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 size_t slab_mgmt_size(size_t nr_objs, size_t align)
661 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
665 * Calculate the number of objects and left-over bytes for a given buffer size.
667 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
668 size_t align, int flags, size_t *left_over,
673 size_t slab_size = PAGE_SIZE << gfporder;
676 * The slab management structure can be either off the slab or
677 * on it. For the latter case, the memory allocated for a
681 * - One kmem_bufctl_t for each object
682 * - Padding to respect alignment of @align
683 * - @buffer_size bytes for each object
685 * If the slab management structure is off the slab, then the
686 * alignment will already be calculated into the size. Because
687 * the slabs are all pages aligned, the objects will be at the
688 * correct alignment when allocated.
690 if (flags & CFLGS_OFF_SLAB) {
692 nr_objs = slab_size / buffer_size;
694 if (nr_objs > SLAB_LIMIT)
695 nr_objs = SLAB_LIMIT;
698 * Ignore padding for the initial guess. The padding
699 * is at most @align-1 bytes, and @buffer_size is at
700 * least @align. In the worst case, this result will
701 * be one greater than the number of objects that fit
702 * into the memory allocation when taking the padding
705 nr_objs = (slab_size - sizeof(struct slab)) /
706 (buffer_size + sizeof(kmem_bufctl_t));
709 * This calculated number will be either the right
710 * amount, or one greater than what we want.
712 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
716 if (nr_objs > SLAB_LIMIT)
717 nr_objs = SLAB_LIMIT;
719 mgmt_size = slab_mgmt_size(nr_objs, align);
722 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
726 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
728 static void __slab_error(const char *function, struct kmem_cache *cachep,
731 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
732 function, cachep->name, msg);
734 add_taint(TAINT_BAD_PAGE);
739 * By default on NUMA we use alien caches to stage the freeing of
740 * objects allocated from other nodes. This causes massive memory
741 * inefficiencies when using fake NUMA setup to split memory into a
742 * large number of small nodes, so it can be disabled on the command
746 static int use_alien_caches __read_mostly = 1;
747 static int __init noaliencache_setup(char *s)
749 use_alien_caches = 0;
752 __setup("noaliencache", noaliencache_setup);
754 static int __init slab_max_order_setup(char *str)
756 get_option(&str, &slab_max_order);
757 slab_max_order = slab_max_order < 0 ? 0 :
758 min(slab_max_order, MAX_ORDER - 1);
759 slab_max_order_set = true;
763 __setup("slab_max_order=", slab_max_order_setup);
767 * Special reaping functions for NUMA systems called from cache_reap().
768 * These take care of doing round robin flushing of alien caches (containing
769 * objects freed on different nodes from which they were allocated) and the
770 * flushing of remote pcps by calling drain_node_pages.
772 static DEFINE_PER_CPU(unsigned long, slab_reap_node);
774 static void init_reap_node(int cpu)
778 node = next_node(cpu_to_mem(cpu), node_online_map);
779 if (node == MAX_NUMNODES)
780 node = first_node(node_online_map);
782 per_cpu(slab_reap_node, cpu) = node;
785 static void next_reap_node(void)
787 int node = __this_cpu_read(slab_reap_node);
789 node = next_node(node, node_online_map);
790 if (unlikely(node >= MAX_NUMNODES))
791 node = first_node(node_online_map);
792 __this_cpu_write(slab_reap_node, node);
796 #define init_reap_node(cpu) do { } while (0)
797 #define next_reap_node(void) do { } while (0)
801 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
802 * via the workqueue/eventd.
803 * Add the CPU number into the expiration time to minimize the possibility of
804 * the CPUs getting into lockstep and contending for the global cache chain
807 static void __cpuinit start_cpu_timer(int cpu)
809 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
812 * When this gets called from do_initcalls via cpucache_init(),
813 * init_workqueues() has already run, so keventd will be setup
816 if (keventd_up() && reap_work->work.func == NULL) {
818 INIT_DEFERRABLE_WORK(reap_work, cache_reap);
819 schedule_delayed_work_on(cpu, reap_work,
820 __round_jiffies_relative(HZ, cpu));
824 static struct array_cache *alloc_arraycache(int node, int entries,
825 int batchcount, gfp_t gfp)
827 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
828 struct array_cache *nc = NULL;
830 nc = kmalloc_node(memsize, gfp, node);
832 * The array_cache structures contain pointers to free object.
833 * However, when such objects are allocated or transferred to another
834 * cache the pointers are not cleared and they could be counted as
835 * valid references during a kmemleak scan. Therefore, kmemleak must
836 * not scan such objects.
838 kmemleak_no_scan(nc);
842 nc->batchcount = batchcount;
844 spin_lock_init(&nc->lock);
849 static inline bool is_slab_pfmemalloc(struct slab *slabp)
851 struct page *page = virt_to_page(slabp->s_mem);
853 return PageSlabPfmemalloc(page);
856 /* Clears pfmemalloc_active if no slabs have pfmalloc set */
857 static void recheck_pfmemalloc_active(struct kmem_cache *cachep,
858 struct array_cache *ac)
860 struct kmem_cache_node *l3 = cachep->node[numa_mem_id()];
864 if (!pfmemalloc_active)
867 spin_lock_irqsave(&l3->list_lock, flags);
868 list_for_each_entry(slabp, &l3->slabs_full, list)
869 if (is_slab_pfmemalloc(slabp))
872 list_for_each_entry(slabp, &l3->slabs_partial, list)
873 if (is_slab_pfmemalloc(slabp))
876 list_for_each_entry(slabp, &l3->slabs_free, list)
877 if (is_slab_pfmemalloc(slabp))
880 pfmemalloc_active = false;
882 spin_unlock_irqrestore(&l3->list_lock, flags);
885 static void *__ac_get_obj(struct kmem_cache *cachep, struct array_cache *ac,
886 gfp_t flags, bool force_refill)
889 void *objp = ac->entry[--ac->avail];
891 /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
892 if (unlikely(is_obj_pfmemalloc(objp))) {
893 struct kmem_cache_node *l3;
895 if (gfp_pfmemalloc_allowed(flags)) {
896 clear_obj_pfmemalloc(&objp);
900 /* The caller cannot use PFMEMALLOC objects, find another one */
901 for (i = 0; i < ac->avail; i++) {
902 /* If a !PFMEMALLOC object is found, swap them */
903 if (!is_obj_pfmemalloc(ac->entry[i])) {
905 ac->entry[i] = ac->entry[ac->avail];
906 ac->entry[ac->avail] = objp;
912 * If there are empty slabs on the slabs_free list and we are
913 * being forced to refill the cache, mark this one !pfmemalloc.
915 l3 = cachep->node[numa_mem_id()];
916 if (!list_empty(&l3->slabs_free) && force_refill) {
917 struct slab *slabp = virt_to_slab(objp);
918 ClearPageSlabPfmemalloc(virt_to_head_page(slabp->s_mem));
919 clear_obj_pfmemalloc(&objp);
920 recheck_pfmemalloc_active(cachep, ac);
924 /* No !PFMEMALLOC objects available */
932 static inline void *ac_get_obj(struct kmem_cache *cachep,
933 struct array_cache *ac, gfp_t flags, bool force_refill)
937 if (unlikely(sk_memalloc_socks()))
938 objp = __ac_get_obj(cachep, ac, flags, force_refill);
940 objp = ac->entry[--ac->avail];
945 static void *__ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
948 if (unlikely(pfmemalloc_active)) {
949 /* Some pfmemalloc slabs exist, check if this is one */
950 struct page *page = virt_to_head_page(objp);
951 if (PageSlabPfmemalloc(page))
952 set_obj_pfmemalloc(&objp);
958 static inline void ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
961 if (unlikely(sk_memalloc_socks()))
962 objp = __ac_put_obj(cachep, ac, objp);
964 ac->entry[ac->avail++] = objp;
968 * Transfer objects in one arraycache to another.
969 * Locking must be handled by the caller.
971 * Return the number of entries transferred.
973 static int transfer_objects(struct array_cache *to,
974 struct array_cache *from, unsigned int max)
976 /* Figure out how many entries to transfer */
977 int nr = min3(from->avail, max, to->limit - to->avail);
982 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
992 #define drain_alien_cache(cachep, alien) do { } while (0)
993 #define reap_alien(cachep, l3) do { } while (0)
995 static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
997 return (struct array_cache **)BAD_ALIEN_MAGIC;
1000 static inline void free_alien_cache(struct array_cache **ac_ptr)
1004 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1009 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
1015 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
1016 gfp_t flags, int nodeid)
1021 #else /* CONFIG_NUMA */
1023 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
1024 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
1026 static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
1028 struct array_cache **ac_ptr;
1029 int memsize = sizeof(void *) * nr_node_ids;
1034 ac_ptr = kzalloc_node(memsize, gfp, node);
1037 if (i == node || !node_online(i))
1039 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
1041 for (i--; i >= 0; i--)
1051 static void free_alien_cache(struct array_cache **ac_ptr)
1062 static void __drain_alien_cache(struct kmem_cache *cachep,
1063 struct array_cache *ac, int node)
1065 struct kmem_cache_node *rl3 = cachep->node[node];
1068 spin_lock(&rl3->list_lock);
1070 * Stuff objects into the remote nodes shared array first.
1071 * That way we could avoid the overhead of putting the objects
1072 * into the free lists and getting them back later.
1075 transfer_objects(rl3->shared, ac, ac->limit);
1077 free_block(cachep, ac->entry, ac->avail, node);
1079 spin_unlock(&rl3->list_lock);
1084 * Called from cache_reap() to regularly drain alien caches round robin.
1086 static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *l3)
1088 int node = __this_cpu_read(slab_reap_node);
1091 struct array_cache *ac = l3->alien[node];
1093 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1094 __drain_alien_cache(cachep, ac, node);
1095 spin_unlock_irq(&ac->lock);
1100 static void drain_alien_cache(struct kmem_cache *cachep,
1101 struct array_cache **alien)
1104 struct array_cache *ac;
1105 unsigned long flags;
1107 for_each_online_node(i) {
1110 spin_lock_irqsave(&ac->lock, flags);
1111 __drain_alien_cache(cachep, ac, i);
1112 spin_unlock_irqrestore(&ac->lock, flags);
1117 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1119 struct slab *slabp = virt_to_slab(objp);
1120 int nodeid = slabp->nodeid;
1121 struct kmem_cache_node *l3;
1122 struct array_cache *alien = NULL;
1125 node = numa_mem_id();
1128 * Make sure we are not freeing a object from another node to the array
1129 * cache on this cpu.
1131 if (likely(slabp->nodeid == node))
1134 l3 = cachep->node[node];
1135 STATS_INC_NODEFREES(cachep);
1136 if (l3->alien && l3->alien[nodeid]) {
1137 alien = l3->alien[nodeid];
1138 spin_lock(&alien->lock);
1139 if (unlikely(alien->avail == alien->limit)) {
1140 STATS_INC_ACOVERFLOW(cachep);
1141 __drain_alien_cache(cachep, alien, nodeid);
1143 ac_put_obj(cachep, alien, objp);
1144 spin_unlock(&alien->lock);
1146 spin_lock(&(cachep->node[nodeid])->list_lock);
1147 free_block(cachep, &objp, 1, nodeid);
1148 spin_unlock(&(cachep->node[nodeid])->list_lock);
1155 * Allocates and initializes node for a node on each slab cache, used for
1156 * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3
1157 * will be allocated off-node since memory is not yet online for the new node.
