e. return objects that are cache-warm
3978 * - reduce the number of spinlock operations.
3979 * - reduce the number of linked list operations on the slab and
3980 * bufctl chains: array operations are cheaper.
3981 * The numbers are guessed, we should auto-tune as described by
3984 if (cachep->size > 131072)
3986 else if (cachep->size > PAGE_SIZE)
3988 else if (cachep->size > 1024)
3990 else if (cachep->size > 256)
3996 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3997 * allocation behaviour: Most allocs on one cpu, most free operations
3998 * on another cpu. For these cases, an efficient object passing between
3999 * cpus is necessary. This is provided by a shared array. The array
4000 * replaces Bonwick's magazine layer.
4001 * On uniprocessor, it's functionally equivalent (but less efficient)
4002 * to a larger limit. Thus disabled by default.
4005 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
4010 * With debugging enabled, large batchcount lead to excessively long
4011 * periods with disabled local interrupts. Limit the batchcount
4016 batchcount = (limit + 1) / 2;
4018 err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
4020 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
4021 cachep->name, -err);
4026 * Drain an array if it contains any elements taking the node lock only if
4027 * necessary. Note that the node listlock also protects the array_cache
4028 * if drain_array() is used on the shared array.
4030 static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
4031 struct array_cache *ac, int force, int node)
4036 if (!ac || !ac->avail)
4038 if (ac->touched && !force) {
4041 spin_lock_irq(&n->list_lock);
4043 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4044 if (tofree > ac->avail)
4045 tofree = (ac->avail + 1) / 2;
4046 free_block(cachep, ac->entry, tofree, node, &list);
4047 ac->avail -= tofree;
4048 memmove(ac->entry, &(ac->entry[tofree]),
4049 sizeof(void *) * ac->avail);
4051 spin_unlock_irq(&n->list_lock);
4052 slabs_destroy(cachep, &list);
4057 * cache_reap - Reclaim memory from caches.
4058 * @w: work descriptor
4060 * Called from workqueue/eventd every few seconds.
4062 * - clear the per-cpu caches for this CPU.
4063 * - return freeable pages to the main free memory pool.
4065 * If we cannot acquire the cache chain mutex then just give up - we'll try
4066 * again on the next iteration.
4068 static void cache_reap(struct work_struct *w)
4070 struct kmem_cache *searchp;
4071 struct kmem_cache_node *n;
4072 int node = numa_mem_id();
4073 struct delayed_work *work = to_delayed_work(w);
4075 if (!mutex_trylock(&slab_mutex))
4076 /* Give up. Setup the next iteration. */
4079 list_for_each_entry(searchp, &slab_caches, list) {
4083 * We only take the node lock if absolutely necessary and we
4084 * have established with reasonable certainty that
4085 * we can do some work if the lock was obtained.
4087 n = get_node(searchp, node);
4089 reap_alien(searchp, n);
4091 drain_array(searchp, n, cpu_cache_get(searchp), 0, node);
4094 * These are racy checks but it does not matter
4095 * if we skip one check or scan twice.
4097 if (time_after(n->next_reap, jiffies))
4100 n->next_reap = jiffies + REAPTIMEOUT_NODE;
4102 drain_array(searchp, n, n->shared, 0, node);
4104 if (n->free_touched)
4105 n->free_touched = 0;
4109 freed = drain_freelist(searchp, n, (n->free_limit +
4110 5 * searchp->num - 1) / (5 * searchp->num));
4111 STATS_ADD_REAPED(searchp, freed);
4117 mutex_unlock(&slab_mutex);
4120 /* Set up the next iteration */
4121 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_AC));
4124 #ifdef CONFIG_SLABINFO
4125 void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
4128 unsigned long active_objs;
4129 unsigned long num_objs;
4130 unsigned long active_slabs = 0;
4131 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4135 struct kmem_cache_node *n;
4139 for_each_kmem_cache_node(cachep, node, n) {
4142 spin_lock_irq(&n->list_lock);
4144 list_for_each_entry(page, &n->slabs_full, lru) {
4145 if (page->active != cachep->num && !error)
4146 error = "slabs_full accounting error";
4147 active_objs += cachep->num;
4150 list_for_each_entry(page, &n->slabs_partial, lru) {
4151 if (page->active == cachep->num && !error)
4152 error = "slabs_partial accounting error";
4153 if (!page->active && !error)
4154 error = "slabs_partial accounting error";
4155 active_objs += page->active;
4158 list_for_each_entry(page, &n->slabs_free, lru) {
4159 if (page->active && !error)
4160 error = "slabs_free accounting error";
4163 free_objects += n->free_objects;
4165 shared_avail += n->shared->avail;
4167 spin_unlock_irq(&n->list_lock);
4169 num_slabs += active_slabs;
4170 num_objs = num_slabs * cachep->num;
4171 if (num_objs - active_objs != free_objects && !error)
4172 error = "free_objects accounting error";
4174 name = cachep->name;
4176 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4178 sinfo->active_objs = active_objs;
4179 sinfo->num_objs = num_objs;
4180 sinfo->active_slabs = active_slabs;
4181 sinfo->num_slabs = num_slabs;
4182 sinfo->shared_avail = shared_avail;
4183 sinfo->limit = cachep->limit;
4184 sinfo->batchcount = cachep->batchcount;
4185 sinfo->shared = cachep->shared;
4186 sinfo->objects_per_slab = cachep->num;
4187 sinfo->cache_order = cachep->gfporder;
4190 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
4194 unsigned long high = cachep->high_mark;
