3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
28 * This means, that your constructor is used only for newly allocated
29 * slabs and you must pass objects with the same initializations to
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
39 * empty slabs with no allocated objects
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
53 * The c_cpuarray may not be read with enabled local interrupts -
54 * it's changed with a smp_call_function().
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
58 * Several members in struct kmem_cache and struct slab never change, they
59 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
68 * Further notes from the original documentation:
70 * 11 April '97. Started multi-threading - markhe
71 * The global cache-chain is protected by the mutex 'slab_mutex'.
72 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
76 * At present, each engine can be growing a cache. This should be blocked.
78 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
89 #include <linux/slab.h>
91 #include <linux/poison.h>
92 #include <linux/swap.h>
93 #include <linux/cache.h>
94 #include <linux/interrupt.h>
95 #include <linux/init.h>
96 #include <linux/compiler.h>
97 #include <linux/cpuset.h>
98 #include <linux/proc_fs.h>
99 #include <linux/seq_file.h>
100 #include <linux/notifier.h>
101 #include <linux/kallsyms.h>
102 #include <linux/cpu.h>
103 #include <linux/sysctl.h>
104 #include <linux/module.h>
105 #include <linux/rcupdate.h>
106 #include <linux/string.h>
107 #include <linux/uaccess.h>
108 #include <linux/nodemask.h>
109 #include <linux/kmemleak.h>
110 #include <linux/mempolicy.h>
111 #include <linux/mutex.h>
112 #include <linux/fault-inject.h>
113 #include <linux/rtmutex.h>
114 #include <linux/reciprocal_div.h>
115 #include <linux/debugobjects.h>
116 #include <linux/kmemcheck.h>
117 #include <linux/memory.h>
118 #include <linux/prefetch.h>
120 #include <net/sock.h>
122 #include <asm/cacheflush.h>
123 #include <asm/tlbflush.h>
124 #include <asm/page.h>
126 #include <trace/events/kmem.h>
128 #include "internal.h"
133 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
134 * 0 for faster, smaller code (especially in the critical paths).
136 * STATS - 1 to collect stats for /proc/slabinfo.
137 * 0 for faster, smaller code (especially in the critical paths).
139 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
142 #ifdef CONFIG_DEBUG_SLAB
145 #define FORCED_DEBUG 1
149 #define FORCED_DEBUG 0
152 /* Shouldn't this be in a header file somewhere? */
153 #define BYTES_PER_WORD sizeof(void *)
154 #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
156 #ifndef ARCH_KMALLOC_FLAGS
157 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
161 * true if a page was allocated from pfmemalloc reserves for network-based
164 static bool pfmemalloc_active __read_mostly;
169 * Bufctl's are used for linking objs within a slab
172 * This implementation relies on "struct page" for locating the cache &
173 * slab an object belongs to.
174 * This allows the bufctl structure to be small (one int), but limits
175 * the number of objects a slab (not a cache) can contain when off-slab
176 * bufctls are used. The limit is the size of the largest general cache
177 * that does not use off-slab slabs.
178 * For 32bit archs with 4 kB pages, is this 56.
179 * This is not serious, as it is only for large objects, when it is unwise
180 * to have too many per slab.
181 * Note: This limit can be raised by introducing a general cache whose size
182 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
185 typedef unsigned int kmem_bufctl_t;
186 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
187 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
188 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
189 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
194 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
195 * arrange for kmem_freepages to be called via RCU. This is useful if
196 * we need to approach a kernel structure obliquely, from its address
197 * obtained without the usual locking. We can lock the structure to
198 * stabilize it and check it's still at the given address, only if we
199 * can be sure that the memory has not been meanwhile reused for some
200 * other kind of object (which our subsystem's lock might corrupt).
202 * rcu_read_lock before reading the address, then rcu_read_unlock after
203 * taking the spinlock within the structure expected at that address.
206 struct rcu_head head;
213 * Manages the objs in a slab. Placed either at the beginning of mem allocated
214 * for a slab, or allocated from an general cache.
215 * Slabs are chained into three list: fully used, partial, fully free slabs.
220 struct list_head list;
221 void *s_mem; /* including colour offset */
222 unsigned int inuse; /* num of objs active in slab */
225 struct slab_rcu __slab_cover_slab_rcu;
233 * - LIFO ordering, to hand out cache-warm objects from _alloc
234 * - reduce the number of linked list operations
235 * - reduce spinlock operations
237 * The limit is stored in the per-cpu structure to reduce the data cache
244 unsigned int batchcount;
245 unsigned int touched;
248 * Must have this definition in here for the proper
249 * alignment of array_cache. Also simplifies accessing
252 * Entries should not be directly dereferenced as
253 * entries belonging to slabs marked pfmemalloc will
254 * have the lower bits set SLAB_OBJ_PFMEMALLOC
258 #define SLAB_OBJ_PFMEMALLOC 1
259 static inline bool is_obj_pfmemalloc(void *objp)
261 return (unsigned long)objp & SLAB_OBJ_PFMEMALLOC;
264 static inline void set_obj_pfmemalloc(void **objp)
266 *objp = (void *)((unsigned long)*objp | SLAB_OBJ_PFMEMALLOC);
270 static inline void clear_obj_pfmemalloc(void **objp)
272 *objp = (void *)((unsigned long)*objp & ~SLAB_OBJ_PFMEMALLOC);
276 * bootstrap: The caches do not work without cpuarrays anymore, but the
277 * cpuarrays are allocated from the generic caches...
279 #define BOOT_CPUCACHE_ENTRIES 1
280 struct arraycache_init {
281 struct array_cache cache;
282 void *entries[BOOT_CPUCACHE_ENTRIES];
286 * Need this for bootstrapping a per node allocator.
288 #define NUM_INIT_LISTS (3 * MAX_NUMNODES)
289 static struct kmem_cache_node __initdata init_kmem_cache_node[NUM_INIT_LISTS];
290 #define CACHE_CACHE 0
291 #define SIZE_AC MAX_NUMNODES
292 #define SIZE_NODE (2 * MAX_NUMNODES)
294 static int drain_freelist(struct kmem_cache *cache,
295 struct kmem_cache_node *n, int tofree);
296 static void free_block(struct kmem_cache *cachep, void **objpp, int len,
298 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
299 static void cache_reap(struct work_struct *unused);
301 static int slab_early_init = 1;
303 #define INDEX_AC kmalloc_index(sizeof(struct arraycache_init))
304 #define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
306 static void kmem_cache_node_init(struct kmem_cache_node *parent)
308 INIT_LIST_HEAD(&parent->slabs_full);
309 INIT_LIST_HEAD(&parent->slabs_partial);
310 INIT_LIST_HEAD(&parent->slabs_free);
311 parent->shared = NULL;
312 parent->alien = NULL;
313 parent->colour_next = 0;
314 spin_lock_init(&parent->list_lock);
315 parent->free_objects = 0;
316 parent->free_touched = 0;
319 #define MAKE_LIST(cachep, listp, slab, nodeid) \
321 INIT_LIST_HEAD(listp); \
322 list_splice(&(cachep->node[nodeid]->slab), listp); \
325 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
327 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
328 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
329 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
332 #define CFLGS_OFF_SLAB (0x80000000UL)
333 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
335 #define BATCHREFILL_LIMIT 16
337 * Optimization question: fewer reaps means less probability for unnessary
338 * cpucache drain/refill cycles.
340 * OTOH the cpuarrays can contain lots of objects,
341 * which could lock up otherwise freeable slabs.
343 #define REAPTIMEOUT_CPUC (2*HZ)
344 #define REAPTIMEOUT_LIST3 (4*HZ)
347 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
348 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
349 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
350 #define STATS_INC_GROWN(x) ((x)->grown++)
351 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
352 #define STATS_SET_HIGH(x) \
354 if ((x)->num_active > (x)->high_mark) \
355 (x)->high_mark = (x)->num_active; \
357 #define STATS_INC_ERR(x) ((x)->errors++)
358 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
359 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
360 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
361 #define STATS_SET_FREEABLE(x, i) \
363 if ((x)->max_freeable < i) \
364 (x)->max_freeable = i; \
366 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
367 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
368 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
369 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
371 #define STATS_INC_ACTIVE(x) do { } while (0)
372 #define STATS_DEC_ACTIVE(x) do { } while (0)
373 #define STATS_INC_ALLOCED(x) do { } while (0)
374 #define STATS_INC_GROWN(x) do { } while (0)
375 #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
376 #define STATS_SET_HIGH(x) do { } while (0)
377 #define STATS_INC_ERR(x) do { } while (0)
378 #define STATS_INC_NODEALLOCS(x) do { } while (0)
379 #define STATS_INC_NODEFREES(x) do { } while (0)
380 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
381 #define STATS_SET_FREEABLE(x, i) do { } while (0)
382 #define STATS_INC_ALLOCHIT(x) do { } while (0)
383 #define STATS_INC_ALLOCMISS(x) do { } while (0)
384 #define STATS_INC_FREEHIT(x) do { } while (0)
385 #define STATS_INC_FREEMISS(x) do { } while (0)
391 * memory layout of objects:
393 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
394 * the end of an object is aligned with the end of the real
395 * allocation. Catches writes behind the end of the allocation.
396 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
398 * cachep->obj_offset: The real object.
399 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
400 * cachep->size - 1* BYTES_PER_WORD: last caller address
401 * [BYTES_PER_WORD long]
403 static int obj_offset(struct kmem_cache *cachep)
405 return cachep->obj_offset;
408 static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
410 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
411 return (unsigned long long*) (objp + obj_offset(cachep) -
412 sizeof(unsigned long long));
415 static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
417 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
418 if (cachep->flags & SLAB_STORE_USER)
419 return (unsigned long long *)(objp + cachep->size -
420 sizeof(unsigned long long) -
422 return (unsigned long long *) (objp + cachep->size -
423 sizeof(unsigned long long));
426 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
428 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
429 return (void **)(objp + cachep->size - BYTES_PER_WORD);
434 #define obj_offset(x) 0
435 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
436 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
437 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
442 * Do not go above this order unless 0 objects fit into the slab or
443 * overridden on the command line.
445 #define SLAB_MAX_ORDER_HI 1
446 #define SLAB_MAX_ORDER_LO 0
447 static int slab_max_order = SLAB_MAX_ORDER_LO;
448 static bool slab_max_order_set __initdata;
450 static inline struct kmem_cache *virt_to_cache(const void *obj)
452 struct page *page = virt_to_head_page(obj);
453 return page->slab_cache;
456 static inline struct slab *virt_to_slab(const void *obj)
458 struct page *page = virt_to_head_page(obj);
460 VM_BUG_ON(!PageSlab(page));
461 return page->slab_page;
464 static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
467 return slab->s_mem + cache->size * idx;
471 * We want to avoid an expensive divide : (offset / cache->size)
472 * Using the fact that size is a constant for a particular cache,
473 * we can replace (offset / cache->size) by
474 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
476 static inline unsigned int obj_to_index(const struct kmem_cache *cache,
477 const struct slab *slab, void *obj)
479 u32 offset = (obj - slab->s_mem);
480 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
483 static struct arraycache_init initarray_generic =
484 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
486 /* internal cache of cache description objs */
487 static struct kmem_cache kmem_cache_boot = {
489 .limit = BOOT_CPUCACHE_ENTRIES,
491 .size = sizeof(struct kmem_cache),
492 .name = "kmem_cache",
495 #define BAD_ALIEN_MAGIC 0x01020304ul
497 #ifdef CONFIG_LOCKDEP
500 * Slab sometimes uses the kmalloc slabs to store the slab headers
501 * for other slabs "off slab".
502 * The locking for this is tricky in that it nests within the locks
503 * of all other slabs in a few places; to deal with this special
504 * locking we put on-slab caches into a separate lock-class.
506 * We set lock class for alien array caches which are up during init.
507 * The lock annotation will be lost if all cpus of a node goes down and
508 * then comes back up during hotplug
510 static struct lock_class_key on_slab_l3_key;
511 static struct lock_class_key on_slab_alc_key;
513 static struct lock_class_key debugobj_l3_key;
514 static struct lock_class_key debugobj_alc_key;
516 static void slab_set_lock_classes(struct kmem_cache *cachep,
517 struct lock_class_key *l3_key, struct lock_class_key *alc_key,
520 struct array_cache **alc;
521 struct kmem_cache_node *n;
528 lockdep_set_class(&n->list_lock, l3_key);
531 * FIXME: This check for BAD_ALIEN_MAGIC
532 * should go away when common slab code is taught to
533 * work even without alien caches.
