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 * Need this for bootstrapping a per node allocator.
291 #define NUM_INIT_LISTS (3 * MAX_NUMNODES)
292 static struct kmem_cache_node __initdata init_kmem_cache_node[NUM_INIT_LISTS];
293 #define CACHE_CACHE 0
294 #define SIZE_AC MAX_NUMNODES
295 #define SIZE_NODE (2 * MAX_NUMNODES)
297 static int drain_freelist(struct kmem_cache *cache,
298 struct kmem_cache_node *n, int tofree);
299 static void free_block(struct kmem_cache *cachep, void **objpp, int len,
301 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
302 static void cache_reap(struct work_struct *unused);
304 static int slab_early_init = 1;
306 #define INDEX_AC kmalloc_index(sizeof(struct arraycache_init))
307 #define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
309 static void kmem_cache_node_init(struct kmem_cache_node *parent)
311 INIT_LIST_HEAD(&parent->slabs_full);
312 INIT_LIST_HEAD(&parent->slabs_partial);
313 INIT_LIST_HEAD(&parent->slabs_free);
314 parent->shared = NULL;
315 parent->alien = NULL;
316 parent->colour_next = 0;
317 spin_lock_init(&parent->list_lock);
318 parent->free_objects = 0;
319 parent->free_touched = 0;
322 #define MAKE_LIST(cachep, listp, slab, nodeid) \
324 INIT_LIST_HEAD(listp); \
325 list_splice(&(cachep->node[nodeid]->slab), listp); \
328 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
330 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
331 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
332 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
335 #define CFLGS_OFF_SLAB (0x80000000UL)
336 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
338 #define BATCHREFILL_LIMIT 16
340 * Optimization question: fewer reaps means less probability for unnessary
341 * cpucache drain/refill cycles.
343 * OTOH the cpuarrays can contain lots of objects,
344 * which could lock up otherwise freeable slabs.
346 #define REAPTIMEOUT_CPUC (2*HZ)
347 #define REAPTIMEOUT_LIST3 (4*HZ)
350 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
351 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
352 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
353 #define STATS_INC_GROWN(x) ((x)->grown++)
354 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
355 #define STATS_SET_HIGH(x) \
357 if ((x)->num_active > (x)->high_mark) \
358 (x)->high_mark = (x)->num_active; \
360 #define STATS_INC_ERR(x) ((x)->errors++)
361 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
362 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
363 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
364 #define STATS_SET_FREEABLE(x, i) \
366 if ((x)->max_freeable < i) \
367 (x)->max_freeable = i; \
369 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
370 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
371 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
372 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
374 #define STATS_INC_ACTIVE(x) do { } while (0)
375 #define STATS_DEC_ACTIVE(x) do { } while (0)
376 #define STATS_INC_ALLOCED(x) do { } while (0)
377 #define STATS_INC_GROWN(x) do { } while (0)
378 #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
379 #define STATS_SET_HIGH(x) do { } while (0)
380 #define STATS_INC_ERR(x) do { } while (0)
381 #define STATS_INC_NODEALLOCS(x) do { } while (0)
382 #define STATS_INC_NODEFREES(x) do { } while (0)
383 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
384 #define STATS_SET_FREEABLE(x, i) do { } while (0)
385 #define STATS_INC_ALLOCHIT(x) do { } while (0)
386 #define STATS_INC_ALLOCMISS(x) do { } while (0)
387 #define STATS_INC_FREEHIT(x) do { } while (0)
388 #define STATS_INC_FREEMISS(x) do { } while (0)
394 * memory layout of objects:
396 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
397 * the end of an object is aligned with the end of the real
398 * allocation. Catches writes behind the end of the allocation.
399 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
401 * cachep->obj_offset: The real object.
402 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
403 * cachep->size - 1* BYTES_PER_WORD: last caller address
404 * [BYTES_PER_WORD long]
406 static int obj_offset(struct kmem_cache *cachep)
408 return cachep->obj_offset;
411 static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
413 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
414 return (unsigned long long*) (objp + obj_offset(cachep) -
415 sizeof(unsigned long long));
418 static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
420 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
421 if (cachep->flags & SLAB_STORE_USER)
422 return (unsigned long long *)(objp + cachep->size -
423 sizeof(unsigned long long) -
425 return (unsigned long long *) (objp + cachep->size -
426 sizeof(unsigned long long));
429 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
431 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
432 return (void **)(objp + cachep->size - BYTES_PER_WORD);
437 #define obj_offset(x) 0
438 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
439 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
440 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
445 * Do not go above this order unless 0 objects fit into the slab or
446 * overridden on the command line.
448 #define SLAB_MAX_ORDER_HI 1
449 #define SLAB_MAX_ORDER_LO 0
450 static int slab_max_order = SLAB_MAX_ORDER_LO;
451 static bool slab_max_order_set __initdata;
453 static inline struct kmem_cache *virt_to_cache(const void *obj)
455 struct page *page = virt_to_head_page(obj);
456 return page->slab_cache;
459 static inline struct slab *virt_to_slab(const void *obj)
461 struct page *page = virt_to_head_page(obj);
463 VM_BUG_ON(!PageSlab(page));
464 return page->slab_page;
467 static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
470 return slab->s_mem + cache->size * idx;
474 * We want to avoid an expensive divide : (offset / cache->size)
475 * Using the fact that size is a constant for a particular cache,
476 * we can replace (offset / cache->size) by
477 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
479 static inline unsigned int obj_to_index(const struct kmem_cache *cache,
480 const struct slab *slab, void *obj)
482 u32 offset = (obj - slab->s_mem);
483 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
486 static struct arraycache_init initarray_generic =
487 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
489 /* internal cache of cache description objs */
490 static struct kmem_cache kmem_cache_boot = {
492 .limit = BOOT_CPUCACHE_ENTRIES,
494 .size = sizeof(struct kmem_cache),
495 .name = "kmem_cache",
498 #define BAD_ALIEN_MAGIC 0x01020304ul
500 #ifdef CONFIG_LOCKDEP
503 * Slab sometimes uses the kmalloc slabs to store the slab headers
504 * for other slabs "off slab".
505 * The locking for this is tricky in that it nests within the locks
506 * of all other slabs in a few places; to deal with this special
507 * locking we put on-slab caches into a separate lock-class.
509 * We set lock class for alien array caches which are up during init.
510 * The lock annotation will be lost if all cpus of a node goes down and
511 * then comes back up during hotplug
513 static struct lock_class_key on_slab_l3_key;
514 static struct lock_class_key on_slab_alc_key;
516 static struct lock_class_key debugobj_l3_key;
517 static struct lock_class_key debugobj_alc_key;
519 static void slab_set_lock_classes(struct kmem_cache *cachep,
520 struct lock_class_key *l3_key, struct lock_class_key *alc_key,
523 struct array_cache **alc;
524 struct kmem_cache_node *n;
531 lockdep_set_class(&n->list_lock, l3_key);
534 * FIXME: This check for BAD_ALIEN_MAGIC
535 * should go away when common slab code is taught to
536 * work even without alien caches.
537 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
538 * for alloc_alien_cache,
540 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
544 lockdep_set_class(&alc[r]->lock, alc_key);
548 static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
550 slab_set_lock_classes(cachep, &debugobj_l3_key, &debugobj_alc_key, node);
553 static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
557 for_each_online_node(node)
558 slab_set_debugobj_lock_classes_node(cachep, node);
561 static void init_node_lock_keys(int q)
568 for (i = 1; i < PAGE_SHIFT + MAX_ORDER; i++) {
569 struct kmem_cache_node *n;
570 struct kmem_cache *cache = kmalloc_caches[i];
576 if (!n || OFF_SLAB(cache))
579 slab_set_lock_classes(cache, &on_slab_l3_key,
580 &on_slab_alc_key, q);
584 static void on_slab_lock_classes_node(struct kmem_cache *cachep, int q)
586 if (!cachep->node[q])
589 slab_set_lock_classes(cachep, &on_slab_l3_key,
590 &on_slab_alc_key, q);
593 static inline void on_slab_lock_classes(struct kmem_cache *cachep)
597 VM_BUG_ON(OFF_SLAB(cachep));
599 on_slab_lock_classes_node(cachep, node);
602 static inline void init_lock_keys(void)
607 init_node_lock_keys(node);
610 static void init_node_lock_keys(int q)
614 static inline void init_lock_keys(void)
618 static inline void on_slab_lock_classes(struct kmem_cache *cachep)
622 static inline void on_slab_lock_classes_node(struct kmem_cache *cachep, int node)
626 static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
630 static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
635 static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
637 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
639 return cachep->array[smp_processor_id()];
642 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
644 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
648 * Calculate the number of objects and left-over bytes for a given buffer size.
650 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
651 size_t align, int flags, size_t *left_over,
656 size_t slab_size = PAGE_SIZE << gfporder;
659 * The slab management structure can be either off the slab or
660 * on it. For the latter case, the memory allocated for a
664 * - One kmem_bufctl_t for each object
665 * - Padding to respect alignment of @align
666 * - @buffer_size bytes for each object
668 * If the slab management structure is off the slab, then the
669 * alignment will already be calculated into the size. Because
670 * the slabs are all pages aligned, the objects will be at the
671 * correct alignment when allocated.