1158 * When hotplugging memory or a cpu, existing node are not replaced if
1161 * Must hold slab_mutex.
1163 static int init_cache_node_node(int node)
1165 struct kmem_cache *cachep;
1166 struct kmem_cache_node *l3;
1167 const int memsize = sizeof(struct kmem_cache_node);
1169 list_for_each_entry(cachep, &slab_caches, list) {
1171 * Set up the size64 kmemlist for cpu before we can
1172 * begin anything. Make sure some other cpu on this
1173 * node has not already allocated this
1175 if (!cachep->node[node]) {
1176 l3 = kmalloc_node(memsize, GFP_KERNEL, node);
1179 kmem_list3_init(l3);
1180 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1181 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1184 * The l3s don't come and go as CPUs come and
1185 * go. slab_mutex is sufficient
1188 cachep->node[node] = l3;
1191 spin_lock_irq(&cachep->node[node]->list_lock);
1192 cachep->node[node]->free_limit =
1193 (1 + nr_cpus_node(node)) *
1194 cachep->batchcount + cachep->num;
1195 spin_unlock_irq(&cachep->node[node]->list_lock);
1200 static void __cpuinit cpuup_canceled(long cpu)
1202 struct kmem_cache *cachep;
1203 struct kmem_cache_node *l3 = NULL;
1204 int node = cpu_to_mem(cpu);
1205 const struct cpumask *mask = cpumask_of_node(node);
1207 list_for_each_entry(cachep, &slab_caches, list) {
1208 struct array_cache *nc;
1209 struct array_cache *shared;
1210 struct array_cache **alien;
1212 /* cpu is dead; no one can alloc from it. */
1213 nc = cachep->array[cpu];
1214 cachep->array[cpu] = NULL;
1215 l3 = cachep->node[node];
1218 goto free_array_cache;
1220 spin_lock_irq(&l3->list_lock);
1222 /* Free limit for this kmem_list3 */
1223 l3->free_limit -= cachep->batchcount;
1225 free_block(cachep, nc->entry, nc->avail, node);
1227 if (!cpumask_empty(mask)) {
1228 spin_unlock_irq(&l3->list_lock);
1229 goto free_array_cache;
1232 shared = l3->shared;
1234 free_block(cachep, shared->entry,
1235 shared->avail, node);
1242 spin_unlock_irq(&l3->list_lock);
1246 drain_alien_cache(cachep, alien);
1247 free_alien_cache(alien);
1253 * In the previous loop, all the objects were freed to
1254 * the respective cache's slabs, now we can go ahead and
1255 * shrink each nodelist to its limit.
1257 list_for_each_entry(cachep, &slab_caches, list) {
1258 l3 = cachep->node[node];
1261 drain_freelist(cachep, l3, l3->free_objects);
1265 static int __cpuinit cpuup_prepare(long cpu)
1267 struct kmem_cache *cachep;
1268 struct kmem_cache_node *l3 = NULL;
1269 int node = cpu_to_mem(cpu);
1273 * We need to do this right in the beginning since
1274 * alloc_arraycache's are going to use this list.
1275 * kmalloc_node allows us to add the slab to the right
1276 * kmem_list3 and not this cpu's kmem_list3
1278 err = init_cache_node_node(node);
1283 * Now we can go ahead with allocating the shared arrays and
1286 list_for_each_entry(cachep, &slab_caches, list) {
1287 struct array_cache *nc;
1288 struct array_cache *shared = NULL;
1289 struct array_cache **alien = NULL;
1291 nc = alloc_arraycache(node, cachep->limit,
1292 cachep->batchcount, GFP_KERNEL);
1295 if (cachep->shared) {
1296 shared = alloc_arraycache(node,
1297 cachep->shared * cachep->batchcount,
1298 0xbaadf00d, GFP_KERNEL);
1304 if (use_alien_caches) {
1305 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
1312 cachep->array[cpu] = nc;
1313 l3 = cachep->node[node];
1316 spin_lock_irq(&l3->list_lock);
1319 * We are serialised from CPU_DEAD or
1320 * CPU_UP_CANCELLED by the cpucontrol lock
1322 l3->shared = shared;
1331 spin_unlock_irq(&l3->list_lock);
1333 free_alien_cache(alien);
1334 if (cachep->flags & SLAB_DEBUG_OBJECTS)
1335 slab_set_debugobj_lock_classes_node(cachep, node);
1336 else if (!OFF_SLAB(cachep) &&
1337 !(cachep->flags & SLAB_DESTROY_BY_RCU))
1338 on_slab_lock_classes_node(cachep, node);
1340 init_node_lock_keys(node);
1344 cpuup_canceled(cpu);
1348 static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1349 unsigned long action, void *hcpu)
1351 long cpu = (long)hcpu;
1355 case CPU_UP_PREPARE:
1356 case CPU_UP_PREPARE_FROZEN:
1357 mutex_lock(&slab_mutex);
1358 err = cpuup_prepare(cpu);
1359 mutex_unlock(&slab_mutex);
1362 case CPU_ONLINE_FROZEN:
1363 start_cpu_timer(cpu);
1365 #ifdef CONFIG_HOTPLUG_CPU
1366 case CPU_DOWN_PREPARE:
1367 case CPU_DOWN_PREPARE_FROZEN:
1369 * Shutdown cache reaper. Note that the slab_mutex is
1370 * held so that if cache_reap() is invoked it cannot do
1371 * anything expensive but will only modify reap_work
1372 * and reschedule the timer.
1374 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1375 /* Now the cache_reaper is guaranteed to be not running. */
1376 per_cpu(slab_reap_work, cpu).work.func = NULL;
1378 case CPU_DOWN_FAILED:
1379 case CPU_DOWN_FAILED_FROZEN:
1380 start_cpu_timer(cpu);
1383 case CPU_DEAD_FROZEN:
1385 * Even if all the cpus of a node are down, we don't free the
1386 * kmem_list3 of any cache. This to avoid a race between
1387 * cpu_down, and a kmalloc allocation from another cpu for
1388 * memory from the node of the cpu going down. The list3
1389 * structure is usually allocated from kmem_cache_create() and
1390 * gets destroyed at kmem_cache_destroy().
1394 case CPU_UP_CANCELED:
1395 case CPU_UP_CANCELED_FROZEN:
1396 mutex_lock(&slab_mutex);
1397 cpuup_canceled(cpu);
1398 mutex_unlock(&slab_mutex);
1401 return notifier_from_errno(err);
1404 static struct notifier_block __cpuinitdata cpucache_notifier = {
1405 &cpuup_callback, NULL, 0
1408 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1410 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1411 * Returns -EBUSY if all objects cannot be drained so that the node is not
1414 * Must hold slab_mutex.
1416 static int __meminit drain_cache_node_node(int node)
1418 struct kmem_cache *cachep;
1421 list_for_each_entry(cachep, &slab_caches, list) {
1422 struct kmem_cache_node *l3;
1424 l3 = cachep->node[node];
1428 drain_freelist(cachep, l3, l3->free_objects);
1430 if (!list_empty(&l3->slabs_full) ||
1431 !list_empty(&l3->slabs_partial)) {
1439 static int __meminit slab_memory_callback(struct notifier_block *self,
1440 unsigned long action, void *arg)
1442 struct memory_notify *mnb = arg;
1446 nid = mnb->status_change_nid;
1451 case MEM_GOING_ONLINE:
1452 mutex_lock(&slab_mutex);
1453 ret = init_cache_node_node(nid);
1454 mutex_unlock(&slab_mutex);
1456 case MEM_GOING_OFFLINE:
1457 mutex_lock(&slab_mutex);
1458 ret = drain_cache_node_node(nid);
1459 mutex_unlock(&slab_mutex);
1463 case MEM_CANCEL_ONLINE:
1464 case MEM_CANCEL_OFFLINE:
1468 return notifier_from_errno(ret);
1470 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1473 * swap the static kmem_list3 with kmalloced memory
1475 static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list,
1478 struct kmem_cache_node *ptr;
1480 ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid);
1483 memcpy(ptr, list, sizeof(struct kmem_cache_node));
1485 * Do not assume that spinlocks can be initialized via memcpy:
1487 spin_lock_init(&ptr->list_lock);
1489 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1490 cachep->node[nodeid] = ptr;
1494 * For setting up all the kmem_list3s for cache whose buffer_size is same as
1495 * size of kmem_list3.
1497 static void __init set_up_list3s(struct kmem_cache *cachep, int index)
1501 for_each_online_node(node) {
1502 cachep->node[node] = &initkmem_list3[index + node];
1503 cachep->node[node]->next_reap = jiffies +
1505 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1510 * The memory after the last cpu cache pointer is used for the
1513 static void setup_node_pointer(struct kmem_cache *cachep)
1515 cachep->node = (struct kmem_cache_node **)&cachep->array[nr_cpu_ids];
1519 * Initialisation. Called after the page allocator have been initialised and
1520 * before smp_init().
1522 void __init kmem_cache_init(void)
1526 kmem_cache = &kmem_cache_boot;
1527 setup_node_pointer(kmem_cache);
1529 if (num_possible_nodes() == 1)
1530 use_alien_caches = 0;
1532 for (i = 0; i < NUM_INIT_LISTS; i++)
1533 kmem_list3_init(&initkmem_list3[i]);
1535 set_up_list3s(kmem_cache, CACHE_CACHE);
1538 * Fragmentation resistance on low memory - only use bigger
1539 * page orders on machines with more than 32MB of memory if
1540 * not overridden on the command line.
1542 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
1543 slab_max_order = SLAB_MAX_ORDER_HI;
1545 /* Bootstrap is tricky, because several objects are allocated
1546 * from caches that do not exist yet:
1547 * 1) initialize the kmem_cache cache: it contains the struct
1548 * kmem_cache structures of all caches, except kmem_cache itself:
1549 * kmem_cache is statically allocated.
1550 * Initially an __init data area is used for the head array and the
1551 * kmem_list3 structures, it's replaced with a kmalloc allocated
1552 * array at the end of the bootstrap.
1553 * 2) Create the first kmalloc cache.
1554 * The struct kmem_cache for the new cache is allocated normally.
1555 * An __init data area is used for the head array.
1556 * 3) Create the remaining kmalloc caches, with minimally sized
1558 * 4) Replace the __init data head arrays for kmem_cache and the first
1559 * kmalloc cache with kmalloc allocated arrays.
1560 * 5) Replace the __init data for kmem_list3 for kmem_cache and
1561 * the other cache's with kmalloc allocated memory.
