3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
28 * This means, that your constructor is used only for newly allocated
29 * slabs and you must pass objects with the same initializations to
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
39 * empty slabs with no allocated objects
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
53 * The c_cpuarray may not be read with enabled local interrupts -
54 * it's changed with a smp_call_function().
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
58 * Several members in struct kmem_cache and struct slab never change, they
59 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
68 * Further notes from the original documentation:
70 * 11 April '97. Started multi-threading - markhe
71 * The global cache-chain is protected by the mutex 'slab_mutex'.
72 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
76 * At present, each engine can be growing a cache. This should be blocked.
78 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
89 #include <linux/slab.h>
91 #include <linux/poison.h>
92 #include <linux/swap.h>
93 #include <linux/cache.h>
94 #include <linux/interrupt.h>
95 #include <linux/init.h>
96 #include <linux/compiler.h>
97 #include <linux/cpuset.h>
98 #include <linux/proc_fs.h>
99 #include <linux/seq_file.h>
100 #include <linux/notifier.h>
101 #include <linux/kallsyms.h>
102 #include <linux/cpu.h>
103 #include <linux/sysctl.h>
104 #include <linux/module.h>
105 #include <linux/rcupdate.h>
106 #include <linux/string.h>
107 #include <linux/uaccess.h>
108 #include <linux/nodemask.h>
109 #include <linux/kmemleak.h>
110 #include <linux/mempolicy.h>
111 #include <linux/mutex.h>
112 #include <linux/fault-inject.h>
113 #include <linux/rtmutex.h>
114 #include <linux/reciprocal_div.h>
115 #include <linux/debugobjects.h>
116 #include <linux/kmemcheck.h>
117 #include <linux/memory.h>
118 #include <linux/prefetch.h>
120 #include <net/sock.h>
122 #include <asm/cacheflush.h>
123 #include <asm/tlbflush.h>
124 #include <asm/page.h>
126 #include <trace/events/kmem.h>
128 #include "internal.h"
133 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
134 * 0 for faster, smaller code (especially in the critical paths).
136 * STATS - 1 to collect stats for /proc/slabinfo.
137 * 0 for faster, smaller code (especially in the critical paths).
139 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
142 #ifdef CONFIG_DEBUG_SLAB
145 #define FORCED_DEBUG 1
149 #define FORCED_DEBUG 0
152 /* Shouldn't this be in a header file somewhere? */
153 #define BYTES_PER_WORD sizeof(void *)
154 #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
156 #ifndef ARCH_KMALLOC_FLAGS
157 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
161 * true if a page was allocated from pfmemalloc reserves for network-based
164 static bool pfmemalloc_active __read_mostly;
169 * Bufctl's are used for linking objs within a slab
172 * This implementation relies on "struct page" for locating the cache &
173 * slab an object belongs to.
174 * This allows the bufctl structure to be small (one int), but limits
175 * the number of objects a slab (not a cache) can contain when off-slab
176 * bufctls are used. The limit is the size of the largest general cache
177 * that does not use off-slab slabs.
178 * For 32bit archs with 4 kB pages, is this 56.
179 * This is not serious, as it is only for large objects, when it is unwise
180 * to have too many per slab.
181 * Note: This limit can be raised by introducing a general cache whose size
182 * is less than 512 (PAGE_SIZE<<3), but greater than 256.
185 typedef unsigned int kmem_bufctl_t;
186 #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0)
187 #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1)
188 #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2)
189 #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3)
194 * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to
195 * arrange for kmem_freepages to be called via RCU. This is useful if
196 * we need to approach a kernel structure obliquely, from its address
197 * obtained without the usual locking. We can lock the structure to
198 * stabilize it and check it's still at the given address, only if we
199 * can be sure that the memory has not been meanwhile reused for some
200 * other kind of object (which our subsystem's lock might corrupt).
202 * rcu_read_lock before reading the address, then rcu_read_unlock after
203 * taking the spinlock within the structure expected at that address.
206 struct rcu_head head;
207 struct kmem_cache *cachep;
214 * Manages the objs in a slab. Placed either at the beginning of mem allocated
215 * for a slab, or allocated from an general cache.
216 * Slabs are chained into three list: fully used, partial, fully free slabs.
221 struct list_head list;
222 unsigned long colouroff;
223 void *s_mem; /* including colour offset */
224 unsigned int inuse; /* num of objs active in slab */
226 unsigned short nodeid;
228 struct slab_rcu __slab_cover_slab_rcu;
236 * - LIFO ordering, to hand out cache-warm objects from _alloc
237 * - reduce the number of linked list operations
238 * - reduce spinlock operations
240 * The limit is stored in the per-cpu structure to reduce the data cache
247 unsigned int batchcount;
248 unsigned int touched;
251 * Must have this definition in here for the proper
252 * alignment of array_cache. Also simplifies accessing
255 * Entries should not be directly dereferenced as
256 * entries belonging to slabs marked pfmemalloc will
257 * have the lower bits set SLAB_OBJ_PFMEMALLOC
261 #define SLAB_OBJ_PFMEMALLOC 1
262 static inline bool is_obj_pfmemalloc(void *objp)
264 return (unsigned long)objp & SLAB_OBJ_PFMEMALLOC;
267 static inline void set_obj_pfmemalloc(void **objp)
269 *objp = (void *)((unsigned long)*objp | SLAB_OBJ_PFMEMALLOC);
273 static inline void clear_obj_pfmemalloc(void **objp)
275 *objp = (void *)((unsigned long)*objp & ~SLAB_OBJ_PFMEMALLOC);
279 * bootstrap: The caches do not work without cpuarrays anymore, but the
280 * cpuarrays are allocated from the generic caches...
282 #define BOOT_CPUCACHE_ENTRIES 1
283 struct arraycache_init {
284 struct array_cache cache;
285 void *entries[BOOT_CPUCACHE_ENTRIES];
289 * The slab lists for all objects.
291 struct kmem_cache_node {
292 struct list_head slabs_partial; /* partial list first, better asm code */
293 struct list_head slabs_full;
294 struct list_head slabs_free;
295 unsigned long free_objects;
296 unsigned int free_limit;
297 unsigned int colour_next; /* Per-node cache coloring */
298 spinlock_t list_lock;
299 struct array_cache *shared; /* shared per node */
300 struct array_cache **alien; /* on other nodes */
301 unsigned long next_reap; /* updated without locking */
302 int free_touched; /* updated without locking */
306 * Need this for bootstrapping a per node allocator.
308 #define NUM_INIT_LISTS (3 * MAX_NUMNODES)
309 static struct kmem_cache_node __initdata initkmem_list3[NUM_INIT_LISTS];
310 #define CACHE_CACHE 0
311 #define SIZE_AC MAX_NUMNODES
312 #define SIZE_L3 (2 * MAX_NUMNODES)
314 static int drain_freelist(struct kmem_cache *cache,
315 struct kmem_cache_node *l3, int tofree);
316 static void free_block(struct kmem_cache *cachep, void **objpp, int len,
318 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
319 static void cache_reap(struct work_struct *unused);
321 static int slab_early_init = 1;
323 #define INDEX_AC kmalloc_index(sizeof(struct arraycache_init))
324 #define INDEX_L3 kmalloc_index(sizeof(struct kmem_cache_node))
326 static void kmem_list3_init(struct kmem_cache_node *parent)
328 INIT_LIST_HEAD(&parent->slabs_full);
329 INIT_LIST_HEAD(&parent->slabs_partial);
330 INIT_LIST_HEAD(&parent->slabs_free);
331 parent->shared = NULL;
332 parent->alien = NULL;
333 parent->colour_next = 0;
334 spin_lock_init(&parent->list_lock);
335 parent->free_objects = 0;
336 parent->free_touched = 0;
339 #define MAKE_LIST(cachep, listp, slab, nodeid) \
341 INIT_LIST_HEAD(listp); \
342 list_splice(&(cachep->node[nodeid]->slab), listp); \
345 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
347 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
348 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
349 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
352 #define CFLGS_OFF_SLAB (0x80000000UL)
353 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
355 #define BATCHREFILL_LIMIT 16
357 * Optimization question: fewer reaps means less probability for unnessary
358 * cpucache drain/refill cycles.
360 * OTOH the cpuarrays can contain lots of objects,
361 * which could lock up otherwise freeable slabs.
363 #define REAPTIMEOUT_CPUC (2*HZ)
364 #define REAPTIMEOUT_LIST3 (4*HZ)
367 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
368 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
369 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
370 #define STATS_INC_GROWN(x) ((x)->grown++)
371 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
372 #define STATS_SET_HIGH(x) \
374 if ((x)->num_active > (x)->high_mark) \
375 (x)->high_mark = (x)->num_active; \
377 #define STATS_INC_ERR(x) ((x)->errors++)
378 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
379 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
380 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
381 #define STATS_SET_FREEABLE(x, i) \
383 if ((x)->max_freeable < i) \
384 (x)->max_freeable = i; \
386 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
387 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
388 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
389 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
391 #define STATS_INC_ACTIVE(x) do { } while (0)
392 #define STATS_DEC_ACTIVE(x) do { } while (0)
393 #define STATS_INC_ALLOCED(x) do { } while (0)
394 #define STATS_INC_GROWN(x) do { } while (0)
395 #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
396 #define STATS_SET_HIGH(x) do { } while (0)
397 #define STATS_INC_ERR(x) do { } while (0)
398 #define STATS_INC_NODEALLOCS(x) do { } while (0)
399 #define STATS_INC_NODEFREES(x) do { } while (0)
400 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
401 #define STATS_SET_FREEABLE(x, i) do { } while (0)
402 #define STATS_INC_ALLOCHIT(x) do { } while (0)
403 #define STATS_INC_ALLOCMISS(x) do { } while (0)
404 #define STATS_INC_FREEHIT(x) do { } while (0)
405 #define STATS_INC_FREEMISS(x) do { } while (0)
411 * memory layout of objects:
413 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
414 * the end of an object is aligned with the end of the real
415 * allocation. Catches writes behind the end of the allocation.
416 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
418 * cachep->obj_offset: The real object.
419 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
420 * cachep->size - 1* BYTES_PER_WORD: last caller address
421 * [BYTES_PER_WORD long]
423 static int obj_offset(struct kmem_cache *cachep)
425 return cachep->obj_offset;
428 static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
430 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
431 return (unsigned long long*) (objp + obj_offset(cachep) -
432 sizeof(unsigned long long));
435 static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
437 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
438 if (cachep->flags & SLAB_STORE_USER)
439 return (unsigned long long *)(objp + cachep->size -
440 sizeof(unsigned long long) -
442 return (unsigned long long *) (objp + cachep->size -
443 sizeof(unsigned long long));
446 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
448 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
449 return (void **)(objp + cachep->size - BYTES_PER_WORD);
454 #define obj_offset(x) 0
455 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
456 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
457 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
462 * Do not go above this order unless 0 objects fit into the slab or
463 * overridden on the command line.
465 #define SLAB_MAX_ORDER_HI 1
466 #define SLAB_MAX_ORDER_LO 0
467 static int slab_max_order = SLAB_MAX_ORDER_LO;
468 static bool slab_max_order_set __initdata;
470 static inline struct kmem_cache *virt_to_cache(const void *obj)
472 struct page *page = virt_to_head_page(obj);
473 return page->slab_cache;
476 static inline struct slab *virt_to_slab(const void *obj)
478 struct page *page = virt_to_head_page(obj);
480 VM_BUG_ON(!PageSlab(page));
481 return page->slab_page;
484 static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
487 return slab->s_mem + cache->size * idx;
491 * We want to avoid an expensive divide : (offset / cache->size)
492 * Using the fact that size is a constant for a particular cache,
493 * we can replace (offset / cache->size) by
494 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
496 static inline unsigned int obj_to_index(const struct kmem_cache *cache,
497 const struct slab *slab, void *obj)
499 u32 offset = (obj - slab->s_mem);
500 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
503 static struct arraycache_init initarray_generic =
504 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
506 /* internal cache of cache description objs */
507 static struct kmem_cache kmem_cache_boot = {
509 .limit = BOOT_CPUCACHE_ENTRIES,
511 .size = sizeof(struct kmem_cache),
512 .name = "kmem_cache",
515 #define BAD_ALIEN_MAGIC 0x01020304ul
517 #ifdef CONFIG_LOCKDEP
520 * Slab sometimes uses the kmalloc slabs to store the slab headers
521 * for other slabs "off slab".
522 * The locking for this is tricky in that it nests within the locks
523 * of all other slabs in a few places; to deal with this special
524 * locking we put on-slab caches into a separate lock-class.
526 * We set lock class for alien array caches which are up during init.
527 * The lock annotation will be lost if all cpus of a node goes down and
528 * then comes back up during hotplug
530 static struct lock_class_key on_slab_l3_key;
531 static struct lock_class_key on_slab_alc_key;
533 static struct lock_class_key debugobj_l3_key;
534 static struct lock_class_key debugobj_alc_key;
536 static void slab_set_lock_classes(struct kmem_cache *cachep,
537 struct lock_class_key *l3_key, struct lock_class_key *alc_key,
540 struct array_cache **alc;
541 struct kmem_cache_node *l3;
544 l3 = cachep->node[q];
548 lockdep_set_class(&l3->list_lock, l3_key);
551 * FIXME: This check for BAD_ALIEN_MAGIC
552 * should go away when common slab code is taught to
553 * work even without alien caches.