534 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
535 * for alloc_alien_cache,
537 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
541 lockdep_set_class(&alc[r]->lock, alc_key);
545 static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
547 slab_set_lock_classes(cachep, &debugobj_l3_key, &debugobj_alc_key, node);
550 static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
554 for_each_online_node(node)
555 slab_set_debugobj_lock_classes_node(cachep, node);
558 static void init_node_lock_keys(int q)
565 for (i = 1; i <= KMALLOC_SHIFT_HIGH; i++) {
566 struct kmem_cache_node *n;
567 struct kmem_cache *cache = kmalloc_caches[i];
573 if (!n || OFF_SLAB(cache))
576 slab_set_lock_classes(cache, &on_slab_l3_key,
577 &on_slab_alc_key, q);
581 static void on_slab_lock_classes_node(struct kmem_cache *cachep, int q)
583 if (!cachep->node[q])
586 slab_set_lock_classes(cachep, &on_slab_l3_key,
587 &on_slab_alc_key, q);
590 static inline void on_slab_lock_classes(struct kmem_cache *cachep)
594 VM_BUG_ON(OFF_SLAB(cachep));
596 on_slab_lock_classes_node(cachep, node);
599 static inline void init_lock_keys(void)
604 init_node_lock_keys(node);
607 static void init_node_lock_keys(int q)
611 static inline void init_lock_keys(void)
615 static inline void on_slab_lock_classes(struct kmem_cache *cachep)
619 static inline void on_slab_lock_classes_node(struct kmem_cache *cachep, int node)
623 static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
627 static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
632 static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
634 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
636 return cachep->array[smp_processor_id()];
639 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
641 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
645 * Calculate the number of objects and left-over bytes for a given buffer size.
647 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
648 size_t align, int flags, size_t *left_over,
653 size_t slab_size = PAGE_SIZE << gfporder;
656 * The slab management structure can be either off the slab or
657 * on it. For the latter case, the memory allocated for a
661 * - One kmem_bufctl_t for each object
662 * - Padding to respect alignment of @align
663 * - @buffer_size bytes for each object
665 * If the slab management structure is off the slab, then the
666 * alignment will already be calculated into the size. Because
667 * the slabs are all pages aligned, the objects will be at the
668 * correct alignment when allocated.
670 if (flags & CFLGS_OFF_SLAB) {
672 nr_objs = slab_size / buffer_size;
674 if (nr_objs > SLAB_LIMIT)
675 nr_objs = SLAB_LIMIT;
678 * Ignore padding for the initial guess. The padding
679 * is at most @align-1 bytes, and @buffer_size is at
680 * least @align. In the worst case, this result will
681 * be one greater than the number of objects that fit
682 * into the memory allocation when taking the padding
685 nr_objs = (slab_size - sizeof(struct slab)) /
686 (buffer_size + sizeof(kmem_bufctl_t));
689 * This calculated number will be either the right
690 * amount, or one greater than what we want.
692 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
696 if (nr_objs > SLAB_LIMIT)
697 nr_objs = SLAB_LIMIT;
699 mgmt_size = slab_mgmt_size(nr_objs, align);
702 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
706 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
708 static void __slab_error(const char *function, struct kmem_cache *cachep,
711 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
712 function, cachep->name, msg);
714 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
719 * By default on NUMA we use alien caches to stage the freeing of
720 * objects allocated from other nodes. This causes massive memory
721 * inefficiencies when using fake NUMA setup to split memory into a
722 * large number of small nodes, so it can be disabled on the command
726 static int use_alien_caches __read_mostly = 1;
727 static int __init noaliencache_setup(char *s)
729 use_alien_caches = 0;
732 __setup("noaliencache", noaliencache_setup);
734 static int __init slab_max_order_setup(char *str)
736 get_option(&str, &slab_max_order);
737 slab_max_order = slab_max_order < 0 ? 0 :
738 min(slab_max_order, MAX_ORDER - 1);
739 slab_max_order_set = true;
743 __setup("slab_max_order=", slab_max_order_setup);
747 * Special reaping functions for NUMA systems called from cache_reap().
748 * These take care of doing round robin flushing of alien caches (containing
749 * objects freed on different nodes from which they were allocated) and the
750 * flushing of remote pcps by calling drain_node_pages.
752 static DEFINE_PER_CPU(unsigned long, slab_reap_node);
754 static void init_reap_node(int cpu)
758 node = next_node(cpu_to_mem(cpu), node_online_map);
759 if (node == MAX_NUMNODES)
760 node = first_node(node_online_map);
762 per_cpu(slab_reap_node, cpu) = node;
765 static void next_reap_node(void)
767 int node = __this_cpu_read(slab_reap_node);
769 node = next_node(node, node_online_map);
770 if (unlikely(node >= MAX_NUMNODES))
771 node = first_node(node_online_map);
772 __this_cpu_write(slab_reap_node, node);
776 #define init_reap_node(cpu) do { } while (0)
777 #define next_reap_node(void) do { } while (0)
781 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
782 * via the workqueue/eventd.
783 * Add the CPU number into the expiration time to minimize the possibility of
784 * the CPUs getting into lockstep and contending for the global cache chain
787 static void start_cpu_timer(int cpu)
789 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
792 * When this gets called from do_initcalls via cpucache_init(),
793 * init_workqueues() has already run, so keventd will be setup
796 if (keventd_up() && reap_work->work.func == NULL) {
798 INIT_DEFERRABLE_WORK(reap_work, cache_reap);
799 schedule_delayed_work_on(cpu, reap_work,
800 __round_jiffies_relative(HZ, cpu));
804 static struct array_cache *alloc_arraycache(int node, int entries,
805 int batchcount, gfp_t gfp)
807 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
808 struct array_cache *nc = NULL;
810 nc = kmalloc_node(memsize, gfp, node);
812 * The array_cache structures contain pointers to free object.
813 * However, when such objects are allocated or transferred to another
814 * cache the pointers are not cleared and they could be counted as
815 * valid references during a kmemleak scan. Therefore, kmemleak must
816 * not scan such objects.
818 kmemleak_no_scan(nc);
822 nc->batchcount = batchcount;
824 spin_lock_init(&nc->lock);
829 static inline bool is_slab_pfmemalloc(struct slab *slabp)
831 struct page *page = virt_to_page(slabp->s_mem);
833 return PageSlabPfmemalloc(page);
836 /* Clears pfmemalloc_active if no slabs have pfmalloc set */
837 static void recheck_pfmemalloc_active(struct kmem_cache *cachep,
838 struct array_cache *ac)
840 struct kmem_cache_node *n = cachep->node[numa_mem_id()];
844 if (!pfmemalloc_active)
847 spin_lock_irqsave(&n->list_lock, flags);
848 list_for_each_entry(slabp, &n->slabs_full, list)
849 if (is_slab_pfmemalloc(slabp))
852 list_for_each_entry(slabp, &n->slabs_partial, list)
853 if (is_slab_pfmemalloc(slabp))
856 list_for_each_entry(slabp, &n->slabs_free, list)
857 if (is_slab_pfmemalloc(slabp))
860 pfmemalloc_active = false;
862 spin_unlock_irqrestore(&n->list_lock, flags);
865 static void *__ac_get_obj(struct kmem_cache *cachep, struct array_cache *ac,
866 gfp_t flags, bool force_refill)
869 void *objp = ac->entry[--ac->avail];
871 /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
872 if (unlikely(is_obj_pfmemalloc(objp))) {
873 struct kmem_cache_node *n;
875 if (gfp_pfmemalloc_allowed(flags)) {
876 clear_obj_pfmemalloc(&objp);
880 /* The caller cannot use PFMEMALLOC objects, find another one */
881 for (i = 0; i < ac->avail; i++) {
882 /* If a !PFMEMALLOC object is found, swap them */
883 if (!is_obj_pfmemalloc(ac->entry[i])) {
885 ac->entry[i] = ac->entry[ac->avail];
886 ac->entry[ac->avail] = objp;
892 * If there are empty slabs on the slabs_free list and we are
893 * being forced to refill the cache, mark this one !pfmemalloc.
895 n = cachep->node[numa_mem_id()];
896 if (!list_empty(&n->slabs_free) && force_refill) {
897 struct slab *slabp = virt_to_slab(objp);
898 ClearPageSlabPfmemalloc(virt_to_head_page(slabp->s_mem));
899 clear_obj_pfmemalloc(&objp);
900 recheck_pfmemalloc_active(cachep, ac);
904 /* No !PFMEMALLOC objects available */
912 static inline void *ac_get_obj(struct kmem_cache *cachep,
913 struct array_cache *ac, gfp_t flags, bool force_refill)
917 if (unlikely(sk_memalloc_socks()))
918 objp = __ac_get_obj(cachep, ac, flags, force_refill);
920 objp = ac->entry[--ac->avail];
925 static void *__ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
928 if (unlikely(pfmemalloc_active)) {
929 /* Some pfmemalloc slabs exist, check if this is one */
930 struct slab *slabp = virt_to_slab(objp);
931 struct page *page = virt_to_head_page(slabp->s_mem);
932 if (PageSlabPfmemalloc(page))
933 set_obj_pfmemalloc(&objp);
939 static inline void ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
942 if (unlikely(sk_memalloc_socks()))
943 objp = __ac_put_obj(cachep, ac, objp);
945 ac->entry[ac->avail++] = objp;
949 * Transfer objects in one arraycache to another.
950 * Locking must be handled by the caller.
952 * Return the number of entries transferred.
954 static int transfer_objects(struct array_cache *to,
955 struct array_cache *from, unsigned int max)
957 /* Figure out how many entries to transfer */
958 int nr = min3(from->avail, max, to->limit - to->avail);
963 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
973 #define drain_alien_cache(cachep, alien) do { } while (0)
974 #define reap_alien(cachep, n) do { } while (0)
976 static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
978 return (struct array_cache **)BAD_ALIEN_MAGIC;
981 static inline void free_alien_cache(struct array_cache **ac_ptr)
985 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
990 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
996 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
997 gfp_t flags, int nodeid)
1002 #else /* CONFIG_NUMA */
1004 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
1005 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
1007 static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
1009 struct array_cache **ac_ptr;
1010 int memsize = sizeof(void *) * nr_node_ids;
1015 ac_ptr = kzalloc_node(memsize, gfp, node);
1018 if (i == node || !node_online(i))
1020 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
1022 for (i--; i >= 0; i--)
1032 static void free_alien_cache(struct array_cache **ac_ptr)
1043 static void __drain_alien_cache(struct kmem_cache *cachep,
1044 struct array_cache *ac, int node)
1046 struct kmem_cache_node *n = cachep->node[node];
1049 spin_lock(&n->list_lock);
1051 * Stuff objects into the remote nodes shared array first.
1052 * That way we could avoid the overhead of putting the objects
1053 * into the free lists and getting them back later.
1056 transfer_objects(n->shared, ac, ac->limit);
1058 free_block(cachep, ac->entry, ac->avail, node);
1060 spin_unlock(&n->list_lock);
1065 * Called from cache_reap() to regularly drain alien caches round robin.
1067 static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *n)
1069 int node = __this_cpu_read(slab_reap_node);
1072 struct array_cache *ac = n->alien[node];
1074 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1075 __drain_alien_cache(cachep, ac, node);
1076 spin_unlock_irq(&ac->lock);
1081 static void drain_alien_cache(struct kmem_cache *cachep,
1082 struct array_cache **alien)
1085 struct array_cache *ac;
1086 unsigned long flags;
1088 for_each_online_node(i) {
1091 spin_lock_irqsave(&ac->lock, flags);
1092 __drain_alien_cache(cachep, ac, i);
1093 spin_unlock_irqrestore(&ac->lock, flags);
1098 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1100 int nodeid = page_to_nid(virt_to_page(objp));
1101 struct kmem_cache_node *n;
1102 struct array_cache *alien = NULL;
1105 node = numa_mem_id();
1108 * Make sure we are not freeing a object from another node to the array
1109 * cache on this cpu.
1111 if (likely(nodeid == node))
1114 n = cachep->node[node];
1115 STATS_INC_NODEFREES(cachep);
1116 if (n->alien && n->alien[nodeid]) {
1117 alien = n->alien[nodeid];
1118 spin_lock(&alien->lock);
1119 if (unlikely(alien->avail == alien->limit)) {
1120 STATS_INC_ACOVERFLOW(cachep);
1121 __drain_alien_cache(cachep, alien, nodeid);
1123 ac_put_obj(cachep, alien, objp);
1124 spin_unlock(&alien->lock);
1126 spin_lock(&(cachep->node[nodeid])->list_lock);
1127 free_block(cachep, &objp, 1, nodeid);
1128 spin_unlock(&(cachep->node[nodeid])->list_lock);
1135 * Allocates and initializes node for a node on each slab cache, used for
1136 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
1137 * will be allocated off-node since memory is not yet online for the new node.
1138 * When hotplugging memory or a cpu, existing node are not replaced if
1141 * Must hold slab_mutex.