673 if (flags & CFLGS_OFF_SLAB) {
675 nr_objs = slab_size / buffer_size;
677 if (nr_objs > SLAB_LIMIT)
678 nr_objs = SLAB_LIMIT;
681 * Ignore padding for the initial guess. The padding
682 * is at most @align-1 bytes, and @buffer_size is at
683 * least @align. In the worst case, this result will
684 * be one greater than the number of objects that fit
685 * into the memory allocation when taking the padding
688 nr_objs = (slab_size - sizeof(struct slab)) /
689 (buffer_size + sizeof(kmem_bufctl_t));
692 * This calculated number will be either the right
693 * amount, or one greater than what we want.
695 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
699 if (nr_objs > SLAB_LIMIT)
700 nr_objs = SLAB_LIMIT;
702 mgmt_size = slab_mgmt_size(nr_objs, align);
705 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
709 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
711 static void __slab_error(const char *function, struct kmem_cache *cachep,
714 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
715 function, cachep->name, msg);
717 add_taint(TAINT_BAD_PAGE);
722 * By default on NUMA we use alien caches to stage the freeing of
723 * objects allocated from other nodes. This causes massive memory
724 * inefficiencies when using fake NUMA setup to split memory into a
725 * large number of small nodes, so it can be disabled on the command
729 static int use_alien_caches __read_mostly = 1;
730 static int __init noaliencache_setup(char *s)
732 use_alien_caches = 0;
735 __setup("noaliencache", noaliencache_setup);
737 static int __init slab_max_order_setup(char *str)
739 get_option(&str, &slab_max_order);
740 slab_max_order = slab_max_order < 0 ? 0 :
741 min(slab_max_order, MAX_ORDER - 1);
742 slab_max_order_set = true;
746 __setup("slab_max_order=", slab_max_order_setup);
750 * Special reaping functions for NUMA systems called from cache_reap().
751 * These take care of doing round robin flushing of alien caches (containing
752 * objects freed on different nodes from which they were allocated) and the
753 * flushing of remote pcps by calling drain_node_pages.
755 static DEFINE_PER_CPU(unsigned long, slab_reap_node);
757 static void init_reap_node(int cpu)
761 node = next_node(cpu_to_mem(cpu), node_online_map);
762 if (node == MAX_NUMNODES)
763 node = first_node(node_online_map);
765 per_cpu(slab_reap_node, cpu) = node;
768 static void next_reap_node(void)
770 int node = __this_cpu_read(slab_reap_node);
772 node = next_node(node, node_online_map);
773 if (unlikely(node >= MAX_NUMNODES))
774 node = first_node(node_online_map);
775 __this_cpu_write(slab_reap_node, node);
779 #define init_reap_node(cpu) do { } while (0)
780 #define next_reap_node(void) do { } while (0)
784 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
785 * via the workqueue/eventd.
786 * Add the CPU number into the expiration time to minimize the possibility of
787 * the CPUs getting into lockstep and contending for the global cache chain
790 static void __cpuinit start_cpu_timer(int cpu)
792 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
795 * When this gets called from do_initcalls via cpucache_init(),
796 * init_workqueues() has already run, so keventd will be setup
799 if (keventd_up() && reap_work->work.func == NULL) {
801 INIT_DEFERRABLE_WORK(reap_work, cache_reap);
802 schedule_delayed_work_on(cpu, reap_work,
803 __round_jiffies_relative(HZ, cpu));
807 static struct array_cache *alloc_arraycache(int node, int entries,
808 int batchcount, gfp_t gfp)
810 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
811 struct array_cache *nc = NULL;
813 nc = kmalloc_node(memsize, gfp, node);
815 * The array_cache structures contain pointers to free object.
816 * However, when such objects are allocated or transferred to another
817 * cache the pointers are not cleared and they could be counted as
818 * valid references during a kmemleak scan. Therefore, kmemleak must
819 * not scan such objects.
821 kmemleak_no_scan(nc);
825 nc->batchcount = batchcount;
827 spin_lock_init(&nc->lock);
832 static inline bool is_slab_pfmemalloc(struct slab *slabp)
834 struct page *page = virt_to_page(slabp->s_mem);
836 return PageSlabPfmemalloc(page);
839 /* Clears pfmemalloc_active if no slabs have pfmalloc set */
840 static void recheck_pfmemalloc_active(struct kmem_cache *cachep,
841 struct array_cache *ac)
843 struct kmem_cache_node *n = cachep->node[numa_mem_id()];
847 if (!pfmemalloc_active)
850 spin_lock_irqsave(&n->list_lock, flags);
851 list_for_each_entry(slabp, &n->slabs_full, list)
852 if (is_slab_pfmemalloc(slabp))
855 list_for_each_entry(slabp, &n->slabs_partial, list)
856 if (is_slab_pfmemalloc(slabp))
859 list_for_each_entry(slabp, &n->slabs_free, list)
860 if (is_slab_pfmemalloc(slabp))
863 pfmemalloc_active = false;
865 spin_unlock_irqrestore(&n->list_lock, flags);
868 static void *__ac_get_obj(struct kmem_cache *cachep, struct array_cache *ac,
869 gfp_t flags, bool force_refill)
872 void *objp = ac->entry[--ac->avail];
874 /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
875 if (unlikely(is_obj_pfmemalloc(objp))) {
876 struct kmem_cache_node *n;
878 if (gfp_pfmemalloc_allowed(flags)) {
879 clear_obj_pfmemalloc(&objp);
883 /* The caller cannot use PFMEMALLOC objects, find another one */
884 for (i = 0; i < ac->avail; i++) {
885 /* If a !PFMEMALLOC object is found, swap them */
886 if (!is_obj_pfmemalloc(ac->entry[i])) {
888 ac->entry[i] = ac->entry[ac->avail];
889 ac->entry[ac->avail] = objp;
895 * If there are empty slabs on the slabs_free list and we are
896 * being forced to refill the cache, mark this one !pfmemalloc.
898 n = cachep->node[numa_mem_id()];
899 if (!list_empty(&n->slabs_free) && force_refill) {
900 struct slab *slabp = virt_to_slab(objp);
901 ClearPageSlabPfmemalloc(virt_to_head_page(slabp->s_mem));
902 clear_obj_pfmemalloc(&objp);
903 recheck_pfmemalloc_active(cachep, ac);
907 /* No !PFMEMALLOC objects available */
915 static inline void *ac_get_obj(struct kmem_cache *cachep,
916 struct array_cache *ac, gfp_t flags, bool force_refill)
920 if (unlikely(sk_memalloc_socks()))
921 objp = __ac_get_obj(cachep, ac, flags, force_refill);
923 objp = ac->entry[--ac->avail];
928 static void *__ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
931 if (unlikely(pfmemalloc_active)) {
932 /* Some pfmemalloc slabs exist, check if this is one */
933 struct page *page = virt_to_head_page(objp);
934 if (PageSlabPfmemalloc(page))
935 set_obj_pfmemalloc(&objp);
941 static inline void ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
944 if (unlikely(sk_memalloc_socks()))
945 objp = __ac_put_obj(cachep, ac, objp);
947 ac->entry[ac->avail++] = objp;
951 * Transfer objects in one arraycache to another.
952 * Locking must be handled by the caller.
954 * Return the number of entries transferred.
956 static int transfer_objects(struct array_cache *to,
957 struct array_cache *from, unsigned int max)
959 /* Figure out how many entries to transfer */
960 int nr = min3(from->avail, max, to->limit - to->avail);
965 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
975 #define drain_alien_cache(cachep, alien) do { } while (0)
976 #define reap_alien(cachep, n) do { } while (0)
978 static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
980 return (struct array_cache **)BAD_ALIEN_MAGIC;
983 static inline void free_alien_cache(struct array_cache **ac_ptr)
987 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
992 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
998 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
999 gfp_t flags, int nodeid)
1004 #else /* CONFIG_NUMA */
1006 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
1007 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
1009 static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
1011 struct array_cache **ac_ptr;
1012 int memsize = sizeof(void *) * nr_node_ids;
1017 ac_ptr = kzalloc_node(memsize, gfp, node);
1020 if (i == node || !node_online(i))
1022 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
1024 for (i--; i >= 0; i--)
1034 static void free_alien_cache(struct array_cache **ac_ptr)
1045 static void __drain_alien_cache(struct kmem_cache *cachep,
1046 struct array_cache *ac, int node)
1048 struct kmem_cache_node *n = cachep->node[node];
1051 spin_lock(&n->list_lock);
1053 * Stuff objects into the remote nodes shared array first.
1054 * That way we could avoid the overhead of putting the objects
1055 * into the free lists and getting them back later.
1058 transfer_objects(n->shared, ac, ac->limit);
1060 free_block(cachep, ac->entry, ac->avail, node);
1062 spin_unlock(&n->list_lock);
1067 * Called from cache_reap() to regularly drain alien caches round robin.