1562 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1565 /* 1) create the kmem_cache */
1568 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1570 create_boot_cache(kmem_cache, "kmem_cache",
1571 offsetof(struct kmem_cache, array[nr_cpu_ids]) +
1572 nr_node_ids * sizeof(struct kmem_cache_node *),
1573 SLAB_HWCACHE_ALIGN);
1574 list_add(&kmem_cache->list, &slab_caches);
1576 /* 2+3) create the kmalloc caches */
1579 * Initialize the caches that provide memory for the array cache and the
1580 * kmem_list3 structures first. Without this, further allocations will
1584 kmalloc_caches[INDEX_AC] = create_kmalloc_cache("kmalloc-ac",
1585 kmalloc_size(INDEX_AC), ARCH_KMALLOC_FLAGS);
1587 if (INDEX_AC != INDEX_L3)
1588 kmalloc_caches[INDEX_L3] =
1589 create_kmalloc_cache("kmalloc-l3",
1590 kmalloc_size(INDEX_L3), ARCH_KMALLOC_FLAGS);
1592 slab_early_init = 0;
1594 /* 4) Replace the bootstrap head arrays */
1596 struct array_cache *ptr;
1598 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1600 memcpy(ptr, cpu_cache_get(kmem_cache),
1601 sizeof(struct arraycache_init));
1603 * Do not assume that spinlocks can be initialized via memcpy:
1605 spin_lock_init(&ptr->lock);
1607 kmem_cache->array[smp_processor_id()] = ptr;
1609 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1611 BUG_ON(cpu_cache_get(kmalloc_caches[INDEX_AC])
1612 != &initarray_generic.cache);
1613 memcpy(ptr, cpu_cache_get(kmalloc_caches[INDEX_AC]),
1614 sizeof(struct arraycache_init));
1616 * Do not assume that spinlocks can be initialized via memcpy:
1618 spin_lock_init(&ptr->lock);
1620 kmalloc_caches[INDEX_AC]->array[smp_processor_id()] = ptr;
1622 /* 5) Replace the bootstrap kmem_list3's */
1626 for_each_online_node(nid) {
1627 init_list(kmem_cache, &initkmem_list3[CACHE_CACHE + nid], nid);
1629 init_list(kmalloc_caches[INDEX_AC],
1630 &initkmem_list3[SIZE_AC + nid], nid);
1632 if (INDEX_AC != INDEX_L3) {
1633 init_list(kmalloc_caches[INDEX_L3],
1634 &initkmem_list3[SIZE_L3 + nid], nid);
1639 create_kmalloc_caches(ARCH_KMALLOC_FLAGS);
1642 void __init kmem_cache_init_late(void)
1644 struct kmem_cache *cachep;
1648 /* 6) resize the head arrays to their final sizes */
1649 mutex_lock(&slab_mutex);
1650 list_for_each_entry(cachep, &slab_caches, list)
1651 if (enable_cpucache(cachep, GFP_NOWAIT))
1653 mutex_unlock(&slab_mutex);
1655 /* Annotate slab for lockdep -- annotate the malloc caches */
1662 * Register a cpu startup notifier callback that initializes
1663 * cpu_cache_get for all new cpus
1665 register_cpu_notifier(&cpucache_notifier);
1669 * Register a memory hotplug callback that initializes and frees
1672 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1676 * The reap timers are started later, with a module init call: That part
1677 * of the kernel is not yet operational.
1681 static int __init cpucache_init(void)
1686 * Register the timers that return unneeded pages to the page allocator
1688 for_each_online_cpu(cpu)
1689 start_cpu_timer(cpu);
1695 __initcall(cpucache_init);
1697 static noinline void
1698 slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1700 struct kmem_cache_node *l3;
1702 unsigned long flags;
1706 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1708 printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n",
1709 cachep->name, cachep->size, cachep->gfporder);
1711 for_each_online_node(node) {
1712 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1713 unsigned long active_slabs = 0, num_slabs = 0;
1715 l3 = cachep->node[node];
1719 spin_lock_irqsave(&l3->list_lock, flags);
1720 list_for_each_entry(slabp, &l3->slabs_full, list) {
1721 active_objs += cachep->num;
1724 list_for_each_entry(slabp, &l3->slabs_partial, list) {
1725 active_objs += slabp->inuse;
1728 list_for_each_entry(slabp, &l3->slabs_free, list)
1731 free_objects += l3->free_objects;
1732 spin_unlock_irqrestore(&l3->list_lock, flags);
1734 num_slabs += active_slabs;
1735 num_objs = num_slabs * cachep->num;
1737 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1738 node, active_slabs, num_slabs, active_objs, num_objs,
1744 * Interface to system's page allocator. No need to hold the cache-lock.
1746 * If we requested dmaable memory, we will get it. Even if we
1747 * did not request dmaable memory, we might get it, but that
1748 * would be relatively rare and ignorable.
1750 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1758 * Nommu uses slab's for process anonymous memory allocations, and thus
1759 * requires __GFP_COMP to properly refcount higher order allocations
1761 flags |= __GFP_COMP;
1764 flags |= cachep->allocflags;
1765 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1766 flags |= __GFP_RECLAIMABLE;
1768 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1770 if (!(flags & __GFP_NOWARN) && printk_ratelimit())
1771 slab_out_of_memory(cachep, flags, nodeid);
1775 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1776 if (unlikely(page->pfmemalloc))
1777 pfmemalloc_active = true;
1779 nr_pages = (1 << cachep->gfporder);
1780 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1781 add_zone_page_state(page_zone(page),
1782 NR_SLAB_RECLAIMABLE, nr_pages);
1784 add_zone_page_state(page_zone(page),
1785 NR_SLAB_UNRECLAIMABLE, nr_pages);
1786 for (i = 0; i < nr_pages; i++) {
1787 __SetPageSlab(page + i);
1789 if (page->pfmemalloc)
1790 SetPageSlabPfmemalloc(page + i);
1792 memcg_bind_pages(cachep, cachep->gfporder);
1794 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1795 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1798 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1800 kmemcheck_mark_unallocated_pages(page, nr_pages);
1803 return page_address(page);
1807 * Interface to system's page release.
1809 static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1811 unsigned long i = (1 << cachep->gfporder);
1812 struct page *page = virt_to_page(addr);
1813 const unsigned long nr_freed = i;
1815 kmemcheck_free_shadow(page, cachep->gfporder);
1817 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1818 sub_zone_page_state(page_zone(page),
1819 NR_SLAB_RECLAIMABLE, nr_freed);
1821 sub_zone_page_state(page_zone(page),
1822 NR_SLAB_UNRECLAIMABLE, nr_freed);
1824 BUG_ON(!PageSlab(page));
1825 __ClearPageSlabPfmemalloc(page);
1826 __ClearPageSlab(page);
1830 memcg_release_pages(cachep, cachep->gfporder);
1831 if (current->reclaim_state)
1832 current->reclaim_state->reclaimed_slab += nr_freed;
1833 free_memcg_kmem_pages((unsigned long)addr, cachep->gfporder);
1836 static void kmem_rcu_free(struct rcu_head *head)
1838 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1839 struct kmem_cache *cachep = slab_rcu->cachep;
1841 kmem_freepages(cachep, slab_rcu->addr);
1842 if (OFF_SLAB(cachep))
1843 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1848 #ifdef CONFIG_DEBUG_PAGEALLOC
1849 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1850 unsigned long caller)
1852 int size = cachep->object_size;
1854 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1856 if (size < 5 * sizeof(unsigned long))
1859 *addr++ = 0x12345678;
1861 *addr++ = smp_processor_id();
1862 size -= 3 * sizeof(unsigned long);
1864 unsigned long *sptr = &caller;
1865 unsigned long svalue;
1867 while (!kstack_end(sptr)) {
1869 if (kernel_text_address(svalue)) {
1871 size -= sizeof(unsigned long);
1872 if (size <= sizeof(unsigned long))
1878 *addr++ = 0x87654321;
1882 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1884 int size = cachep->object_size;
1885 addr = &((char *)addr)[obj_offset(cachep)];
1887 memset(addr, val, size);
1888 *(unsigned char *)(addr + size - 1) = POISON_END;
1891 static void dump_line(char *data, int offset, int limit)
1894 unsigned char error = 0;
1897 printk(KERN_ERR "%03x: ", offset);
1898 for (i = 0; i < limit; i++) {
1899 if (data[offset + i] != POISON_FREE) {
1900 error = data[offset + i];
1904 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1905 &data[offset], limit, 1);
1907 if (bad_count == 1) {
1908 error ^= POISON_FREE;
1909 if (!(error & (error - 1))) {
1910 printk(KERN_ERR "Single bit error detected. Probably "
1913 printk(KERN_ERR "Run memtest86+ or a similar memory "
1916 printk(KERN_ERR "Run a memory test tool.\n");
1925 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1930 if (cachep->flags & SLAB_RED_ZONE) {
1931 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
1932 *dbg_redzone1(cachep, objp),
1933 *dbg_redzone2(cachep, objp));
1936 if (cachep->flags & SLAB_STORE_USER) {
1937 printk(KERN_ERR "Last user: [<%p>]",
1938 *dbg_userword(cachep, objp));
1939 print_symbol("(%s)",
1940 (unsigned long)*dbg_userword(cachep, objp));
1943 realobj = (char *)objp + obj_offset(cachep);
1944 size = cachep->object_size;
1945 for (i = 0; i < size && lines; i += 16, lines--) {
1948 if (i + limit > size)
1950 dump_line(realobj, i, limit);
1954 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1960 realobj = (char *)objp + obj_offset(cachep);
1961 size = cachep->object_size;
1963 for (i = 0; i < size; i++) {
1964 char exp = POISON_FREE;
1967 if (realobj[i] != exp) {
1973 "Slab corruption (%s): %s start=%p, len=%d\n",
1974 print_tainted(), cachep->name, realobj, size);
1975 print_objinfo(cachep, objp, 0);
1977 /* Hexdump the affected line */
1980 if (i + limit > size)
1982 dump_line(realobj, i, limit);
1985 /* Limit to 5 lines */
1991 /* Print some data about the neighboring objects, if they
1994 struct slab *slabp = virt_to_slab(objp);
1997 objnr = obj_to_index(cachep, slabp, objp);
1999 objp = index_to_obj(cachep, slabp, objnr - 1);
2000 realobj = (char *)objp + obj_offset(cachep);
2001 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
2003 print_objinfo(cachep, objp, 2);
2005 if (objnr + 1 < cachep->num) {
2006 objp = index_to_obj(cachep, slabp, objnr + 1);
2007 realobj = (char *)objp + obj_offset(cachep);
2008 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
2010 print_objinfo(cachep, objp, 2);
2017 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
2020 for (i = 0; i < cachep->num; i++) {
2021 void *objp = index_to_obj(cachep, slabp, i);
2023 if (cachep->flags & SLAB_POISON) {
2024 #ifdef CONFIG_DEBUG_PAGEALLOC
2025 if (cachep->size % PAGE_SIZE == 0 &&
2027 kernel_map_pages(virt_to_page(objp),
2028 cachep->size / PAGE_SIZE, 1);
2030 check_poison_obj(cachep, objp);
2032 check_poison_obj(cachep, objp);
2035 if (cachep->flags & SLAB_RED_ZONE) {
2036 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2037 slab_error(cachep, "start of a freed object "
2039 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2040 slab_error(cachep, "end of a freed object "
2046 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
2052 * slab_destroy - destroy and release all objects in a slab
2053 * @cachep: cache pointer being destroyed
2054 * @slabp: slab pointer being destroyed
2056 * Destroy all the objs in a slab, and release the mem back to the system.
2057 * Before calling the slab must have been unlinked from the cache. The
2058 * cache-lock is not held/needed.
2060 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
2062 void *addr = slabp->s_mem - slabp->colouroff;
2064 slab_destroy_debugcheck(cachep, slabp);
2065 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
2066 struct slab_rcu *slab_rcu;
2068 slab_rcu = (struct slab_rcu *)slabp;
2069 slab_rcu->cachep = cachep;
2070 slab_rcu->addr = addr;
2071 call_rcu(&slab_rcu->head, kmem_rcu_free);
2073 kmem_freepages(cachep, addr);
2074 if (OFF_SLAB(cachep))
2075 kmem_cache_free(cachep->slabp_cache, slabp);
2080 * calculate_slab_order - calculate size (page order) of slabs
2081 * @cachep: pointer to the cache that is being created
2082 * @size: size of objects to be created in this cache.