554 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
555 * for alloc_alien_cache,
557 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
561 lockdep_set_class(&alc[r]->lock, alc_key);
565 static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
567 slab_set_lock_classes(cachep, &debugobj_l3_key, &debugobj_alc_key, node);
570 static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
574 for_each_online_node(node)
575 slab_set_debugobj_lock_classes_node(cachep, node);
578 static void init_node_lock_keys(int q)
585 for (i = 1; i < PAGE_SHIFT + MAX_ORDER; i++) {
586 struct kmem_cache_node *l3;
587 struct kmem_cache *cache = kmalloc_caches[i];
593 if (!l3 || OFF_SLAB(cache))
596 slab_set_lock_classes(cache, &on_slab_l3_key,
597 &on_slab_alc_key, q);
601 static void on_slab_lock_classes_node(struct kmem_cache *cachep, int q)
603 if (!cachep->node[q])
606 slab_set_lock_classes(cachep, &on_slab_l3_key,
607 &on_slab_alc_key, q);
610 static inline void on_slab_lock_classes(struct kmem_cache *cachep)
614 VM_BUG_ON(OFF_SLAB(cachep));
616 on_slab_lock_classes_node(cachep, node);
619 static inline void init_lock_keys(void)
624 init_node_lock_keys(node);
627 static void init_node_lock_keys(int q)
631 static inline void init_lock_keys(void)
635 static inline void on_slab_lock_classes(struct kmem_cache *cachep)
639 static inline void on_slab_lock_classes_node(struct kmem_cache *cachep, int node)
643 static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
647 static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
652 static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
654 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
656 return cachep->array[smp_processor_id()];
659 static inline struct kmem_cache *__find_general_cachep(size_t size,
665 /* This happens if someone tries to call
666 * kmem_cache_create(), or __kmalloc(), before
667 * the generic caches are initialized.
669 BUG_ON(kmalloc_caches[INDEX_AC] == NULL);
672 return ZERO_SIZE_PTR;
674 i = kmalloc_index(size);
677 * Really subtle: The last entry with cs->cs_size==ULONG_MAX
678 * has cs_{dma,}cachep==NULL. Thus no special case
679 * for large kmalloc calls required.
681 #ifdef CONFIG_ZONE_DMA
682 if (unlikely(gfpflags & GFP_DMA))
683 return kmalloc_dma_caches[i];
685 return kmalloc_caches[i];
688 static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags)
690 return __find_general_cachep(size, gfpflags);
693 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
695 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
699 * Calculate the number of objects and left-over bytes for a given buffer size.
701 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
702 size_t align, int flags, size_t *left_over,
707 size_t slab_size = PAGE_SIZE << gfporder;
710 * The slab management structure can be either off the slab or
711 * on it. For the latter case, the memory allocated for a
715 * - One kmem_bufctl_t for each object
716 * - Padding to respect alignment of @align
717 * - @buffer_size bytes for each object
719 * If the slab management structure is off the slab, then the
720 * alignment will already be calculated into the size. Because
721 * the slabs are all pages aligned, the objects will be at the
722 * correct alignment when allocated.
724 if (flags & CFLGS_OFF_SLAB) {
726 nr_objs = slab_size / buffer_size;
728 if (nr_objs > SLAB_LIMIT)
729 nr_objs = SLAB_LIMIT;
732 * Ignore padding for the initial guess. The padding
733 * is at most @align-1 bytes, and @buffer_size is at
734 * least @align. In the worst case, this result will
735 * be one greater than the number of objects that fit
736 * into the memory allocation when taking the padding
739 nr_objs = (slab_size - sizeof(struct slab)) /
740 (buffer_size + sizeof(kmem_bufctl_t));
743 * This calculated number will be either the right
744 * amount, or one greater than what we want.
746 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
750 if (nr_objs > SLAB_LIMIT)
751 nr_objs = SLAB_LIMIT;
753 mgmt_size = slab_mgmt_size(nr_objs, align);
756 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
760 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
762 static void __slab_error(const char *function, struct kmem_cache *cachep,
765 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
766 function, cachep->name, msg);
768 add_taint(TAINT_BAD_PAGE);
773 * By default on NUMA we use alien caches to stage the freeing of
774 * objects allocated from other nodes. This causes massive memory
775 * inefficiencies when using fake NUMA setup to split memory into a
776 * large number of small nodes, so it can be disabled on the command
780 static int use_alien_caches __read_mostly = 1;
781 static int __init noaliencache_setup(char *s)
783 use_alien_caches = 0;
786 __setup("noaliencache", noaliencache_setup);
788 static int __init slab_max_order_setup(char *str)
790 get_option(&str, &slab_max_order);
791 slab_max_order = slab_max_order < 0 ? 0 :
792 min(slab_max_order, MAX_ORDER - 1);
793 slab_max_order_set = true;
797 __setup("slab_max_order=", slab_max_order_setup);
801 * Special reaping functions for NUMA systems called from cache_reap().
802 * These take care of doing round robin flushing of alien caches (containing
803 * objects freed on different nodes from which they were allocated) and the
804 * flushing of remote pcps by calling drain_node_pages.
806 static DEFINE_PER_CPU(unsigned long, slab_reap_node);
808 static void init_reap_node(int cpu)
812 node = next_node(cpu_to_mem(cpu), node_online_map);
813 if (node == MAX_NUMNODES)
814 node = first_node(node_online_map);
816 per_cpu(slab_reap_node, cpu) = node;
819 static void next_reap_node(void)
821 int node = __this_cpu_read(slab_reap_node);
823 node = next_node(node, node_online_map);
824 if (unlikely(node >= MAX_NUMNODES))
825 node = first_node(node_online_map);
826 __this_cpu_write(slab_reap_node, node);
830 #define init_reap_node(cpu) do { } while (0)
831 #define next_reap_node(void) do { } while (0)
835 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
836 * via the workqueue/eventd.
837 * Add the CPU number into the expiration time to minimize the possibility of
838 * the CPUs getting into lockstep and contending for the global cache chain
841 static void __cpuinit start_cpu_timer(int cpu)
843 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
846 * When this gets called from do_initcalls via cpucache_init(),
847 * init_workqueues() has already run, so keventd will be setup
850 if (keventd_up() && reap_work->work.func == NULL) {
852 INIT_DEFERRABLE_WORK(reap_work, cache_reap);
853 schedule_delayed_work_on(cpu, reap_work,
854 __round_jiffies_relative(HZ, cpu));
858 static struct array_cache *alloc_arraycache(int node, int entries,
859 int batchcount, gfp_t gfp)
861 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
862 struct array_cache *nc = NULL;
864 nc = kmalloc_node(memsize, gfp, node);
866 * The array_cache structures contain pointers to free object.
867 * However, when such objects are allocated or transferred to another
868 * cache the pointers are not cleared and they could be counted as
869 * valid references during a kmemleak scan. Therefore, kmemleak must
870 * not scan such objects.
872 kmemleak_no_scan(nc);
876 nc->batchcount = batchcount;
878 spin_lock_init(&nc->lock);
883 static inline bool is_slab_pfmemalloc(struct slab *slabp)
885 struct page *page = virt_to_page(slabp->s_mem);
887 return PageSlabPfmemalloc(page);
890 /* Clears pfmemalloc_active if no slabs have pfmalloc set */
891 static void recheck_pfmemalloc_active(struct kmem_cache *cachep,
892 struct array_cache *ac)
894 struct kmem_cache_node *l3 = cachep->node[numa_mem_id()];
898 if (!pfmemalloc_active)
901 spin_lock_irqsave(&l3->list_lock, flags);
902 list_for_each_entry(slabp, &l3->slabs_full, list)
903 if (is_slab_pfmemalloc(slabp))
906 list_for_each_entry(slabp, &l3->slabs_partial, list)
907 if (is_slab_pfmemalloc(slabp))
910 list_for_each_entry(slabp, &l3->slabs_free, list)
911 if (is_slab_pfmemalloc(slabp))
914 pfmemalloc_active = false;
916 spin_unlock_irqrestore(&l3->list_lock, flags);
919 static void *__ac_get_obj(struct kmem_cache *cachep, struct array_cache *ac,
920 gfp_t flags, bool force_refill)
923 void *objp = ac->entry[--ac->avail];
925 /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
926 if (unlikely(is_obj_pfmemalloc(objp))) {
927 struct kmem_cache_node *l3;
929 if (gfp_pfmemalloc_allowed(flags)) {
930 clear_obj_pfmemalloc(&objp);
934 /* The caller cannot use PFMEMALLOC objects, find another one */
935 for (i = 0; i < ac->avail; i++) {
936 /* If a !PFMEMALLOC object is found, swap them */
937 if (!is_obj_pfmemalloc(ac->entry[i])) {
939 ac->entry[i] = ac->entry[ac->avail];
940 ac->entry[ac->avail] = objp;
946 * If there are empty slabs on the slabs_free list and we are
947 * being forced to refill the cache, mark this one !pfmemalloc.
949 l3 = cachep->node[numa_mem_id()];
950 if (!list_empty(&l3->slabs_free) && force_refill) {
951 struct slab *slabp = virt_to_slab(objp);
952 ClearPageSlabPfmemalloc(virt_to_head_page(slabp->s_mem));
953 clear_obj_pfmemalloc(&objp);
954 recheck_pfmemalloc_active(cachep, ac);
958 /* No !PFMEMALLOC objects available */
966 static inline void *ac_get_obj(struct kmem_cache *cachep,
967 struct array_cache *ac, gfp_t flags, bool force_refill)
971 if (unlikely(sk_memalloc_socks()))
972 objp = __ac_get_obj(cachep, ac, flags, force_refill);
974 objp = ac->entry[--ac->avail];
979 static void *__ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
982 if (unlikely(pfmemalloc_active)) {
983 /* Some pfmemalloc slabs exist, check if this is one */
984 struct page *page = virt_to_head_page(objp);
985 if (PageSlabPfmemalloc(page))
986 set_obj_pfmemalloc(&objp);
992 static inline void ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
995 if (unlikely(sk_memalloc_socks()))
996 objp = __ac_put_obj(cachep, ac, objp);
998 ac->entry[ac->avail++] = objp;
1002 * Transfer objects in one arraycache to another.
1003 * Locking must be handled by the caller.
1005 * Return the number of entries transferred.
1007 static int transfer_objects(struct array_cache *to,
1008 struct array_cache *from, unsigned int max)
1010 /* Figure out how many entries to transfer */
1011 int nr = min3(from->avail, max, to->limit - to->avail);
1016 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
1017 sizeof(void *) *nr);
1026 #define drain_alien_cache(cachep, alien) do { } while (0)
1027 #define reap_alien(cachep, l3) do { } while (0)
1029 static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
1031 return (struct array_cache **)BAD_ALIEN_MAGIC;
1034 static inline void free_alien_cache(struct array_cache **ac_ptr)
1038 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1043 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
1049 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
1050 gfp_t flags, int nodeid)
1055 #else /* CONFIG_NUMA */
1057 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
1058 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
1060 static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
1062 struct array_cache **ac_ptr;
1063 int memsize = sizeof(void *) * nr_node_ids;
1068 ac_ptr = kzalloc_node(memsize, gfp, node);
1071 if (i == node || !node_online(i))
1073 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
1075 for (i--; i >= 0; i--)
1085 static void free_alien_cache(struct array_cache **ac_ptr)
1096 static void __drain_alien_cache(struct kmem_cache *cachep,
1097 struct array_cache *ac, int node)
1099 struct kmem_cache_node *rl3 = cachep->node[node];
1102 spin_lock(&rl3->list_lock);
1104 * Stuff objects into the remote nodes shared array first.
1105 * That way we could avoid the overhead of putting the objects
1106 * into the free lists and getting them back later.
1109 transfer_objects(rl3->shared, ac, ac->limit);
1111 free_block(cachep, ac->entry, ac->avail, node);
1113 spin_unlock(&rl3->list_lock);
1118 * Called from cache_reap() to regularly drain alien caches round robin.
1120 static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *l3)
1122 int node = __this_cpu_read(slab_reap_node);
1125 struct array_cache *ac = l3->alien[node];
1127 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1128 __drain_alien_cache(cachep, ac, node);
1129 spin_unlock_irq(&ac->lock);
1134 static void drain_alien_cache(struct kmem_cache *cachep,
1135 struct array_cache **alien)
1138 struct array_cache *ac;
1139 unsigned long flags;
1141 for_each_online_node(i) {
1144 spin_lock_irqsave(&ac->lock, flags);
1145 __drain_alien_cache(cachep, ac, i);
1146 spin_unlock_irqrestore(&ac->lock, flags);
1151 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1153 struct slab *slabp = virt_to_slab(objp);
1154 int nodeid = slabp->nodeid;
1155 struct kmem_cache_node *l3;
1156 struct array_cache *alien = NULL;
1159 node = numa_mem_id();
1162 * Make sure we are not freeing a object from another node to the array
1163 * cache on this cpu.