1069 static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *n)
1071 int node = __this_cpu_read(slab_reap_node);
1074 struct array_cache *ac = n->alien[node];
1076 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1077 __drain_alien_cache(cachep, ac, node);
1078 spin_unlock_irq(&ac->lock);
1083 static void drain_alien_cache(struct kmem_cache *cachep,
1084 struct array_cache **alien)
1087 struct array_cache *ac;
1088 unsigned long flags;
1090 for_each_online_node(i) {
1093 spin_lock_irqsave(&ac->lock, flags);
1094 __drain_alien_cache(cachep, ac, i);
1095 spin_unlock_irqrestore(&ac->lock, flags);
1100 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1102 struct slab *slabp = virt_to_slab(objp);
1103 int nodeid = slabp->nodeid;
1104 struct kmem_cache_node *n;
1105 struct array_cache *alien = NULL;
1108 node = numa_mem_id();
1111 * Make sure we are not freeing a object from another node to the array
1112 * cache on this cpu.
1114 if (likely(slabp->nodeid == node))
1117 n = cachep->node[node];
1118 STATS_INC_NODEFREES(cachep);
1119 if (n->alien && n->alien[nodeid]) {
1120 alien = n->alien[nodeid];
1121 spin_lock(&alien->lock);
1122 if (unlikely(alien->avail == alien->limit)) {
1123 STATS_INC_ACOVERFLOW(cachep);
1124 __drain_alien_cache(cachep, alien, nodeid);
1126 ac_put_obj(cachep, alien, objp);
1127 spin_unlock(&alien->lock);
1129 spin_lock(&(cachep->node[nodeid])->list_lock);
1130 free_block(cachep, &objp, 1, nodeid);
1131 spin_unlock(&(cachep->node[nodeid])->list_lock);
1138 * Allocates and initializes node for a node on each slab cache, used for
1139 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
1140 * will be allocated off-node since memory is not yet online for the new node.
1141 * When hotplugging memory or a cpu, existing node are not replaced if
1144 * Must hold slab_mutex.
1146 static int init_cache_node_node(int node)
1148 struct kmem_cache *cachep;
1149 struct kmem_cache_node *n;
1150 const int memsize = sizeof(struct kmem_cache_node);
1152 list_for_each_entry(cachep, &slab_caches, list) {
1154 * Set up the size64 kmemlist for cpu before we can
1155 * begin anything. Make sure some other cpu on this
1156 * node has not already allocated this
1158 if (!cachep->node[node]) {
1159 n = kmalloc_node(memsize, GFP_KERNEL, node);
1162 kmem_cache_node_init(n);
1163 n->next_reap = jiffies + REAPTIMEOUT_LIST3 +
1164 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1167 * The l3s don't come and go as CPUs come and
1168 * go. slab_mutex is sufficient
1171 cachep->node[node] = n;
1174 spin_lock_irq(&cachep->node[node]->list_lock);
1175 cachep->node[node]->free_limit =
1176 (1 + nr_cpus_node(node)) *
1177 cachep->batchcount + cachep->num;
1178 spin_unlock_irq(&cachep->node[node]->list_lock);
1183 static void __cpuinit cpuup_canceled(long cpu)
1185 struct kmem_cache *cachep;
1186 struct kmem_cache_node *n = NULL;
1187 int node = cpu_to_mem(cpu);
1188 const struct cpumask *mask = cpumask_of_node(node);
1190 list_for_each_entry(cachep, &slab_caches, list) {
1191 struct array_cache *nc;
1192 struct array_cache *shared;
1193 struct array_cache **alien;
1195 /* cpu is dead; no one can alloc from it. */
1196 nc = cachep->array[cpu];
1197 cachep->array[cpu] = NULL;
1198 n = cachep->node[node];
1201 goto free_array_cache;
1203 spin_lock_irq(&n->list_lock);
1205 /* Free limit for this kmem_cache_node */
1206 n->free_limit -= cachep->batchcount;
1208 free_block(cachep, nc->entry, nc->avail, node);
1210 if (!cpumask_empty(mask)) {
1211 spin_unlock_irq(&n->list_lock);
1212 goto free_array_cache;
1217 free_block(cachep, shared->entry,
1218 shared->avail, node);
1225 spin_unlock_irq(&n->list_lock);
1229 drain_alien_cache(cachep, alien);
1230 free_alien_cache(alien);
1236 * In the previous loop, all the objects were freed to
1237 * the respective cache's slabs, now we can go ahead and
1238 * shrink each nodelist to its limit.
1240 list_for_each_entry(cachep, &slab_caches, list) {
1241 n = cachep->node[node];
1244 drain_freelist(cachep, n, n->free_objects);
1248 static int __cpuinit cpuup_prepare(long cpu)
1250 struct kmem_cache *cachep;
1251 struct kmem_cache_node *n = NULL;
1252 int node = cpu_to_mem(cpu);
1256 * We need to do this right in the beginning since
1257 * alloc_arraycache's are going to use this list.
1258 * kmalloc_node allows us to add the slab to the right
1259 * kmem_cache_node and not this cpu's kmem_cache_node
1261 err = init_cache_node_node(node);
1266 * Now we can go ahead with allocating the shared arrays and
1269 list_for_each_entry(cachep, &slab_caches, list) {
1270 struct array_cache *nc;
1271 struct array_cache *shared = NULL;
1272 struct array_cache **alien = NULL;
1274 nc = alloc_arraycache(node, cachep->limit,
1275 cachep->batchcount, GFP_KERNEL);
1278 if (cachep->shared) {
1279 shared = alloc_arraycache(node,
1280 cachep->shared * cachep->batchcount,
1281 0xbaadf00d, GFP_KERNEL);
1287 if (use_alien_caches) {
1288 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
1295 cachep->array[cpu] = nc;
1296 n = cachep->node[node];
1299 spin_lock_irq(&n->list_lock);
1302 * We are serialised from CPU_DEAD or
1303 * CPU_UP_CANCELLED by the cpucontrol lock
1314 spin_unlock_irq(&n->list_lock);
1316 free_alien_cache(alien);
1317 if (cachep->flags & SLAB_DEBUG_OBJECTS)
1318 slab_set_debugobj_lock_classes_node(cachep, node);
1319 else if (!OFF_SLAB(cachep) &&
1320 !(cachep->flags & SLAB_DESTROY_BY_RCU))
1321 on_slab_lock_classes_node(cachep, node);
1323 init_node_lock_keys(node);
1327 cpuup_canceled(cpu);
1331 static int __cpuinit cpuup_callback(struct notifier_block *nfb,
1332 unsigned long action, void *hcpu)
1334 long cpu = (long)hcpu;
1338 case CPU_UP_PREPARE:
1339 case CPU_UP_PREPARE_FROZEN:
1340 mutex_lock(&slab_mutex);
1341 err = cpuup_prepare(cpu);
1342 mutex_unlock(&slab_mutex);
1345 case CPU_ONLINE_FROZEN:
1346 start_cpu_timer(cpu);
1348 #ifdef CONFIG_HOTPLUG_CPU
1349 case CPU_DOWN_PREPARE:
1350 case CPU_DOWN_PREPARE_FROZEN:
1352 * Shutdown cache reaper. Note that the slab_mutex is
1353 * held so that if cache_reap() is invoked it cannot do
1354 * anything expensive but will only modify reap_work
1355 * and reschedule the timer.
1357 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1358 /* Now the cache_reaper is guaranteed to be not running. */
1359 per_cpu(slab_reap_work, cpu).work.func = NULL;
1361 case CPU_DOWN_FAILED:
1362 case CPU_DOWN_FAILED_FROZEN:
1363 start_cpu_timer(cpu);
1366 case CPU_DEAD_FROZEN:
1368 * Even if all the cpus of a node are down, we don't free the
1369 * kmem_cache_node of any cache. This to avoid a race between
1370 * cpu_down, and a kmalloc allocation from another cpu for
1371 * memory from the node of the cpu going down. The node
1372 * structure is usually allocated from kmem_cache_create() and
1373 * gets destroyed at kmem_cache_destroy().
1377 case CPU_UP_CANCELED:
1378 case CPU_UP_CANCELED_FROZEN:
1379 mutex_lock(&slab_mutex);
1380 cpuup_canceled(cpu);
1381 mutex_unlock(&slab_mutex);
1384 return notifier_from_errno(err);
1387 static struct notifier_block __cpuinitdata cpucache_notifier = {
1388 &cpuup_callback, NULL, 0
1391 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1393 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1394 * Returns -EBUSY if all objects cannot be drained so that the node is not
1397 * Must hold slab_mutex.