1143 static int init_cache_node_node(int node)
1145 struct kmem_cache *cachep;
1146 struct kmem_cache_node *n;
1147 const int memsize = sizeof(struct kmem_cache_node);
1149 list_for_each_entry(cachep, &slab_caches, list) {
1151 * Set up the size64 kmemlist for cpu before we can
1152 * begin anything. Make sure some other cpu on this
1153 * node has not already allocated this
1155 if (!cachep->node[node]) {
1156 n = kmalloc_node(memsize, GFP_KERNEL, node);
1159 kmem_cache_node_init(n);
1160 n->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1161 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1164 * The l3s don't come and go as CPUs come and
1165 * go. slab_mutex is sufficient
1168 cachep->node[node] = n;
1171 spin_lock_irq(&cachep->node[node]->list_lock);
1172 cachep->node[node]->free_limit =
1173 (1 + nr_cpus_node(node)) *
1174 cachep->batchcount + cachep->num;
1175 spin_unlock_irq(&cachep->node[node]->list_lock);
1180 static inline int slabs_tofree(struct kmem_cache *cachep,
1181 struct kmem_cache_node *n)
1183 return (n->free_objects + cachep->num - 1) / cachep->num;
1186 static void cpuup_canceled(long cpu)
1188 struct kmem_cache *cachep;
1189 struct kmem_cache_node *n = NULL;
1190 int node = cpu_to_mem(cpu);
1191 const struct cpumask *mask = cpumask_of_node(node);
1193 list_for_each_entry(cachep, &slab_caches, list) {
1194 struct array_cache *nc;
1195 struct array_cache *shared;
1196 struct array_cache **alien;
1198 /* cpu is dead; no one can alloc from it. */
1199 nc = cachep->array[cpu];
1200 cachep->array[cpu] = NULL;
1201 n = cachep->node[node];
1204 goto free_array_cache;
1206 spin_lock_irq(&n->list_lock);
1208 /* Free limit for this kmem_cache_node */
1209 n->free_limit -= cachep->batchcount;
1211 free_block(cachep, nc->entry, nc->avail, node);
1213 if (!cpumask_empty(mask)) {
1214 spin_unlock_irq(&n->list_lock);
1215 goto free_array_cache;
1220 free_block(cachep, shared->entry,
1221 shared->avail, node);
1228 spin_unlock_irq(&n->list_lock);
1232 drain_alien_cache(cachep, alien);
1233 free_alien_cache(alien);
1239 * In the previous loop, all the objects were freed to
1240 * the respective cache's slabs, now we can go ahead and
1241 * shrink each nodelist to its limit.
1243 list_for_each_entry(cachep, &slab_caches, list) {
1244 n = cachep->node[node];
1247 drain_freelist(cachep, n, slabs_tofree(cachep, n));
1251 static int cpuup_prepare(long cpu)
1253 struct kmem_cache *cachep;
1254 struct kmem_cache_node *n = NULL;
1255 int node = cpu_to_mem(cpu);
1259 * We need to do this right in the beginning since
1260 * alloc_arraycache's are going to use this list.
1261 * kmalloc_node allows us to add the slab to the right
1262 * kmem_cache_node and not this cpu's kmem_cache_node
1264 err = init_cache_node_node(node);
1269 * Now we can go ahead with allocating the shared arrays and
1272 list_for_each_entry(cachep, &slab_caches, list) {
1273 struct array_cache *nc;
1274 struct array_cache *shared = NULL;
1275 struct array_cache **alien = NULL;
1277 nc = alloc_arraycache(node, cachep->limit,
1278 cachep->batchcount, GFP_KERNEL);
1281 if (cachep->shared) {
1282 shared = alloc_arraycache(node,
1283 cachep->shared * cachep->batchcount,
1284 0xbaadf00d, GFP_KERNEL);
1290 if (use_alien_caches) {
1291 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
1298 cachep->array[cpu] = nc;
1299 n = cachep->node[node];
1302 spin_lock_irq(&n->list_lock);
1305 * We are serialised from CPU_DEAD or
1306 * CPU_UP_CANCELLED by the cpucontrol lock
1317 spin_unlock_irq(&n->list_lock);
1319 free_alien_cache(alien);
1320 if (cachep->flags & SLAB_DEBUG_OBJECTS)
1321 slab_set_debugobj_lock_classes_node(cachep, node);
1322 else if (!OFF_SLAB(cachep) &&
1323 !(cachep->flags & SLAB_DESTROY_BY_RCU))
1324 on_slab_lock_classes_node(cachep, node);
1326 init_node_lock_keys(node);
1330 cpuup_canceled(cpu);
1334 static int cpuup_callback(struct notifier_block *nfb,
1335 unsigned long action, void *hcpu)
1337 long cpu = (long)hcpu;
1341 case CPU_UP_PREPARE:
1342 case CPU_UP_PREPARE_FROZEN:
1343 mutex_lock(&slab_mutex);
1344 err = cpuup_prepare(cpu);
1345 mutex_unlock(&slab_mutex);
1348 case CPU_ONLINE_FROZEN:
1349 start_cpu_timer(cpu);
1351 #ifdef CONFIG_HOTPLUG_CPU
1352 case CPU_DOWN_PREPARE:
1353 case CPU_DOWN_PREPARE_FROZEN:
1355 * Shutdown cache reaper. Note that the slab_mutex is
1356 * held so that if cache_reap() is invoked it cannot do
1357 * anything expensive but will only modify reap_work
1358 * and reschedule the timer.
1360 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1361 /* Now the cache_reaper is guaranteed to be not running. */
1362 per_cpu(slab_reap_work, cpu).work.func = NULL;
1364 case CPU_DOWN_FAILED:
1365 case CPU_DOWN_FAILED_FROZEN:
1366 start_cpu_timer(cpu);
1369 case CPU_DEAD_FROZEN:
1371 * Even if all the cpus of a node are down, we don't free the
1372 * kmem_cache_node of any cache. This to avoid a race between
1373 * cpu_down, and a kmalloc allocation from another cpu for
1374 * memory from the node of the cpu going down. The node
1375 * structure is usually allocated from kmem_cache_create() and
1376 * gets destroyed at kmem_cache_destroy().
1380 case CPU_UP_CANCELED:
1381 case CPU_UP_CANCELED_FROZEN:
1382 mutex_lock(&slab_mutex);
1383 cpuup_canceled(cpu);
1384 mutex_unlock(&slab_mutex);
1387 return notifier_from_errno(err);
1390 static struct notifier_block cpucache_notifier = {
1391 &cpuup_callback, NULL, 0
1394 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1396 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1397 * Returns -EBUSY if all objects cannot be drained so that the node is not
1400 * Must hold slab_mutex.
1402 static int __meminit drain_cache_node_node(int node)
1404 struct kmem_cache *cachep;
1407 list_for_each_entry(cachep, &slab_caches, list) {
1408 struct kmem_cache_node *n;
1410 n = cachep->node[node];
1414 drain_freelist(cachep, n, slabs_tofree(cachep, n));
1416 if (!list_empty(&n->slabs_full) ||
1417 !list_empty(&n->slabs_partial)) {
1425 static int __meminit slab_memory_callback(struct notifier_block *self,
1426 unsigned long action, void *arg)
1428 struct memory_notify *mnb = arg;
1432 nid = mnb->status_change_nid;
1437 case MEM_GOING_ONLINE:
1438 mutex_lock(&slab_mutex);
1439 ret = init_cache_node_node(nid);
1440 mutex_unlock(&slab_mutex);
1442 case MEM_GOING_OFFLINE:
1443 mutex_lock(&slab_mutex);
1444 ret = drain_cache_node_node(nid);
1445 mutex_unlock(&slab_mutex);
1449 case MEM_CANCEL_ONLINE:
1450 case MEM_CANCEL_OFFLINE:
1454 return notifier_from_errno(ret);
1456 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1459 * swap the static kmem_cache_node with kmalloced memory
1461 static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list,
1464 struct kmem_cache_node *ptr;
1466 ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid);
1469 memcpy(ptr, list, sizeof(struct kmem_cache_node));
1471 * Do not assume that spinlocks can be initialized via memcpy:
1473 spin_lock_init(&ptr->list_lock);
1475 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1476 cachep->node[nodeid] = ptr;
1480 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1481 * size of kmem_cache_node.
1483 static void __init set_up_node(struct kmem_cache *cachep, int index)
1487 for_each_online_node(node) {
1488 cachep->node[node] = &init_kmem_cache_node[index + node];
1489 cachep->node[node]->next_reap = jiffies +
1491 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1496 * The memory after the last cpu cache pointer is used for the
1499 static void setup_node_pointer(struct kmem_cache *cachep)
1501 cachep->node = (struct kmem_cache_node **)&cachep->array[nr_cpu_ids];
1505 * Initialisation. Called after the page allocator have been initialised and
1506 * before smp_init().
1508 void __init kmem_cache_init(void)
1512 kmem_cache = &kmem_cache_boot;
1513 setup_node_pointer(kmem_cache);
1515 if (num_possible_nodes() == 1)
1516 use_alien_caches = 0;
1518 for (i = 0; i < NUM_INIT_LISTS; i++)
1519 kmem_cache_node_init(&init_kmem_cache_node[i]);
1521 set_up_node(kmem_cache, CACHE_CACHE);
1524 * Fragmentation resistance on low memory - only use bigger
1525 * page orders on machines with more than 32MB of memory if
1526 * not overridden on the command line.
1528 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
1529 slab_max_order = SLAB_MAX_ORDER_HI;
1531 /* Bootstrap is tricky, because several objects are allocated
1532 * from caches that do not exist yet:
1533 * 1) initialize the kmem_cache cache: it contains the struct
1534 * kmem_cache structures of all caches, except kmem_cache itself:
1535 * kmem_cache is statically allocated.
1536 * Initially an __init data area is used for the head array and the
1537 * kmem_cache_node structures, it's replaced with a kmalloc allocated
1538 * array at the end of the bootstrap.
1539 * 2) Create the first kmalloc cache.
1540 * The struct kmem_cache for the new cache is allocated normally.
1541 * An __init data area is used for the head array.
1542 * 3) Create the remaining kmalloc caches, with minimally sized
1544 * 4) Replace the __init data head arrays for kmem_cache and the first
1545 * kmalloc cache with kmalloc allocated arrays.
1546 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1547 * the other cache's with kmalloc allocated memory.
1548 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1551 /* 1) create the kmem_cache */
1554 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1556 create_boot_cache(kmem_cache, "kmem_cache",
1557 offsetof(struct kmem_cache, array[nr_cpu_ids]) +
1558 nr_node_ids * sizeof(struct kmem_cache_node *),
1559 SLAB_HWCACHE_ALIGN);
1560 list_add(&kmem_cache->list, &slab_caches);
1562 /* 2+3) create the kmalloc caches */
1565 * Initialize the caches that provide memory for the array cache and the
1566 * kmem_cache_node structures first. Without this, further allocations will
1570 kmalloc_caches[INDEX_AC] = create_kmalloc_cache("kmalloc-ac",
1571 kmalloc_size(INDEX_AC), ARCH_KMALLOC_FLAGS);
1573 if (INDEX_AC != INDEX_NODE)
1574 kmalloc_caches[INDEX_NODE] =
1575 create_kmalloc_cache("kmalloc-node",
1576 kmalloc_size(INDEX_NODE), ARCH_KMALLOC_FLAGS);
1578 slab_early_init = 0;
1580 /* 4) Replace the bootstrap head arrays */
1582 struct array_cache *ptr;
1584 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1586 memcpy(ptr, cpu_cache_get(kmem_cache),
1587 sizeof(struct arraycache_init));
1589 * Do not assume that spinlocks can be initialized via memcpy:
1591 spin_lock_init(&ptr->lock);
1593 kmem_cache->array[smp_processor_id()] = ptr;
1595 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1597 BUG_ON(cpu_cache_get(kmalloc_caches[INDEX_AC])
1598 != &initarray_generic.cache);
1599 memcpy(ptr, cpu_cache_get(kmalloc_caches[INDEX_AC]),
1600 sizeof(struct arraycache_init));
1602 * Do not assume that spinlocks can be initialized via memcpy:
1604 spin_lock_init(&ptr->lock);
1606 kmalloc_caches[INDEX_AC]->array[smp_processor_id()] = ptr;
1608 /* 5) Replace the bootstrap kmem_cache_node */
1612 for_each_online_node(nid) {
1613 init_list(kmem_cache, &init_kmem_cache_node[CACHE_CACHE + nid], nid);
1615 init_list(kmalloc_caches[INDEX_AC],
1616 &init_kmem_cache_node[SIZE_AC + nid], nid);
1618 if (INDEX_AC != INDEX_NODE) {
1619 init_list(kmalloc_caches[INDEX_NODE],
1620 &init_kmem_cache_node[SIZE_NODE + nid], nid);
1625 create_kmalloc_caches(ARCH_KMALLOC_FLAGS);
1628 void __init kmem_cache_init_late(void)
1630 struct kmem_cache *cachep;
1634 /* 6) resize the head arrays to their final sizes */
1635 mutex_lock(&slab_mutex);
1636 list_for_each_entry(cachep, &slab_caches, list)
1637 if (enable_cpucache(cachep, GFP_NOWAIT))
1639 mutex_unlock(&slab_mutex);
1641 /* Annotate slab for lockdep -- annotate the malloc caches */
1648 * Register a cpu startup notifier callback that initializes
1649 * cpu_cache_get for all new cpus
1651 register_cpu_notifier(&cpucache_notifier);
1655 * Register a memory hotplug callback that initializes and frees
1658 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1662 * The reap timers are started later, with a module init call: That part
1663 * of the kernel is not yet operational.