1165 if (likely(slabp->nodeid == node))
1168 l3 = cachep->node[node];
1169 STATS_INC_NODEFREES(cachep);
1170 if (l3->alien && l3->alien[nodeid]) {
1171 alien = l3->alien[nodeid];
1172 spin_lock(&alien->lock);
1173 if (unlikely(alien->avail == alien->limit)) {
1174 STATS_INC_ACOVERFLOW(cachep);
1175 __drain_alien_cache(cachep, alien, nodeid);
1177 ac_put_obj(cachep, alien, objp);
1178 spin_unlock(&alien->lock);
1180 spin_lock(&(cachep->node[nodeid])->list_lock);
1181 free_block(cachep, &objp, 1, nodeid);
1182 spin_unlock(&(cachep->node[nodeid])->list_lock);
1189 * Allocates and initializes node for a node on each slab cache, used for
1190 * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3
1191 * will be allocated off-node since memory is not yet online for the new node.
1192 * When hotplugging memory or a cpu, existing node are not replaced if
1195 * Must hold slab_mutex.
1197 static int init_cache_node_node(int node)
1199 struct kmem_cache *cachep;
1200 struct kmem_cache_node *l3;
1201 const int memsize = sizeof(struct kmem_cache_node);
1203 list_for_each_entry(cachep, &slab_caches, list) {
1205 * Set up the size64 kmemlist for cpu before we can
1206 * begin anything. Make sure some other cpu on this
1207 * node has not already allocated this
1209 if (!cachep->node[node]) {
1210 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1213 kmem_list3_init(l3);
1214 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1215 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1218 * The l3s don't come and go as CPUs come and
1219 * go. slab_mutex is sufficient
1222 cachep->node[node] = l3;
1225 spin_lock_irq(&cachep->node[node]->list_lock);
1226 cachep->node[node]->free_limit =
1227 (1 + nr_cpus_node(node)) *
1228 cachep->batchcount + cachep->num;
1229 spin_unlock_irq(&cachep->node[node]->list_lock);
1234 static void __cpuinit cpuup_canceled(long cpu)
1236 struct kmem_cache *cachep;
1237 struct kmem_cache_node *l3 = NULL;
1238 int node = cpu_to_mem(cpu);
1239 const struct cpumask *mask = cpumask_of_node(node);
1241 list_for_each_entry(cachep, &slab_caches, list) {
1242 struct array_cache *nc;
1243 struct array_cache *shared;
1244 struct array_cache **alien;
1246 /* cpu is dead; no one can alloc from it. */
1247 nc = cachep->array[cpu];
1248 cachep->array[cpu] = NULL;
1249 l3 = cachep->node[node];
1252 goto free_array_cache;
1254 spin_lock_irq(&l3->list_lock);
1256 /* Free limit for this kmem_list3 */
1257 l3->free_limit -= cachep->batchcount;
1259 free_block(cachep, nc->entry, nc->avail, node);
1261 if (!cpumask_empty(mask)) {
1262 spin_unlock_irq(&l3->list_lock);
1263 goto free_array_cache;
1266 shared = l3->shared;
1268 free_block(cachep, shared->entry,
1269 shared->avail, node);
1276 spin_unlock_irq(&l3->list_lock);
1280 drain_alien_cache(cachep, alien);
1281 free_alien_cache(alien);
1287 * In the previous loop, all the objects were freed to
1288 * the respective cache's slabs, now we can go ahead and
1289 * shrink each nodelist to its limit.
1291 list_for_each_entry(cachep, &slab_caches, list) {
1292 l3 = cachep->node[node];
1295 drain_freelist(cachep, l3, l3->free_objects);
1299 static int __cpuinit cpuup_prepare(long cpu)
1301 struct kmem_cache *cachep;
1302 struct kmem_cache_node *l3 = NULL;
1303 int node = cpu_to_mem(cpu);
1307 * We need to do this right in the beginning since
1308 * alloc_arraycache's are going to use this list.
1309 * kmalloc_node allows us to add the slab to the right
1310 * kmem_list3 and not this cpu's kmem_list3
1312 err = init_cache_node_node(node);
1317 * Now we can go ahead with allocating the shared arrays and
1320 list_for_each_entry(cachep, &slab_caches, list) {
1321 struct array_cache *nc;
1322 struct array_cache *shared = NULL;
1323 struct array_cache **alien = NULL;
1325 nc = alloc_arraycache(node, cachep->limit,
1326 cachep->batchcount, GFP_KERNEL);
1329 if (cachep->shared) {
1330 shared = alloc_arraycache(node,
1331 cachep->shared * cachep->batchcount,
1332 0xbaadf00d, GFP_KERNEL);
1338 if (use_alien_caches) {
1339 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
1346 cachep->array[cpu] = nc;
1347 l3 = cachep->node[node];
1350 spin_lock_irq(&l3->list_lock);
1353 * We are serialised from CPU_DEAD or
1354 * CPU_UP_CANCELLED by the cpucontrol lock
1356 l3->shared = shared;
1365 spin_unlock_irq(&l3->list_lock);
1367 free_alien_cache(alien);
1368 if (cachep->flags & SLAB_DEBUG_OBJECTS)
1369 slab_set_debugobj_lock_classes_node(cachep, node);
1370 else if (!OFF_SLAB(cachep) &&
1371 !(cachep->flags & SLAB_DESTROY_BY_RCU))
1372 on_slab_lock_classes_node(cachep, node);
1374 init_node_lock_keys(node);
1378 cpuup_canceled(cpu);
1382 static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1383 unsigned long action, void *hcpu)
1385 long cpu = (long)hcpu;
1389 case CPU_UP_PREPARE:
1390 case CPU_UP_PREPARE_FROZEN:
1391 mutex_lock(&slab_mutex);
1392 err = cpuup_prepare(cpu);
1393 mutex_unlock(&slab_mutex);
1396 case CPU_ONLINE_FROZEN:
1397 start_cpu_timer(cpu);
1399 #ifdef CONFIG_HOTPLUG_CPU
1400 case CPU_DOWN_PREPARE:
1401 case CPU_DOWN_PREPARE_FROZEN:
1403 * Shutdown cache reaper. Note that the slab_mutex is
1404 * held so that if cache_reap() is invoked it cannot do
1405 * anything expensive but will only modify reap_work
1406 * and reschedule the timer.
1408 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1409 /* Now the cache_reaper is guaranteed to be not running. */
1410 per_cpu(slab_reap_work, cpu).work.func = NULL;
1412 case CPU_DOWN_FAILED:
1413 case CPU_DOWN_FAILED_FROZEN:
1414 start_cpu_timer(cpu);
1417 case CPU_DEAD_FROZEN:
1419 * Even if all the cpus of a node are down, we don't free the
1420 * kmem_list3 of any cache. This to avoid a race between
1421 * cpu_down, and a kmalloc allocation from another cpu for
1422 * memory from the node of the cpu going down. The list3
1423 * structure is usually allocated from kmem_cache_create() and
1424 * gets destroyed at kmem_cache_destroy().
1428 case CPU_UP_CANCELED:
1429 case CPU_UP_CANCELED_FROZEN:
1430 mutex_lock(&slab_mutex);
1431 cpuup_canceled(cpu);
1432 mutex_unlock(&slab_mutex);
1435 return notifier_from_errno(err);
1438 static struct notifier_block __cpuinitdata cpucache_notifier = {
1439 &cpuup_callback, NULL, 0
1442 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1444 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1445 * Returns -EBUSY if all objects cannot be drained so that the node is not
1448 * Must hold slab_mutex.
1450 static int __meminit drain_cache_node_node(int node)
1452 struct kmem_cache *cachep;
1455 list_for_each_entry(cachep, &slab_caches, list) {
1456 struct kmem_cache_node *l3;
1458 l3 = cachep->node[node];
1462 drain_freelist(cachep, l3, l3->free_objects);
1464 if (!list_empty(&l3->slabs_full) ||
1465 !list_empty(&l3->slabs_partial)) {
1473 static int __meminit slab_memory_callback(struct notifier_block *self,
1474 unsigned long action, void *arg)
1476 struct memory_notify *mnb = arg;
1480 nid = mnb->status_change_nid;
1485 case MEM_GOING_ONLINE:
1486 mutex_lock(&slab_mutex);
1487 ret = init_cache_node_node(nid);
1488 mutex_unlock(&slab_mutex);
1490 case MEM_GOING_OFFLINE:
1491 mutex_lock(&slab_mutex);
1492 ret = drain_cache_node_node(nid);
1493 mutex_unlock(&slab_mutex);
1497 case MEM_CANCEL_ONLINE:
1498 case MEM_CANCEL_OFFLINE:
1502 return notifier_from_errno(ret);
1504 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1507 * swap the static kmem_list3 with kmalloced memory
1509 static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list,
1512 struct kmem_cache_node *ptr;
1514 ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid);
1517 memcpy(ptr, list, sizeof(struct kmem_cache_node));
1519 * Do not assume that spinlocks can be initialized via memcpy:
1521 spin_lock_init(&ptr->list_lock);
1523 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1524 cachep->node[nodeid] = ptr;
1528 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1529 * size of kmem_list3.
1531 static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1535 for_each_online_node(node) {
1536 cachep->node[node] = &initkmem_list3[index + node];
1537 cachep->node[node]->next_reap = jiffies +
1539 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1544 * The memory after the last cpu cache pointer is used for the
1547 static void setup_node_pointer(struct kmem_cache *cachep)
1549 cachep->node = (struct kmem_cache_node **)&cachep->array[nr_cpu_ids];
1553 * Initialisation. Called after the page allocator have been initialised and
1554 * before smp_init().
1556 void __init kmem_cache_init(void)
1560 kmem_cache = &kmem_cache_boot;
1561 setup_node_pointer(kmem_cache);
1563 if (num_possible_nodes() == 1)
1564 use_alien_caches = 0;
1566 for (i = 0; i < NUM_INIT_LISTS; i++)
1567 kmem_list3_init(&initkmem_list3[i]);
1569 set_up_list3s(kmem_cache, CACHE_CACHE);
1572 * Fragmentation resistance on low memory - only use bigger
1573 * page orders on machines with more than 32MB of memory if
1574 * not overridden on the command line.
1576 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
1577 slab_max_order = SLAB_MAX_ORDER_HI;
1579 /* Bootstrap is tricky, because several objects are allocated
1580 * from caches that do not exist yet:
1581 * 1) initialize the kmem_cache cache: it contains the struct
1582 * kmem_cache structures of all caches, except kmem_cache itself:
1583 * kmem_cache is statically allocated.
1584 * Initially an __init data area is used for the head array and the
1585 * kmem_list3 structures, it's replaced with a kmalloc allocated
1586 * array at the end of the bootstrap.
1587 * 2) Create the first kmalloc cache.
1588 * The struct kmem_cache for the new cache is allocated normally.
1589 * An __init data area is used for the head array.
1590 * 3) Create the remaining kmalloc caches, with minimally sized
1592 * 4) Replace the __init data head arrays for kmem_cache and the first
1593 * kmalloc cache with kmalloc allocated arrays.
1594 * 5) Replace the __init data for kmem_list3 for kmem_cache and
1595 * the other cache's with kmalloc allocated memory.
1596 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1599 /* 1) create the kmem_cache */
1602 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1604 create_boot_cache(kmem_cache, "kmem_cache",
1605 offsetof(struct kmem_cache, array[nr_cpu_ids]) +
1606 nr_node_ids * sizeof(struct kmem_cache_node *),
1607 SLAB_HWCACHE_ALIGN);
1608 list_add(&kmem_cache->list, &slab_caches);
1610 /* 2+3) create the kmalloc caches */
1613 * Initialize the caches that provide memory for the array cache and the
1614 * kmem_list3 structures first. Without this, further allocations will
1618 kmalloc_caches[INDEX_AC] = create_kmalloc_cache("kmalloc-ac",
1619 kmalloc_size(INDEX_AC), ARCH_KMALLOC_FLAGS);
1621 if (INDEX_AC != INDEX_L3)
1622 kmalloc_caches[INDEX_L3] =
1623 create_kmalloc_cache("kmalloc-l3",
1624 kmalloc_size(INDEX_L3), ARCH_KMALLOC_FLAGS);
1626 slab_early_init = 0;
1628 /* 4) Replace the bootstrap head arrays */
1630 struct array_cache *ptr;
1632 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1634 memcpy(ptr, cpu_cache_get(kmem_cache),
1635 sizeof(struct arraycache_init));
1637 * Do not assume that spinlocks can be initialized via memcpy:
1639 spin_lock_init(&ptr->lock);
1641 kmem_cache->array[smp_processor_id()] = ptr;
1643 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1645 BUG_ON(cpu_cache_get(kmalloc_caches[INDEX_AC])
1646 != &initarray_generic.cache);
1647 memcpy(ptr, cpu_cache_get(kmalloc_caches[INDEX_AC]),
1648 sizeof(struct arraycache_init));
1650 * Do not assume that spinlocks can be initialized via memcpy:
1652 spin_lock_init(&ptr->lock);
1654 kmalloc_caches[INDEX_AC]->array[smp_processor_id()] = ptr;
1656 /* 5) Replace the bootstrap kmem_list3's */
1660 for_each_online_node(nid) {
1661 init_list(kmem_cache, &initkmem_list3[CACHE_CACHE + nid], nid);
1663 init_list(kmalloc_caches[INDEX_AC],
1664 &initkmem_list3[SIZE_AC + nid], nid);
1666 if (INDEX_AC != INDEX_L3) {
1667 init_list(kmalloc_caches[INDEX_L3],
1668 &initkmem_list3[SIZE_L3 + nid], nid);
1673 create_kmalloc_caches(ARCH_KMALLOC_FLAGS);
1676 void __init kmem_cache_init_late(void)
1678 struct kmem_cache *cachep;
1682 /* 6) resize the head arrays to their final sizes */
1683 mutex_lock(&slab_mutex);
1684 list_for_each_entry(cachep, &slab_caches, list)
1685 if (enable_cpucache(cachep, GFP_NOWAIT))
1687 mutex_unlock(&slab_mutex);
1689 /* Annotate slab for lockdep -- annotate the malloc caches */
1696 * Register a cpu startup notifier callback that initializes
1697 * cpu_cache_get for all new cpus
1699 register_cpu_notifier(&cpucache_notifier);
1703 * Register a memory hotplug callback that initializes and frees
1706 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1710 * The reap timers are started later, with a module init call: That part
1711 * of the kernel is not yet operational.