1399 static int __meminit drain_cache_node_node(int node)
1401 struct kmem_cache *cachep;
1404 list_for_each_entry(cachep, &slab_caches, list) {
1405 struct kmem_cache_node *n;
1407 n = cachep->node[node];
1411 drain_freelist(cachep, n, n->free_objects);
1413 if (!list_empty(&n->slabs_full) ||
1414 !list_empty(&n->slabs_partial)) {
1422 static int __meminit slab_memory_callback(struct notifier_block *self,
1423 unsigned long action, void *arg)
1425 struct memory_notify *mnb = arg;
1429 nid = mnb->status_change_nid;
1434 case MEM_GOING_ONLINE:
1435 mutex_lock(&slab_mutex);
1436 ret = init_cache_node_node(nid);
1437 mutex_unlock(&slab_mutex);
1439 case MEM_GOING_OFFLINE:
1440 mutex_lock(&slab_mutex);
1441 ret = drain_cache_node_node(nid);
1442 mutex_unlock(&slab_mutex);
1446 case MEM_CANCEL_ONLINE:
1447 case MEM_CANCEL_OFFLINE:
1451 return notifier_from_errno(ret);
1453 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1456 * swap the static kmem_cache_node with kmalloced memory
1458 static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list,
1461 struct kmem_cache_node *ptr;
1463 ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid);
1466 memcpy(ptr, list, sizeof(struct kmem_cache_node));
1468 * Do not assume that spinlocks can be initialized via memcpy:
1470 spin_lock_init(&ptr->list_lock);
1472 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1473 cachep->node[nodeid] = ptr;
1477 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1478 * size of kmem_cache_node.
1480 static void __init set_up_node(struct kmem_cache *cachep, int index)
1484 for_each_online_node(node) {
1485 cachep->node[node] = &init_kmem_cache_node[index + node];
1486 cachep->node[node]->next_reap = jiffies +
1488 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1493 * The memory after the last cpu cache pointer is used for the
1496 static void setup_node_pointer(struct kmem_cache *cachep)
1498 cachep->node = (struct kmem_cache_node **)&cachep->array[nr_cpu_ids];
1502 * Initialisation. Called after the page allocator have been initialised and
1503 * before smp_init().
1505 void __init kmem_cache_init(void)
1509 kmem_cache = &kmem_cache_boot;
1510 setup_node_pointer(kmem_cache);
1512 if (num_possible_nodes() == 1)
1513 use_alien_caches = 0;
1515 for (i = 0; i < NUM_INIT_LISTS; i++)
1516 kmem_cache_node_init(&init_kmem_cache_node[i]);
1518 set_up_node(kmem_cache, CACHE_CACHE);
1521 * Fragmentation resistance on low memory - only use bigger
1522 * page orders on machines with more than 32MB of memory if
1523 * not overridden on the command line.
1525 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
1526 slab_max_order = SLAB_MAX_ORDER_HI;
1528 /* Bootstrap is tricky, because several objects are allocated
1529 * from caches that do not exist yet:
1530 * 1) initialize the kmem_cache cache: it contains the struct
1531 * kmem_cache structures of all caches, except kmem_cache itself:
1532 * kmem_cache is statically allocated.
1533 * Initially an __init data area is used for the head array and the
1534 * kmem_cache_node structures, it's replaced with a kmalloc allocated
1535 * array at the end of the bootstrap.
1536 * 2) Create the first kmalloc cache.
1537 * The struct kmem_cache for the new cache is allocated normally.
1538 * An __init data area is used for the head array.
1539 * 3) Create the remaining kmalloc caches, with minimally sized
1541 * 4) Replace the __init data head arrays for kmem_cache and the first
1542 * kmalloc cache with kmalloc allocated arrays.
1543 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1544 * the other cache's with kmalloc allocated memory.
1545 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1548 /* 1) create the kmem_cache */
1551 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1553 create_boot_cache(kmem_cache, "kmem_cache",
1554 offsetof(struct kmem_cache, array[nr_cpu_ids]) +
1555 nr_node_ids * sizeof(struct kmem_cache_node *),
1556 SLAB_HWCACHE_ALIGN);
1557 list_add(&kmem_cache->list, &slab_caches);
1559 /* 2+3) create the kmalloc caches */
1562 * Initialize the caches that provide memory for the array cache and the
1563 * kmem_cache_node structures first. Without this, further allocations will
1567 kmalloc_caches[INDEX_AC] = create_kmalloc_cache("kmalloc-ac",
1568 kmalloc_size(INDEX_AC), ARCH_KMALLOC_FLAGS);
1570 if (INDEX_AC != INDEX_NODE)
1571 kmalloc_caches[INDEX_NODE] =
1572 create_kmalloc_cache("kmalloc-node",
1573 kmalloc_size(INDEX_NODE), ARCH_KMALLOC_FLAGS);
1575 slab_early_init = 0;
1577 /* 4) Replace the bootstrap head arrays */
1579 struct array_cache *ptr;
1581 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1583 memcpy(ptr, cpu_cache_get(kmem_cache),
1584 sizeof(struct arraycache_init));
1586 * Do not assume that spinlocks can be initialized via memcpy:
1588 spin_lock_init(&ptr->lock);
1590 kmem_cache->array[smp_processor_id()] = ptr;
1592 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1594 BUG_ON(cpu_cache_get(kmalloc_caches[INDEX_AC])
1595 != &initarray_generic.cache);
1596 memcpy(ptr, cpu_cache_get(kmalloc_caches[INDEX_AC]),
1597 sizeof(struct arraycache_init));
1599 * Do not assume that spinlocks can be initialized via memcpy:
1601 spin_lock_init(&ptr->lock);
1603 kmalloc_caches[INDEX_AC]->array[smp_processor_id()] = ptr;
1605 /* 5) Replace the bootstrap kmem_cache_node */
1609 for_each_online_node(nid) {
1610 init_list(kmem_cache, &init_kmem_cache_node[CACHE_CACHE + nid], nid);
1612 init_list(kmalloc_caches[INDEX_AC],
1613 &init_kmem_cache_node[SIZE_AC + nid], nid);
1615 if (INDEX_AC != INDEX_NODE) {
1616 init_list(kmalloc_caches[INDEX_NODE],
1617 &init_kmem_cache_node[SIZE_NODE + nid], nid);
1622 create_kmalloc_caches(ARCH_KMALLOC_FLAGS);
1625 void __init kmem_cache_init_late(void)
1627 struct kmem_cache *cachep;
1631 /* 6) resize the head arrays to their final sizes */
1632 mutex_lock(&slab_mutex);
1633 list_for_each_entry(cachep, &slab_caches, list)
1634 if (enable_cpucache(cachep, GFP_NOWAIT))
1636 mutex_unlock(&slab_mutex);
1638 /* Annotate slab for lockdep -- annotate the malloc caches */
1645 * Register a cpu startup notifier callback that initializes
1646 * cpu_cache_get for all new cpus
1648 register_cpu_notifier(&cpucache_notifier);
1652 * Register a memory hotplug callback that initializes and frees
1655 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1659 * The reap timers are started later, with a module init call: That part
1660 * of the kernel is not yet operational.
1664 static int __init cpucache_init(void)
1669 * Register the timers that return unneeded pages to the page allocator
1671 for_each_online_cpu(cpu)
1672 start_cpu_timer(cpu);
1678 __initcall(cpucache_init);
1680 static noinline void
1681 slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1683 struct kmem_cache_node *n;
1685 unsigned long flags;
1689 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1691 printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n",
1692 cachep->name, cachep->size, cachep->gfporder);
1694 for_each_online_node(node) {
1695 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1696 unsigned long active_slabs = 0, num_slabs = 0;
1698 n = cachep->node[node];
1702 spin_lock_irqsave(&n->list_lock, flags);
1703 list_for_each_entry(slabp, &n->slabs_full, list) {
1704 active_objs += cachep->num;
1707 list_for_each_entry(slabp, &n->slabs_partial, list) {
1708 active_objs += slabp->inuse;
1711 list_for_each_entry(slabp, &n->slabs_free, list)
1714 free_objects += n->free_objects;
1715 spin_unlock_irqrestore(&n->list_lock, flags);
1717 num_slabs += active_slabs;
1718 num_objs = num_slabs * cachep->num;
1720 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1721 node, active_slabs, num_slabs, active_objs, num_objs,
1727 * Interface to system's page allocator. No need to hold the cache-lock.
1729 * If we requested dmaable memory, we will get it. Even if we
1730 * did not request dmaable memory, we might get it, but that
1731 * would be relatively rare and ignorable.
1733 static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid)
1741 * Nommu uses slab's for process anonymous memory allocations, and thus
1742 * requires __GFP_COMP to properly refcount higher order allocations
1744 flags |= __GFP_COMP;
1747 flags |= cachep->allocflags;
1748 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1749 flags |= __GFP_RECLAIMABLE;
1751 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1753 if (!(flags & __GFP_NOWARN) && printk_ratelimit())
1754 slab_out_of_memory(cachep, flags, nodeid);
1758 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1759 if (unlikely(page->pfmemalloc))
1760 pfmemalloc_active = true;
1762 nr_pages = (1 << cachep->gfporder);
1763 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1764 add_zone_page_state(page_zone(page),
1765 NR_SLAB_RECLAIMABLE, nr_pages);
1767 add_zone_page_state(page_zone(page),
1768 NR_SLAB_UNRECLAIMABLE, nr_pages);
1769 for (i = 0; i < nr_pages; i++) {
1770 __SetPageSlab(page + i);
1772 if (page->pfmemalloc)
1773 SetPageSlabPfmemalloc(page + i);
1775 memcg_bind_pages(cachep, cachep->gfporder);
1777 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1778 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1781 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1783 kmemcheck_mark_unallocated_pages(page, nr_pages);
1786 return page_address(page);
1790 * Interface to system's page release.