1667 static int __init cpucache_init(void)
1672 * Register the timers that return unneeded pages to the page allocator
1674 for_each_online_cpu(cpu)
1675 start_cpu_timer(cpu);
1681 __initcall(cpucache_init);
1683 static noinline void
1684 slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1686 struct kmem_cache_node *n;
1688 unsigned long flags;
1692 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1694 printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n",
1695 cachep->name, cachep->size, cachep->gfporder);
1697 for_each_online_node(node) {
1698 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1699 unsigned long active_slabs = 0, num_slabs = 0;
1701 n = cachep->node[node];
1705 spin_lock_irqsave(&n->list_lock, flags);
1706 list_for_each_entry(slabp, &n->slabs_full, list) {
1707 active_objs += cachep->num;
1710 list_for_each_entry(slabp, &n->slabs_partial, list) {
1711 active_objs += slabp->inuse;
1714 list_for_each_entry(slabp, &n->slabs_free, list)
1717 free_objects += n->free_objects;
1718 spin_unlock_irqrestore(&n->list_lock, flags);
1720 num_slabs += active_slabs;
1721 num_objs = num_slabs * cachep->num;
1723 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1724 node, active_slabs, num_slabs, active_objs, num_objs,
1730 * Interface to system's page allocator. No need to hold the cache-lock.
1732 * If we requested dmaable memory, we will get it. Even if we
1733 * did not request dmaable memory, we might get it, but that
1734 * would be relatively rare and ignorable.
1736 static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags,
1745 * Nommu uses slab's for process anonymous memory allocations, and thus
1746 * requires __GFP_COMP to properly refcount higher order allocations
1748 flags |= __GFP_COMP;
1751 flags |= cachep->allocflags;
1752 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1753 flags |= __GFP_RECLAIMABLE;
1755 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1757 if (!(flags & __GFP_NOWARN) && printk_ratelimit())
1758 slab_out_of_memory(cachep, flags, nodeid);
1762 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1763 if (unlikely(page->pfmemalloc))
1764 pfmemalloc_active = true;
1766 nr_pages = (1 << cachep->gfporder);
1767 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1768 add_zone_page_state(page_zone(page),
1769 NR_SLAB_RECLAIMABLE, nr_pages);
1771 add_zone_page_state(page_zone(page),
1772 NR_SLAB_UNRECLAIMABLE, nr_pages);
1773 for (i = 0; i < nr_pages; i++) {
1774 __SetPageSlab(page + i);
1776 if (page->pfmemalloc)
1777 SetPageSlabPfmemalloc(page);
1779 memcg_bind_pages(cachep, cachep->gfporder);
1781 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1782 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1785 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1787 kmemcheck_mark_unallocated_pages(page, nr_pages);
1794 * Interface to system's page release.
1796 static void kmem_freepages(struct kmem_cache *cachep, struct page *page)
1798 unsigned long i = (1 << cachep->gfporder);
1799 const unsigned long nr_freed = i;
1801 kmemcheck_free_shadow(page, cachep->gfporder);
1803 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1804 sub_zone_page_state(page_zone(page),
1805 NR_SLAB_RECLAIMABLE, nr_freed);
1807 sub_zone_page_state(page_zone(page),
1808 NR_SLAB_UNRECLAIMABLE, nr_freed);
1810 __ClearPageSlabPfmemalloc(page);
1812 BUG_ON(!PageSlab(page));
1813 __ClearPageSlab(page);
1817 memcg_release_pages(cachep, cachep->gfporder);
1818 if (current->reclaim_state)
1819 current->reclaim_state->reclaimed_slab += nr_freed;
1820 __free_memcg_kmem_pages(page, cachep->gfporder);
1823 static void kmem_rcu_free(struct rcu_head *head)
1825 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1826 struct kmem_cache *cachep = slab_rcu->page->slab_cache;
1828 kmem_freepages(cachep, slab_rcu->page);
1829 if (OFF_SLAB(cachep))
1830 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1835 #ifdef CONFIG_DEBUG_PAGEALLOC
1836 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1837 unsigned long caller)
1839 int size = cachep->object_size;
1841 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1843 if (size < 5 * sizeof(unsigned long))
1846 *addr++ = 0x12345678;
1848 *addr++ = smp_processor_id();
1849 size -= 3 * sizeof(unsigned long);
1851 unsigned long *sptr = &caller;
1852 unsigned long svalue;
1854 while (!kstack_end(sptr)) {
1856 if (kernel_text_address(svalue)) {
1858 size -= sizeof(unsigned long);
1859 if (size <= sizeof(unsigned long))
1865 *addr++ = 0x87654321;
1869 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1871 int size = cachep->object_size;
1872 addr = &((char *)addr)[obj_offset(cachep)];
1874 memset(addr, val, size);
1875 *(unsigned char *)(addr + size - 1) = POISON_END;
1878 static void dump_line(char *data, int offset, int limit)
1881 unsigned char error = 0;
1884 printk(KERN_ERR "%03x: ", offset);
1885 for (i = 0; i < limit; i++) {
1886 if (data[offset + i] != POISON_FREE) {
1887 error = data[offset + i];
1891 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1892 &data[offset], limit, 1);
1894 if (bad_count == 1) {
1895 error ^= POISON_FREE;
1896 if (!(error & (error - 1))) {
1897 printk(KERN_ERR "Single bit error detected. Probably "
1900 printk(KERN_ERR "Run memtest86+ or a similar memory "
1903 printk(KERN_ERR "Run a memory test tool.\n");
1912 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1917 if (cachep->flags & SLAB_RED_ZONE) {
1918 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
1919 *dbg_redzone1(cachep, objp),
1920 *dbg_redzone2(cachep, objp));
1923 if (cachep->flags & SLAB_STORE_USER) {
1924 printk(KERN_ERR "Last user: [<%p>](%pSR)\n",
1925 *dbg_userword(cachep, objp),
1926 *dbg_userword(cachep, objp));
1928 realobj = (char *)objp + obj_offset(cachep);
1929 size = cachep->object_size;
1930 for (i = 0; i < size && lines; i += 16, lines--) {
1933 if (i + limit > size)
1935 dump_line(realobj, i, limit);
1939 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1945 realobj = (char *)objp + obj_offset(cachep);
1946 size = cachep->object_size;
1948 for (i = 0; i < size; i++) {
1949 char exp = POISON_FREE;
1952 if (realobj[i] != exp) {
1958 "Slab corruption (%s): %s start=%p, len=%d\n",
1959 print_tainted(), cachep->name, realobj, size);
1960 print_objinfo(cachep, objp, 0);
1962 /* Hexdump the affected line */
1965 if (i + limit > size)
1967 dump_line(realobj, i, limit);
1970 /* Limit to 5 lines */
1976 /* Print some data about the neighboring objects, if they
1979 struct slab *slabp = virt_to_slab(objp);
1982 objnr = obj_to_index(cachep, slabp, objp);
1984 objp = index_to_obj(cachep, slabp, objnr - 1);
1985 realobj = (char *)objp + obj_offset(cachep);
1986 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1988 print_objinfo(cachep, objp, 2);
1990 if (objnr + 1 < cachep->num) {
1991 objp = index_to_obj(cachep, slabp, objnr + 1);
1992 realobj = (char *)objp + obj_offset(cachep);
1993 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1995 print_objinfo(cachep, objp, 2);
2002 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
2005 for (i = 0; i < cachep->num; i++) {
2006 void *objp = index_to_obj(cachep, slabp, i);
2008 if (cachep->flags & SLAB_POISON) {
2009 #ifdef CONFIG_DEBUG_PAGEALLOC
2010 if (cachep->size % PAGE_SIZE == 0 &&
2012 kernel_map_pages(virt_to_page(objp),
2013 cachep->size / PAGE_SIZE, 1);
2015 check_poison_obj(cachep, objp);
2017 check_poison_obj(cachep, objp);
2020 if (cachep->flags & SLAB_RED_ZONE) {
2021 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2022 slab_error(cachep, "start of a freed object "
2024 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2025 slab_error(cachep, "end of a freed object "
2031 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
2037 * slab_destroy - destroy and release all objects in a slab
2038 * @cachep: cache pointer being destroyed
2039 * @slabp: slab pointer being destroyed
2041 * Destroy all the objs in a slab, and release the mem back to the system.
2042 * Before calling the slab must have been unlinked from the cache. The
2043 * cache-lock is not held/needed.
2045 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
2047 struct page *page = virt_to_head_page(slabp->s_mem);
2049 slab_destroy_debugcheck(cachep, slabp);
2050 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
2051 struct slab_rcu *slab_rcu;
2053 slab_rcu = (struct slab_rcu *)slabp;
2054 slab_rcu->page = page;
2055 call_rcu(&slab_rcu->head, kmem_rcu_free);
2057 kmem_freepages(cachep, page);
2058 if (OFF_SLAB(cachep))
2059 kmem_cache_free(cachep->slabp_cache, slabp);
2064 * calculate_slab_order - calculate size (page order) of slabs
2065 * @cachep: pointer to the cache that is being created
2066 * @size: size of objects to be created in this cache.
2067 * @align: required alignment for the objects.
2068 * @flags: slab allocation flags
2070 * Also calculates the number of objects per slab.
2072 * This could be made much more intelligent. For now, try to avoid using
2073 * high order pages for slabs. When the gfp() functions are more friendly
2074 * towards high-order requests, this should be changed.
2076 static size_t calculate_slab_order(struct kmem_cache *cachep,
2077 size_t size, size_t align, unsigned long flags)
2079 unsigned long offslab_limit;
2080 size_t left_over = 0;
2083 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
2087 cache_estimate(gfporder, size, align, flags, &remainder, &num);
2091 if (flags & CFLGS_OFF_SLAB) {
2093 * Max number of objs-per-slab for caches which
2094 * use off-slab slabs. Needed to avoid a possible
2095 * looping condition in cache_grow().
2097 offslab_limit = size - sizeof(struct slab);
2098 offslab_limit /= sizeof(kmem_bufctl_t);
2100 if (num > offslab_limit)
2104 /* Found something acceptable - save it away */
2106 cachep->gfporder = gfporder;
2107 left_over = remainder;
2110 * A VFS-reclaimable slab tends to have most allocations
2111 * as GFP_NOFS and we really don't want to have to be allocating
2112 * higher-order pages when we are unable to shrink dcache.
2114 if (flags & SLAB_RECLAIM_ACCOUNT)
2118 * Large number of objects is good, but very large slabs are
2119 * currently bad for the gfp()s.
2121 if (gfporder >= slab_max_order)
2125 * Acceptable internal fragmentation?
2127 if (left_over * 8 <= (PAGE_SIZE << gfporder))
2133 static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
2135 if (slab_state >= FULL)
2136 return enable_cpucache(cachep, gfp);
2138 if (slab_state == DOWN) {
2140 * Note: Creation of first cache (kmem_cache).
2141 * The setup_node is taken care
2142 * of by the caller of __kmem_cache_create
2144 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2145 slab_state = PARTIAL;
2146 } else if (slab_state == PARTIAL) {
2148 * Note: the second kmem_cache_create must create the cache
2149 * that's used by kmalloc(24), otherwise the creation of
2150 * further caches will BUG().
2152 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2155 * If the cache that's used by kmalloc(sizeof(kmem_cache_node)) is
2156 * the second cache, then we need to set up all its node/,
2157 * otherwise the creation of further caches will BUG().
2159 set_up_node(cachep, SIZE_AC);
2160 if (INDEX_AC == INDEX_NODE)
2161 slab_state = PARTIAL_NODE;
2163 slab_state = PARTIAL_ARRAYCACHE;
2165 /* Remaining boot caches */
2166 cachep->array[smp_processor_id()] =
2167 kmalloc(sizeof(struct arraycache_init), gfp);
2169 if (slab_state == PARTIAL_ARRAYCACHE) {
2170 set_up_node(cachep, SIZE_NODE);
2171 slab_state = PARTIAL_NODE;
2174 for_each_online_node(node) {
2175 cachep->node[node] =
2176 kmalloc_node(sizeof(struct kmem_cache_node),
2178 BUG_ON(!cachep->node[node]);
2179 kmem_cache_node_init(cachep->node[node]);
2183 cachep->node[numa_mem_id()]->next_reap =
2184 jiffies + REAPTIMEOUT_LIST3 +
2185 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2187 cpu_cache_get(cachep)->avail = 0;
2188 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2189 cpu_cache_get(cachep)->batchcount = 1;
2190 cpu_cache_get(cachep)->touched = 0;
2191 cachep->batchcount = 1;
2192 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2197 * __kmem_cache_create - Create a cache.
2198 * @cachep: cache management descriptor
2199 * @flags: SLAB flags
2201 * Returns a ptr to the cache on success, NULL on failure.
2202 * Cannot be called within a int, but can be interrupted.
2203 * The @ctor is run when new pages are allocated by the cache.
2207 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2208 * to catch references to uninitialised memory.
2210 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2211 * for buffer overruns.