1715 static int __init cpucache_init(void)
1720 * Register the timers that return unneeded pages to the page allocator
1722 for_each_online_cpu(cpu)
1723 start_cpu_timer(cpu);
1729 __initcall(cpucache_init);
1731 static noinline void
1732 slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1734 struct kmem_cache_node *l3;
1736 unsigned long flags;
1740 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1742 printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n",
1743 cachep->name, cachep->size, cachep->gfporder);
1745 for_each_online_node(node) {
1746 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1747 unsigned long active_slabs = 0, num_slabs = 0;
1749 l3 = cachep->node[node];
1753 spin_lock_irqsave(&l3->list_lock, flags);
1754 list_for_each_entry(slabp, &l3->slabs_full, list) {
1755 active_objs += cachep->num;
1758 list_for_each_entry(slabp, &l3->slabs_partial, list) {
1759 active_objs += slabp->inuse;
1762 list_for_each_entry(slabp, &l3->slabs_free, list)
1765 free_objects += l3->free_objects;
1766 spin_unlock_irqrestore(&l3->list_lock, flags);
1768 num_slabs += active_slabs;
1769 num_objs = num_slabs * cachep->num;
1771 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1772 node, active_slabs, num_slabs, active_objs, num_objs,
1778 * Interface to system's page allocator. No need to hold the cache-lock.
1780 * If we requested dmaable memory, we will get it. Even if we
1781 * did not request dmaable memory, we might get it, but that
1782 * would be relatively rare and ignorable.
1784 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1792 * Nommu uses slab's for process anonymous memory allocations, and thus
1793 * requires __GFP_COMP to properly refcount higher order allocations
1795 flags |= __GFP_COMP;
1798 flags |= cachep->allocflags;
1799 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1800 flags |= __GFP_RECLAIMABLE;
1802 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1804 if (!(flags & __GFP_NOWARN) && printk_ratelimit())
1805 slab_out_of_memory(cachep, flags, nodeid);
1809 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1810 if (unlikely(page->pfmemalloc))
1811 pfmemalloc_active = true;
1813 nr_pages = (1 << cachep->gfporder);
1814 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1815 add_zone_page_state(page_zone(page),
1816 NR_SLAB_RECLAIMABLE, nr_pages);
1818 add_zone_page_state(page_zone(page),
1819 NR_SLAB_UNRECLAIMABLE, nr_pages);
1820 for (i = 0; i < nr_pages; i++) {
1821 __SetPageSlab(page + i);
1823 if (page->pfmemalloc)
1824 SetPageSlabPfmemalloc(page + i);
1826 memcg_bind_pages(cachep, cachep->gfporder);
1828 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1829 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1832 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1834 kmemcheck_mark_unallocated_pages(page, nr_pages);
1837 return page_address(page);
1841 * Interface to system's page release.
1843 static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1845 unsigned long i = (1 << cachep->gfporder);
1846 struct page *page = virt_to_page(addr);
1847 const unsigned long nr_freed = i;
1849 kmemcheck_free_shadow(page, cachep->gfporder);
1851 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1852 sub_zone_page_state(page_zone(page),
1853 NR_SLAB_RECLAIMABLE, nr_freed);
1855 sub_zone_page_state(page_zone(page),
1856 NR_SLAB_UNRECLAIMABLE, nr_freed);
1858 BUG_ON(!PageSlab(page));
1859 __ClearPageSlabPfmemalloc(page);
1860 __ClearPageSlab(page);
1864 memcg_release_pages(cachep, cachep->gfporder);
1865 if (current->reclaim_state)
1866 current->reclaim_state->reclaimed_slab += nr_freed;
1867 free_memcg_kmem_pages((unsigned long)addr, cachep->gfporder);
1870 static void kmem_rcu_free(struct rcu_head *head)
1872 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1873 struct kmem_cache *cachep = slab_rcu->cachep;
1875 kmem_freepages(cachep, slab_rcu->addr);
1876 if (OFF_SLAB(cachep))
1877 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1882 #ifdef CONFIG_DEBUG_PAGEALLOC
1883 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1884 unsigned long caller)
1886 int size = cachep->object_size;
1888 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1890 if (size < 5 * sizeof(unsigned long))
1893 *addr++ = 0x12345678;
1895 *addr++ = smp_processor_id();
1896 size -= 3 * sizeof(unsigned long);
1898 unsigned long *sptr = &caller;
1899 unsigned long svalue;
1901 while (!kstack_end(sptr)) {
1903 if (kernel_text_address(svalue)) {
1905 size -= sizeof(unsigned long);
1906 if (size <= sizeof(unsigned long))
1912 *addr++ = 0x87654321;
1916 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1918 int size = cachep->object_size;
1919 addr = &((char *)addr)[obj_offset(cachep)];
1921 memset(addr, val, size);
1922 *(unsigned char *)(addr + size - 1) = POISON_END;
1925 static void dump_line(char *data, int offset, int limit)
1928 unsigned char error = 0;
1931 printk(KERN_ERR "%03x: ", offset);
1932 for (i = 0; i < limit; i++) {
1933 if (data[offset + i] != POISON_FREE) {
1934 error = data[offset + i];
1938 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1939 &data[offset], limit, 1);
1941 if (bad_count == 1) {
1942 error ^= POISON_FREE;
1943 if (!(error & (error - 1))) {
1944 printk(KERN_ERR "Single bit error detected. Probably "
1947 printk(KERN_ERR "Run memtest86+ or a similar memory "
1950 printk(KERN_ERR "Run a memory test tool.\n");
1959 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1964 if (cachep->flags & SLAB_RED_ZONE) {
1965 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
1966 *dbg_redzone1(cachep, objp),
1967 *dbg_redzone2(cachep, objp));
1970 if (cachep->flags & SLAB_STORE_USER) {
1971 printk(KERN_ERR "Last user: [<%p>]",
1972 *dbg_userword(cachep, objp));
1973 print_symbol("(%s)",
1974 (unsigned long)*dbg_userword(cachep, objp));
1977 realobj = (char *)objp + obj_offset(cachep);
1978 size = cachep->object_size;
1979 for (i = 0; i < size && lines; i += 16, lines--) {
1982 if (i + limit > size)
1984 dump_line(realobj, i, limit);
1988 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1994 realobj = (char *)objp + obj_offset(cachep);
1995 size = cachep->object_size;
1997 for (i = 0; i < size; i++) {
1998 char exp = POISON_FREE;
2001 if (realobj[i] != exp) {
2007 "Slab corruption (%s): %s start=%p, len=%d\n",
2008 print_tainted(), cachep->name, realobj, size);
2009 print_objinfo(cachep, objp, 0);
2011 /* Hexdump the affected line */
2014 if (i + limit > size)
2016 dump_line(realobj, i, limit);
2019 /* Limit to 5 lines */
2025 /* Print some data about the neighboring objects, if they
2028 struct slab *slabp = virt_to_slab(objp);
2031 objnr = obj_to_index(cachep, slabp, objp);
2033 objp = index_to_obj(cachep, slabp, objnr - 1);
2034 realobj = (char *)objp + obj_offset(cachep);
2035 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
2037 print_objinfo(cachep, objp, 2);
2039 if (objnr + 1 < cachep->num) {
2040 objp = index_to_obj(cachep, slabp, objnr + 1);
2041 realobj = (char *)objp + obj_offset(cachep);
2042 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
2044 print_objinfo(cachep, objp, 2);
2051 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
2054 for (i = 0; i < cachep->num; i++) {
2055 void *objp = index_to_obj(cachep, slabp, i);
2057 if (cachep->flags & SLAB_POISON) {
2058 #ifdef CONFIG_DEBUG_PAGEALLOC
2059 if (cachep->size % PAGE_SIZE == 0 &&
2061 kernel_map_pages(virt_to_page(objp),
2062 cachep->size / PAGE_SIZE, 1);
2064 check_poison_obj(cachep, objp);
2066 check_poison_obj(cachep, objp);
2069 if (cachep->flags & SLAB_RED_ZONE) {
2070 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2071 slab_error(cachep, "start of a freed object "
2073 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2074 slab_error(cachep, "end of a freed object "
2080 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
2086 * slab_destroy - destroy and release all objects in a slab
2087 * @cachep: cache pointer being destroyed
2088 * @slabp: slab pointer being destroyed
2090 * Destroy all the objs in a slab, and release the mem back to the system.
2091 * Before calling the slab must have been unlinked from the cache. The
2092 * cache-lock is not held/needed.
2094 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
2096 void *addr = slabp->s_mem - slabp->colouroff;
2098 slab_destroy_debugcheck(cachep, slabp);
2099 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
2100 struct slab_rcu *slab_rcu;
2102 slab_rcu = (struct slab_rcu *)slabp;
2103 slab_rcu->cachep = cachep;
2104 slab_rcu->addr = addr;
2105 call_rcu(&slab_rcu->head, kmem_rcu_free);
2107 kmem_freepages(cachep, addr);
2108 if (OFF_SLAB(cachep))
2109 kmem_cache_free(cachep->slabp_cache, slabp);
2114 * calculate_slab_order - calculate size (page order) of slabs
2115 * @cachep: pointer to the cache that is being created
2116 * @size: size of objects to be created in this cache.
2117 * @align: required alignment for the objects.
2118 * @flags: slab allocation flags
2120 * Also calculates the number of objects per slab.
2122 * This could be made much more intelligent. For now, try to avoid using
2123 * high order pages for slabs. When the gfp() functions are more friendly
2124 * towards high-order requests, this should be changed.
2126 static size_t calculate_slab_order(struct kmem_cache *cachep,
2127 size_t size, size_t align, unsigned long flags)
2129 unsigned long offslab_limit;
2130 size_t left_over = 0;
2133 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
2137 cache_estimate(gfporder, size, align, flags, &remainder, &num);
2141 if (flags & CFLGS_OFF_SLAB) {
2143 * Max number of objs-per-slab for caches which
2144 * use off-slab slabs. Needed to avoid a possible
2145 * looping condition in cache_grow().
2147 offslab_limit = size - sizeof(struct slab);
2148 offslab_limit /= sizeof(kmem_bufctl_t);
2150 if (num > offslab_limit)
2154 /* Found something acceptable - save it away */
2156 cachep->gfporder = gfporder;
2157 left_over = remainder;
2160 * A VFS-reclaimable slab tends to have most allocations
2161 * as GFP_NOFS and we really don't want to have to be allocating
2162 * higher-order pages when we are unable to shrink dcache.
2164 if (flags & SLAB_RECLAIM_ACCOUNT)
2168 * Large number of objects is good, but very large slabs are
2169 * currently bad for the gfp()s.
2171 if (gfporder >= slab_max_order)
2175 * Acceptable internal fragmentation?
2177 if (left_over * 8 <= (PAGE_SIZE << gfporder))
2183 static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
2185 if (slab_state >= FULL)
2186 return enable_cpucache(cachep, gfp);
2188 if (slab_state == DOWN) {
2190 * Note: Creation of first cache (kmem_cache).
2191 * The setup_list3s is taken care
2192 * of by the caller of __kmem_cache_create
2194 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2195 slab_state = PARTIAL;
2196 } else if (slab_state == PARTIAL) {
2198 * Note: the second kmem_cache_create must create the cache
2199 * that's used by kmalloc(24), otherwise the creation of
2200 * further caches will BUG().
2202 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2205 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2206 * the second cache, then we need to set up all its list3s,
2207 * otherwise the creation of further caches will BUG().
2209 set_up_list3s(cachep, SIZE_AC);
2210 if (INDEX_AC == INDEX_L3)
2211 slab_state = PARTIAL_L3;
2213 slab_state = PARTIAL_ARRAYCACHE;
2215 /* Remaining boot caches */
2216 cachep->array[smp_processor_id()] =
2217 kmalloc(sizeof(struct arraycache_init), gfp);
2219 if (slab_state == PARTIAL_ARRAYCACHE) {
2220 set_up_list3s(cachep, SIZE_L3);
2221 slab_state = PARTIAL_L3;
2224 for_each_online_node(node) {
2225 cachep->node[node] =
2226 kmalloc_node(sizeof(struct kmem_cache_node),
2228 BUG_ON(!cachep->node[node]);
2229 kmem_list3_init(cachep->node[node]);
2233 cachep->node[numa_mem_id()]->next_reap =
2234 jiffies + REAPTIMEOUT_LIST3 +
2235 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2237 cpu_cache_get(cachep)->avail = 0;
2238 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2239 cpu_cache_get(cachep)->batchcount = 1;
2240 cpu_cache_get(cachep)->touched = 0;
2241 cachep->batchcount = 1;
2242 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2247 * __kmem_cache_create - Create a cache.