1792 static void kmem_freepages(struct kmem_cache *cachep, void *addr)
1794 unsigned long i = (1 << cachep->gfporder);
1795 struct page *page = virt_to_page(addr);
1796 const unsigned long nr_freed = i;
1798 kmemcheck_free_shadow(page, cachep->gfporder);
1800 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1801 sub_zone_page_state(page_zone(page),
1802 NR_SLAB_RECLAIMABLE, nr_freed);
1804 sub_zone_page_state(page_zone(page),
1805 NR_SLAB_UNRECLAIMABLE, nr_freed);
1807 BUG_ON(!PageSlab(page));
1808 __ClearPageSlabPfmemalloc(page);
1809 __ClearPageSlab(page);
1813 memcg_release_pages(cachep, cachep->gfporder);
1814 if (current->reclaim_state)
1815 current->reclaim_state->reclaimed_slab += nr_freed;
1816 free_memcg_kmem_pages((unsigned long)addr, cachep->gfporder);
1819 static void kmem_rcu_free(struct rcu_head *head)
1821 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1822 struct kmem_cache *cachep = slab_rcu->cachep;
1824 kmem_freepages(cachep, slab_rcu->addr);
1825 if (OFF_SLAB(cachep))
1826 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1831 #ifdef CONFIG_DEBUG_PAGEALLOC
1832 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1833 unsigned long caller)
1835 int size = cachep->object_size;
1837 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1839 if (size < 5 * sizeof(unsigned long))
1842 *addr++ = 0x12345678;
1844 *addr++ = smp_processor_id();
1845 size -= 3 * sizeof(unsigned long);
1847 unsigned long *sptr = &caller;
1848 unsigned long svalue;
1850 while (!kstack_end(sptr)) {
1852 if (kernel_text_address(svalue)) {
1854 size -= sizeof(unsigned long);
1855 if (size <= sizeof(unsigned long))
1861 *addr++ = 0x87654321;
1865 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1867 int size = cachep->object_size;
1868 addr = &((char *)addr)[obj_offset(cachep)];
1870 memset(addr, val, size);
1871 *(unsigned char *)(addr + size - 1) = POISON_END;
1874 static void dump_line(char *data, int offset, int limit)
1877 unsigned char error = 0;
1880 printk(KERN_ERR "%03x: ", offset);
1881 for (i = 0; i < limit; i++) {
1882 if (data[offset + i] != POISON_FREE) {
1883 error = data[offset + i];
1887 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1888 &data[offset], limit, 1);
1890 if (bad_count == 1) {
1891 error ^= POISON_FREE;
1892 if (!(error & (error - 1))) {
1893 printk(KERN_ERR "Single bit error detected. Probably "
1896 printk(KERN_ERR "Run memtest86+ or a similar memory "
1899 printk(KERN_ERR "Run a memory test tool.\n");
1908 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1913 if (cachep->flags & SLAB_RED_ZONE) {
1914 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
1915 *dbg_redzone1(cachep, objp),
1916 *dbg_redzone2(cachep, objp));
1919 if (cachep->flags & SLAB_STORE_USER) {
1920 printk(KERN_ERR "Last user: [<%p>]",
1921 *dbg_userword(cachep, objp));
1922 print_symbol("(%s)",
1923 (unsigned long)*dbg_userword(cachep, objp));
1926 realobj = (char *)objp + obj_offset(cachep);
1927 size = cachep->object_size;
1928 for (i = 0; i < size && lines; i += 16, lines--) {
1931 if (i + limit > size)
1933 dump_line(realobj, i, limit);
1937 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1943 realobj = (char *)objp + obj_offset(cachep);
1944 size = cachep->object_size;
1946 for (i = 0; i < size; i++) {
1947 char exp = POISON_FREE;
1950 if (realobj[i] != exp) {
1956 "Slab corruption (%s): %s start=%p, len=%d\n",
1957 print_tainted(), cachep->name, realobj, size);
1958 print_objinfo(cachep, objp, 0);
1960 /* Hexdump the affected line */
1963 if (i + limit > size)
1965 dump_line(realobj, i, limit);
1968 /* Limit to 5 lines */
1974 /* Print some data about the neighboring objects, if they
1977 struct slab *slabp = virt_to_slab(objp);
1980 objnr = obj_to_index(cachep, slabp, objp);
1982 objp = index_to_obj(cachep, slabp, objnr - 1);
1983 realobj = (char *)objp + obj_offset(cachep);
1984 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1986 print_objinfo(cachep, objp, 2);
1988 if (objnr + 1 < cachep->num) {
1989 objp = index_to_obj(cachep, slabp, objnr + 1);
1990 realobj = (char *)objp + obj_offset(cachep);
1991 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1993 print_objinfo(cachep, objp, 2);
2000 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
2003 for (i = 0; i < cachep->num; i++) {
2004 void *objp = index_to_obj(cachep, slabp, i);
2006 if (cachep->flags & SLAB_POISON) {
2007 #ifdef CONFIG_DEBUG_PAGEALLOC
2008 if (cachep->size % PAGE_SIZE == 0 &&
2010 kernel_map_pages(virt_to_page(objp),
2011 cachep->size / PAGE_SIZE, 1);
2013 check_poison_obj(cachep, objp);
2015 check_poison_obj(cachep, objp);
2018 if (cachep->flags & SLAB_RED_ZONE) {
2019 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2020 slab_error(cachep, "start of a freed object "
2022 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2023 slab_error(cachep, "end of a freed object "
2029 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
2035 * slab_destroy - destroy and release all objects in a slab
2036 * @cachep: cache pointer being destroyed
2037 * @slabp: slab pointer being destroyed
2039 * Destroy all the objs in a slab, and release the mem back to the system.
2040 * Before calling the slab must have been unlinked from the cache. The
2041 * cache-lock is not held/needed.
2043 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
2045 void *addr = slabp->s_mem - slabp->colouroff;
2047 slab_destroy_debugcheck(cachep, slabp);
2048 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
2049 struct slab_rcu *slab_rcu;
2051 slab_rcu = (struct slab_rcu *)slabp;
2052 slab_rcu->cachep = cachep;
2053 slab_rcu->addr = addr;
2054 call_rcu(&slab_rcu->head, kmem_rcu_free);
2056 kmem_freepages(cachep, addr);
2057 if (OFF_SLAB(cachep))
2058 kmem_cache_free(cachep->slabp_cache, slabp);
2063 * calculate_slab_order - calculate size (page order) of slabs
2064 * @cachep: pointer to the cache that is being created
2065 * @size: size of objects to be created in this cache.
2066 * @align: required alignment for the objects.
2067 * @flags: slab allocation flags
2069 * Also calculates the number of objects per slab.
2071 * This could be made much more intelligent. For now, try to avoid using
2072 * high order pages for slabs. When the gfp() functions are more friendly
2073 * towards high-order requests, this should be changed.
2075 static size_t calculate_slab_order(struct kmem_cache *cachep,
2076 size_t size, size_t align, unsigned long flags)
2078 unsigned long offslab_limit;
2079 size_t left_over = 0;
2082 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
2086 cache_estimate(gfporder, size, align, flags, &remainder, &num);
2090 if (flags & CFLGS_OFF_SLAB) {
2092 * Max number of objs-per-slab for caches which
2093 * use off-slab slabs. Needed to avoid a possible
2094 * looping condition in cache_grow().
2096 offslab_limit = size - sizeof(struct slab);
2097 offslab_limit /= sizeof(kmem_bufctl_t);
2099 if (num > offslab_limit)
2103 /* Found something acceptable - save it away */
2105 cachep->gfporder = gfporder;
2106 left_over = remainder;
2109 * A VFS-reclaimable slab tends to have most allocations
2110 * as GFP_NOFS and we really don't want to have to be allocating
2111 * higher-order pages when we are unable to shrink dcache.
2113 if (flags & SLAB_RECLAIM_ACCOUNT)
2117 * Large number of objects is good, but very large slabs are
2118 * currently bad for the gfp()s.
2120 if (gfporder >= slab_max_order)
2124 * Acceptable internal fragmentation?
2126 if (left_over * 8 <= (PAGE_SIZE << gfporder))
2132 static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
2134 if (slab_state >= FULL)
2135 return enable_cpucache(cachep, gfp);
2137 if (slab_state == DOWN) {
2139 * Note: Creation of first cache (kmem_cache).