2213 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2214 * cacheline. This can be beneficial if you're counting cycles as closely
2218 __kmem_cache_create (struct kmem_cache *cachep, unsigned long flags)
2220 size_t left_over, slab_size, ralign;
2223 size_t size = cachep->size;
2228 * Enable redzoning and last user accounting, except for caches with
2229 * large objects, if the increased size would increase the object size
2230 * above the next power of two: caches with object sizes just above a
2231 * power of two have a significant amount of internal fragmentation.
2233 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2234 2 * sizeof(unsigned long long)))
2235 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2236 if (!(flags & SLAB_DESTROY_BY_RCU))
2237 flags |= SLAB_POISON;
2239 if (flags & SLAB_DESTROY_BY_RCU)
2240 BUG_ON(flags & SLAB_POISON);
2244 * Check that size is in terms of words. This is needed to avoid
2245 * unaligned accesses for some archs when redzoning is used, and makes
2246 * sure any on-slab bufctl's are also correctly aligned.
2248 if (size & (BYTES_PER_WORD - 1)) {
2249 size += (BYTES_PER_WORD - 1);
2250 size &= ~(BYTES_PER_WORD - 1);
2254 * Redzoning and user store require word alignment or possibly larger.
2255 * Note this will be overridden by architecture or caller mandated
2256 * alignment if either is greater than BYTES_PER_WORD.
2258 if (flags & SLAB_STORE_USER)
2259 ralign = BYTES_PER_WORD;
2261 if (flags & SLAB_RED_ZONE) {
2262 ralign = REDZONE_ALIGN;
2263 /* If redzoning, ensure that the second redzone is suitably
2264 * aligned, by adjusting the object size accordingly. */
2265 size += REDZONE_ALIGN - 1;
2266 size &= ~(REDZONE_ALIGN - 1);
2269 /* 3) caller mandated alignment */
2270 if (ralign < cachep->align) {
2271 ralign = cachep->align;
2273 /* disable debug if necessary */
2274 if (ralign > __alignof__(unsigned long long))
2275 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2279 cachep->align = ralign;
2281 if (slab_is_available())
2286 setup_node_pointer(cachep);
2290 * Both debugging options require word-alignment which is calculated
2293 if (flags & SLAB_RED_ZONE) {
2294 /* add space for red zone words */
2295 cachep->obj_offset += sizeof(unsigned long long);
2296 size += 2 * sizeof(unsigned long long);
2298 if (flags & SLAB_STORE_USER) {
2299 /* user store requires one word storage behind the end of
2300 * the real object. But if the second red zone needs to be
2301 * aligned to 64 bits, we must allow that much space.
2303 if (flags & SLAB_RED_ZONE)
2304 size += REDZONE_ALIGN;
2306 size += BYTES_PER_WORD;
2308 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2309 if (size >= kmalloc_size(INDEX_NODE + 1)
2310 && cachep->object_size > cache_line_size()
2311 && ALIGN(size, cachep->align) < PAGE_SIZE) {
2312 cachep->obj_offset += PAGE_SIZE - ALIGN(size, cachep->align);
2319 * Determine if the slab management is 'on' or 'off' slab.
2320 * (bootstrapping cannot cope with offslab caches so don't do
2321 * it too early on. Always use on-slab management when
2322 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2324 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
2325 !(flags & SLAB_NOLEAKTRACE))
2327 * Size is large, assume best to place the slab management obj
2328 * off-slab (should allow better packing of objs).
2330 flags |= CFLGS_OFF_SLAB;
2332 size = ALIGN(size, cachep->align);
2334 left_over = calculate_slab_order(cachep, size, cachep->align, flags);
2339 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2340 + sizeof(struct slab), cachep->align);
2343 * If the slab has been placed off-slab, and we have enough space then
2344 * move it on-slab. This is at the expense of any extra colouring.
2346 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2347 flags &= ~CFLGS_OFF_SLAB;
2348 left_over -= slab_size;
2351 if (flags & CFLGS_OFF_SLAB) {
2352 /* really off slab. No need for manual alignment */
2354 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2356 #ifdef CONFIG_PAGE_POISONING
2357 /* If we're going to use the generic kernel_map_pages()
2358 * poisoning, then it's going to smash the contents of
2359 * the redzone and userword anyhow, so switch them off.
2361 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2362 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2366 cachep->colour_off = cache_line_size();
2367 /* Offset must be a multiple of the alignment. */
2368 if (cachep->colour_off < cachep->align)
2369 cachep->colour_off = cachep->align;
2370 cachep->colour = left_over / cachep->colour_off;
2371 cachep->slab_size = slab_size;
2372 cachep->flags = flags;
2373 cachep->allocflags = 0;
2374 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2375 cachep->allocflags |= GFP_DMA;
2376 cachep->size = size;
2377 cachep->reciprocal_buffer_size = reciprocal_value(size);
2379 if (flags & CFLGS_OFF_SLAB) {
2380 cachep->slabp_cache = kmalloc_slab(slab_size, 0u);
2382 * This is a possibility for one of the malloc_sizes caches.
2383 * But since we go off slab only for object size greater than
2384 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2385 * this should not happen at all.
2386 * But leave a BUG_ON for some lucky dude.
2388 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
2391 err = setup_cpu_cache(cachep, gfp);
2393 __kmem_cache_shutdown(cachep);
2397 if (flags & SLAB_DEBUG_OBJECTS) {
2399 * Would deadlock through slab_destroy()->call_rcu()->
2400 * debug_object_activate()->kmem_cache_alloc().
2402 WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU);
2404 slab_set_debugobj_lock_classes(cachep);
2405 } else if (!OFF_SLAB(cachep) && !(flags & SLAB_DESTROY_BY_RCU))
2406 on_slab_lock_classes(cachep);
2412 static void check_irq_off(void)
2414 BUG_ON(!irqs_disabled());
2417 static void check_irq_on(void)
2419 BUG_ON(irqs_disabled());
2422 static void check_spinlock_acquired(struct kmem_cache *cachep)
2426 assert_spin_locked(&cachep->node[numa_mem_id()]->list_lock);
2430 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2434 assert_spin_locked(&cachep->node[node]->list_lock);
2439 #define check_irq_off() do { } while(0)
2440 #define check_irq_on() do { } while(0)
2441 #define check_spinlock_acquired(x) do { } while(0)
2442 #define check_spinlock_acquired_node(x, y) do { } while(0)
2445 static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
2446 struct array_cache *ac,
2447 int force, int node);
2449 static void do_drain(void *arg)
2451 struct kmem_cache *cachep = arg;
2452 struct array_cache *ac;
2453 int node = numa_mem_id();
2456 ac = cpu_cache_get(cachep);
2457 spin_lock(&cachep->node[node]->list_lock);
2458 free_block(cachep, ac->entry, ac->avail, node);
2459 spin_unlock(&cachep->node[node]->list_lock);
2463 static void drain_cpu_caches(struct kmem_cache *cachep)
2465 struct kmem_cache_node *n;
2468 on_each_cpu(do_drain, cachep, 1);
2470 for_each_online_node(node) {
2471 n = cachep->node[node];
2473 drain_alien_cache(cachep, n->alien);
2476 for_each_online_node(node) {
2477 n = cachep->node[node];
2479 drain_array(cachep, n, n->shared, 1, node);
2484 * Remove slabs from the list of free slabs.
2485 * Specify the number of slabs to drain in tofree.
2487 * Returns the actual number of slabs released.
2489 static int drain_freelist(struct kmem_cache *cache,
2490 struct kmem_cache_node *n, int tofree)
2492 struct list_head *p;
2497 while (nr_freed < tofree && !list_empty(&n->slabs_free)) {
2499 spin_lock_irq(&n->list_lock);
2500 p = n->slabs_free.prev;
2501 if (p == &n->slabs_free) {
2502 spin_unlock_irq(&n->list_lock);
2506 slabp = list_entry(p, struct slab, list);
2508 BUG_ON(slabp->inuse);
2510 list_del(&slabp->list);
2512 * Safe to drop the lock. The slab is no longer linked
2515 n->free_objects -= cache->num;
2516 spin_unlock_irq(&n->list_lock);
2517 slab_destroy(cache, slabp);
2524 /* Called with slab_mutex held to protect against cpu hotplug */
2525 static int __cache_shrink(struct kmem_cache *cachep)
2528 struct kmem_cache_node *n;
2530 drain_cpu_caches(cachep);
2533 for_each_online_node(i) {
2534 n = cachep->node[i];
2538 drain_freelist(cachep, n, slabs_tofree(cachep, n));
2540 ret += !list_empty(&n->slabs_full) ||
2541 !list_empty(&n->slabs_partial);
2543 return (ret ? 1 : 0);
2547 * kmem_cache_shrink - Shrink a cache.
2548 * @cachep: The cache to shrink.
2550 * Releases as many slabs as possible for a cache.
2551 * To help debugging, a zero exit status indicates all slabs were released.
2553 int kmem_cache_shrink(struct kmem_cache *cachep)
2556 BUG_ON(!cachep || in_interrupt());
2559 mutex_lock(&slab_mutex);
2560 ret = __cache_shrink(cachep);
2561 mutex_unlock(&slab_mutex);
2565 EXPORT_SYMBOL(kmem_cache_shrink);
2567 int __kmem_cache_shutdown(struct kmem_cache *cachep)
2570 struct kmem_cache_node *n;
2571 int rc = __cache_shrink(cachep);
2576 for_each_online_cpu(i)
2577 kfree(cachep->array[i]);
2579 /* NUMA: free the node structures */
2580 for_each_online_node(i) {
2581 n = cachep->node[i];
2584 free_alien_cache(n->alien);
2592 * Get the memory for a slab management obj.
2593 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2594 * always come from malloc_sizes caches. The slab descriptor cannot
2595 * come from the same cache which is getting created because,
2596 * when we are searching for an appropriate cache for these
2597 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2598 * If we are creating a malloc_sizes cache here it would not be visible to
2599 * kmem_find_general_cachep till the initialization is complete.
2600 * Hence we cannot have slabp_cache same as the original cache.
2602 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep,
2603 struct page *page, int colour_off,
2604 gfp_t local_flags, int nodeid)
2607 void *addr = page_address(page);
2609 if (OFF_SLAB(cachep)) {
2610 /* Slab management obj is off-slab. */
2611 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2612 local_flags, nodeid);
2614 * If the first object in the slab is leaked (it's allocated
2615 * but no one has a reference to it), we want to make sure
2616 * kmemleak does not treat the ->s_mem pointer as a reference
2617 * to the object. Otherwise we will not report the leak.
2619 kmemleak_scan_area(&slabp->list, sizeof(struct list_head),
2624 slabp = addr + colour_off;
2625 colour_off += cachep->slab_size;
2628 slabp->s_mem = addr + colour_off;
2633 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2635 return (kmem_bufctl_t *) (slabp + 1);
2638 static void cache_init_objs(struct kmem_cache *cachep,
2643 for (i = 0; i < cachep->num; i++) {
2644 void *objp = index_to_obj(cachep, slabp, i);
2646 /* need to poison the objs? */
2647 if (cachep->flags & SLAB_POISON)
2648 poison_obj(cachep, objp, POISON_FREE);
2649 if (cachep->flags & SLAB_STORE_USER)
2650 *dbg_userword(cachep, objp) = NULL;
2652 if (cachep->flags & SLAB_RED_ZONE) {
2653 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2654 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2657 * Constructors are not allowed to allocate memory from the same
2658 * cache which they are a constructor for. Otherwise, deadlock.
2659 * They must also be threaded.
2661 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2662 cachep->ctor(objp + obj_offset(cachep));
2664 if (cachep->flags & SLAB_RED_ZONE) {
2665 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2666 slab_error(cachep, "constructor overwrote the"
2667 " end of an object");
2668 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2669 slab_error(cachep, "constructor overwrote the"
2670 " start of an object");
2672 if ((cachep->size % PAGE_SIZE) == 0 &&
2673 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2674 kernel_map_pages(virt_to_page(objp),
2675 cachep->size / PAGE_SIZE, 0);
2680 slab_bufctl(slabp)[i] = i + 1;
2682 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2685 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2687 if (CONFIG_ZONE_DMA_FLAG) {
2688 if (flags & GFP_DMA)
2689 BUG_ON(!(cachep->allocflags & GFP_DMA));
2691 BUG_ON(cachep->allocflags & GFP_DMA);
2695 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2698 void *objp = index_to_obj(cachep, slabp, slabp->free);
2702 next = slab_bufctl(slabp)[slabp->free];
2704 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2705 WARN_ON(page_to_nid(virt_to_page(objp)) != nodeid);
2712 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2713 void *objp, int nodeid)
2715 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2718 /* Verify that the slab belongs to the intended node */
2719 WARN_ON(page_to_nid(virt_to_page(objp)) != nodeid);
2721 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2722 printk(KERN_ERR "slab: double free detected in cache "
2723 "'%s', objp %p\n", cachep->name, objp);
2727 slab_bufctl(slabp)[objnr] = slabp->free;
2728 slabp->free = objnr;
2733 * Map pages beginning at addr to the given cache and slab. This is required
2734 * for the slab allocator to be able to lookup the cache and slab of a
2735 * virtual address for kfree, ksize, and slab debugging.