2248 * @cachep: cache management descriptor
2249 * @flags: SLAB flags
2251 * Returns a ptr to the cache on success, NULL on failure.
2252 * Cannot be called within a int, but can be interrupted.
2253 * The @ctor is run when new pages are allocated by the cache.
2257 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2258 * to catch references to uninitialised memory.
2260 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2261 * for buffer overruns.
2263 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2264 * cacheline. This can be beneficial if you're counting cycles as closely
2268 __kmem_cache_create (struct kmem_cache *cachep, unsigned long flags)
2270 size_t left_over, slab_size, ralign;
2273 size_t size = cachep->size;
2278 * Enable redzoning and last user accounting, except for caches with
2279 * large objects, if the increased size would increase the object size
2280 * above the next power of two: caches with object sizes just above a
2281 * power of two have a significant amount of internal fragmentation.
2283 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2284 2 * sizeof(unsigned long long)))
2285 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2286 if (!(flags & SLAB_DESTROY_BY_RCU))
2287 flags |= SLAB_POISON;
2289 if (flags & SLAB_DESTROY_BY_RCU)
2290 BUG_ON(flags & SLAB_POISON);
2294 * Check that size is in terms of words. This is needed to avoid
2295 * unaligned accesses for some archs when redzoning is used, and makes
2296 * sure any on-slab bufctl's are also correctly aligned.
2298 if (size & (BYTES_PER_WORD - 1)) {
2299 size += (BYTES_PER_WORD - 1);
2300 size &= ~(BYTES_PER_WORD - 1);
2304 * Redzoning and user store require word alignment or possibly larger.
2305 * Note this will be overridden by architecture or caller mandated
2306 * alignment if either is greater than BYTES_PER_WORD.
2308 if (flags & SLAB_STORE_USER)
2309 ralign = BYTES_PER_WORD;
2311 if (flags & SLAB_RED_ZONE) {
2312 ralign = REDZONE_ALIGN;
2313 /* If redzoning, ensure that the second redzone is suitably
2314 * aligned, by adjusting the object size accordingly. */
2315 size += REDZONE_ALIGN - 1;
2316 size &= ~(REDZONE_ALIGN - 1);
2319 /* 3) caller mandated alignment */
2320 if (ralign < cachep->align) {
2321 ralign = cachep->align;
2323 /* disable debug if necessary */
2324 if (ralign > __alignof__(unsigned long long))
2325 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2329 cachep->align = ralign;
2331 if (slab_is_available())
2336 setup_node_pointer(cachep);
2340 * Both debugging options require word-alignment which is calculated
2343 if (flags & SLAB_RED_ZONE) {
2344 /* add space for red zone words */
2345 cachep->obj_offset += sizeof(unsigned long long);
2346 size += 2 * sizeof(unsigned long long);
2348 if (flags & SLAB_STORE_USER) {
2349 /* user store requires one word storage behind the end of
2350 * the real object. But if the second red zone needs to be
2351 * aligned to 64 bits, we must allow that much space.
2353 if (flags & SLAB_RED_ZONE)
2354 size += REDZONE_ALIGN;
2356 size += BYTES_PER_WORD;
2358 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2359 if (size >= kmalloc_size(INDEX_L3 + 1)
2360 && cachep->object_size > cache_line_size() && ALIGN(size, align) < PAGE_SIZE) {
2361 cachep->obj_offset += PAGE_SIZE - ALIGN(size, align);
2368 * Determine if the slab management is 'on' or 'off' slab.
2369 * (bootstrapping cannot cope with offslab caches so don't do
2370 * it too early on. Always use on-slab management when
2371 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2373 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
2374 !(flags & SLAB_NOLEAKTRACE))
2376 * Size is large, assume best to place the slab management obj
2377 * off-slab (should allow better packing of objs).
2379 flags |= CFLGS_OFF_SLAB;
2381 size = ALIGN(size, cachep->align);
2383 left_over = calculate_slab_order(cachep, size, cachep->align, flags);
2388 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2389 + sizeof(struct slab), cachep->align);
2392 * If the slab has been placed off-slab, and we have enough space then
2393 * move it on-slab. This is at the expense of any extra colouring.
2395 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2396 flags &= ~CFLGS_OFF_SLAB;
2397 left_over -= slab_size;
2400 if (flags & CFLGS_OFF_SLAB) {
2401 /* really off slab. No need for manual alignment */
2403 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2405 #ifdef CONFIG_PAGE_POISONING
2406 /* If we're going to use the generic kernel_map_pages()
2407 * poisoning, then it's going to smash the contents of
2408 * the redzone and userword anyhow, so switch them off.
2410 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2411 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2415 cachep->colour_off = cache_line_size();
2416 /* Offset must be a multiple of the alignment. */
2417 if (cachep->colour_off < cachep->align)
2418 cachep->colour_off = cachep->align;
2419 cachep->colour = left_over / cachep->colour_off;
2420 cachep->slab_size = slab_size;
2421 cachep->flags = flags;
2422 cachep->allocflags = 0;
2423 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2424 cachep->allocflags |= GFP_DMA;
2425 cachep->size = size;
2426 cachep->reciprocal_buffer_size = reciprocal_value(size);
2428 if (flags & CFLGS_OFF_SLAB) {
2429 cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u);
2431 * This is a possibility for one of the malloc_sizes caches.
2432 * But since we go off slab only for object size greater than
2433 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2434 * this should not happen at all.
2435 * But leave a BUG_ON for some lucky dude.
2437 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
2440 err = setup_cpu_cache(cachep, gfp);
2442 __kmem_cache_shutdown(cachep);
2446 if (flags & SLAB_DEBUG_OBJECTS) {
2448 * Would deadlock through slab_destroy()->call_rcu()->
2449 * debug_object_activate()->kmem_cache_alloc().
2451 WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU);
2453 slab_set_debugobj_lock_classes(cachep);
2454 } else if (!OFF_SLAB(cachep) && !(flags & SLAB_DESTROY_BY_RCU))
2455 on_slab_lock_classes(cachep);
2461 static void check_irq_off(void)
2463 BUG_ON(!irqs_disabled());
2466 static void check_irq_on(void)
2468 BUG_ON(irqs_disabled());
2471 static void check_spinlock_acquired(struct kmem_cache *cachep)
2475 assert_spin_locked(&cachep->node[numa_mem_id()]->list_lock);
2479 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2483 assert_spin_locked(&cachep->node[node]->list_lock);
2488 #define check_irq_off() do { } while(0)
2489 #define check_irq_on() do { } while(0)
2490 #define check_spinlock_acquired(x) do { } while(0)
2491 #define check_spinlock_acquired_node(x, y) do { } while(0)
2494 static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *l3,
2495 struct array_cache *ac,
2496 int force, int node);
2498 static void do_drain(void *arg)
2500 struct kmem_cache *cachep = arg;
2501 struct array_cache *ac;
2502 int node = numa_mem_id();
2505 ac = cpu_cache_get(cachep);
2506 spin_lock(&cachep->node[node]->list_lock);
2507 free_block(cachep, ac->entry, ac->avail, node);
2508 spin_unlock(&cachep->node[node]->list_lock);
2512 static void drain_cpu_caches(struct kmem_cache *cachep)
2514 struct kmem_cache_node *l3;
2517 on_each_cpu(do_drain, cachep, 1);
2519 for_each_online_node(node) {
2520 l3 = cachep->node[node];
2521 if (l3 && l3->alien)
2522 drain_alien_cache(cachep, l3->alien);
2525 for_each_online_node(node) {
2526 l3 = cachep->node[node];
2528 drain_array(cachep, l3, l3->shared, 1, node);
2533 * Remove slabs from the list of free slabs.
2534 * Specify the number of slabs to drain in tofree.
2536 * Returns the actual number of slabs released.
2538 static int drain_freelist(struct kmem_cache *cache,
2539 struct kmem_cache_node *l3, int tofree)
2541 struct list_head *p;
2546 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
2548 spin_lock_irq(&l3->list_lock);
2549 p = l3->slabs_free.prev;
2550 if (p == &l3->slabs_free) {
2551 spin_unlock_irq(&l3->list_lock);
2555 slabp = list_entry(p, struct slab, list);
2557 BUG_ON(slabp->inuse);
2559 list_del(&slabp->list);
2561 * Safe to drop the lock. The slab is no longer linked
2564 l3->free_objects -= cache->num;
2565 spin_unlock_irq(&l3->list_lock);
2566 slab_destroy(cache, slabp);
2573 /* Called with slab_mutex held to protect against cpu hotplug */
2574 static int __cache_shrink(struct kmem_cache *cachep)
2577 struct kmem_cache_node *l3;
2579 drain_cpu_caches(cachep);
2582 for_each_online_node(i) {
2583 l3 = cachep->node[i];
2587 drain_freelist(cachep, l3, l3->free_objects);
2589 ret += !list_empty(&l3->slabs_full) ||
2590 !list_empty(&l3->slabs_partial);
2592 return (ret ? 1 : 0);
2596 * kmem_cache_shrink - Shrink a cache.
2597 * @cachep: The cache to shrink.
2599 * Releases as many slabs as possible for a cache.
2600 * To help debugging, a zero exit status indicates all slabs were released.
2602 int kmem_cache_shrink(struct kmem_cache *cachep)
2605 BUG_ON(!cachep || in_interrupt());
2608 mutex_lock(&slab_mutex);
2609 ret = __cache_shrink(cachep);
2610 mutex_unlock(&slab_mutex);
2614 EXPORT_SYMBOL(kmem_cache_shrink);
2616 int __kmem_cache_shutdown(struct kmem_cache *cachep)
2619 struct kmem_cache_node *l3;
2620 int rc = __cache_shrink(cachep);
2625 for_each_online_cpu(i)
2626 kfree(cachep->array[i]);
2628 /* NUMA: free the list3 structures */
2629 for_each_online_node(i) {
2630 l3 = cachep->node[i];
2633 free_alien_cache(l3->alien);
2641 * Get the memory for a slab management obj.
2642 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2643 * always come from malloc_sizes caches. The slab descriptor cannot
2644 * come from the same cache which is getting created because,
2645 * when we are searching for an appropriate cache for these
2646 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2647 * If we are creating a malloc_sizes cache here it would not be visible to
2648 * kmem_find_general_cachep till the initialization is complete.
2649 * Hence we cannot have slabp_cache same as the original cache.
2651 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2652 int colour_off, gfp_t local_flags,
2657 if (OFF_SLAB(cachep)) {
2658 /* Slab management obj is off-slab. */
2659 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2660 local_flags, nodeid);
2662 * If the first object in the slab is leaked (it's allocated
2663 * but no one has a reference to it), we want to make sure
2664 * kmemleak does not treat the ->s_mem pointer as a reference
2665 * to the object. Otherwise we will not report the leak.
2667 kmemleak_scan_area(&slabp->list, sizeof(struct list_head),
2672 slabp = objp + colour_off;
2673 colour_off += cachep->slab_size;
2676 slabp->colouroff = colour_off;
2677 slabp->s_mem = objp + colour_off;
2678 slabp->nodeid = nodeid;
2683 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2685 return (kmem_bufctl_t *) (slabp + 1);
2688 static void cache_init_objs(struct kmem_cache *cachep,
2693 for (i = 0; i < cachep->num; i++) {
2694 void *objp = index_to_obj(cachep, slabp, i);
2696 /* need to poison the objs? */
2697 if (cachep->flags & SLAB_POISON)
2698 poison_obj(cachep, objp, POISON_FREE);
2699 if (cachep->flags & SLAB_STORE_USER)
2700 *dbg_userword(cachep, objp) = NULL;
2702 if (cachep->flags & SLAB_RED_ZONE) {
2703 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2704 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2707 * Constructors are not allowed to allocate memory from the same
2708 * cache which they are a constructor for. Otherwise, deadlock.
2709 * They must also be threaded.
2711 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2712 cachep->ctor(objp + obj_offset(cachep));
2714 if (cachep->flags & SLAB_RED_ZONE) {
2715 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2716 slab_error(cachep, "constructor overwrote the"
2717 " end of an object");
2718 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2719 slab_error(cachep, "constructor overwrote the"
2720 " start of an object");
2722 if ((cachep->size % PAGE_SIZE) == 0 &&
2723 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2724 kernel_map_pages(virt_to_page(objp),
2725 cachep->size / PAGE_SIZE, 0);
2730 slab_bufctl(slabp)[i] = i + 1;
2732 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2735 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2737 if (CONFIG_ZONE_DMA_FLAG) {
2738 if (flags & GFP_DMA)
2739 BUG_ON(!(cachep->allocflags & GFP_DMA));
2741 BUG_ON(cachep->allocflags & GFP_DMA);
2745 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2748 void *objp = index_to_obj(cachep, slabp, slabp->free);
2752 next = slab_bufctl(slabp)[slabp->free];
2754 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2755 WARN_ON(slabp->nodeid != nodeid);
2762 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2763 void *objp, int nodeid)
2765 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2768 /* Verify that the slab belongs to the intended node */
2769 WARN_ON(slabp->nodeid != nodeid);
2771 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2772 printk(KERN_ERR "slab: double free detected in cache "
2773 "'%s', objp %p\n", cachep->name, objp);
2777 slab_bufctl(slabp)[objnr] = slabp->free;
2778 slabp->free = objnr;
2783 * Map pages beginning at addr to the given cache and slab. This is required
2784 * for the slab allocator to be able to lookup the cache and slab of a
2785 * virtual address for kfree, ksize, and slab debugging.