2140 * The setup_node is taken care
2141 * of by the caller of __kmem_cache_create
2143 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2144 slab_state = PARTIAL;
2145 } else if (slab_state == PARTIAL) {
2147 * Note: the second kmem_cache_create must create the cache
2148 * that's used by kmalloc(24), otherwise the creation of
2149 * further caches will BUG().
2151 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2154 * If the cache that's used by kmalloc(sizeof(kmem_cache_node)) is
2155 * the second cache, then we need to set up all its node/,
2156 * otherwise the creation of further caches will BUG().
2158 set_up_node(cachep, SIZE_AC);
2159 if (INDEX_AC == INDEX_NODE)
2160 slab_state = PARTIAL_NODE;
2162 slab_state = PARTIAL_ARRAYCACHE;
2164 /* Remaining boot caches */
2165 cachep->array[smp_processor_id()] =
2166 kmalloc(sizeof(struct arraycache_init), gfp);
2168 if (slab_state == PARTIAL_ARRAYCACHE) {
2169 set_up_node(cachep, SIZE_NODE);
2170 slab_state = PARTIAL_NODE;
2173 for_each_online_node(node) {
2174 cachep->node[node] =
2175 kmalloc_node(sizeof(struct kmem_cache_node),
2177 BUG_ON(!cachep->node[node]);
2178 kmem_cache_node_init(cachep->node[node]);
2182 cachep->node[numa_mem_id()]->next_reap =
2183 jiffies + REAPTIMEOUT_LIST3 +
2184 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2186 cpu_cache_get(cachep)->avail = 0;
2187 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2188 cpu_cache_get(cachep)->batchcount = 1;
2189 cpu_cache_get(cachep)->touched = 0;
2190 cachep->batchcount = 1;
2191 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2196 * __kmem_cache_create - Create a cache.
2197 * @cachep: cache management descriptor
2198 * @flags: SLAB flags
2200 * Returns a ptr to the cache on success, NULL on failure.
2201 * Cannot be called within a int, but can be interrupted.
2202 * The @ctor is run when new pages are allocated by the cache.
2206 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2207 * to catch references to uninitialised memory.
2209 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2210 * for buffer overruns.
2212 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2213 * cacheline. This can be beneficial if you're counting cycles as closely
2217 __kmem_cache_create (struct kmem_cache *cachep, unsigned long flags)
2219 size_t left_over, slab_size, ralign;
2222 size_t size = cachep->size;
2227 * Enable redzoning and last user accounting, except for caches with
2228 * large objects, if the increased size would increase the object size
2229 * above the next power of two: caches with object sizes just above a
2230 * power of two have a significant amount of internal fragmentation.
2232 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2233 2 * sizeof(unsigned long long)))
2234 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2235 if (!(flags & SLAB_DESTROY_BY_RCU))
2236 flags |= SLAB_POISON;
2238 if (flags & SLAB_DESTROY_BY_RCU)
2239 BUG_ON(flags & SLAB_POISON);
2243 * Check that size is in terms of words. This is needed to avoid
2244 * unaligned accesses for some archs when redzoning is used, and makes
2245 * sure any on-slab bufctl's are also correctly aligned.
2247 if (size & (BYTES_PER_WORD - 1)) {
2248 size += (BYTES_PER_WORD - 1);
2249 size &= ~(BYTES_PER_WORD - 1);
2253 * Redzoning and user store require word alignment or possibly larger.
2254 * Note this will be overridden by architecture or caller mandated
2255 * alignment if either is greater than BYTES_PER_WORD.
2257 if (flags & SLAB_STORE_USER)
2258 ralign = BYTES_PER_WORD;
2260 if (flags & SLAB_RED_ZONE) {
2261 ralign = REDZONE_ALIGN;
2262 /* If redzoning, ensure that the second redzone is suitably
2263 * aligned, by adjusting the object size accordingly. */
2264 size += REDZONE_ALIGN - 1;
2265 size &= ~(REDZONE_ALIGN - 1);
2268 /* 3) caller mandated alignment */
2269 if (ralign < cachep->align) {
2270 ralign = cachep->align;
2272 /* disable debug if necessary */
2273 if (ralign > __alignof__(unsigned long long))
2274 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2278 cachep->align = ralign;
2280 if (slab_is_available())
2285 setup_node_pointer(cachep);
2289 * Both debugging options require word-alignment which is calculated
2292 if (flags & SLAB_RED_ZONE) {
2293 /* add space for red zone words */
2294 cachep->obj_offset += sizeof(unsigned long long);
2295 size += 2 * sizeof(unsigned long long);
2297 if (flags & SLAB_STORE_USER) {
2298 /* user store requires one word storage behind the end of
2299 * the real object. But if the second red zone needs to be
2300 * aligned to 64 bits, we must allow that much space.
2302 if (flags & SLAB_RED_ZONE)
2303 size += REDZONE_ALIGN;
2305 size += BYTES_PER_WORD;
2307 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2308 if (size >= kmalloc_size(INDEX_NODE + 1)
2309 && cachep->object_size > cache_line_size()
2310 && ALIGN(size, cachep->align) < PAGE_SIZE) {
2311 cachep->obj_offset += PAGE_SIZE - ALIGN(size, cachep->align);
2318 * Determine if the slab management is 'on' or 'off' slab.
2319 * (bootstrapping cannot cope with offslab caches so don't do
2320 * it too early on. Always use on-slab management when
2321 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2323 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
2324 !(flags & SLAB_NOLEAKTRACE))
2326 * Size is large, assume best to place the slab management obj
2327 * off-slab (should allow better packing of objs).
2329 flags |= CFLGS_OFF_SLAB;
2331 size = ALIGN(size, cachep->align);
2333 left_over = calculate_slab_order(cachep, size, cachep->align, flags);
2338 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2339 + sizeof(struct slab), cachep->align);
2342 * If the slab has been placed off-slab, and we have enough space then
2343 * move it on-slab. This is at the expense of any extra colouring.
2345 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2346 flags &= ~CFLGS_OFF_SLAB;
2347 left_over -= slab_size;
2350 if (flags & CFLGS_OFF_SLAB) {
2351 /* really off slab. No need for manual alignment */
2353 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2355 #ifdef CONFIG_PAGE_POISONING
2356 /* If we're going to use the generic kernel_map_pages()
2357 * poisoning, then it's going to smash the contents of
2358 * the redzone and userword anyhow, so switch them off.
2360 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2361 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2365 cachep->colour_off = cache_line_size();
2366 /* Offset must be a multiple of the alignment. */
2367 if (cachep->colour_off < cachep->align)
2368 cachep->colour_off = cachep->align;
2369 cachep->colour = left_over / cachep->colour_off;
2370 cachep->slab_size = slab_size;
2371 cachep->flags = flags;
2372 cachep->allocflags = 0;
2373 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2374 cachep->allocflags |= GFP_DMA;
2375 cachep->size = size;
2376 cachep->reciprocal_buffer_size = reciprocal_value(size);
2378 if (flags & CFLGS_OFF_SLAB) {
2379 cachep->slabp_cache = kmalloc_slab(slab_size, 0u);
2381 * This is a possibility for one of the malloc_sizes caches.
2382 * But since we go off slab only for object size greater than
2383 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2384 * this should not happen at all.
2385 * But leave a BUG_ON for some lucky dude.
2387 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
2390 err = setup_cpu_cache(cachep, gfp);
2392 __kmem_cache_shutdown(cachep);
2396 if (flags & SLAB_DEBUG_OBJECTS) {
2398 * Would deadlock through slab_destroy()->call_rcu()->
2399 * debug_object_activate()->kmem_cache_alloc().
2401 WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU);
2403 slab_set_debugobj_lock_classes(cachep);
2404 } else if (!OFF_SLAB(cachep) && !(flags & SLAB_DESTROY_BY_RCU))
2405 on_slab_lock_classes(cachep);
2411 static void check_irq_off(void)
2413 BUG_ON(!irqs_disabled());
2416 static void check_irq_on(void)
2418 BUG_ON(irqs_disabled());
2421 static void check_spinlock_acquired(struct kmem_cache *cachep)
2425 assert_spin_locked(&cachep->node[numa_mem_id()]->list_lock);
2429 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2433 assert_spin_locked(&cachep->node[node]->list_lock);
2438 #define check_irq_off() do { } while(0)
2439 #define check_irq_on() do { } while(0)
2440 #define check_spinlock_acquired(x) do { } while(0)
2441 #define check_spinlock_acquired_node(x, y) do { } while(0)
2444 static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
2445 struct array_cache *ac,
2446 int force, int node);
2448 static void do_drain(void *arg)
2450 struct kmem_cache *cachep = arg;
2451 struct array_cache *ac;
2452 int node = numa_mem_id();
2455 ac = cpu_cache_get(cachep);
2456 spin_lock(&cachep->node[node]->list_lock);
2457 free_block(cachep, ac->entry, ac->avail, node);
2458 spin_unlock(&cachep->node[node]->list_lock);
2462 static void drain_cpu_caches(struct kmem_cache *cachep)
2464 struct kmem_cache_node *n;
2467 on_each_cpu(do_drain, cachep, 1);
2469 for_each_online_node(node) {
2470 n = cachep->node[node];
2472 drain_alien_cache(cachep, n->alien);
2475 for_each_online_node(node) {
2476 n = cachep->node[node];
2478 drain_array(cachep, n, n->shared, 1, node);
2483 * Remove slabs from the list of free slabs.