2083 * @align: required alignment for the objects.
2084 * @flags: slab allocation flags
2086 * Also calculates the number of objects per slab.
2088 * This could be made much more intelligent. For now, try to avoid using
2089 * high order pages for slabs. When the gfp() functions are more friendly
2090 * towards high-order requests, this should be changed.
2092 static size_t calculate_slab_order(struct kmem_cache *cachep,
2093 size_t size, size_t align, unsigned long flags)
2095 unsigned long offslab_limit;
2096 size_t left_over = 0;
2099 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
2103 cache_estimate(gfporder, size, align, flags, &remainder, &num);
2107 if (flags & CFLGS_OFF_SLAB) {
2109 * Max number of objs-per-slab for caches which
2110 * use off-slab slabs. Needed to avoid a possible
2111 * looping condition in cache_grow().
2113 offslab_limit = size - sizeof(struct slab);
2114 offslab_limit /= sizeof(kmem_bufctl_t);
2116 if (num > offslab_limit)
2120 /* Found something acceptable - save it away */
2122 cachep->gfporder = gfporder;
2123 left_over = remainder;
2126 * A VFS-reclaimable slab tends to have most allocations
2127 * as GFP_NOFS and we really don't want to have to be allocating
2128 * higher-order pages when we are unable to shrink dcache.
2130 if (flags & SLAB_RECLAIM_ACCOUNT)
2134 * Large number of objects is good, but very large slabs are
2135 * currently bad for the gfp()s.
2137 if (gfporder >= slab_max_order)
2141 * Acceptable internal fragmentation?
2143 if (left_over * 8 <= (PAGE_SIZE << gfporder))
2149 static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
2151 if (slab_state >= FULL)
2152 return enable_cpucache(cachep, gfp);
2154 if (slab_state == DOWN) {
2156 * Note: Creation of first cache (kmem_cache).
2157 * The setup_list3s is taken care
2158 * of by the caller of __kmem_cache_create
2160 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2161 slab_state = PARTIAL;
2162 } else if (slab_state == PARTIAL) {
2164 * Note: the second kmem_cache_create must create the cache
2165 * that's used by kmalloc(24), otherwise the creation of
2166 * further caches will BUG().
2168 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2171 * If the cache that's used by kmalloc(sizeof(kmem_list3)) is
2172 * the second cache, then we need to set up all its list3s,
2173 * otherwise the creation of further caches will BUG().
2175 set_up_list3s(cachep, SIZE_AC);
2176 if (INDEX_AC == INDEX_L3)
2177 slab_state = PARTIAL_L3;
2179 slab_state = PARTIAL_ARRAYCACHE;
2181 /* Remaining boot caches */
2182 cachep->array[smp_processor_id()] =
2183 kmalloc(sizeof(struct arraycache_init), gfp);
2185 if (slab_state == PARTIAL_ARRAYCACHE) {
2186 set_up_list3s(cachep, SIZE_L3);
2187 slab_state = PARTIAL_L3;
2190 for_each_online_node(node) {
2191 cachep->node[node] =
2192 kmalloc_node(sizeof(struct kmem_cache_node),
2194 BUG_ON(!cachep->node[node]);
2195 kmem_list3_init(cachep->node[node]);
2199 cachep->node[numa_mem_id()]->next_reap =
2200 jiffies + REAPTIMEOUT_LIST3 +
2201 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2203 cpu_cache_get(cachep)->avail = 0;
2204 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2205 cpu_cache_get(cachep)->batchcount = 1;
2206 cpu_cache_get(cachep)->touched = 0;
2207 cachep->batchcount = 1;
2208 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2213 * __kmem_cache_create - Create a cache.
2214 * @cachep: cache management descriptor
2215 * @flags: SLAB flags
2217 * Returns a ptr to the cache on success, NULL on failure.
2218 * Cannot be called within a int, but can be interrupted.
2219 * The @ctor is run when new pages are allocated by the cache.
2223 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2224 * to catch references to uninitialised memory.
2226 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2227 * for buffer overruns.
2229 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2230 * cacheline. This can be beneficial if you're counting cycles as closely
2234 __kmem_cache_create (struct kmem_cache *cachep, unsigned long flags)
2236 size_t left_over, slab_size, ralign;
2239 size_t size = cachep->size;
2244 * Enable redzoning and last user accounting, except for caches with
2245 * large objects, if the increased size would increase the object size
2246 * above the next power of two: caches with object sizes just above a
2247 * power of two have a significant amount of internal fragmentation.
2249 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2250 2 * sizeof(unsigned long long)))
2251 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2252 if (!(flags & SLAB_DESTROY_BY_RCU))
2253 flags |= SLAB_POISON;
2255 if (flags & SLAB_DESTROY_BY_RCU)
2256 BUG_ON(flags & SLAB_POISON);
2260 * Check that size is in terms of words. This is needed to avoid
2261 * unaligned accesses for some archs when redzoning is used, and makes
2262 * sure any on-slab bufctl's are also correctly aligned.
2264 if (size & (BYTES_PER_WORD - 1)) {
2265 size += (BYTES_PER_WORD - 1);
2266 size &= ~(BYTES_PER_WORD - 1);
2270 * Redzoning and user store require word alignment or possibly larger.
2271 * Note this will be overridden by architecture or caller mandated
2272 * alignment if either is greater than BYTES_PER_WORD.
2274 if (flags & SLAB_STORE_USER)
2275 ralign = BYTES_PER_WORD;
2277 if (flags & SLAB_RED_ZONE) {
2278 ralign = REDZONE_ALIGN;
2279 /* If redzoning, ensure that the second redzone is suitably
2280 * aligned, by adjusting the object size accordingly. */
2281 size += REDZONE_ALIGN - 1;
2282 size &= ~(REDZONE_ALIGN - 1);
2285 /* 3) caller mandated alignment */
2286 if (ralign < cachep->align) {
2287 ralign = cachep->align;
2289 /* disable debug if necessary */
2290 if (ralign > __alignof__(unsigned long long))
2291 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2295 cachep->align = ralign;
2297 if (slab_is_available())
2302 setup_node_pointer(cachep);
2306 * Both debugging options require word-alignment which is calculated
2309 if (flags & SLAB_RED_ZONE) {
2310 /* add space for red zone words */
2311 cachep->obj_offset += sizeof(unsigned long long);
2312 size += 2 * sizeof(unsigned long long);
2314 if (flags & SLAB_STORE_USER) {
2315 /* user store requires one word storage behind the end of
2316 * the real object. But if the second red zone needs to be
2317 * aligned to 64 bits, we must allow that much space.
2319 if (flags & SLAB_RED_ZONE)
2320 size += REDZONE_ALIGN;
2322 size += BYTES_PER_WORD;
2324 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2325 if (size >= kmalloc_size(INDEX_L3 + 1)
2326 && cachep->object_size > cache_line_size() && ALIGN(size, align) < PAGE_SIZE) {
2327 cachep->obj_offset += PAGE_SIZE - ALIGN(size, align);
2334 * Determine if the slab management is 'on' or 'off' slab.
2335 * (bootstrapping cannot cope with offslab caches so don't do
2336 * it too early on. Always use on-slab management when
2337 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2339 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
2340 !(flags & SLAB_NOLEAKTRACE))
2342 * Size is large, assume best to place the slab management obj
2343 * off-slab (should allow better packing of objs).
2345 flags |= CFLGS_OFF_SLAB;
2347 size = ALIGN(size, cachep->align);
2349 left_over = calculate_slab_order(cachep, size, cachep->align, flags);
2354 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2355 + sizeof(struct slab), cachep->align);
2358 * If the slab has been placed off-slab, and we have enough space then
2359 * move it on-slab. This is at the expense of any extra colouring.
2361 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2362 flags &= ~CFLGS_OFF_SLAB;
2363 left_over -= slab_size;
2366 if (flags & CFLGS_OFF_SLAB) {
2367 /* really off slab. No need for manual alignment */
2369 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2371 #ifdef CONFIG_PAGE_POISONING
2372 /* If we're going to use the generic kernel_map_pages()
2373 * poisoning, then it's going to smash the contents of
2374 * the redzone and userword anyhow, so switch them off.
2376 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2377 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2381 cachep->colour_off = cache_line_size();
2382 /* Offset must be a multiple of the alignment. */
2383 if (cachep->colour_off < cachep->align)
2384 cachep->colour_off = cachep->align;
2385 cachep->colour = left_over / cachep->colour_off;
2386 cachep->slab_size = slab_size;
2387 cachep->flags = flags;
2388 cachep->allocflags = 0;
2389 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2390 cachep->allocflags |= GFP_DMA;
2391 cachep->size = size;
2392 cachep->reciprocal_buffer_size = reciprocal_value(size);
2394 if (flags & CFLGS_OFF_SLAB) {
2395 cachep->slabp_cache = kmalloc_slab(slab_size, 0u);
2397 * This is a possibility for one of the malloc_sizes caches.
2398 * But since we go off slab only for object size greater than
2399 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2400 * this should not happen at all.
2401 * But leave a BUG_ON for some lucky dude.
2403 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
2406 err = setup_cpu_cache(cachep, gfp);
2408 __kmem_cache_shutdown(cachep);
2412 if (flags & SLAB_DEBUG_OBJECTS) {
2414 * Would deadlock through slab_destroy()->call_rcu()->
2415 * debug_object_activate()->kmem_cache_alloc().
2417 WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU);
2419 slab_set_debugobj_lock_classes(cachep);
2420 } else if (!OFF_SLAB(cachep) && !(flags & SLAB_DESTROY_BY_RCU))
2421 on_slab_lock_classes(cachep);
2427 static void check_irq_off(void)
2429 BUG_ON(!irqs_disabled());
2432 static void check_irq_on(void)
2434 BUG_ON(irqs_disabled());
2437 static void check_spinlock_acquired(struct kmem_cache *cachep)
2441 assert_spin_locked(&cachep->node[numa_mem_id()]->list_lock);
2445 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2449 assert_spin_locked(&cachep->node[node]->list_lock);
2454 #define check_irq_off() do { } while(0)
2455 #define check_irq_on() do { } while(0)
2456 #define check_spinlock_acquired(x) do { } while(0)
2457 #define check_spinlock_acquired_node(x, y) do { } while(0)
2460 static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *l3,
2461 struct array_cache *ac,
2462 int force, int node);
2464 static void do_drain(void *arg)
2466 struct kmem_cache *cachep = arg;
2467 struct array_cache *ac;
2468 int node = numa_mem_id();
2471 ac = cpu_cache_get(cachep);
2472 spin_lock(&cachep->node[node]->list_lock);
2473 free_block(cachep, ac->entry, ac->avail, node);
2474 spin_unlock(&cachep->node[node]->list_lock);
2478 static void drain_cpu_caches(struct kmem_cache *cachep)
2480 struct kmem_cache_node *l3;
2483 on_each_cpu(do_drain, cachep, 1);
2485 for_each_online_node(node) {
2486 l3 = cachep->node[node];
2487 if (l3 && l3->alien)
2488 drain_alien_cache(cachep, l3->alien);
2491 for_each_online_node(node) {
2492 l3 = cachep->node[node];
2494 drain_array(cachep, l3, l3->shared, 1, node);
2499 * Remove slabs from the list of free slabs.