2787 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2793 page = virt_to_page(addr);
2796 if (likely(!PageCompound(page)))
2797 nr_pages <<= cache->gfporder;
2800 page->slab_cache = cache;
2801 page->slab_page = slab;
2803 } while (--nr_pages);
2807 * Grow (by 1) the number of slabs within a cache. This is called by
2808 * kmem_cache_alloc() when there are no active objs left in a cache.
2810 static int cache_grow(struct kmem_cache *cachep,
2811 gfp_t flags, int nodeid, void *objp)
2816 struct kmem_cache_node *l3;
2819 * Be lazy and only check for valid flags here, keeping it out of the
2820 * critical path in kmem_cache_alloc().
2822 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2823 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2825 /* Take the l3 list lock to change the colour_next on this node */
2827 l3 = cachep->node[nodeid];
2828 spin_lock(&l3->list_lock);
2830 /* Get colour for the slab, and cal the next value. */
2831 offset = l3->colour_next;
2833 if (l3->colour_next >= cachep->colour)
2834 l3->colour_next = 0;
2835 spin_unlock(&l3->list_lock);
2837 offset *= cachep->colour_off;
2839 if (local_flags & __GFP_WAIT)
2843 * The test for missing atomic flag is performed here, rather than
2844 * the more obvious place, simply to reduce the critical path length
2845 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2846 * will eventually be caught here (where it matters).
2848 kmem_flagcheck(cachep, flags);
2851 * Get mem for the objs. Attempt to allocate a physical page from
2855 objp = kmem_getpages(cachep, local_flags, nodeid);
2859 /* Get slab management. */
2860 slabp = alloc_slabmgmt(cachep, objp, offset,
2861 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2865 slab_map_pages(cachep, slabp, objp);
2867 cache_init_objs(cachep, slabp);
2869 if (local_flags & __GFP_WAIT)
2870 local_irq_disable();
2872 spin_lock(&l3->list_lock);
2874 /* Make slab active. */
2875 list_add_tail(&slabp->list, &(l3->slabs_free));
2876 STATS_INC_GROWN(cachep);
2877 l3->free_objects += cachep->num;
2878 spin_unlock(&l3->list_lock);
2881 kmem_freepages(cachep, objp);
2883 if (local_flags & __GFP_WAIT)
2884 local_irq_disable();
2891 * Perform extra freeing checks:
2892 * - detect bad pointers.
2893 * - POISON/RED_ZONE checking
2895 static void kfree_debugcheck(const void *objp)
2897 if (!virt_addr_valid(objp)) {
2898 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2899 (unsigned long)objp);
2904 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2906 unsigned long long redzone1, redzone2;
2908 redzone1 = *dbg_redzone1(cache, obj);
2909 redzone2 = *dbg_redzone2(cache, obj);
2914 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2917 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2918 slab_error(cache, "double free detected");
2920 slab_error(cache, "memory outside object was overwritten");
2922 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2923 obj, redzone1, redzone2);
2926 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2927 unsigned long caller)
2933 BUG_ON(virt_to_cache(objp) != cachep);
2935 objp -= obj_offset(cachep);
2936 kfree_debugcheck(objp);
2937 page = virt_to_head_page(objp);
2939 slabp = page->slab_page;
2941 if (cachep->flags & SLAB_RED_ZONE) {
2942 verify_redzone_free(cachep, objp);
2943 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2944 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2946 if (cachep->flags & SLAB_STORE_USER)
2947 *dbg_userword(cachep, objp) = (void *)caller;
2949 objnr = obj_to_index(cachep, slabp, objp);
2951 BUG_ON(objnr >= cachep->num);
2952 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
2954 #ifdef CONFIG_DEBUG_SLAB_LEAK
2955 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2957 if (cachep->flags & SLAB_POISON) {
2958 #ifdef CONFIG_DEBUG_PAGEALLOC
2959 if ((cachep->size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
2960 store_stackinfo(cachep, objp, caller);
2961 kernel_map_pages(virt_to_page(objp),
2962 cachep->size / PAGE_SIZE, 0);
2964 poison_obj(cachep, objp, POISON_FREE);
2967 poison_obj(cachep, objp, POISON_FREE);
2973 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
2978 /* Check slab's freelist to see if this obj is there. */
2979 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2981 if (entries > cachep->num || i >= cachep->num)
2984 if (entries != cachep->num - slabp->inuse) {
2986 printk(KERN_ERR "slab: Internal list corruption detected in "
2987 "cache '%s'(%d), slabp %p(%d). Tainted(%s). Hexdump:\n",
2988 cachep->name, cachep->num, slabp, slabp->inuse,
2990 print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 16, 1, slabp,
2991 sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t),
2997 #define kfree_debugcheck(x) do { } while(0)
2998 #define cache_free_debugcheck(x,objp,z) (objp)
2999 #define check_slabp(x,y) do { } while(0)
3002 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags,
3006 struct kmem_cache_node *l3;
3007 struct array_cache *ac;
3011 node = numa_mem_id();
3012 if (unlikely(force_refill))
3015 ac = cpu_cache_get(cachep);
3016 batchcount = ac->batchcount;
3017 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
3019 * If there was little recent activity on this cache, then
3020 * perform only a partial refill. Otherwise we could generate
3023 batchcount = BATCHREFILL_LIMIT;
3025 l3 = cachep->node[node];
3027 BUG_ON(ac->avail > 0 || !l3);
3028 spin_lock(&l3->list_lock);
3030 /* See if we can refill from the shared array */
3031 if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) {
3032 l3->shared->touched = 1;
3036 while (batchcount > 0) {
3037 struct list_head *entry;
3039 /* Get slab alloc is to come from. */
3040 entry = l3->slabs_partial.next;
3041 if (entry == &l3->slabs_partial) {
3042 l3->free_touched = 1;
3043 entry = l3->slabs_free.next;
3044 if (entry == &l3->slabs_free)
3048 slabp = list_entry(entry, struct slab, list);
3049 check_slabp(cachep, slabp);
3050 check_spinlock_acquired(cachep);
3053 * The slab was either on partial or free list so
3054 * there must be at least one object available for
3057 BUG_ON(slabp->inuse >= cachep->num);
3059 while (slabp->inuse < cachep->num && batchcount--) {
3060 STATS_INC_ALLOCED(cachep);
3061 STATS_INC_ACTIVE(cachep);
3062 STATS_SET_HIGH(cachep);
3064 ac_put_obj(cachep, ac, slab_get_obj(cachep, slabp,
3067 check_slabp(cachep, slabp);
3069 /* move slabp to correct slabp list: */
3070 list_del(&slabp->list);
3071 if (slabp->free == BUFCTL_END)
3072 list_add(&slabp->list, &l3->slabs_full);
3074 list_add(&slabp->list, &l3->slabs_partial);
3078 l3->free_objects -= ac->avail;
3080 spin_unlock(&l3->list_lock);
3082 if (unlikely(!ac->avail)) {
3085 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
3087 /* cache_grow can reenable interrupts, then ac could change. */
3088 ac = cpu_cache_get(cachep);
3089 node = numa_mem_id();
3091 /* no objects in sight? abort */
3092 if (!x && (ac->avail == 0 || force_refill))
3095 if (!ac->avail) /* objects refilled by interrupt? */
3100 return ac_get_obj(cachep, ac, flags, force_refill);
3103 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3106 might_sleep_if(flags & __GFP_WAIT);
3108 kmem_flagcheck(cachep, flags);
3113 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3114 gfp_t flags, void *objp, unsigned long caller)
3118 if (cachep->flags & SLAB_POISON) {
3119 #ifdef CONFIG_DEBUG_PAGEALLOC
3120 if ((cachep->size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
3121 kernel_map_pages(virt_to_page(objp),
3122 cachep->size / PAGE_SIZE, 1);
3124 check_poison_obj(cachep, objp);
3126 check_poison_obj(cachep, objp);
3128 poison_obj(cachep, objp, POISON_INUSE);
3130 if (cachep->flags & SLAB_STORE_USER)
3131 *dbg_userword(cachep, objp) = (void *)caller;
3133 if (cachep->flags & SLAB_RED_ZONE) {
3134 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3135 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3136 slab_error(cachep, "double free, or memory outside"
3137 " object was overwritten");
3139 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3140 objp, *dbg_redzone1(cachep, objp),
3141 *dbg_redzone2(cachep, objp));
3143 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3144 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3146 #ifdef CONFIG_DEBUG_SLAB_LEAK
3151 slabp = virt_to_head_page(objp)->slab_page;
3152 objnr = (unsigned)(objp - slabp->s_mem) / cachep->size;
3153 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3156 objp += obj_offset(cachep);
3157 if (cachep->ctor && cachep->flags & SLAB_POISON)
3159 if (ARCH_SLAB_MINALIGN &&
3160 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
3161 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3162 objp, (int)ARCH_SLAB_MINALIGN);
3167 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3170 static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
3172 if (cachep == kmem_cache)
3175 return should_failslab(cachep->object_size, flags, cachep->flags);
3178 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3181 struct array_cache *ac;
3182 bool force_refill = false;
3186 ac = cpu_cache_get(cachep);
3187 if (likely(ac->avail)) {
3189 objp = ac_get_obj(cachep, ac, flags, false);
3192 * Allow for the possibility all avail objects are not allowed
3193 * by the current flags
3196 STATS_INC_ALLOCHIT(cachep);
3199 force_refill = true;
3202 STATS_INC_ALLOCMISS(cachep);
3203 objp = cache_alloc_refill(cachep, flags, force_refill);
3205 * the 'ac' may be updated by cache_alloc_refill(),
3206 * and kmemleak_erase() requires its correct value.
3208 ac = cpu_cache_get(cachep);
3212 * To avoid a false negative, if an object that is in one of the
3213 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3214 * treat the array pointers as a reference to the object.
3217 kmemleak_erase(&ac->entry[ac->avail]);
3223 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3225 * If we are in_interrupt, then process context, including cpusets and
3226 * mempolicy, may not apply and should not be used for allocation policy.
3228 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3230 int nid_alloc, nid_here;
3232 if (in_interrupt() || (flags & __GFP_THISNODE))
3234 nid_alloc = nid_here = numa_mem_id();
3235 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3236 nid_alloc = cpuset_slab_spread_node();
3237 else if (current->mempolicy)
3238 nid_alloc = slab_node();
3239 if (nid_alloc != nid_here)
3240 return ____cache_alloc_node(cachep, flags, nid_alloc);
3245 * Fallback function if there was no memory available and no objects on a
3246 * certain node and fall back is permitted. First we scan all the
3247 * available node for available objects. If that fails then we
3248 * perform an allocation without specifying a node. This allows the page
3249 * allocator to do its reclaim / fallback magic. We then insert the
3250 * slab into the proper nodelist and then allocate from it.
3252 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3254 struct zonelist *zonelist;
3258 enum zone_type high_zoneidx = gfp_zone(flags);
3261 unsigned int cpuset_mems_cookie;
3263 if (flags & __GFP_THISNODE)
3266 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3269 cpuset_mems_cookie = get_mems_allowed();
3270 zonelist = node_zonelist(slab_node(), flags);
3274 * Look through allowed nodes for objects available
3275 * from existing per node queues.
3277 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3278 nid = zone_to_nid(zone);
3280 if (cpuset_zone_allowed_hardwall(zone, flags) &&
3282 cache->node[nid]->free_objects) {
3283 obj = ____cache_alloc_node(cache,
3284 flags | GFP_THISNODE, nid);
3292 * This allocation will be performed within the constraints
3293 * of the current cpuset / memory policy requirements.
3294 * We may trigger various forms of reclaim on the allowed
3295 * set and go into memory reserves if necessary.
3297 if (local_flags & __GFP_WAIT)
3299 kmem_flagcheck(cache, flags);
3300 obj = kmem_getpages(cache, local_flags, numa_mem_id());
3301 if (local_flags & __GFP_WAIT)
3302 local_irq_disable();
3305 * Insert into the appropriate per node queues
3307 nid = page_to_nid(virt_to_page(obj));
3308 if (cache_grow(cache, flags, nid, obj)) {
3309 obj = ____cache_alloc_node(cache,
3310 flags | GFP_THISNODE, nid);
3313 * Another processor may allocate the
3314 * objects in the slab since we are
3315 * not holding any locks.