2737 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2743 if (likely(!PageCompound(page)))
2744 nr_pages <<= cache->gfporder;
2747 page->slab_cache = cache;
2748 page->slab_page = slab;
2750 } while (--nr_pages);
2754 * Grow (by 1) the number of slabs within a cache. This is called by
2755 * kmem_cache_alloc() when there are no active objs left in a cache.
2757 static int cache_grow(struct kmem_cache *cachep,
2758 gfp_t flags, int nodeid, struct page *page)
2763 struct kmem_cache_node *n;
2766 * Be lazy and only check for valid flags here, keeping it out of the
2767 * critical path in kmem_cache_alloc().
2769 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2770 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2772 /* Take the node list lock to change the colour_next on this node */
2774 n = cachep->node[nodeid];
2775 spin_lock(&n->list_lock);
2777 /* Get colour for the slab, and cal the next value. */
2778 offset = n->colour_next;
2780 if (n->colour_next >= cachep->colour)
2782 spin_unlock(&n->list_lock);
2784 offset *= cachep->colour_off;
2786 if (local_flags & __GFP_WAIT)
2790 * The test for missing atomic flag is performed here, rather than
2791 * the more obvious place, simply to reduce the critical path length
2792 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2793 * will eventually be caught here (where it matters).
2795 kmem_flagcheck(cachep, flags);
2798 * Get mem for the objs. Attempt to allocate a physical page from
2802 page = kmem_getpages(cachep, local_flags, nodeid);
2806 /* Get slab management. */
2807 slabp = alloc_slabmgmt(cachep, page, offset,
2808 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2812 slab_map_pages(cachep, slabp, page);
2814 cache_init_objs(cachep, slabp);
2816 if (local_flags & __GFP_WAIT)
2817 local_irq_disable();
2819 spin_lock(&n->list_lock);
2821 /* Make slab active. */
2822 list_add_tail(&slabp->list, &(n->slabs_free));
2823 STATS_INC_GROWN(cachep);
2824 n->free_objects += cachep->num;
2825 spin_unlock(&n->list_lock);
2828 kmem_freepages(cachep, page);
2830 if (local_flags & __GFP_WAIT)
2831 local_irq_disable();
2838 * Perform extra freeing checks:
2839 * - detect bad pointers.
2840 * - POISON/RED_ZONE checking
2842 static void kfree_debugcheck(const void *objp)
2844 if (!virt_addr_valid(objp)) {
2845 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2846 (unsigned long)objp);
2851 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2853 unsigned long long redzone1, redzone2;
2855 redzone1 = *dbg_redzone1(cache, obj);
2856 redzone2 = *dbg_redzone2(cache, obj);
2861 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2864 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2865 slab_error(cache, "double free detected");
2867 slab_error(cache, "memory outside object was overwritten");
2869 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2870 obj, redzone1, redzone2);
2873 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2874 unsigned long caller)
2880 BUG_ON(virt_to_cache(objp) != cachep);
2882 objp -= obj_offset(cachep);
2883 kfree_debugcheck(objp);
2884 page = virt_to_head_page(objp);
2886 slabp = page->slab_page;
2888 if (cachep->flags & SLAB_RED_ZONE) {
2889 verify_redzone_free(cachep, objp);
2890 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2891 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2893 if (cachep->flags & SLAB_STORE_USER)
2894 *dbg_userword(cachep, objp) = (void *)caller;
2896 objnr = obj_to_index(cachep, slabp, objp);
2898 BUG_ON(objnr >= cachep->num);
2899 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
2901 #ifdef CONFIG_DEBUG_SLAB_LEAK
2902 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2904 if (cachep->flags & SLAB_POISON) {
2905 #ifdef CONFIG_DEBUG_PAGEALLOC
2906 if ((cachep->size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
2907 store_stackinfo(cachep, objp, caller);
2908 kernel_map_pages(virt_to_page(objp),
2909 cachep->size / PAGE_SIZE, 0);
2911 poison_obj(cachep, objp, POISON_FREE);
2914 poison_obj(cachep, objp, POISON_FREE);
2920 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
2925 /* Check slab's freelist to see if this obj is there. */
2926 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2928 if (entries > cachep->num || i >= cachep->num)
2931 if (entries != cachep->num - slabp->inuse) {
2933 printk(KERN_ERR "slab: Internal list corruption detected in "
2934 "cache '%s'(%d), slabp %p(%d). Tainted(%s). Hexdump:\n",
2935 cachep->name, cachep->num, slabp, slabp->inuse,
2937 print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 16, 1, slabp,
2938 sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t),
2944 #define kfree_debugcheck(x) do { } while(0)
2945 #define cache_free_debugcheck(x,objp,z) (objp)
2946 #define check_slabp(x,y) do { } while(0)
2949 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags,
2953 struct kmem_cache_node *n;
2954 struct array_cache *ac;
2958 node = numa_mem_id();
2959 if (unlikely(force_refill))
2962 ac = cpu_cache_get(cachep);
2963 batchcount = ac->batchcount;
2964 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2966 * If there was little recent activity on this cache, then
2967 * perform only a partial refill. Otherwise we could generate
2970 batchcount = BATCHREFILL_LIMIT;
2972 n = cachep->node[node];
2974 BUG_ON(ac->avail > 0 || !n);
2975 spin_lock(&n->list_lock);
2977 /* See if we can refill from the shared array */
2978 if (n->shared && transfer_objects(ac, n->shared, batchcount)) {
2979 n->shared->touched = 1;
2983 while (batchcount > 0) {
2984 struct list_head *entry;
2986 /* Get slab alloc is to come from. */
2987 entry = n->slabs_partial.next;
2988 if (entry == &n->slabs_partial) {
2989 n->free_touched = 1;
2990 entry = n->slabs_free.next;
2991 if (entry == &n->slabs_free)
2995 slabp = list_entry(entry, struct slab, list);
2996 check_slabp(cachep, slabp);
2997 check_spinlock_acquired(cachep);
3000 * The slab was either on partial or free list so
3001 * there must be at least one object available for
3004 BUG_ON(slabp->inuse >= cachep->num);
3006 while (slabp->inuse < cachep->num && batchcount--) {
3007 STATS_INC_ALLOCED(cachep);
3008 STATS_INC_ACTIVE(cachep);
3009 STATS_SET_HIGH(cachep);
3011 ac_put_obj(cachep, ac, slab_get_obj(cachep, slabp,
3014 check_slabp(cachep, slabp);
3016 /* move slabp to correct slabp list: */
3017 list_del(&slabp->list);
3018 if (slabp->free == BUFCTL_END)
3019 list_add(&slabp->list, &n->slabs_full);
3021 list_add(&slabp->list, &n->slabs_partial);
3025 n->free_objects -= ac->avail;
3027 spin_unlock(&n->list_lock);
3029 if (unlikely(!ac->avail)) {
3032 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
3034 /* cache_grow can reenable interrupts, then ac could change. */
3035 ac = cpu_cache_get(cachep);
3036 node = numa_mem_id();
3038 /* no objects in sight? abort */
3039 if (!x && (ac->avail == 0 || force_refill))
3042 if (!ac->avail) /* objects refilled by interrupt? */
3047 return ac_get_obj(cachep, ac, flags, force_refill);
3050 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3053 might_sleep_if(flags & __GFP_WAIT);
3055 kmem_flagcheck(cachep, flags);
3060 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3061 gfp_t flags, void *objp, unsigned long caller)
3065 if (cachep->flags & SLAB_POISON) {
3066 #ifdef CONFIG_DEBUG_PAGEALLOC
3067 if ((cachep->size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
3068 kernel_map_pages(virt_to_page(objp),
3069 cachep->size / PAGE_SIZE, 1);
3071 check_poison_obj(cachep, objp);
3073 check_poison_obj(cachep, objp);
3075 poison_obj(cachep, objp, POISON_INUSE);
3077 if (cachep->flags & SLAB_STORE_USER)
3078 *dbg_userword(cachep, objp) = (void *)caller;
3080 if (cachep->flags & SLAB_RED_ZONE) {
3081 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3082 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3083 slab_error(cachep, "double free, or memory outside"
3084 " object was overwritten");
3086 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3087 objp, *dbg_redzone1(cachep, objp),
3088 *dbg_redzone2(cachep, objp));
3090 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3091 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3093 #ifdef CONFIG_DEBUG_SLAB_LEAK
3098 slabp = virt_to_head_page(objp)->slab_page;
3099 objnr = (unsigned)(objp - slabp->s_mem) / cachep->size;
3100 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3103 objp += obj_offset(cachep);
3104 if (cachep->ctor && cachep->flags & SLAB_POISON)
3106 if (ARCH_SLAB_MINALIGN &&
3107 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
3108 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3109 objp, (int)ARCH_SLAB_MINALIGN);
3114 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3117 static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
3119 if (cachep == kmem_cache)
3122 return should_failslab(cachep->object_size, flags, cachep->flags);
3125 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3128 struct array_cache *ac;
3129 bool force_refill = false;
3133 ac = cpu_cache_get(cachep);
3134 if (likely(ac->avail)) {
3136 objp = ac_get_obj(cachep, ac, flags, false);
3139 * Allow for the possibility all avail objects are not allowed
3140 * by the current flags
3143 STATS_INC_ALLOCHIT(cachep);
3146 force_refill = true;
3149 STATS_INC_ALLOCMISS(cachep);
3150 objp = cache_alloc_refill(cachep, flags, force_refill);
3152 * the 'ac' may be updated by cache_alloc_refill(),
3153 * and kmemleak_erase() requires its correct value.
3155 ac = cpu_cache_get(cachep);
3159 * To avoid a false negative, if an object that is in one of the
3160 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3161 * treat the array pointers as a reference to the object.
3164 kmemleak_erase(&ac->entry[ac->avail]);
3170 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3172 * If we are in_interrupt, then process context, including cpusets and
3173 * mempolicy, may not apply and should not be used for allocation policy.
3175 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3177 int nid_alloc, nid_here;
3179 if (in_interrupt() || (flags & __GFP_THISNODE))
3181 nid_alloc = nid_here = numa_mem_id();
3182 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3183 nid_alloc = cpuset_slab_spread_node();
3184 else if (current->mempolicy)
3185 nid_alloc = slab_node();
3186 if (nid_alloc != nid_here)
3187 return ____cache_alloc_node(cachep, flags, nid_alloc);
3192 * Fallback function if there was no memory available and no objects on a
3193 * certain node and fall back is permitted. First we scan all the
3194 * available node for available objects. If that fails then we
3195 * perform an allocation without specifying a node. This allows the page
3196 * allocator to do its reclaim / fallback magic. We then insert the
3197 * slab into the proper nodelist and then allocate from it.
3199 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3201 struct zonelist *zonelist;
3205 enum zone_type high_zoneidx = gfp_zone(flags);
3208 unsigned int cpuset_mems_cookie;
3210 if (flags & __GFP_THISNODE)
3213 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3216 cpuset_mems_cookie = get_mems_allowed();
3217 zonelist = node_zonelist(slab_node(), flags);
3221 * Look through allowed nodes for objects available
3222 * from existing per node queues.
3224 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3225 nid = zone_to_nid(zone);
3227 if (cpuset_zone_allowed_hardwall(zone, flags) &&
3229 cache->node[nid]->free_objects) {
3230 obj = ____cache_alloc_node(cache,
3231 flags | GFP_THISNODE, nid);
3239 * This allocation will be performed within the constraints
3240 * of the current cpuset / memory policy requirements.
3241 * We may trigger various forms of reclaim on the allowed
3242 * set and go into memory reserves if necessary.
3246 if (local_flags & __GFP_WAIT)
3248 kmem_flagcheck(cache, flags);
3249 page = kmem_getpages(cache, local_flags, numa_mem_id());
3250 if (local_flags & __GFP_WAIT)
3251 local_irq_disable();
3254 * Insert into the appropriate per node queues
3256 nid = page_to_nid(page);
3257 if (cache_grow(cache, flags, nid, page)) {
3258 obj = ____cache_alloc_node(cache,
3259 flags | GFP_THISNODE, nid);
3262 * Another processor may allocate the
3263 * objects in the slab since we are
3264 * not holding any locks.