2484 * Specify the number of slabs to drain in tofree.
2486 * Returns the actual number of slabs released.
2488 static int drain_freelist(struct kmem_cache *cache,
2489 struct kmem_cache_node *n, int tofree)
2491 struct list_head *p;
2496 while (nr_freed < tofree && !list_empty(&n->slabs_free)) {
2498 spin_lock_irq(&n->list_lock);
2499 p = n->slabs_free.prev;
2500 if (p == &n->slabs_free) {
2501 spin_unlock_irq(&n->list_lock);
2505 slabp = list_entry(p, struct slab, list);
2507 BUG_ON(slabp->inuse);
2509 list_del(&slabp->list);
2511 * Safe to drop the lock. The slab is no longer linked
2514 n->free_objects -= cache->num;
2515 spin_unlock_irq(&n->list_lock);
2516 slab_destroy(cache, slabp);
2523 /* Called with slab_mutex held to protect against cpu hotplug */
2524 static int __cache_shrink(struct kmem_cache *cachep)
2527 struct kmem_cache_node *n;
2529 drain_cpu_caches(cachep);
2532 for_each_online_node(i) {
2533 n = cachep->node[i];
2537 drain_freelist(cachep, n, n->free_objects);
2539 ret += !list_empty(&n->slabs_full) ||
2540 !list_empty(&n->slabs_partial);
2542 return (ret ? 1 : 0);
2546 * kmem_cache_shrink - Shrink a cache.
2547 * @cachep: The cache to shrink.
2549 * Releases as many slabs as possible for a cache.
2550 * To help debugging, a zero exit status indicates all slabs were released.
2552 int kmem_cache_shrink(struct kmem_cache *cachep)
2555 BUG_ON(!cachep || in_interrupt());
2558 mutex_lock(&slab_mutex);
2559 ret = __cache_shrink(cachep);
2560 mutex_unlock(&slab_mutex);
2564 EXPORT_SYMBOL(kmem_cache_shrink);
2566 int __kmem_cache_shutdown(struct kmem_cache *cachep)
2569 struct kmem_cache_node *n;
2570 int rc = __cache_shrink(cachep);
2575 for_each_online_cpu(i)
2576 kfree(cachep->array[i]);
2578 /* NUMA: free the node structures */
2579 for_each_online_node(i) {
2580 n = cachep->node[i];
2583 free_alien_cache(n->alien);
2591 * Get the memory for a slab management obj.
2592 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2593 * always come from malloc_sizes caches. The slab descriptor cannot
2594 * come from the same cache which is getting created because,
2595 * when we are searching for an appropriate cache for these
2596 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2597 * If we are creating a malloc_sizes cache here it would not be visible to
2598 * kmem_find_general_cachep till the initialization is complete.
2599 * Hence we cannot have slabp_cache same as the original cache.
2601 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp,
2602 int colour_off, gfp_t local_flags,
2607 if (OFF_SLAB(cachep)) {
2608 /* Slab management obj is off-slab. */
2609 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2610 local_flags, nodeid);
2612 * If the first object in the slab is leaked (it's allocated
2613 * but no one has a reference to it), we want to make sure
2614 * kmemleak does not treat the ->s_mem pointer as a reference
2615 * to the object. Otherwise we will not report the leak.
2617 kmemleak_scan_area(&slabp->list, sizeof(struct list_head),
2622 slabp = objp + colour_off;
2623 colour_off += cachep->slab_size;
2626 slabp->colouroff = colour_off;
2627 slabp->s_mem = objp + colour_off;
2628 slabp->nodeid = nodeid;
2633 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2635 return (kmem_bufctl_t *) (slabp + 1);
2638 static void cache_init_objs(struct kmem_cache *cachep,
2643 for (i = 0; i < cachep->num; i++) {
2644 void *objp = index_to_obj(cachep, slabp, i);
2646 /* need to poison the objs? */
2647 if (cachep->flags & SLAB_POISON)
2648 poison_obj(cachep, objp, POISON_FREE);
2649 if (cachep->flags & SLAB_STORE_USER)
2650 *dbg_userword(cachep, objp) = NULL;
2652 if (cachep->flags & SLAB_RED_ZONE) {
2653 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2654 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2657 * Constructors are not allowed to allocate memory from the same
2658 * cache which they are a constructor for. Otherwise, deadlock.
2659 * They must also be threaded.
2661 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2662 cachep->ctor(objp + obj_offset(cachep));
2664 if (cachep->flags & SLAB_RED_ZONE) {
2665 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2666 slab_error(cachep, "constructor overwrote the"
2667 " end of an object");
2668 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2669 slab_error(cachep, "constructor overwrote the"
2670 " start of an object");
2672 if ((cachep->size % PAGE_SIZE) == 0 &&
2673 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2674 kernel_map_pages(virt_to_page(objp),
2675 cachep->size / PAGE_SIZE, 0);
2680 slab_bufctl(slabp)[i] = i + 1;
2682 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2685 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2687 if (CONFIG_ZONE_DMA_FLAG) {
2688 if (flags & GFP_DMA)
2689 BUG_ON(!(cachep->allocflags & GFP_DMA));
2691 BUG_ON(cachep->allocflags & GFP_DMA);
2695 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2698 void *objp = index_to_obj(cachep, slabp, slabp->free);
2702 next = slab_bufctl(slabp)[slabp->free];
2704 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2705 WARN_ON(slabp->nodeid != nodeid);
2712 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2713 void *objp, int nodeid)
2715 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2718 /* Verify that the slab belongs to the intended node */
2719 WARN_ON(slabp->nodeid != nodeid);
2721 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2722 printk(KERN_ERR "slab: double free detected in cache "
2723 "'%s', objp %p\n", cachep->name, objp);
2727 slab_bufctl(slabp)[objnr] = slabp->free;
2728 slabp->free = objnr;
2733 * Map pages beginning at addr to the given cache and slab. This is required
2734 * for the slab allocator to be able to lookup the cache and slab of a
2735 * virtual address for kfree, ksize, and slab debugging.
2737 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2743 page = virt_to_page(addr);
2746 if (likely(!PageCompound(page)))
2747 nr_pages <<= cache->gfporder;
2750 page->slab_cache = cache;
2751 page->slab_page = slab;
2753 } while (--nr_pages);
2757 * Grow (by 1) the number of slabs within a cache. This is called by
2758 * kmem_cache_alloc() when there are no active objs left in a cache.
2760 static int cache_grow(struct kmem_cache *cachep,
2761 gfp_t flags, int nodeid, void *objp)
2766 struct kmem_cache_node *n;
2769 * Be lazy and only check for valid flags here, keeping it out of the
2770 * critical path in kmem_cache_alloc().
2772 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2773 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2775 /* Take the node list lock to change the colour_next on this node */
2777 n = cachep->node[nodeid];
2778 spin_lock(&n->list_lock);
2780 /* Get colour for the slab, and cal the next value. */
2781 offset = n->colour_next;
2783 if (n->colour_next >= cachep->colour)
2785 spin_unlock(&n->list_lock);
2787 offset *= cachep->colour_off;
2789 if (local_flags & __GFP_WAIT)
2793 * The test for missing atomic flag is performed here, rather than
2794 * the more obvious place, simply to reduce the critical path length
2795 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2796 * will eventually be caught here (where it matters).
2798 kmem_flagcheck(cachep, flags);
2801 * Get mem for the objs. Attempt to allocate a physical page from
2805 objp = kmem_getpages(cachep, local_flags, nodeid);
2809 /* Get slab management. */
2810 slabp = alloc_slabmgmt(cachep, objp, offset,
2811 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2815 slab_map_pages(cachep, slabp, objp);
2817 cache_init_objs(cachep, slabp);
2819 if (local_flags & __GFP_WAIT)
2820 local_irq_disable();
2822 spin_lock(&n->list_lock);
2824 /* Make slab active. */
2825 list_add_tail(&slabp->list, &(n->slabs_free));
2826 STATS_INC_GROWN(cachep);
2827 n->free_objects += cachep->num;
2828 spin_unlock(&n->list_lock);
2831 kmem_freepages(cachep, objp);
2833 if (local_flags & __GFP_WAIT)
2834 local_irq_disable();
2841 * Perform extra freeing checks:
2842 * - detect bad pointers.