2500 * Specify the number of slabs to drain in tofree.
2502 * Returns the actual number of slabs released.
2504 static int drain_freelist(struct kmem_cache *cache,
2505 struct kmem_cache_node *l3, int tofree)
2507 struct list_head *p;
2512 while (nr_freed < tofree && !list_empty(&l3->slabs_free)) {
2514 spin_lock_irq(&l3->list_lock);
2515 p = l3->slabs_free.prev;
2516 if (p == &l3->slabs_free) {
2517 spin_unlock_irq(&l3->list_lock);
2521 slabp = list_entry(p, struct slab, list);
2523 BUG_ON(slabp->inuse);
2525 list_del(&slabp->list);
2527 * Safe to drop the lock. The slab is no longer linked
2530 l3->free_objects -= cache->num;
2531 spin_unlock_irq(&l3->list_lock);
2532 slab_destroy(cache, slabp);
2539 /* Called with slab_mutex held to protect against cpu hotplug */
2540 static int __cache_shrink(struct kmem_cache *cachep)
2543 struct kmem_cache_node *l3;
2545 drain_cpu_caches(cachep);
2548 for_each_online_node(i) {
2549 l3 = cachep->node[i];
2553 drain_freelist(cachep, l3, l3->free_objects);
2555 ret += !list_empty(&l3->slabs_full) ||
2556 !list_empty(&l3->slabs_partial);
2558 return (ret ? 1 : 0);
2562 * kmem_cache_shrink - Shrink a cache.
2563 * @cachep: The cache to shrink.
2565 * Releases as many slabs as possible for a cache.
2566 * To help debugging, a zero exit status indicates all slabs were released.
2568 int kmem_cache_shrink(struct kmem_cache *cachep)
2571 BUG_ON(!cachep || in_interrupt());
2574 mutex_lock(&slab_mutex);
2575 ret = __cache_shrink(cachep);
2576 mutex_unlock(&slab_mutex);
2580 EXPORT_SYMBOL(kmem_cache_shrink);
2582 int __kmem_cache_shutdown(struct kmem_cache *cachep)
2585 struct kmem_cache_node *l3;
2586 int rc = __cache_shrink(cachep);
2591 for_each_online_cpu(i)
2592 kfree(cachep->array[i]);
2594 /* NUMA: free the list3 structures */
2595 for_each_online_node(i) {
2596 l3 = cachep->node[i];
2599 free_alien_cache(l3->alien);
2607 * Get the memory for a slab management obj.
2608 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2609 * always come from malloc_sizes caches. The slab descriptor cannot
2610 * come from the same cache which is getting created because,
2611 * when we are searching for an appropriate cache for these
2612 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2613 * If we are creating a malloc_sizes cache here it would not be visible to
2614 * kmem_find_general_cachep till the initialization is complete.
2615 * Hence we cannot have slabp_cache same as the original cache.
2617 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2618 int colour_off, gfp_t local_flags,
2623 if (OFF_SLAB(cachep)) {
2624 /* Slab management obj is off-slab. */
2625 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2626 local_flags, nodeid);
2628 * If the first object in the slab is leaked (it's allocated
2629 * but no one has a reference to it), we want to make sure
2630 * kmemleak does not treat the ->s_mem pointer as a reference
2631 * to the object. Otherwise we will not report the leak.
2633 kmemleak_scan_area(&slabp->list, sizeof(struct list_head),
2638 slabp = objp + colour_off;
2639 colour_off += cachep->slab_size;
2642 slabp->colouroff = colour_off;
2643 slabp->s_mem = objp + colour_off;
2644 slabp->nodeid = nodeid;
2649 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2651 return (kmem_bufctl_t *) (slabp + 1);
2654 static void cache_init_objs(struct kmem_cache *cachep,
2659 for (i = 0; i < cachep->num; i++) {
2660 void *objp = index_to_obj(cachep, slabp, i);
2662 /* need to poison the objs? */
2663 if (cachep->flags & SLAB_POISON)
2664 poison_obj(cachep, objp, POISON_FREE);
2665 if (cachep->flags & SLAB_STORE_USER)
2666 *dbg_userword(cachep, objp) = NULL;
2668 if (cachep->flags & SLAB_RED_ZONE) {
2669 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2670 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2673 * Constructors are not allowed to allocate memory from the same
2674 * cache which they are a constructor for. Otherwise, deadlock.
2675 * They must also be threaded.
2677 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2678 cachep->ctor(objp + obj_offset(cachep));
2680 if (cachep->flags & SLAB_RED_ZONE) {
2681 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2682 slab_error(cachep, "constructor overwrote the"
2683 " end of an object");
2684 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2685 slab_error(cachep, "constructor overwrote the"
2686 " start of an object");
2688 if ((cachep->size % PAGE_SIZE) == 0 &&
2689 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2690 kernel_map_pages(virt_to_page(objp),
2691 cachep->size / PAGE_SIZE, 0);
2696 slab_bufctl(slabp)[i] = i + 1;
2698 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2701 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2703 if (CONFIG_ZONE_DMA_FLAG) {
2704 if (flags & GFP_DMA)
2705 BUG_ON(!(cachep->allocflags & GFP_DMA));
2707 BUG_ON(cachep->allocflags & GFP_DMA);
2711 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2714 void *objp = index_to_obj(cachep, slabp, slabp->free);
2718 next = slab_bufctl(slabp)[slabp->free];
2720 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2721 WARN_ON(slabp->nodeid != nodeid);
2728 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2729 void *objp, int nodeid)
2731 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2734 /* Verify that the slab belongs to the intended node */
2735 WARN_ON(slabp->nodeid != nodeid);
2737 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2738 printk(KERN_ERR "slab: double free detected in cache "
2739 "'%s', objp %p\n", cachep->name, objp);
2743 slab_bufctl(slabp)[objnr] = slabp->free;
2744 slabp->free = objnr;
2749 * Map pages beginning at addr to the given cache and slab. This is required
2750 * for the slab allocator to be able to lookup the cache and slab of a
2751 * virtual address for kfree, ksize, and slab debugging.
2753 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2759 page = virt_to_page(addr);
2762 if (likely(!PageCompound(page)))
2763 nr_pages <<= cache->gfporder;
2766 page->slab_cache = cache;
2767 page->slab_page = slab;
2769 } while (--nr_pages);
2773 * Grow (by 1) the number of slabs within a cache. This is called by
2774 * kmem_cache_alloc() when there are no active objs left in a cache.
2776 static int cache_grow(struct kmem_cache *cachep,
2777 gfp_t flags, int nodeid, void *objp)
2782 struct kmem_cache_node *l3;
2785 * Be lazy and only check for valid flags here, keeping it out of the
2786 * critical path in kmem_cache_alloc().
2788 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2789 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2791 /* Take the l3 list lock to change the colour_next on this node */
2793 l3 = cachep->node[nodeid];
2794 spin_lock(&l3->list_lock);
2796 /* Get colour for the slab, and cal the next value. */
2797 offset = l3->colour_next;
2799 if (l3->colour_next >= cachep->colour)
2800 l3->colour_next = 0;
2801 spin_unlock(&l3->list_lock);
2803 offset *= cachep->colour_off;
2805 if (local_flags & __GFP_WAIT)
2809 * The test for missing atomic flag is performed here, rather than
2810 * the more obvious place, simply to reduce the critical path length
2811 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2812 * will eventually be caught here (where it matters).
2814 kmem_flagcheck(cachep, flags);
2817 * Get mem for the objs. Attempt to allocate a physical page from
2821 objp = kmem_getpages(cachep, local_flags, nodeid);
2825 /* Get slab management. */
2826 slabp = alloc_slabmgmt(cachep, objp, offset,
2827 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2831 slab_map_pages(cachep, slabp, objp);
2833 cache_init_objs(cachep, slabp);
2835 if (local_flags & __GFP_WAIT)
2836 local_irq_disable();
2838 spin_lock(&l3->list_lock);
2840 /* Make slab active. */
2841 list_add_tail(&slabp->list, &(l3->slabs_free));
2842 STATS_INC_GROWN(cachep);
2843 l3->free_objects += cachep->num;
2844 spin_unlock(&l3->list_lock);
2847 kmem_freepages(cachep, objp);
2849 if (local_flags & __GFP_WAIT)
2850 local_irq_disable();
2857 * Perform extra freeing checks:
2858 * - detect bad pointers.