4195 unsigned long allocs = cachep->num_allocations;
4196 unsigned long grown = cachep->grown;
4197 unsigned long reaped = cachep->reaped;
4198 unsigned long errors = cachep->errors;
4199 unsigned long max_freeable = cachep->max_freeable;
4200 unsigned long node_allocs = cachep->node_allocs;
4201 unsigned long node_frees = cachep->node_frees;
4202 unsigned long overflows = cachep->node_overflow;
4204 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4205 "%4lu %4lu %4lu %4lu %4lu",
4206 allocs, high, grown,
4207 reaped, errors, max_freeable, node_allocs,
4208 node_frees, overflows);
4212 unsigned long allochit = atomic_read(&cachep->allochit);
4213 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4214 unsigned long freehit = atomic_read(&cachep->freehit);
4215 unsigned long freemiss = atomic_read(&cachep->freemiss);
4217 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4218 allochit, allocmiss, freehit, freemiss);
4223 #define MAX_SLABINFO_WRITE 128
4225 * slabinfo_write - Tuning for the slab allocator
4227 * @buffer: user buffer
4228 * @count: data length
4231 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4232 size_t count, loff_t *ppos)
4234 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4235 int limit, batchcount, shared, res;
4236 struct kmem_cache *cachep;
4238 if (count > MAX_SLABINFO_WRITE)
4240 if (copy_from_user(&kbuf, buffer, count))
4242 kbuf[MAX_SLABINFO_WRITE] = '\0';
4244 tmp = strchr(kbuf, ' ');
4249 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4252 /* Find the cache in the chain of caches. */
4253 mutex_lock(&slab_mutex);
4255 list_for_each_entry(cachep, &slab_caches, list) {
4256 if (!strcmp(cachep->name, kbuf)) {
4257 if (limit < 1 || batchcount < 1 ||
4258 batchcount > limit || shared < 0) {
4261 res = do_tune_cpucache(cachep, limit,
4268 mutex_unlock(&slab_mutex);
4274 #ifdef CONFIG_DEBUG_SLAB_LEAK
4276 static void *leaks_start(struct seq_file *m, loff_t *pos)
4278 mutex_lock(&slab_mutex);
4279 return seq_list_start(&slab_caches, *pos);
4282 static inline int add_caller(unsigned long *n, unsigned long v)
4292 unsigned long *q = p + 2 * i;
4306 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4312 static void handle_slab(unsigned long *n, struct kmem_cache *c,
4320 for (i = 0, p = page->s_mem; i < c->num; i++, p += c->size) {
4321 if (get_obj_status(page, i) != OBJECT_ACTIVE)
4324 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4329 static void show_symbol(struct seq_file *m, unsigned long address)
4331 #ifdef CONFIG_KALLSYMS
4332 unsigned long offset, size;
4333 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4335 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4336 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4338 seq_printf(m, " [%s]", modname);
4342 seq_printf(m, "%p", (void *)address);
4345 static int leaks_show(struct seq_file *m, void *p)
4347 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
4349 struct kmem_cache_node *n;
4351 unsigned long *x = m->private;
4355 if (!(cachep->flags & SLAB_STORE_USER))
4357 if (!(cachep->flags & SLAB_RED_ZONE))
4360 /* OK, we can do it */
4364 for_each_kmem_cache_node(cachep, node, n) {
4367 spin_lock_irq(&n->list_lock);
4369 list_for_each_entry(page, &n->slabs_full, lru)
4370 handle_slab(x, cachep, page);
4371 list_for_each_entry(page, &n->slabs_partial, lru)
4372 handle_slab(x, cachep, page);
4373 spin_unlock_irq(&n->list_lock);
4375 name = cachep->name;
4377 /* Increase the buffer size */
4378 mutex_unlock(&slab_mutex);
4379 m->private = kzalloc(x[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4381 /* Too bad, we are really out */
4383 mutex_lock(&slab_mutex);
4386 *(unsigned long *)m->private = x[0] * 2;
4388 mutex_lock(&slab_mutex);
4389 /* Now make sure this entry will be retried */
4393 for (i = 0; i < x[1]; i++) {
4394 seq_printf(m, "%s: %lu ", name, x[2*i+3]);
4395 show_symbol(m, x[2*i+2]);
4402 static const struct seq_operations slabstats_op = {
4403 .start = leaks_start,
4409 static int slabstats_open(struct inode *inode, struct file *file)
4411 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4414 ret = seq_open(file, &slabstats_op);
4416 struct seq_file *m = file->private_data;
4417 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4426 static const struct file_operations proc_slabstats_operations = {
4427 .open = slabstats_open,
4429 .llseek = seq_lseek,
4430 .release = seq_release_private,
4434 static int __init slab_proc_init(void)
4436 #ifdef CONFIG_DEBUG_SLAB_LEAK
4437 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4441 module_init(slab_proc_init);
4445 * ksize - get the actual amount of memory allocated for a given object
4446 * @objp: Pointer to the object
4448 * kmalloc may internally round up allocations and return more memory
4449 * than requested. ksize() can be used to determine the actual amount of
4450 * memory allocated. The caller may use this additional memory, even though
4451 * a smaller amount of memory was initially specified with the kmalloc call.
4452 * The caller must guarantee that objp points to a valid object previously
4453 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4454 * must not be freed during the duration of the call.
4456 size_t ksize(const void *objp)
4459 if (unlikely(objp == ZERO_SIZE_PTR))
4462 return virt_to_cache(objp)->object_size;
4464 EXPORT_SYMBOL(ksize);