3319 /* cache_grow already freed obj */
3325 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !obj))
3331 * A interface to enable slab creation on nodeid
3333 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3336 struct list_head *entry;
3338 struct kmem_cache_node *l3;
3342 l3 = cachep->node[nodeid];
3347 spin_lock(&l3->list_lock);
3348 entry = l3->slabs_partial.next;
3349 if (entry == &l3->slabs_partial) {
3350 l3->free_touched = 1;
3351 entry = l3->slabs_free.next;
3352 if (entry == &l3->slabs_free)
3356 slabp = list_entry(entry, struct slab, list);
3357 check_spinlock_acquired_node(cachep, nodeid);
3358 check_slabp(cachep, slabp);
3360 STATS_INC_NODEALLOCS(cachep);
3361 STATS_INC_ACTIVE(cachep);
3362 STATS_SET_HIGH(cachep);
3364 BUG_ON(slabp->inuse == cachep->num);
3366 obj = slab_get_obj(cachep, slabp, nodeid);
3367 check_slabp(cachep, slabp);
3369 /* move slabp to correct slabp list: */
3370 list_del(&slabp->list);
3372 if (slabp->free == BUFCTL_END)
3373 list_add(&slabp->list, &l3->slabs_full);
3375 list_add(&slabp->list, &l3->slabs_partial);
3377 spin_unlock(&l3->list_lock);
3381 spin_unlock(&l3->list_lock);
3382 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
3386 return fallback_alloc(cachep, flags);
3393 * kmem_cache_alloc_node - Allocate an object on the specified node
3394 * @cachep: The cache to allocate from.
3395 * @flags: See kmalloc().
3396 * @nodeid: node number of the target node.
3397 * @caller: return address of caller, used for debug information
3399 * Identical to kmem_cache_alloc but it will allocate memory on the given
3400 * node, which can improve the performance for cpu bound structures.
3402 * Fallback to other node is possible if __GFP_THISNODE is not set.
3404 static __always_inline void *
3405 slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3406 unsigned long caller)
3408 unsigned long save_flags;
3410 int slab_node = numa_mem_id();
3412 flags &= gfp_allowed_mask;
3414 lockdep_trace_alloc(flags);
3416 if (slab_should_failslab(cachep, flags))
3419 cachep = memcg_kmem_get_cache(cachep, flags);
3421 cache_alloc_debugcheck_before(cachep, flags);
3422 local_irq_save(save_flags);
3424 if (nodeid == NUMA_NO_NODE)
3427 if (unlikely(!cachep->node[nodeid])) {
3428 /* Node not bootstrapped yet */
3429 ptr = fallback_alloc(cachep, flags);
3433 if (nodeid == slab_node) {
3435 * Use the locally cached objects if possible.
3436 * However ____cache_alloc does not allow fallback
3437 * to other nodes. It may fail while we still have
3438 * objects on other nodes available.
3440 ptr = ____cache_alloc(cachep, flags);
3444 /* ___cache_alloc_node can fall back to other nodes */
3445 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3447 local_irq_restore(save_flags);
3448 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3449 kmemleak_alloc_recursive(ptr, cachep->object_size, 1, cachep->flags,
3453 kmemcheck_slab_alloc(cachep, flags, ptr, cachep->object_size);
3455 if (unlikely((flags & __GFP_ZERO) && ptr))
3456 memset(ptr, 0, cachep->object_size);
3461 static __always_inline void *
3462 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3466 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3467 objp = alternate_node_alloc(cache, flags);
3471 objp = ____cache_alloc(cache, flags);
3474 * We may just have run out of memory on the local node.
3475 * ____cache_alloc_node() knows how to locate memory on other nodes
3478 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3485 static __always_inline void *
3486 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3488 return ____cache_alloc(cachep, flags);
3491 #endif /* CONFIG_NUMA */
3493 static __always_inline void *
3494 slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
3496 unsigned long save_flags;
3499 flags &= gfp_allowed_mask;
3501 lockdep_trace_alloc(flags);
3503 if (slab_should_failslab(cachep, flags))
3506 cachep = memcg_kmem_get_cache(cachep, flags);
3508 cache_alloc_debugcheck_before(cachep, flags);
3509 local_irq_save(save_flags);
3510 objp = __do_cache_alloc(cachep, flags);
3511 local_irq_restore(save_flags);
3512 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3513 kmemleak_alloc_recursive(objp, cachep->object_size, 1, cachep->flags,
3518 kmemcheck_slab_alloc(cachep, flags, objp, cachep->object_size);
3520 if (unlikely((flags & __GFP_ZERO) && objp))
3521 memset(objp, 0, cachep->object_size);
3527 * Caller needs to acquire correct kmem_list's list_lock
3529 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3533 struct kmem_cache_node *l3;
3535 for (i = 0; i < nr_objects; i++) {
3539 clear_obj_pfmemalloc(&objpp[i]);
3542 slabp = virt_to_slab(objp);
3543 l3 = cachep->node[node];
3544 list_del(&slabp->list);
3545 check_spinlock_acquired_node(cachep, node);
3546 check_slabp(cachep, slabp);
3547 slab_put_obj(cachep, slabp, objp, node);
3548 STATS_DEC_ACTIVE(cachep);
3550 check_slabp(cachep, slabp);
3552 /* fixup slab chains */
3553 if (slabp->inuse == 0) {
3554 if (l3->free_objects > l3->free_limit) {
3555 l3->free_objects -= cachep->num;
3556 /* No need to drop any previously held
3557 * lock here, even if we have a off-slab slab
3558 * descriptor it is guaranteed to come from
3559 * a different cache, refer to comments before
3562 slab_destroy(cachep, slabp);
3564 list_add(&slabp->list, &l3->slabs_free);
3567 /* Unconditionally move a slab to the end of the
3568 * partial list on free - maximum time for the
3569 * other objects to be freed, too.
3571 list_add_tail(&slabp->list, &l3->slabs_partial);
3576 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3579 struct kmem_cache_node *l3;
3580 int node = numa_mem_id();
3582 batchcount = ac->batchcount;
3584 BUG_ON(!batchcount || batchcount > ac->avail);
3587 l3 = cachep->node[node];
3588 spin_lock(&l3->list_lock);
3590 struct array_cache *shared_array = l3->shared;
3591 int max = shared_array->limit - shared_array->avail;
3593 if (batchcount > max)
3595 memcpy(&(shared_array->entry[shared_array->avail]),
3596 ac->entry, sizeof(void *) * batchcount);
3597 shared_array->avail += batchcount;
3602 free_block(cachep, ac->entry, batchcount, node);
3607 struct list_head *p;
3609 p = l3->slabs_free.next;
3610 while (p != &(l3->slabs_free)) {
3613 slabp = list_entry(p, struct slab, list);
3614 BUG_ON(slabp->inuse);
3619 STATS_SET_FREEABLE(cachep, i);
3622 spin_unlock(&l3->list_lock);
3623 ac->avail -= batchcount;
3624 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3628 * Release an obj back to its cache. If the obj has a constructed state, it must
3629 * be in this state _before_ it is released. Called with disabled ints.
3631 static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3632 unsigned long caller)
3634 struct array_cache *ac = cpu_cache_get(cachep);
3637 kmemleak_free_recursive(objp, cachep->flags);
3638 objp = cache_free_debugcheck(cachep, objp, caller);
3640 kmemcheck_slab_free(cachep, objp, cachep->object_size);
3643 * Skip calling cache_free_alien() when the platform is not numa.
3644 * This will avoid cache misses that happen while accessing slabp (which
3645 * is per page memory reference) to get nodeid. Instead use a global
3646 * variable to skip the call, which is mostly likely to be present in
3649 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3652 if (likely(ac->avail < ac->limit)) {
3653 STATS_INC_FREEHIT(cachep);
3655 STATS_INC_FREEMISS(cachep);
3656 cache_flusharray(cachep, ac);
3659 ac_put_obj(cachep, ac, objp);
3663 * kmem_cache_alloc - Allocate an object
3664 * @cachep: The cache to allocate from.
3665 * @flags: See kmalloc().
3667 * Allocate an object from this cache. The flags are only relevant
3668 * if the cache has no available objects.
3670 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3672 void *ret = slab_alloc(cachep, flags, _RET_IP_);
3674 trace_kmem_cache_alloc(_RET_IP_, ret,
3675 cachep->object_size, cachep->size, flags);
3679 EXPORT_SYMBOL(kmem_cache_alloc);
3681 #ifdef CONFIG_TRACING
3683 kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
3687 ret = slab_alloc(cachep, flags, _RET_IP_);
3689 trace_kmalloc(_RET_IP_, ret,
3690 size, cachep->size, flags);
3693 EXPORT_SYMBOL(kmem_cache_alloc_trace);
3697 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3699 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3701 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3702 cachep->object_size, cachep->size,
3707 EXPORT_SYMBOL(kmem_cache_alloc_node);
3709 #ifdef CONFIG_TRACING
3710 void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
3717 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3719 trace_kmalloc_node(_RET_IP_, ret,
3724 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3727 static __always_inline void *
3728 __do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
3730 struct kmem_cache *cachep;
3732 cachep = kmem_find_general_cachep(size, flags);
3733 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3735 return kmem_cache_alloc_node_trace(cachep, flags, node, size);
3738 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3739 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3741 return __do_kmalloc_node(size, flags, node, _RET_IP_);
3743 EXPORT_SYMBOL(__kmalloc_node);
3745 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3746 int node, unsigned long caller)
3748 return __do_kmalloc_node(size, flags, node, caller);
3750 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3752 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3754 return __do_kmalloc_node(size, flags, node, 0);
3756 EXPORT_SYMBOL(__kmalloc_node);
3757 #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3758 #endif /* CONFIG_NUMA */
3761 * __do_kmalloc - allocate memory
3762 * @size: how many bytes of memory are required.
3763 * @flags: the type of memory to allocate (see kmalloc).
3764 * @caller: function caller for debug tracking of the caller
3766 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3767 unsigned long caller)
3769 struct kmem_cache *cachep;
3772 /* If you want to save a few bytes .text space: replace
3774 * Then kmalloc uses the uninlined functions instead of the inline
3777 cachep = __find_general_cachep(size, flags);
3778 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3780 ret = slab_alloc(cachep, flags, caller);
3782 trace_kmalloc(caller, ret,
3783 size, cachep->size, flags);
3789 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3790 void *__kmalloc(size_t size, gfp_t flags)
3792 return __do_kmalloc(size, flags, _RET_IP_);
3794 EXPORT_SYMBOL(__kmalloc);
3796 void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3798 return __do_kmalloc(size, flags, caller);
3800 EXPORT_SYMBOL(__kmalloc_track_caller);
3803 void *__kmalloc(size_t size, gfp_t flags)
3805 return __do_kmalloc(size, flags, 0);
3807 EXPORT_SYMBOL(__kmalloc);
3811 * kmem_cache_free - Deallocate an object
3812 * @cachep: The cache the allocation was from.
3813 * @objp: The previously allocated object.
3815 * Free an object which was previously allocated from this
3818 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3820 unsigned long flags;
3821 cachep = cache_from_obj(cachep, objp);
3825 local_irq_save(flags);
3826 debug_check_no_locks_freed(objp, cachep->object_size);
3827 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3828 debug_check_no_obj_freed(objp, cachep->object_size);
3829 __cache_free(cachep, objp, _RET_IP_);
3830 local_irq_restore(flags);
3832 trace_kmem_cache_free(_RET_IP_, objp);
3834 EXPORT_SYMBOL(kmem_cache_free);
3837 * kfree - free previously allocated memory
3838 * @objp: pointer returned by kmalloc.
3840 * If @objp is NULL, no operation is performed.
3842 * Don't free memory not originally allocated by kmalloc()
3843 * or you will run into trouble.
3845 void kfree(const void *objp)
3847 struct kmem_cache *c;
3848 unsigned long flags;
3850 trace_kfree(_RET_IP_, objp);
3852 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3854 local_irq_save(flags);
3855 kfree_debugcheck(objp);
3856 c = virt_to_cache(objp);
3857 debug_check_no_locks_freed(objp, c->object_size);
3859 debug_check_no_obj_freed(objp, c->object_size);
3860 __cache_free(c, (void *)objp, _RET_IP_);
3861 local_irq_restore(flags);
3863 EXPORT_SYMBOL(kfree);
3866 * This initializes kmem_list3 or resizes various caches for all nodes.