2859 * - POISON/RED_ZONE checking
2861 static void kfree_debugcheck(const void *objp)
2863 if (!virt_addr_valid(objp)) {
2864 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2865 (unsigned long)objp);
2870 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2872 unsigned long long redzone1, redzone2;
2874 redzone1 = *dbg_redzone1(cache, obj);
2875 redzone2 = *dbg_redzone2(cache, obj);
2880 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2883 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2884 slab_error(cache, "double free detected");
2886 slab_error(cache, "memory outside object was overwritten");
2888 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2889 obj, redzone1, redzone2);
2892 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2893 unsigned long caller)
2899 BUG_ON(virt_to_cache(objp) != cachep);
2901 objp -= obj_offset(cachep);
2902 kfree_debugcheck(objp);
2903 page = virt_to_head_page(objp);
2905 slabp = page->slab_page;
2907 if (cachep->flags & SLAB_RED_ZONE) {
2908 verify_redzone_free(cachep, objp);
2909 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2910 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2912 if (cachep->flags & SLAB_STORE_USER)
2913 *dbg_userword(cachep, objp) = (void *)caller;
2915 objnr = obj_to_index(cachep, slabp, objp);
2917 BUG_ON(objnr >= cachep->num);
2918 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
2920 #ifdef CONFIG_DEBUG_SLAB_LEAK
2921 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2923 if (cachep->flags & SLAB_POISON) {
2924 #ifdef CONFIG_DEBUG_PAGEALLOC
2925 if ((cachep->size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
2926 store_stackinfo(cachep, objp, caller);
2927 kernel_map_pages(virt_to_page(objp),
2928 cachep->size / PAGE_SIZE, 0);
2930 poison_obj(cachep, objp, POISON_FREE);
2933 poison_obj(cachep, objp, POISON_FREE);
2939 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
2944 /* Check slab's freelist to see if this obj is there. */
2945 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2947 if (entries > cachep->num || i >= cachep->num)
2950 if (entries != cachep->num - slabp->inuse) {
2952 printk(KERN_ERR "slab: Internal list corruption detected in "
2953 "cache '%s'(%d), slabp %p(%d). Tainted(%s). Hexdump:\n",
2954 cachep->name, cachep->num, slabp, slabp->inuse,
2956 print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 16, 1, slabp,
2957 sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t),
2963 #define kfree_debugcheck(x) do { } while(0)
2964 #define cache_free_debugcheck(x,objp,z) (objp)
2965 #define check_slabp(x,y) do { } while(0)
2968 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags,
2972 struct kmem_cache_node *l3;
2973 struct array_cache *ac;
2977 node = numa_mem_id();
2978 if (unlikely(force_refill))
2981 ac = cpu_cache_get(cachep);
2982 batchcount = ac->batchcount;
2983 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2985 * If there was little recent activity on this cache, then
2986 * perform only a partial refill. Otherwise we could generate
2989 batchcount = BATCHREFILL_LIMIT;
2991 l3 = cachep->node[node];
2993 BUG_ON(ac->avail > 0 || !l3);
2994 spin_lock(&l3->list_lock);
2996 /* See if we can refill from the shared array */
2997 if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) {
2998 l3->shared->touched = 1;
3002 while (batchcount > 0) {
3003 struct list_head *entry;
3005 /* Get slab alloc is to come from. */
3006 entry = l3->slabs_partial.next;
3007 if (entry == &l3->slabs_partial) {
3008 l3->free_touched = 1;
3009 entry = l3->slabs_free.next;
3010 if (entry == &l3->slabs_free)
3014 slabp = list_entry(entry, struct slab, list);
3015 check_slabp(cachep, slabp);
3016 check_spinlock_acquired(cachep);
3019 * The slab was either on partial or free list so
3020 * there must be at least one object available for
3023 BUG_ON(slabp->inuse >= cachep->num);
3025 while (slabp->inuse < cachep->num && batchcount--) {
3026 STATS_INC_ALLOCED(cachep);
3027 STATS_INC_ACTIVE(cachep);
3028 STATS_SET_HIGH(cachep);
3030 ac_put_obj(cachep, ac, slab_get_obj(cachep, slabp,
3033 check_slabp(cachep, slabp);
3035 /* move slabp to correct slabp list: */
3036 list_del(&slabp->list);
3037 if (slabp->free == BUFCTL_END)
3038 list_add(&slabp->list, &l3->slabs_full);
3040 list_add(&slabp->list, &l3->slabs_partial);
3044 l3->free_objects -= ac->avail;
3046 spin_unlock(&l3->list_lock);
3048 if (unlikely(!ac->avail)) {
3051 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
3053 /* cache_grow can reenable interrupts, then ac could change. */
3054 ac = cpu_cache_get(cachep);
3055 node = numa_mem_id();
3057 /* no objects in sight? abort */
3058 if (!x && (ac->avail == 0 || force_refill))
3061 if (!ac->avail) /* objects refilled by interrupt? */
3066 return ac_get_obj(cachep, ac, flags, force_refill);
3069 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3072 might_sleep_if(flags & __GFP_WAIT);
3074 kmem_flagcheck(cachep, flags);
3079 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3080 gfp_t flags, void *objp, unsigned long caller)
3084 if (cachep->flags & SLAB_POISON) {
3085 #ifdef CONFIG_DEBUG_PAGEALLOC
3086 if ((cachep->size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
3087 kernel_map_pages(virt_to_page(objp),
3088 cachep->size / PAGE_SIZE, 1);
3090 check_poison_obj(cachep, objp);
3092 check_poison_obj(cachep, objp);
3094 poison_obj(cachep, objp, POISON_INUSE);
3096 if (cachep->flags & SLAB_STORE_USER)
3097 *dbg_userword(cachep, objp) = (void *)caller;
3099 if (cachep->flags & SLAB_RED_ZONE) {
3100 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3101 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3102 slab_error(cachep, "double free, or memory outside"
3103 " object was overwritten");
3105 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3106 objp, *dbg_redzone1(cachep, objp),
3107 *dbg_redzone2(cachep, objp));
3109 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3110 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3112 #ifdef CONFIG_DEBUG_SLAB_LEAK
3117 slabp = virt_to_head_page(objp)->slab_page;
3118 objnr = (unsigned)(objp - slabp->s_mem) / cachep->size;
3119 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3122 objp += obj_offset(cachep);
3123 if (cachep->ctor && cachep->flags & SLAB_POISON)
3125 if (ARCH_SLAB_MINALIGN &&
3126 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
3127 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3128 objp, (int)ARCH_SLAB_MINALIGN);
3133 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3136 static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
3138 if (cachep == kmem_cache)
3141 return should_failslab(cachep->object_size, flags, cachep->flags);
3144 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3147 struct array_cache *ac;
3148 bool force_refill = false;
3152 ac = cpu_cache_get(cachep);
3153 if (likely(ac->avail)) {
3155 objp = ac_get_obj(cachep, ac, flags, false);
3158 * Allow for the possibility all avail objects are not allowed
3159 * by the current flags
3162 STATS_INC_ALLOCHIT(cachep);
3165 force_refill = true;
3168 STATS_INC_ALLOCMISS(cachep);
3169 objp = cache_alloc_refill(cachep, flags, force_refill);
3171 * the 'ac' may be updated by cache_alloc_refill(),
3172 * and kmemleak_erase() requires its correct value.
3174 ac = cpu_cache_get(cachep);
3178 * To avoid a false negative, if an object that is in one of the
3179 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3180 * treat the array pointers as a reference to the object.
3183 kmemleak_erase(&ac->entry[ac->avail]);
3189 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3191 * If we are in_interrupt, then process context, including cpusets and
3192 * mempolicy, may not apply and should not be used for allocation policy.
3194 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3196 int nid_alloc, nid_here;
3198 if (in_interrupt() || (flags & __GFP_THISNODE))
3200 nid_alloc = nid_here = numa_mem_id();
3201 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3202 nid_alloc = cpuset_slab_spread_node();
3203 else if (current->mempolicy)
3204 nid_alloc = slab_node();
3205 if (nid_alloc != nid_here)
3206 return ____cache_alloc_node(cachep, flags, nid_alloc);
3211 * Fallback function if there was no memory available and no objects on a
3212 * certain node and fall back is permitted. First we scan all the
3213 * available node for available objects. If that fails then we
3214 * perform an allocation without specifying a node. This allows the page
3215 * allocator to do its reclaim / fallback magic. We then insert the
3216 * slab into the proper nodelist and then allocate from it.
3218 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3220 struct zonelist *zonelist;
3224 enum zone_type high_zoneidx = gfp_zone(flags);
3227 unsigned int cpuset_mems_cookie;
3229 if (flags & __GFP_THISNODE)
3232 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3235 cpuset_mems_cookie = get_mems_allowed();
3236 zonelist = node_zonelist(slab_node(), flags);
3240 * Look through allowed nodes for objects available
3241 * from existing per node queues.
3243 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3244 nid = zone_to_nid(zone);
3246 if (cpuset_zone_allowed_hardwall(zone, flags) &&
3248 cache->node[nid]->free_objects) {
3249 obj = ____cache_alloc_node(cache,
3250 flags | GFP_THISNODE, nid);
3258 * This allocation will be performed within the constraints
3259 * of the current cpuset / memory policy requirements.
3260 * We may trigger various forms of reclaim on the allowed
3261 * set and go into memory reserves if necessary.
3263 if (local_flags & __GFP_WAIT)
3265 kmem_flagcheck(cache, flags);
3266 obj = kmem_getpages(cache, local_flags, numa_mem_id());
3267 if (local_flags & __GFP_WAIT)
3268 local_irq_disable();
3271 * Insert into the appropriate per node queues
3273 nid = page_to_nid(virt_to_page(obj));
3274 if (cache_grow(cache, flags, nid, obj)) {
3275 obj = ____cache_alloc_node(cache,
3276 flags | GFP_THISNODE, nid);
3279 * Another processor may allocate the
3280 * objects in the slab since we are
3281 * not holding any locks.
3285 /* cache_grow already freed obj */
3291 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !obj))
3297 * A interface to enable slab creation on nodeid
3299 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3302 struct list_head *entry;
3304 struct kmem_cache_node *l3;
3308 l3 = cachep->node[nodeid];
3313 spin_lock(&l3->list_lock);
3314 entry = l3->slabs_partial.next;
3315 if (entry == &l3->slabs_partial) {
3316 l3->free_touched = 1;
3317 entry = l3->slabs_free.next;
3318 if (entry == &l3->slabs_free)
3322 slabp = list_entry(entry, struct slab, list);
3323 check_spinlock_acquired_node(cachep, nodeid);
3324 check_slabp(cachep, slabp);
3326 STATS_INC_NODEALLOCS(cachep);
3327 STATS_INC_ACTIVE(cachep);
3328 STATS_SET_HIGH(cachep);
3330 BUG_ON(slabp->inuse == cachep->num);
3332 obj = slab_get_obj(cachep, slabp, nodeid);
3333 check_slabp(cachep, slabp);
3335 /* move slabp to correct slabp list: */
3336 list_del(&slabp->list);
3338 if (slabp->free == BUFCTL_END)
3339 list_add(&slabp->list, &l3->slabs_full);
3341 list_add(&slabp->list, &l3->slabs_partial);
3343 spin_unlock(&l3->list_lock);
3347 spin_unlock(&l3->list_lock);
3348 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
3352 return fallback_alloc(cachep, flags);
3359 * kmem_cache_alloc_node - Allocate an object on the specified node
3360 * @cachep: The cache to allocate from.
3361 * @flags: See kmalloc().
3362 * @nodeid: node number of the target node.
3363 * @caller: return address of caller, used for debug information
3365 * Identical to kmem_cache_alloc but it will allocate memory on the given
3366 * node, which can improve the performance for cpu bound structures.
3368 * Fallback to other node is possible if __GFP_THISNODE is not set.
3370 static __always_inline void *
3371 slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3372 unsigned long caller)
3374 unsigned long save_flags;
3376 int slab_node = numa_mem_id();
3378 flags &= gfp_allowed_mask;
3380 lockdep_trace_alloc(flags);
3382 if (slab_should_failslab(cachep, flags))
3385 cachep = memcg_kmem_get_cache(cachep, flags);
3387 cache_alloc_debugcheck_before(cachep, flags);
3388 local_irq_save(save_flags);
3390 if (nodeid == NUMA_NO_NODE)
3393 if (unlikely(!cachep->node[nodeid])) {
3394 /* Node not bootstrapped yet */
3395 ptr = fallback_alloc(cachep, flags);
3399 if (nodeid == slab_node) {
3401 * Use the locally cached objects if possible.
3402 * However ____cache_alloc does not allow fallback
3403 * to other nodes. It may fail while we still have
3404 * objects on other nodes available.
3406 ptr = ____cache_alloc(cachep, flags);
3410 /* ___cache_alloc_node can fall back to other nodes */
3411 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3413 local_irq_restore(save_flags);
3414 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3415 kmemleak_alloc_recursive(ptr, cachep->object_size, 1, cachep->flags,
3419 kmemcheck_slab_alloc(cachep, flags, ptr, cachep->object_size);
3421 if (unlikely((flags & __GFP_ZERO) && ptr))
3422 memset(ptr, 0, cachep->object_size);
3427 static __always_inline void *
3428 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3432 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3433 objp = alternate_node_alloc(cache, flags);
3437 objp = ____cache_alloc(cache, flags);
3440 * We may just have run out of memory on the local node.