3268 /* cache_grow already freed obj */
3274 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !obj))
3280 * A interface to enable slab creation on nodeid
3282 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3285 struct list_head *entry;
3287 struct kmem_cache_node *n;
3291 VM_BUG_ON(nodeid > num_online_nodes());
3292 n = cachep->node[nodeid];
3297 spin_lock(&n->list_lock);
3298 entry = n->slabs_partial.next;
3299 if (entry == &n->slabs_partial) {
3300 n->free_touched = 1;
3301 entry = n->slabs_free.next;
3302 if (entry == &n->slabs_free)
3306 slabp = list_entry(entry, struct slab, list);
3307 check_spinlock_acquired_node(cachep, nodeid);
3308 check_slabp(cachep, slabp);
3310 STATS_INC_NODEALLOCS(cachep);
3311 STATS_INC_ACTIVE(cachep);
3312 STATS_SET_HIGH(cachep);
3314 BUG_ON(slabp->inuse == cachep->num);
3316 obj = slab_get_obj(cachep, slabp, nodeid);
3317 check_slabp(cachep, slabp);
3319 /* move slabp to correct slabp list: */
3320 list_del(&slabp->list);
3322 if (slabp->free == BUFCTL_END)
3323 list_add(&slabp->list, &n->slabs_full);
3325 list_add(&slabp->list, &n->slabs_partial);
3327 spin_unlock(&n->list_lock);
3331 spin_unlock(&n->list_lock);
3332 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
3336 return fallback_alloc(cachep, flags);
3342 static __always_inline void *
3343 slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3344 unsigned long caller)
3346 unsigned long save_flags;
3348 int slab_node = numa_mem_id();
3350 flags &= gfp_allowed_mask;
3352 lockdep_trace_alloc(flags);
3354 if (slab_should_failslab(cachep, flags))
3357 cachep = memcg_kmem_get_cache(cachep, flags);
3359 cache_alloc_debugcheck_before(cachep, flags);
3360 local_irq_save(save_flags);
3362 if (nodeid == NUMA_NO_NODE)
3365 if (unlikely(!cachep->node[nodeid])) {
3366 /* Node not bootstrapped yet */
3367 ptr = fallback_alloc(cachep, flags);
3371 if (nodeid == slab_node) {
3373 * Use the locally cached objects if possible.
3374 * However ____cache_alloc does not allow fallback
3375 * to other nodes. It may fail while we still have
3376 * objects on other nodes available.
3378 ptr = ____cache_alloc(cachep, flags);
3382 /* ___cache_alloc_node can fall back to other nodes */
3383 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3385 local_irq_restore(save_flags);
3386 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3387 kmemleak_alloc_recursive(ptr, cachep->object_size, 1, cachep->flags,
3391 kmemcheck_slab_alloc(cachep, flags, ptr, cachep->object_size);
3393 if (unlikely((flags & __GFP_ZERO) && ptr))
3394 memset(ptr, 0, cachep->object_size);
3399 static __always_inline void *
3400 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3404 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3405 objp = alternate_node_alloc(cache, flags);
3409 objp = ____cache_alloc(cache, flags);
3412 * We may just have run out of memory on the local node.
3413 * ____cache_alloc_node() knows how to locate memory on other nodes
3416 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3423 static __always_inline void *
3424 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3426 return ____cache_alloc(cachep, flags);
3429 #endif /* CONFIG_NUMA */
3431 static __always_inline void *
3432 slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
3434 unsigned long save_flags;
3437 flags &= gfp_allowed_mask;
3439 lockdep_trace_alloc(flags);
3441 if (slab_should_failslab(cachep, flags))
3444 cachep = memcg_kmem_get_cache(cachep, flags);
3446 cache_alloc_debugcheck_before(cachep, flags);
3447 local_irq_save(save_flags);
3448 objp = __do_cache_alloc(cachep, flags);
3449 local_irq_restore(save_flags);
3450 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3451 kmemleak_alloc_recursive(objp, cachep->object_size, 1, cachep->flags,
3456 kmemcheck_slab_alloc(cachep, flags, objp, cachep->object_size);
3458 if (unlikely((flags & __GFP_ZERO) && objp))
3459 memset(objp, 0, cachep->object_size);
3465 * Caller needs to acquire correct kmem_list's list_lock
3467 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3471 struct kmem_cache_node *n;
3473 for (i = 0; i < nr_objects; i++) {
3477 clear_obj_pfmemalloc(&objpp[i]);
3480 slabp = virt_to_slab(objp);
3481 n = cachep->node[node];
3482 list_del(&slabp->list);
3483 check_spinlock_acquired_node(cachep, node);
3484 check_slabp(cachep, slabp);
3485 slab_put_obj(cachep, slabp, objp, node);
3486 STATS_DEC_ACTIVE(cachep);
3488 check_slabp(cachep, slabp);
3490 /* fixup slab chains */
3491 if (slabp->inuse == 0) {
3492 if (n->free_objects > n->free_limit) {
3493 n->free_objects -= cachep->num;
3494 /* No need to drop any previously held
3495 * lock here, even if we have a off-slab slab
3496 * descriptor it is guaranteed to come from
3497 * a different cache, refer to comments before
3500 slab_destroy(cachep, slabp);
3502 list_add(&slabp->list, &n->slabs_free);
3505 /* Unconditionally move a slab to the end of the
3506 * partial list on free - maximum time for the
3507 * other objects to be freed, too.
3509 list_add_tail(&slabp->list, &n->slabs_partial);
3514 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3517 struct kmem_cache_node *n;
3518 int node = numa_mem_id();
3520 batchcount = ac->batchcount;
3522 BUG_ON(!batchcount || batchcount > ac->avail);
3525 n = cachep->node[node];
3526 spin_lock(&n->list_lock);
3528 struct array_cache *shared_array = n->shared;
3529 int max = shared_array->limit - shared_array->avail;
3531 if (batchcount > max)
3533 memcpy(&(shared_array->entry[shared_array->avail]),
3534 ac->entry, sizeof(void *) * batchcount);
3535 shared_array->avail += batchcount;
3540 free_block(cachep, ac->entry, batchcount, node);
3545 struct list_head *p;
3547 p = n->slabs_free.next;
3548 while (p != &(n->slabs_free)) {
3551 slabp = list_entry(p, struct slab, list);
3552 BUG_ON(slabp->inuse);
3557 STATS_SET_FREEABLE(cachep, i);
3560 spin_unlock(&n->list_lock);
3561 ac->avail -= batchcount;
3562 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3566 * Release an obj back to its cache. If the obj has a constructed state, it must
3567 * be in this state _before_ it is released. Called with disabled ints.
3569 static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3570 unsigned long caller)
3572 struct array_cache *ac = cpu_cache_get(cachep);
3575 kmemleak_free_recursive(objp, cachep->flags);
3576 objp = cache_free_debugcheck(cachep, objp, caller);
3578 kmemcheck_slab_free(cachep, objp, cachep->object_size);
3581 * Skip calling cache_free_alien() when the platform is not numa.
3582 * This will avoid cache misses that happen while accessing slabp (which
3583 * is per page memory reference) to get nodeid. Instead use a global
3584 * variable to skip the call, which is mostly likely to be present in
3587 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3590 if (likely(ac->avail < ac->limit)) {
3591 STATS_INC_FREEHIT(cachep);
3593 STATS_INC_FREEMISS(cachep);
3594 cache_flusharray(cachep, ac);
3597 ac_put_obj(cachep, ac, objp);
3601 * kmem_cache_alloc - Allocate an object
3602 * @cachep: The cache to allocate from.
3603 * @flags: See kmalloc().
3605 * Allocate an object from this cache. The flags are only relevant
3606 * if the cache has no available objects.
3608 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3610 void *ret = slab_alloc(cachep, flags, _RET_IP_);
3612 trace_kmem_cache_alloc(_RET_IP_, ret,
3613 cachep->object_size, cachep->size, flags);
3617 EXPORT_SYMBOL(kmem_cache_alloc);
3619 #ifdef CONFIG_TRACING
3621 kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
3625 ret = slab_alloc(cachep, flags, _RET_IP_);
3627 trace_kmalloc(_RET_IP_, ret,
3628 size, cachep->size, flags);
3631 EXPORT_SYMBOL(kmem_cache_alloc_trace);
3636 * kmem_cache_alloc_node - Allocate an object on the specified node
3637 * @cachep: The cache to allocate from.
3638 * @flags: See kmalloc().
3639 * @nodeid: node number of the target node.
3641 * Identical to kmem_cache_alloc but it will allocate memory on the given
3642 * node, which can improve the performance for cpu bound structures.
3644 * Fallback to other node is possible if __GFP_THISNODE is not set.
3646 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3648 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3650 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3651 cachep->object_size, cachep->size,
3656 EXPORT_SYMBOL(kmem_cache_alloc_node);
3658 #ifdef CONFIG_TRACING
3659 void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
3666 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3668 trace_kmalloc_node(_RET_IP_, ret,
3673 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3676 static __always_inline void *
3677 __do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
3679 struct kmem_cache *cachep;
3681 cachep = kmalloc_slab(size, flags);
3682 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3684 return kmem_cache_alloc_node_trace(cachep, flags, node, size);
3687 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3688 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3690 return __do_kmalloc_node(size, flags, node, _RET_IP_);
3692 EXPORT_SYMBOL(__kmalloc_node);
3694 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3695 int node, unsigned long caller)
3697 return __do_kmalloc_node(size, flags, node, caller);
3699 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3701 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3703 return __do_kmalloc_node(size, flags, node, 0);
3705 EXPORT_SYMBOL(__kmalloc_node);
3706 #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3707 #endif /* CONFIG_NUMA */
3710 * __do_kmalloc - allocate memory
3711 * @size: how many bytes of memory are required.
3712 * @flags: the type of memory to allocate (see kmalloc).
3713 * @caller: function caller for debug tracking of the caller
3715 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3716 unsigned long caller)
3718 struct kmem_cache *cachep;
3721 /* If you want to save a few bytes .text space: replace
3723 * Then kmalloc uses the uninlined functions instead of the inline
3726 cachep = kmalloc_slab(size, flags);
3727 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3729 ret = slab_alloc(cachep, flags, caller);
3731 trace_kmalloc(caller, ret,
3732 size, cachep->size, flags);
3738 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3739 void *__kmalloc(size_t size, gfp_t flags)
3741 return __do_kmalloc(size, flags, _RET_IP_);
3743 EXPORT_SYMBOL(__kmalloc);
3745 void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3747 return __do_kmalloc(size, flags, caller);
3749 EXPORT_SYMBOL(__kmalloc_track_caller);
3752 void *__kmalloc(size_t size, gfp_t flags)
3754 return __do_kmalloc(size, flags, 0);
3756 EXPORT_SYMBOL(__kmalloc);
3760 * kmem_cache_free - Deallocate an object
3761 * @cachep: The cache the allocation was from.
3762 * @objp: The previously allocated object.
3764 * Free an object which was previously allocated from this
3767 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3769 unsigned long flags;
3770 cachep = cache_from_obj(cachep, objp);
3774 local_irq_save(flags);
3775 debug_check_no_locks_freed(objp, cachep->object_size);
3776 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3777 debug_check_no_obj_freed(objp, cachep->object_size);
3778 __cache_free(cachep, objp, _RET_IP_);
3779 local_irq_restore(flags);
3781 trace_kmem_cache_free(_RET_IP_, objp);
3783 EXPORT_SYMBOL(kmem_cache_free);
3786 * kfree - free previously allocated memory
3787 * @objp: pointer returned by kmalloc.
3789 * If @objp is NULL, no operation is performed.
3791 * Don't free memory not originally allocated by kmalloc()
3792 * or you will run into trouble.
3794 void kfree(const void *objp)
3796 struct kmem_cache *c;
3797 unsigned long flags;
3799 trace_kfree(_RET_IP_, objp);
3801 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3803 local_irq_save(flags);
3804 kfree_debugcheck(objp);
3805 c = virt_to_cache(objp);
3806 debug_check_no_locks_freed(objp, c->object_size);
3808 debug_check_no_obj_freed(objp, c->object_size);
3809 __cache_free(c, (void *)objp, _RET_IP_);
3810 local_irq_restore(flags);
3812 EXPORT_SYMBOL(kfree);
3815 * This initializes kmem_cache_node or resizes various caches for all nodes.