3868 static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
3871 struct kmem_cache_node *l3;
3872 struct array_cache *new_shared;
3873 struct array_cache **new_alien = NULL;
3875 for_each_online_node(node) {
3877 if (use_alien_caches) {
3878 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
3884 if (cachep->shared) {
3885 new_shared = alloc_arraycache(node,
3886 cachep->shared*cachep->batchcount,
3889 free_alien_cache(new_alien);
3894 l3 = cachep->node[node];
3896 struct array_cache *shared = l3->shared;
3898 spin_lock_irq(&l3->list_lock);
3901 free_block(cachep, shared->entry,
3902 shared->avail, node);
3904 l3->shared = new_shared;
3906 l3->alien = new_alien;
3909 l3->free_limit = (1 + nr_cpus_node(node)) *
3910 cachep->batchcount + cachep->num;
3911 spin_unlock_irq(&l3->list_lock);
3913 free_alien_cache(new_alien);
3916 l3 = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
3918 free_alien_cache(new_alien);
3923 kmem_list3_init(l3);
3924 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3925 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3926 l3->shared = new_shared;
3927 l3->alien = new_alien;
3928 l3->free_limit = (1 + nr_cpus_node(node)) *
3929 cachep->batchcount + cachep->num;
3930 cachep->node[node] = l3;
3935 if (!cachep->list.next) {
3936 /* Cache is not active yet. Roll back what we did */
3939 if (cachep->node[node]) {
3940 l3 = cachep->node[node];
3943 free_alien_cache(l3->alien);
3945 cachep->node[node] = NULL;
3953 struct ccupdate_struct {
3954 struct kmem_cache *cachep;
3955 struct array_cache *new[0];
3958 static void do_ccupdate_local(void *info)
3960 struct ccupdate_struct *new = info;
3961 struct array_cache *old;
3964 old = cpu_cache_get(new->cachep);
3966 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3967 new->new[smp_processor_id()] = old;
3970 /* Always called with the slab_mutex held */
3971 static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
3972 int batchcount, int shared, gfp_t gfp)
3974 struct ccupdate_struct *new;
3977 new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *),
3982 for_each_online_cpu(i) {
3983 new->new[i] = alloc_arraycache(cpu_to_mem(i), limit,
3986 for (i--; i >= 0; i--)
3992 new->cachep = cachep;
3994 on_each_cpu(do_ccupdate_local, (void *)new, 1);
3997 cachep->batchcount = batchcount;
3998 cachep->limit = limit;
3999 cachep->shared = shared;
4001 for_each_online_cpu(i) {
4002 struct array_cache *ccold = new->new[i];
4005 spin_lock_irq(&cachep->node[cpu_to_mem(i)]->list_lock);
4006 free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i));
4007 spin_unlock_irq(&cachep->node[cpu_to_mem(i)]->list_lock);
4011 return alloc_kmemlist(cachep, gfp);
4014 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
4015 int batchcount, int shared, gfp_t gfp)
4018 struct kmem_cache *c = NULL;
4021 ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
4023 if (slab_state < FULL)
4026 if ((ret < 0) || !is_root_cache(cachep))
4029 VM_BUG_ON(!mutex_is_locked(&slab_mutex));
4030 for_each_memcg_cache_index(i) {
4031 c = cache_from_memcg(cachep, i);
4033 /* return value determined by the parent cache only */
4034 __do_tune_cpucache(c, limit, batchcount, shared, gfp);
4040 /* Called with slab_mutex held always */
4041 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
4048 if (!is_root_cache(cachep)) {
4049 struct kmem_cache *root = memcg_root_cache(cachep);
4050 limit = root->limit;
4051 shared = root->shared;
4052 batchcount = root->batchcount;
4055 if (limit && shared && batchcount)
4058 * The head array serves three purposes:
4059 * - create a LIFO ordering, i.e. return objects that are cache-warm
4060 * - reduce the number of spinlock operations.
4061 * - reduce the number of linked list operations on the slab and
4062 * bufctl chains: array operations are cheaper.
4063 * The numbers are guessed, we should auto-tune as described by
4066 if (cachep->size > 131072)
4068 else if (cachep->size > PAGE_SIZE)
4070 else if (cachep->size > 1024)
4072 else if (cachep->size > 256)
4078 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
4079 * allocation behaviour: Most allocs on one cpu, most free operations
4080 * on another cpu. For these cases, an efficient object passing between
4081 * cpus is necessary. This is provided by a shared array. The array
4082 * replaces Bonwick's magazine layer.
4083 * On uniprocessor, it's functionally equivalent (but less efficient)
4084 * to a larger limit. Thus disabled by default.
4087 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
4092 * With debugging enabled, large batchcount lead to excessively long
4093 * periods with disabled local interrupts. Limit the batchcount
4098 batchcount = (limit + 1) / 2;
4100 err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
4102 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
4103 cachep->name, -err);
4108 * Drain an array if it contains any elements taking the l3 lock only if
4109 * necessary. Note that the l3 listlock also protects the array_cache
4110 * if drain_array() is used on the shared array.
4112 static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *l3,
4113 struct array_cache *ac, int force, int node)
4117 if (!ac || !ac->avail)
4119 if (ac->touched && !force) {
4122 spin_lock_irq(&l3->list_lock);
4124 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4125 if (tofree > ac->avail)
4126 tofree = (ac->avail + 1) / 2;
4127 free_block(cachep, ac->entry, tofree, node);
4128 ac->avail -= tofree;
4129 memmove(ac->entry, &(ac->entry[tofree]),
4130 sizeof(void *) * ac->avail);
4132 spin_unlock_irq(&l3->list_lock);
4137 * cache_reap - Reclaim memory from caches.
4138 * @w: work descriptor
4140 * Called from workqueue/eventd every few seconds.
4142 * - clear the per-cpu caches for this CPU.
4143 * - return freeable pages to the main free memory pool.
4145 * If we cannot acquire the cache chain mutex then just give up - we'll try
4146 * again on the next iteration.
4148 static void cache_reap(struct work_struct *w)
4150 struct kmem_cache *searchp;
4151 struct kmem_cache_node *l3;
4152 int node = numa_mem_id();
4153 struct delayed_work *work = to_delayed_work(w);
4155 if (!mutex_trylock(&slab_mutex))
4156 /* Give up. Setup the next iteration. */
4159 list_for_each_entry(searchp, &slab_caches, list) {
4163 * We only take the l3 lock if absolutely necessary and we
4164 * have established with reasonable certainty that
4165 * we can do some work if the lock was obtained.
4167 l3 = searchp->node[node];
4169 reap_alien(searchp, l3);
4171 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
4174 * These are racy checks but it does not matter
4175 * if we skip one check or scan twice.
4177 if (time_after(l3->next_reap, jiffies))
4180 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
4182 drain_array(searchp, l3, l3->shared, 0, node);
4184 if (l3->free_touched)
4185 l3->free_touched = 0;
4189 freed = drain_freelist(searchp, l3, (l3->free_limit +
4190 5 * searchp->num - 1) / (5 * searchp->num));
4191 STATS_ADD_REAPED(searchp, freed);
4197 mutex_unlock(&slab_mutex);
4200 /* Set up the next iteration */
4201 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
4204 #ifdef CONFIG_SLABINFO
4205 void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
4208 unsigned long active_objs;
4209 unsigned long num_objs;
4210 unsigned long active_slabs = 0;
4211 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4215 struct kmem_cache_node *l3;
4219 for_each_online_node(node) {
4220 l3 = cachep->node[node];
4225 spin_lock_irq(&l3->list_lock);
4227 list_for_each_entry(slabp, &l3->slabs_full, list) {
4228 if (slabp->inuse != cachep->num && !error)
4229 error = "slabs_full accounting error";
4230 active_objs += cachep->num;
4233 list_for_each_entry(slabp, &l3->slabs_partial, list) {
4234 if (slabp->inuse == cachep->num && !error)
4235 error = "slabs_partial inuse accounting error";
4236 if (!slabp->inuse && !error)
4237 error = "slabs_partial/inuse accounting error";
4238 active_objs += slabp->inuse;
4241 list_for_each_entry(slabp, &l3->slabs_free, list) {
4242 if (slabp->inuse && !error)
4243 error = "slabs_free/inuse accounting error";
4246 free_objects += l3->free_objects;
4248 shared_avail += l3->shared->avail;
4250 spin_unlock_irq(&l3->list_lock);
4252 num_slabs += active_slabs;
4253 num_objs = num_slabs * cachep->num;
4254 if (num_objs - active_objs != free_objects && !error)
4255 error = "free_objects accounting error";
4257 name = cachep->name;
4259 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4261 sinfo->active_objs = active_objs;
4262 sinfo->num_objs = num_objs;
4263 sinfo->active_slabs = active_slabs;
4264 sinfo->num_slabs = num_slabs;
4265 sinfo->shared_avail = shared_avail;
4266 sinfo->limit = cachep->limit;
4267 sinfo->batchcount = cachep->batchcount;
4268 sinfo->shared = cachep->shared;
4269 sinfo->objects_per_slab = cachep->num;
4270 sinfo->cache_order = cachep->gfporder;
4273 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
4277 unsigned long high = cachep->high_mark;
4278 unsigned long allocs = cachep->num_allocations;
4279 unsigned long grown = cachep->grown;
4280 unsigned long reaped = cachep->reaped;
4281 unsigned long errors = cachep->errors;
4282 unsigned long max_freeable = cachep->max_freeable;
4283 unsigned long node_allocs = cachep->node_allocs;
4284 unsigned long node_frees = cachep->node_frees;
4285 unsigned long overflows = cachep->node_overflow;
4287 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4288 "%4lu %4lu %4lu %4lu %4lu",
4289 allocs, high, grown,
4290 reaped, errors, max_freeable, node_allocs,
4291 node_frees, overflows);
4295 unsigned long allochit = atomic_read(&cachep->allochit);
4296 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4297 unsigned long freehit = atomic_read(&cachep->freehit);
4298 unsigned long freemiss = atomic_read(&cachep->freemiss);
4300 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4301 allochit, allocmiss, freehit, freemiss);
4306 #define MAX_SLABINFO_WRITE 128
4308 * slabinfo_write - Tuning for the slab allocator
4310 * @buffer: user buffer
4311 * @count: data length
4314 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4315 size_t count, loff_t *ppos)
4317 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4318 int limit, batchcount, shared, res;
4319 struct kmem_cache *cachep;
4321 if (count > MAX_SLABINFO_WRITE)
4323 if (copy_from_user(&kbuf, buffer, count))
4325 kbuf[MAX_SLABINFO_WRITE] = '\0';
4327 tmp = strchr(kbuf, ' ');
4332 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4335 /* Find the cache in the chain of caches. */
4336 mutex_lock(&slab_mutex);
4338 list_for_each_entry(cachep, &slab_caches, list) {
4339 if (!strcmp(cachep->name, kbuf)) {
4340 if (limit < 1 || batchcount < 1 ||
4341 batchcount > limit || shared < 0) {
4344 res = do_tune_cpucache(cachep, limit,
4351 mutex_unlock(&slab_mutex);
4357 #ifdef CONFIG_DEBUG_SLAB_LEAK
4359 static void *leaks_start(struct seq_file *m, loff_t *pos)
4361 mutex_lock(&slab_mutex);
4362 return seq_list_start(&slab_caches, *pos);
4365 static inline int add_caller(unsigned long *n, unsigned long v)
4375 unsigned long *q = p + 2 * i;
4389 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4395 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4401 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->size) {
4402 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4404 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4409 static void show_symbol(struct seq_file *m, unsigned long address)
4411 #ifdef CONFIG_KALLSYMS
4412 unsigned long offset, size;
4413 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4415 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4416 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4418 seq_printf(m, " [%s]", modname);
4422 seq_printf(m, "%p", (void *)address);
4425 static int leaks_show(struct seq_file *m, void *p)
4427 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
4429 struct kmem_cache_node *l3;
4431 unsigned long *n = m->private;
4435 if (!(cachep->flags & SLAB_STORE_USER))
4437 if (!(cachep->flags & SLAB_RED_ZONE))
4440 /* OK, we can do it */
4444 for_each_online_node(node) {
4445 l3 = cachep->node[node];
4450 spin_lock_irq(&l3->list_lock);
4452 list_for_each_entry(slabp, &l3->slabs_full, list)
4453 handle_slab(n, cachep, slabp);
4454 list_for_each_entry(slabp, &l3->slabs_partial, list)
4455 handle_slab(n, cachep, slabp);
4456 spin_unlock_irq(&l3->list_lock);
4458 name = cachep->name;
4460 /* Increase the buffer size */
4461 mutex_unlock(&slab_mutex);
4462 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4464 /* Too bad, we are really out */
4466 mutex_lock(&slab_mutex);
4469 *(unsigned long *)m->private = n[0] * 2;
4471 mutex_lock(&slab_mutex);
4472 /* Now make sure this entry will be retried */
4476 for (i = 0; i < n[1]; i++) {
4477 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4478 show_symbol(m, n[2*i+2]);
4485 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4487 return seq_list_next(p, &slab_caches, pos);
4490 static void s_stop(struct seq_file *m, void *p)
4492 mutex_unlock(&slab_mutex);
4495 static const struct seq_operations slabstats_op = {
4496 .start = leaks_start,
4502 static int slabstats_open(struct inode *inode, struct file *file)
4504 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4507 ret = seq_open(file, &slabstats_op);
4509 struct seq_file *m = file->private_data;
4510 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4519 static const struct file_operations proc_slabstats_operations = {
4520 .open = slabstats_open,
4522 .llseek = seq_lseek,
4523 .release = seq_release_private,
4527 static int __init slab_proc_init(void)
4529 #ifdef CONFIG_DEBUG_SLAB_LEAK
4530 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4534 module_init(slab_proc_init);
4538 * ksize - get the actual amount of memory allocated for a given object
4539 * @objp: Pointer to the object
4541 * kmalloc may internally round up allocations and return more memory
4542 * than requested. ksize() can be used to determine the actual amount of
4543 * memory allocated. The caller may use this additional memory, even though
4544 * a smaller amount of memory was initially specified with the kmalloc call.
4545 * The caller must guarantee that objp points to a valid object previously
4546 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4547 * must not be freed during the duration of the call.
4549 size_t ksize(const void *objp)
4552 if (unlikely(objp == ZERO_SIZE_PTR))
4555 return virt_to_cache(objp)->object_size;
4557 EXPORT_SYMBOL(ksize);