3441 * ____cache_alloc_node() knows how to locate memory on other nodes
3444 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3451 static __always_inline void *
3452 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3454 return ____cache_alloc(cachep, flags);
3457 #endif /* CONFIG_NUMA */
3459 static __always_inline void *
3460 slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
3462 unsigned long save_flags;
3465 flags &= gfp_allowed_mask;
3467 lockdep_trace_alloc(flags);
3469 if (slab_should_failslab(cachep, flags))
3472 cachep = memcg_kmem_get_cache(cachep, flags);
3474 cache_alloc_debugcheck_before(cachep, flags);
3475 local_irq_save(save_flags);
3476 objp = __do_cache_alloc(cachep, flags);
3477 local_irq_restore(save_flags);
3478 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3479 kmemleak_alloc_recursive(objp, cachep->object_size, 1, cachep->flags,
3484 kmemcheck_slab_alloc(cachep, flags, objp, cachep->object_size);
3486 if (unlikely((flags & __GFP_ZERO) && objp))
3487 memset(objp, 0, cachep->object_size);
3493 * Caller needs to acquire correct kmem_list's list_lock
3495 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3499 struct kmem_cache_node *l3;
3501 for (i = 0; i < nr_objects; i++) {
3505 clear_obj_pfmemalloc(&objpp[i]);
3508 slabp = virt_to_slab(objp);
3509 l3 = cachep->node[node];
3510 list_del(&slabp->list);
3511 check_spinlock_acquired_node(cachep, node);
3512 check_slabp(cachep, slabp);
3513 slab_put_obj(cachep, slabp, objp, node);
3514 STATS_DEC_ACTIVE(cachep);
3516 check_slabp(cachep, slabp);
3518 /* fixup slab chains */
3519 if (slabp->inuse == 0) {
3520 if (l3->free_objects > l3->free_limit) {
3521 l3->free_objects -= cachep->num;
3522 /* No need to drop any previously held
3523 * lock here, even if we have a off-slab slab
3524 * descriptor it is guaranteed to come from
3525 * a different cache, refer to comments before
3528 slab_destroy(cachep, slabp);
3530 list_add(&slabp->list, &l3->slabs_free);
3533 /* Unconditionally move a slab to the end of the
3534 * partial list on free - maximum time for the
3535 * other objects to be freed, too.
3537 list_add_tail(&slabp->list, &l3->slabs_partial);
3542 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3545 struct kmem_cache_node *l3;
3546 int node = numa_mem_id();
3548 batchcount = ac->batchcount;
3550 BUG_ON(!batchcount || batchcount > ac->avail);
3553 l3 = cachep->node[node];
3554 spin_lock(&l3->list_lock);
3556 struct array_cache *shared_array = l3->shared;
3557 int max = shared_array->limit - shared_array->avail;
3559 if (batchcount > max)
3561 memcpy(&(shared_array->entry[shared_array->avail]),
3562 ac->entry, sizeof(void *) * batchcount);
3563 shared_array->avail += batchcount;
3568 free_block(cachep, ac->entry, batchcount, node);
3573 struct list_head *p;
3575 p = l3->slabs_free.next;
3576 while (p != &(l3->slabs_free)) {
3579 slabp = list_entry(p, struct slab, list);
3580 BUG_ON(slabp->inuse);
3585 STATS_SET_FREEABLE(cachep, i);
3588 spin_unlock(&l3->list_lock);
3589 ac->avail -= batchcount;
3590 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3594 * Release an obj back to its cache. If the obj has a constructed state, it must
3595 * be in this state _before_ it is released. Called with disabled ints.
3597 static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3598 unsigned long caller)
3600 struct array_cache *ac = cpu_cache_get(cachep);
3603 kmemleak_free_recursive(objp, cachep->flags);
3604 objp = cache_free_debugcheck(cachep, objp, caller);
3606 kmemcheck_slab_free(cachep, objp, cachep->object_size);
3609 * Skip calling cache_free_alien() when the platform is not numa.
3610 * This will avoid cache misses that happen while accessing slabp (which
3611 * is per page memory reference) to get nodeid. Instead use a global
3612 * variable to skip the call, which is mostly likely to be present in
3615 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3618 if (likely(ac->avail < ac->limit)) {
3619 STATS_INC_FREEHIT(cachep);
3621 STATS_INC_FREEMISS(cachep);
3622 cache_flusharray(cachep, ac);
3625 ac_put_obj(cachep, ac, objp);
3629 * kmem_cache_alloc - Allocate an object
3630 * @cachep: The cache to allocate from.
3631 * @flags: See kmalloc().
3633 * Allocate an object from this cache. The flags are only relevant
3634 * if the cache has no available objects.
3636 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3638 void *ret = slab_alloc(cachep, flags, _RET_IP_);
3640 trace_kmem_cache_alloc(_RET_IP_, ret,
3641 cachep->object_size, cachep->size, flags);
3645 EXPORT_SYMBOL(kmem_cache_alloc);
3647 #ifdef CONFIG_TRACING
3649 kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
3653 ret = slab_alloc(cachep, flags, _RET_IP_);
3655 trace_kmalloc(_RET_IP_, ret,
3656 size, cachep->size, flags);
3659 EXPORT_SYMBOL(kmem_cache_alloc_trace);
3663 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3665 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3667 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3668 cachep->object_size, cachep->size,
3673 EXPORT_SYMBOL(kmem_cache_alloc_node);
3675 #ifdef CONFIG_TRACING
3676 void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
3683 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3685 trace_kmalloc_node(_RET_IP_, ret,
3690 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3693 static __always_inline void *
3694 __do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
3696 struct kmem_cache *cachep;
3698 cachep = kmalloc_slab(size, flags);
3699 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3701 return kmem_cache_alloc_node_trace(cachep, flags, node, size);
3704 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3705 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3707 return __do_kmalloc_node(size, flags, node, _RET_IP_);
3709 EXPORT_SYMBOL(__kmalloc_node);
3711 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3712 int node, unsigned long caller)
3714 return __do_kmalloc_node(size, flags, node, caller);
3716 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3718 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3720 return __do_kmalloc_node(size, flags, node, 0);
3722 EXPORT_SYMBOL(__kmalloc_node);
3723 #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3724 #endif /* CONFIG_NUMA */
3727 * __do_kmalloc - allocate memory
3728 * @size: how many bytes of memory are required.
3729 * @flags: the type of memory to allocate (see kmalloc).
3730 * @caller: function caller for debug tracking of the caller
3732 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3733 unsigned long caller)
3735 struct kmem_cache *cachep;
3738 /* If you want to save a few bytes .text space: replace
3740 * Then kmalloc uses the uninlined functions instead of the inline
3743 cachep = kmalloc_slab(size, flags);
3744 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3746 ret = slab_alloc(cachep, flags, caller);
3748 trace_kmalloc(caller, ret,
3749 size, cachep->size, flags);
3755 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3756 void *__kmalloc(size_t size, gfp_t flags)
3758 return __do_kmalloc(size, flags, _RET_IP_);
3760 EXPORT_SYMBOL(__kmalloc);
3762 void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3764 return __do_kmalloc(size, flags, caller);
3766 EXPORT_SYMBOL(__kmalloc_track_caller);
3769 void *__kmalloc(size_t size, gfp_t flags)
3771 return __do_kmalloc(size, flags, 0);
3773 EXPORT_SYMBOL(__kmalloc);
3777 * kmem_cache_free - Deallocate an object
3778 * @cachep: The cache the allocation was from.
3779 * @objp: The previously allocated object.
3781 * Free an object which was previously allocated from this
3784 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3786 unsigned long flags;
3787 cachep = cache_from_obj(cachep, objp);
3791 local_irq_save(flags);
3792 debug_check_no_locks_freed(objp, cachep->object_size);
3793 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3794 debug_check_no_obj_freed(objp, cachep->object_size);
3795 __cache_free(cachep, objp, _RET_IP_);
3796 local_irq_restore(flags);
3798 trace_kmem_cache_free(_RET_IP_, objp);
3800 EXPORT_SYMBOL(kmem_cache_free);
3803 * kfree - free previously allocated memory
3804 * @objp: pointer returned by kmalloc.
3806 * If @objp is NULL, no operation is performed.
3808 * Don't free memory not originally allocated by kmalloc()
3809 * or you will run into trouble.
3811 void kfree(const void *objp)
3813 struct kmem_cache *c;
3814 unsigned long flags;
3816 trace_kfree(_RET_IP_, objp);
3818 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3820 local_irq_save(flags);
3821 kfree_debugcheck(objp);
3822 c = virt_to_cache(objp);
3823 debug_check_no_locks_freed(objp, c->object_size);
3825 debug_check_no_obj_freed(objp, c->object_size);
3826 __cache_free(c, (void *)objp, _RET_IP_);
3827 local_irq_restore(flags);
3829 EXPORT_SYMBOL(kfree);
3832 * This initializes kmem_list3 or resizes various caches for all nodes.
3834 static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
3837 struct kmem_cache_node *l3;
3838 struct array_cache *new_shared;
3839 struct array_cache **new_alien = NULL;
3841 for_each_online_node(node) {
3843 if (use_alien_caches) {
3844 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
3850 if (cachep->shared) {
3851 new_shared = alloc_arraycache(node,
3852 cachep->shared*cachep->batchcount,
3855 free_alien_cache(new_alien);
3860 l3 = cachep->node[node];
3862 struct array_cache *shared = l3->shared;
3864 spin_lock_irq(&l3->list_lock);
3867 free_block(cachep, shared->entry,
3868 shared->avail, node);
3870 l3->shared = new_shared;
3872 l3->alien = new_alien;
3875 l3->free_limit = (1 + nr_cpus_node(node)) *
3876 cachep->batchcount + cachep->num;
3877 spin_unlock_irq(&l3->list_lock);
3879 free_alien_cache(new_alien);
3882 l3 = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
3884 free_alien_cache(new_alien);
3889 kmem_list3_init(l3);
3890 l3->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3891 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3892 l3->shared = new_shared;
3893 l3->alien = new_alien;
3894 l3->free_limit = (1 + nr_cpus_node(node)) *
3895 cachep->batchcount + cachep->num;
3896 cachep->node[node] = l3;
3901 if (!cachep->list.next) {
3902 /* Cache is not active yet. Roll back what we did */
3905 if (cachep->node[node]) {
3906 l3 = cachep->node[node];
3909 free_alien_cache(l3->alien);
3911 cachep->node[node] = NULL;
3919 struct ccupdate_struct {
3920 struct kmem_cache *cachep;
3921 struct array_cache *new[0];
3924 static void do_ccupdate_local(void *info)
3926 struct ccupdate_struct *new = info;
3927 struct array_cache *old;
3930 old = cpu_cache_get(new->cachep);
3932 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3933 new->new[smp_processor_id()] = old;
3936 /* Always called with the slab_mutex held */
3937 static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
3938 int batchcount, int shared, gfp_t gfp)
3940 struct ccupdate_struct *new;
3943 new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *),
3948 for_each_online_cpu(i) {
3949 new->new[i] = alloc_arraycache(cpu_to_mem(i), limit,
3952 for (i--; i >= 0; i--)
3958 new->cachep = cachep;
3960 on_each_cpu(do_ccupdate_local, (void *)new, 1);
3963 cachep->batchcount = batchcount;
3964 cachep->limit = limit;
3965 cachep->shared = shared;
3967 for_each_online_cpu(i) {
3968 struct array_cache *ccold = new->new[i];
3971 spin_lock_irq(&cachep->node[cpu_to_mem(i)]->list_lock);
3972 free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i));
3973 spin_unlock_irq(&cachep->node[cpu_to_mem(i)]->list_lock);
3977 return alloc_kmemlist(cachep, gfp);
3980 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3981 int batchcount, int shared, gfp_t gfp)
3984 struct kmem_cache *c = NULL;
3987 ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3989 if (slab_state < FULL)
3992 if ((ret < 0) || !is_root_cache(cachep))
3995 VM_BUG_ON(!mutex_is_locked(&slab_mutex));
3996 for_each_memcg_cache_index(i) {
3997 c = cache_from_memcg(cachep, i);
3999 /* return value determined by the parent cache only */
4000 __do_tune_cpucache(c, limit, batchcount, shared, gfp);
4006 /* Called with slab_mutex held always */
4007 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
4014 if (!is_root_cache(cachep)) {
4015 struct kmem_cache *root = memcg_root_cache(cachep);
4016 limit = root->limit;
4017 shared = root->shared;
4018 batchcount = root->batchcount;
4021 if (limit && shared && batchcount)
4024 * The head array serves three purposes:
4025 * - create a LIFO ordering, i.e. return objects that are cache-warm
4026 * - reduce the number of spinlock operations.