3817 static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
3820 struct kmem_cache_node *n;
3821 struct array_cache *new_shared;
3822 struct array_cache **new_alien = NULL;
3824 for_each_online_node(node) {
3826 if (use_alien_caches) {
3827 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
3833 if (cachep->shared) {
3834 new_shared = alloc_arraycache(node,
3835 cachep->shared*cachep->batchcount,
3838 free_alien_cache(new_alien);
3843 n = cachep->node[node];
3845 struct array_cache *shared = n->shared;
3847 spin_lock_irq(&n->list_lock);
3850 free_block(cachep, shared->entry,
3851 shared->avail, node);
3853 n->shared = new_shared;
3855 n->alien = new_alien;
3858 n->free_limit = (1 + nr_cpus_node(node)) *
3859 cachep->batchcount + cachep->num;
3860 spin_unlock_irq(&n->list_lock);
3862 free_alien_cache(new_alien);
3865 n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
3867 free_alien_cache(new_alien);
3872 kmem_cache_node_init(n);
3873 n->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3874 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3875 n->shared = new_shared;
3876 n->alien = new_alien;
3877 n->free_limit = (1 + nr_cpus_node(node)) *
3878 cachep->batchcount + cachep->num;
3879 cachep->node[node] = n;
3884 if (!cachep->list.next) {
3885 /* Cache is not active yet. Roll back what we did */
3888 if (cachep->node[node]) {
3889 n = cachep->node[node];
3892 free_alien_cache(n->alien);
3894 cachep->node[node] = NULL;
3902 struct ccupdate_struct {
3903 struct kmem_cache *cachep;
3904 struct array_cache *new[0];
3907 static void do_ccupdate_local(void *info)
3909 struct ccupdate_struct *new = info;
3910 struct array_cache *old;
3913 old = cpu_cache_get(new->cachep);
3915 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3916 new->new[smp_processor_id()] = old;
3919 /* Always called with the slab_mutex held */
3920 static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
3921 int batchcount, int shared, gfp_t gfp)
3923 struct ccupdate_struct *new;
3926 new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *),
3931 for_each_online_cpu(i) {
3932 new->new[i] = alloc_arraycache(cpu_to_mem(i), limit,
3935 for (i--; i >= 0; i--)
3941 new->cachep = cachep;
3943 on_each_cpu(do_ccupdate_local, (void *)new, 1);
3946 cachep->batchcount = batchcount;
3947 cachep->limit = limit;
3948 cachep->shared = shared;
3950 for_each_online_cpu(i) {
3951 struct array_cache *ccold = new->new[i];
3954 spin_lock_irq(&cachep->node[cpu_to_mem(i)]->list_lock);
3955 free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i));
3956 spin_unlock_irq(&cachep->node[cpu_to_mem(i)]->list_lock);
3960 return alloc_kmemlist(cachep, gfp);
3963 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3964 int batchcount, int shared, gfp_t gfp)
3967 struct kmem_cache *c = NULL;
3970 ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3972 if (slab_state < FULL)
3975 if ((ret < 0) || !is_root_cache(cachep))
3978 VM_BUG_ON(!mutex_is_locked(&slab_mutex));
3979 for_each_memcg_cache_index(i) {
3980 c = cache_from_memcg(cachep, i);
3982 /* return value determined by the parent cache only */
3983 __do_tune_cpucache(c, limit, batchcount, shared, gfp);
3989 /* Called with slab_mutex held always */
3990 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
3997 if (!is_root_cache(cachep)) {
3998 struct kmem_cache *root = memcg_root_cache(cachep);
3999 limit = root->limit;
4000 shared = root->shared;
4001 batchcount = root->batchcount;
4004 if (limit && shared && batchcount)
4007 * The head array serves three purposes:
4008 * - create a LIFO ordering, i.e. return objects that are cache-warm
4009 * - reduce the number of spinlock operations.
4010 * - reduce the number of linked list operations on the slab and
4011 * bufctl chains: array operations are cheaper.
4012 * The numbers are guessed, we should auto-tune as described by
4015 if (cachep->size > 131072)
4017 else if (cachep->size > PAGE_SIZE)
4019 else if (cachep->size > 1024)
4021 else if (cachep->size > 256)
4027 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
4028 * allocation behaviour: Most allocs on one cpu, most free operations
4029 * on another cpu. For these cases, an efficient object passing between
4030 * cpus is necessary. This is provided by a shared array. The array
4031 * replaces Bonwick's magazine layer.
4032 * On uniprocessor, it's functionally equivalent (but less efficient)
4033 * to a larger limit. Thus disabled by default.
4036 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
4041 * With debugging enabled, large batchcount lead to excessively long
4042 * periods with disabled local interrupts. Limit the batchcount
4047 batchcount = (limit + 1) / 2;
4049 err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
4051 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
4052 cachep->name, -err);
4057 * Drain an array if it contains any elements taking the node lock only if
4058 * necessary. Note that the node listlock also protects the array_cache
4059 * if drain_array() is used on the shared array.
4061 static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
4062 struct array_cache *ac, int force, int node)
4066 if (!ac || !ac->avail)
4068 if (ac->touched && !force) {
4071 spin_lock_irq(&n->list_lock);
4073 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4074 if (tofree > ac->avail)
4075 tofree = (ac->avail + 1) / 2;
4076 free_block(cachep, ac->entry, tofree, node);
4077 ac->avail -= tofree;
4078 memmove(ac->entry, &(ac->entry[tofree]),
4079 sizeof(void *) * ac->avail);
4081 spin_unlock_irq(&n->list_lock);
4086 * cache_reap - Reclaim memory from caches.
4087 * @w: work descriptor
4089 * Called from workqueue/eventd every few seconds.
4091 * - clear the per-cpu caches for this CPU.
4092 * - return freeable pages to the main free memory pool.
4094 * If we cannot acquire the cache chain mutex then just give up - we'll try
4095 * again on the next iteration.
4097 static void cache_reap(struct work_struct *w)
4099 struct kmem_cache *searchp;
4100 struct kmem_cache_node *n;
4101 int node = numa_mem_id();
4102 struct delayed_work *work = to_delayed_work(w);
4104 if (!mutex_trylock(&slab_mutex))
4105 /* Give up. Setup the next iteration. */
4108 list_for_each_entry(searchp, &slab_caches, list) {
4112 * We only take the node lock if absolutely necessary and we
4113 * have established with reasonable certainty that
4114 * we can do some work if the lock was obtained.
4116 n = searchp->node[node];
4118 reap_alien(searchp, n);
4120 drain_array(searchp, n, cpu_cache_get(searchp), 0, node);
4123 * These are racy checks but it does not matter
4124 * if we skip one check or scan twice.
4126 if (time_after(n->next_reap, jiffies))
4129 n->next_reap = jiffies + REAPTIMEOUT_LIST3;
4131 drain_array(searchp, n, n->shared, 0, node);
4133 if (n->free_touched)
4134 n->free_touched = 0;
4138 freed = drain_freelist(searchp, n, (n->free_limit +
4139 5 * searchp->num - 1) / (5 * searchp->num));
4140 STATS_ADD_REAPED(searchp, freed);
4146 mutex_unlock(&slab_mutex);
4149 /* Set up the next iteration */
4150 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
4153 #ifdef CONFIG_SLABINFO
4154 void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
4157 unsigned long active_objs;
4158 unsigned long num_objs;
4159 unsigned long active_slabs = 0;
4160 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4164 struct kmem_cache_node *n;
4168 for_each_online_node(node) {
4169 n = cachep->node[node];
4174 spin_lock_irq(&n->list_lock);
4176 list_for_each_entry(slabp, &n->slabs_full, list) {
4177 if (slabp->inuse != cachep->num && !error)
4178 error = "slabs_full accounting error";
4179 active_objs += cachep->num;
4182 list_for_each_entry(slabp, &n->slabs_partial, list) {
4183 if (slabp->inuse == cachep->num && !error)
4184 error = "slabs_partial inuse accounting error";
4185 if (!slabp->inuse && !error)
4186 error = "slabs_partial/inuse accounting error";
4187 active_objs += slabp->inuse;
4190 list_for_each_entry(slabp, &n->slabs_free, list) {
4191 if (slabp->inuse && !error)
4192 error = "slabs_free/inuse accounting error";
4195 free_objects += n->free_objects;
4197 shared_avail += n->shared->avail;
4199 spin_unlock_irq(&n->list_lock);
4201 num_slabs += active_slabs;
4202 num_objs = num_slabs * cachep->num;
4203 if (num_objs - active_objs != free_objects && !error)
4204 error = "free_objects accounting error";
4206 name = cachep->name;
4208 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4210 sinfo->active_objs = active_objs;
4211 sinfo->num_objs = num_objs;
4212 sinfo->active_slabs = active_slabs;
4213 sinfo->num_slabs = num_slabs;
4214 sinfo->shared_avail = shared_avail;
4215 sinfo->limit = cachep->limit;
4216 sinfo->batchcount = cachep->batchcount;
4217 sinfo->shared = cachep->shared;
4218 sinfo->objects_per_slab = cachep->num;
4219 sinfo->cache_order = cachep->gfporder;
4222 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
4226 unsigned long high = cachep->high_mark;
4227 unsigned long allocs = cachep->num_allocations;
4228 unsigned long grown = cachep->grown;
4229 unsigned long reaped = cachep->reaped;
4230 unsigned long errors = cachep->errors;
4231 unsigned long max_freeable = cachep->max_freeable;
4232 unsigned long node_allocs = cachep->node_allocs;
4233 unsigned long node_frees = cachep->node_frees;
4234 unsigned long overflows = cachep->node_overflow;
4236 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4237 "%4lu %4lu %4lu %4lu %4lu",
4238 allocs, high, grown,
4239 reaped, errors, max_freeable, node_allocs,
4240 node_frees, overflows);
4244 unsigned long allochit = atomic_read(&cachep->allochit);
4245 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4246 unsigned long freehit = atomic_read(&cachep->freehit);
4247 unsigned long freemiss = atomic_read(&cachep->freemiss);
4249 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4250 allochit, allocmiss, freehit, freemiss);
4255 #define MAX_SLABINFO_WRITE 128
4257 * slabinfo_write - Tuning for the slab allocator
4259 * @buffer: user buffer
4260 * @count: data length
4263 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4264 size_t count, loff_t *ppos)
4266 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4267 int limit, batchcount, shared, res;
4268 struct kmem_cache *cachep;
4270 if (count > MAX_SLABINFO_WRITE)
4272 if (copy_from_user(&kbuf, buffer, count))
4274 kbuf[MAX_SLABINFO_WRITE] = '\0';
4276 tmp = strchr(kbuf, ' ');
4281 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4284 /* Find the cache in the chain of caches. */
4285 mutex_lock(&slab_mutex);
4287 list_for_each_entry(cachep, &slab_caches, list) {
4288 if (!strcmp(cachep->name, kbuf)) {
4289 if (limit < 1 || batchcount < 1 ||
4290 batchcount > limit || shared < 0) {
4293 res = do_tune_cpucache(cachep, limit,
4300 mutex_unlock(&slab_mutex);
4306 #ifdef CONFIG_DEBUG_SLAB_LEAK
4308 static void *leaks_start(struct seq_file *m, loff_t *pos)
4310 mutex_lock(&slab_mutex);
4311 return seq_list_start(&slab_caches, *pos);
4314 static inline int add_caller(unsigned long *n, unsigned long v)
4324 unsigned long *q = p + 2 * i;
4338 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4344 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4350 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->size) {
4351 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4353 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4358 static void show_symbol(struct seq_file *m, unsigned long address)
4360 #ifdef CONFIG_KALLSYMS
4361 unsigned long offset, size;
4362 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4364 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4365 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4367 seq_printf(m, " [%s]", modname);
4371 seq_printf(m, "%p", (void *)address);
4374 static int leaks_show(struct seq_file *m, void *p)
4376 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
4378 struct kmem_cache_node *n;
4380 unsigned long *x = m->private;
4384 if (!(cachep->flags & SLAB_STORE_USER))
4386 if (!(cachep->flags & SLAB_RED_ZONE))
4389 /* OK, we can do it */
4393 for_each_online_node(node) {
4394 n = cachep->node[node];
4399 spin_lock_irq(&n->list_lock);
4401 list_for_each_entry(slabp, &n->slabs_full, list)
4402 handle_slab(x, cachep, slabp);
4403 list_for_each_entry(slabp, &n->slabs_partial, list)
4404 handle_slab(x, cachep, slabp);
4405 spin_unlock_irq(&n->list_lock);
4407 name = cachep->name;
4409 /* Increase the buffer size */
4410 mutex_unlock(&slab_mutex);
4411 m->private = kzalloc(x[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4413 /* Too bad, we are really out */
4415 mutex_lock(&slab_mutex);
4418 *(unsigned long *)m->private = x[0] * 2;
4420 mutex_lock(&slab_mutex);
4421 /* Now make sure this entry will be retried */
4425 for (i = 0; i < x[1]; i++) {
4426 seq_printf(m, "%s: %lu ", name, x[2*i+3]);
4427 show_symbol(m, x[2*i+2]);
4434 static const struct seq_operations slabstats_op = {
4435 .start = leaks_start,
4441 static int slabstats_open(struct inode *inode, struct file *file)
4443 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4446 ret = seq_open(file, &slabstats_op);
4448 struct seq_file *m = file->private_data;
4449 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4458 static const struct file_operations proc_slabstats_operations = {
4459 .open = slabstats_open,
4461 .llseek = seq_lseek,
4462 .release = seq_release_private,
4466 static int __init slab_proc_init(void)
4468 #ifdef CONFIG_DEBUG_SLAB_LEAK
4469 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4473 module_init(slab_proc_init);
4477 * ksize - get the actual amount of memory allocated for a given object
4478 * @objp: Pointer to the object
4480 * kmalloc may internally round up allocations and return more memory
4481 * than requested. ksize() can be used to determine the actual amount of
4482 * memory allocated. The caller may use this additional memory, even though
4483 * a smaller amount of memory was initially specified with the kmalloc call.
4484 * The caller must guarantee that objp points to a valid object previously
4485 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4486 * must not be freed during the duration of the call.
4488 size_t ksize(const void *objp)
4491 if (unlikely(objp == ZERO_SIZE_PTR))
4494 return virt_to_cache(objp)->object_size;
4496 EXPORT_SYMBOL(ksize);