4027 * - reduce the number of linked list operations on the slab and
4028 * bufctl chains: array operations are cheaper.
4029 * The numbers are guessed, we should auto-tune as described by
4032 if (cachep->size > 131072)
4034 else if (cachep->size > PAGE_SIZE)
4036 else if (cachep->size > 1024)
4038 else if (cachep->size > 256)
4044 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
4045 * allocation behaviour: Most allocs on one cpu, most free operations
4046 * on another cpu. For these cases, an efficient object passing between
4047 * cpus is necessary. This is provided by a shared array. The array
4048 * replaces Bonwick's magazine layer.
4049 * On uniprocessor, it's functionally equivalent (but less efficient)
4050 * to a larger limit. Thus disabled by default.
4053 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
4058 * With debugging enabled, large batchcount lead to excessively long
4059 * periods with disabled local interrupts. Limit the batchcount
4064 batchcount = (limit + 1) / 2;
4066 err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
4068 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
4069 cachep->name, -err);
4074 * Drain an array if it contains any elements taking the l3 lock only if
4075 * necessary. Note that the l3 listlock also protects the array_cache
4076 * if drain_array() is used on the shared array.
4078 static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *l3,
4079 struct array_cache *ac, int force, int node)
4083 if (!ac || !ac->avail)
4085 if (ac->touched && !force) {
4088 spin_lock_irq(&l3->list_lock);
4090 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4091 if (tofree > ac->avail)
4092 tofree = (ac->avail + 1) / 2;
4093 free_block(cachep, ac->entry, tofree, node);
4094 ac->avail -= tofree;
4095 memmove(ac->entry, &(ac->entry[tofree]),
4096 sizeof(void *) * ac->avail);
4098 spin_unlock_irq(&l3->list_lock);
4103 * cache_reap - Reclaim memory from caches.
4104 * @w: work descriptor
4106 * Called from workqueue/eventd every few seconds.
4108 * - clear the per-cpu caches for this CPU.
4109 * - return freeable pages to the main free memory pool.
4111 * If we cannot acquire the cache chain mutex then just give up - we'll try
4112 * again on the next iteration.
4114 static void cache_reap(struct work_struct *w)
4116 struct kmem_cache *searchp;
4117 struct kmem_cache_node *l3;
4118 int node = numa_mem_id();
4119 struct delayed_work *work = to_delayed_work(w);
4121 if (!mutex_trylock(&slab_mutex))
4122 /* Give up. Setup the next iteration. */
4125 list_for_each_entry(searchp, &slab_caches, list) {
4129 * We only take the l3 lock if absolutely necessary and we
4130 * have established with reasonable certainty that
4131 * we can do some work if the lock was obtained.
4133 l3 = searchp->node[node];
4135 reap_alien(searchp, l3);
4137 drain_array(searchp, l3, cpu_cache_get(searchp), 0, node);
4140 * These are racy checks but it does not matter
4141 * if we skip one check or scan twice.
4143 if (time_after(l3->next_reap, jiffies))
4146 l3->next_reap = jiffies + REAPTIMEOUT_LIST3;
4148 drain_array(searchp, l3, l3->shared, 0, node);
4150 if (l3->free_touched)
4151 l3->free_touched = 0;
4155 freed = drain_freelist(searchp, l3, (l3->free_limit +
4156 5 * searchp->num - 1) / (5 * searchp->num));
4157 STATS_ADD_REAPED(searchp, freed);
4163 mutex_unlock(&slab_mutex);
4166 /* Set up the next iteration */
4167 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
4170 #ifdef CONFIG_SLABINFO
4171 void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
4174 unsigned long active_objs;
4175 unsigned long num_objs;
4176 unsigned long active_slabs = 0;
4177 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4181 struct kmem_cache_node *l3;
4185 for_each_online_node(node) {
4186 l3 = cachep->node[node];
4191 spin_lock_irq(&l3->list_lock);
4193 list_for_each_entry(slabp, &l3->slabs_full, list) {
4194 if (slabp->inuse != cachep->num && !error)
4195 error = "slabs_full accounting error";
4196 active_objs += cachep->num;
4199 list_for_each_entry(slabp, &l3->slabs_partial, list) {
4200 if (slabp->inuse == cachep->num && !error)
4201 error = "slabs_partial inuse accounting error";
4202 if (!slabp->inuse && !error)
4203 error = "slabs_partial/inuse accounting error";
4204 active_objs += slabp->inuse;
4207 list_for_each_entry(slabp, &l3->slabs_free, list) {
4208 if (slabp->inuse && !error)
4209 error = "slabs_free/inuse accounting error";
4212 free_objects += l3->free_objects;
4214 shared_avail += l3->shared->avail;
4216 spin_unlock_irq(&l3->list_lock);
4218 num_slabs += active_slabs;
4219 num_objs = num_slabs * cachep->num;
4220 if (num_objs - active_objs != free_objects && !error)
4221 error = "free_objects accounting error";
4223 name = cachep->name;
4225 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4227 sinfo->active_objs = active_objs;
4228 sinfo->num_objs = num_objs;
4229 sinfo->active_slabs = active_slabs;
4230 sinfo->num_slabs = num_slabs;
4231 sinfo->shared_avail = shared_avail;
4232 sinfo->limit = cachep->limit;
4233 sinfo->batchcount = cachep->batchcount;
4234 sinfo->shared = cachep->shared;
4235 sinfo->objects_per_slab = cachep->num;
4236 sinfo->cache_order = cachep->gfporder;
4239 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
4243 unsigned long high = cachep->high_mark;
4244 unsigned long allocs = cachep->num_allocations;
4245 unsigned long grown = cachep->grown;
4246 unsigned long reaped = cachep->reaped;
4247 unsigned long errors = cachep->errors;
4248 unsigned long max_freeable = cachep->max_freeable;
4249 unsigned long node_allocs = cachep->node_allocs;
4250 unsigned long node_frees = cachep->node_frees;
4251 unsigned long overflows = cachep->node_overflow;
4253 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4254 "%4lu %4lu %4lu %4lu %4lu",
4255 allocs, high, grown,
4256 reaped, errors, max_freeable, node_allocs,
4257 node_frees, overflows);
4261 unsigned long allochit = atomic_read(&cachep->allochit);
4262 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4263 unsigned long freehit = atomic_read(&cachep->freehit);
4264 unsigned long freemiss = atomic_read(&cachep->freemiss);
4266 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4267 allochit, allocmiss, freehit, freemiss);
4272 #define MAX_SLABINFO_WRITE 128
4274 * slabinfo_write - Tuning for the slab allocator
4276 * @buffer: user buffer
4277 * @count: data length
4280 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4281 size_t count, loff_t *ppos)
4283 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4284 int limit, batchcount, shared, res;
4285 struct kmem_cache *cachep;
4287 if (count > MAX_SLABINFO_WRITE)
4289 if (copy_from_user(&kbuf, buffer, count))
4291 kbuf[MAX_SLABINFO_WRITE] = '\0';
4293 tmp = strchr(kbuf, ' ');
4298 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4301 /* Find the cache in the chain of caches. */
4302 mutex_lock(&slab_mutex);
4304 list_for_each_entry(cachep, &slab_caches, list) {
4305 if (!strcmp(cachep->name, kbuf)) {
4306 if (limit < 1 || batchcount < 1 ||
4307 batchcount > limit || shared < 0) {
4310 res = do_tune_cpucache(cachep, limit,
4317 mutex_unlock(&slab_mutex);
4323 #ifdef CONFIG_DEBUG_SLAB_LEAK
4325 static void *leaks_start(struct seq_file *m, loff_t *pos)
4327 mutex_lock(&slab_mutex);
4328 return seq_list_start(&slab_caches, *pos);
4331 static inline int add_caller(unsigned long *n, unsigned long v)
4341 unsigned long *q = p + 2 * i;
4355 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4361 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4367 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->size) {
4368 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4370 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4375 static void show_symbol(struct seq_file *m, unsigned long address)
4377 #ifdef CONFIG_KALLSYMS
4378 unsigned long offset, size;
4379 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4381 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4382 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4384 seq_printf(m, " [%s]", modname);
4388 seq_printf(m, "%p", (void *)address);
4391 static int leaks_show(struct seq_file *m, void *p)
4393 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
4395 struct kmem_cache_node *l3;
4397 unsigned long *n = m->private;
4401 if (!(cachep->flags & SLAB_STORE_USER))
4403 if (!(cachep->flags & SLAB_RED_ZONE))
4406 /* OK, we can do it */
4410 for_each_online_node(node) {
4411 l3 = cachep->node[node];
4416 spin_lock_irq(&l3->list_lock);
4418 list_for_each_entry(slabp, &l3->slabs_full, list)
4419 handle_slab(n, cachep, slabp);
4420 list_for_each_entry(slabp, &l3->slabs_partial, list)
4421 handle_slab(n, cachep, slabp);
4422 spin_unlock_irq(&l3->list_lock);
4424 name = cachep->name;
4426 /* Increase the buffer size */
4427 mutex_unlock(&slab_mutex);
4428 m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4430 /* Too bad, we are really out */
4432 mutex_lock(&slab_mutex);
4435 *(unsigned long *)m->private = n[0] * 2;
4437 mutex_lock(&slab_mutex);
4438 /* Now make sure this entry will be retried */
4442 for (i = 0; i < n[1]; i++) {
4443 seq_printf(m, "%s: %lu ", name, n[2*i+3]);
4444 show_symbol(m, n[2*i+2]);
4451 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4453 return seq_list_next(p, &slab_caches, pos);
4456 static void s_stop(struct seq_file *m, void *p)
4458 mutex_unlock(&slab_mutex);
4461 static const struct seq_operations slabstats_op = {
4462 .start = leaks_start,
4468 static int slabstats_open(struct inode *inode, struct file *file)
4470 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4473 ret = seq_open(file, &slabstats_op);
4475 struct seq_file *m = file->private_data;
4476 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4485 static const struct file_operations proc_slabstats_operations = {
4486 .open = slabstats_open,
4488 .llseek = seq_lseek,
4489 .release = seq_release_private,
4493 static int __init slab_proc_init(void)
4495 #ifdef CONFIG_DEBUG_SLAB_LEAK
4496 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4500 module_init(slab_proc_init);
4504 * ksize - get the actual amount of memory allocated for a given object
4505 * @objp: Pointer to the object
4507 * kmalloc may internally round up allocations and return more memory
4508 * than requested. ksize() can be used to determine the actual amount of
4509 * memory allocated. The caller may use this additional memory, even though
4510 * a smaller amount of memory was initially specified with the kmalloc call.
4511 * The caller must guarantee that objp points to a valid object previously
4512 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4513 * must not be freed during the duration of the call.
4515 size_t ksize(const void *objp)
4518 if (unlikely(objp == ZERO_SIZE_PTR))
4521 return virt_to_cache(objp)->object_size;
4523 EXPORT_SYMBOL(ksize);