2843 * - POISON/RED_ZONE checking
2845 static void kfree_debugcheck(const void *objp)
2847 if (!virt_addr_valid(objp)) {
2848 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2849 (unsigned long)objp);
2854 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2856 unsigned long long redzone1, redzone2;
2858 redzone1 = *dbg_redzone1(cache, obj);
2859 redzone2 = *dbg_redzone2(cache, obj);
2864 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2867 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2868 slab_error(cache, "double free detected");
2870 slab_error(cache, "memory outside object was overwritten");
2872 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2873 obj, redzone1, redzone2);
2876 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2877 unsigned long caller)
2883 BUG_ON(virt_to_cache(objp) != cachep);
2885 objp -= obj_offset(cachep);
2886 kfree_debugcheck(objp);
2887 page = virt_to_head_page(objp);
2889 slabp = page->slab_page;
2891 if (cachep->flags & SLAB_RED_ZONE) {
2892 verify_redzone_free(cachep, objp);
2893 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2894 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2896 if (cachep->flags & SLAB_STORE_USER)
2897 *dbg_userword(cachep, objp) = (void *)caller;
2899 objnr = obj_to_index(cachep, slabp, objp);
2901 BUG_ON(objnr >= cachep->num);
2902 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
2904 #ifdef CONFIG_DEBUG_SLAB_LEAK
2905 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2907 if (cachep->flags & SLAB_POISON) {
2908 #ifdef CONFIG_DEBUG_PAGEALLOC
2909 if ((cachep->size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
2910 store_stackinfo(cachep, objp, caller);
2911 kernel_map_pages(virt_to_page(objp),
2912 cachep->size / PAGE_SIZE, 0);
2914 poison_obj(cachep, objp, POISON_FREE);
2917 poison_obj(cachep, objp, POISON_FREE);
2923 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
2928 /* Check slab's freelist to see if this obj is there. */
2929 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2931 if (entries > cachep->num || i >= cachep->num)
2934 if (entries != cachep->num - slabp->inuse) {
2936 printk(KERN_ERR "slab: Internal list corruption detected in "
2937 "cache '%s'(%d), slabp %p(%d). Tainted(%s). Hexdump:\n",
2938 cachep->name, cachep->num, slabp, slabp->inuse,
2940 print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 16, 1, slabp,
2941 sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t),
2947 #define kfree_debugcheck(x) do { } while(0)
2948 #define cache_free_debugcheck(x,objp,z) (objp)
2949 #define check_slabp(x,y) do { } while(0)
2952 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags,
2956 struct kmem_cache_node *n;
2957 struct array_cache *ac;
2961 node = numa_mem_id();
2962 if (unlikely(force_refill))
2965 ac = cpu_cache_get(cachep);
2966 batchcount = ac->batchcount;
2967 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2969 * If there was little recent activity on this cache, then
2970 * perform only a partial refill. Otherwise we could generate
2973 batchcount = BATCHREFILL_LIMIT;
2975 n = cachep->node[node];
2977 BUG_ON(ac->avail > 0 || !n);
2978 spin_lock(&n->list_lock);
2980 /* See if we can refill from the shared array */
2981 if (n->shared && transfer_objects(ac, n->shared, batchcount)) {
2982 n->shared->touched = 1;
2986 while (batchcount > 0) {
2987 struct list_head *entry;
2989 /* Get slab alloc is to come from. */
2990 entry = n->slabs_partial.next;
2991 if (entry == &n->slabs_partial) {
2992 n->free_touched = 1;
2993 entry = n->slabs_free.next;
2994 if (entry == &n->slabs_free)
2998 slabp = list_entry(entry, struct slab, list);
2999 check_slabp(cachep, slabp);
3000 check_spinlock_acquired(cachep);
3003 * The slab was either on partial or free list so
3004 * there must be at least one object available for
3007 BUG_ON(slabp->inuse >= cachep->num);
3009 while (slabp->inuse < cachep->num && batchcount--) {
3010 STATS_INC_ALLOCED(cachep);
3011 STATS_INC_ACTIVE(cachep);
3012 STATS_SET_HIGH(cachep);
3014 ac_put_obj(cachep, ac, slab_get_obj(cachep, slabp,
3017 check_slabp(cachep, slabp);
3019 /* move slabp to correct slabp list: */
3020 list_del(&slabp->list);
3021 if (slabp->free == BUFCTL_END)
3022 list_add(&slabp->list, &n->slabs_full);
3024 list_add(&slabp->list, &n->slabs_partial);
3028 n->free_objects -= ac->avail;
3030 spin_unlock(&n->list_lock);
3032 if (unlikely(!ac->avail)) {
3035 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
3037 /* cache_grow can reenable interrupts, then ac could change. */
3038 ac = cpu_cache_get(cachep);
3039 node = numa_mem_id();
3041 /* no objects in sight? abort */
3042 if (!x && (ac->avail == 0 || force_refill))
3045 if (!ac->avail) /* objects refilled by interrupt? */
3050 return ac_get_obj(cachep, ac, flags, force_refill);
3053 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3056 might_sleep_if(flags & __GFP_WAIT);
3058 kmem_flagcheck(cachep, flags);
3063 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3064 gfp_t flags, void *objp, unsigned long caller)
3068 if (cachep->flags & SLAB_POISON) {
3069 #ifdef CONFIG_DEBUG_PAGEALLOC
3070 if ((cachep->size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
3071 kernel_map_pages(virt_to_page(objp),
3072 cachep->size / PAGE_SIZE, 1);
3074 check_poison_obj(cachep, objp);
3076 check_poison_obj(cachep, objp);
3078 poison_obj(cachep, objp, POISON_INUSE);
3080 if (cachep->flags & SLAB_STORE_USER)
3081 *dbg_userword(cachep, objp) = (void *)caller;
3083 if (cachep->flags & SLAB_RED_ZONE) {
3084 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3085 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3086 slab_error(cachep, "double free, or memory outside"
3087 " object was overwritten");
3089 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3090 objp, *dbg_redzone1(cachep, objp),
3091 *dbg_redzone2(cachep, objp));
3093 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3094 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3096 #ifdef CONFIG_DEBUG_SLAB_LEAK
3101 slabp = virt_to_head_page(objp)->slab_page;
3102 objnr = (unsigned)(objp - slabp->s_mem) / cachep->size;
3103 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3106 objp += obj_offset(cachep);
3107 if (cachep->ctor && cachep->flags & SLAB_POISON)
3109 if (ARCH_SLAB_MINALIGN &&
3110 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
3111 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3112 objp, (int)ARCH_SLAB_MINALIGN);
3117 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3120 static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
3122 if (cachep == kmem_cache)
3125 return should_failslab(cachep->object_size, flags, cachep->flags);
3128 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3131 struct array_cache *ac;
3132 bool force_refill = false;
3136 ac = cpu_cache_get(cachep);
3137 if (likely(ac->avail)) {
3139 objp = ac_get_obj(cachep, ac, flags, false);
3142 * Allow for the possibility all avail objects are not allowed
3143 * by the current flags
3146 STATS_INC_ALLOCHIT(cachep);
3149 force_refill = true;
3152 STATS_INC_ALLOCMISS(cachep);
3153 objp = cache_alloc_refill(cachep, flags, force_refill);
3155 * the 'ac' may be updated by cache_alloc_refill(),
3156 * and kmemleak_erase() requires its correct value.
3158 ac = cpu_cache_get(cachep);
3162 * To avoid a false negative, if an object that is in one of the
3163 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3164 * treat the array pointers as a reference to the object.
3167 kmemleak_erase(&ac->entry[ac->avail]);
3173 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3175 * If we are in_interrupt, then process context, including cpusets and
3176 * mempolicy, may not apply and should not be used for allocation policy.
3178 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3180 int nid_alloc, nid_here;
3182 if (in_interrupt() || (flags & __GFP_THISNODE))
3184 nid_alloc = nid_here = numa_mem_id();
3185 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3186 nid_alloc = cpuset_slab_spread_node();
3187 else if (current->mempolicy)
3188 nid_alloc = slab_node();
3189 if (nid_alloc != nid_here)
3190 return ____cache_alloc_node(cachep, flags, nid_alloc);
3195 * Fallback function if there was no memory available and no objects on a
3196 * certain node and fall back is permitted. First we scan all the
3197 * available node for available objects. If that fails then we
3198 * perform an allocation without specifying a node. This allows the page
3199 * allocator to do its reclaim / fallback magic. We then insert the
3200 * slab into the proper nodelist and then allocate from it.
3202 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3204 struct zonelist *zonelist;
3208 enum zone_type high_zoneidx = gfp_zone(flags);
3211 unsigned int cpuset_mems_cookie;
3213 if (flags & __GFP_THISNODE)
3216 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3219 cpuset_mems_cookie = get_mems_allowed();
3220 zonelist = node_zonelist(slab_node(), flags);
3224 * Look through allowed nodes for objects available
3225 * from existing per node queues.
3227 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3228 nid = zone_to_nid(zone);
3230 if (cpuset_zone_allowed_hardwall(zone, flags) &&
3232 cache->node[nid]->free_objects) {
3233 obj = ____cache_alloc_node(cache,
3234 flags | GFP_THISNODE, nid);
3242 * This allocation will be performed within the constraints
3243 * of the current cpuset / memory policy requirements.
3244 * We may trigger various forms of reclaim on the allowed
3245 * set and go into memory reserves if necessary.
3247 if (local_flags & __GFP_WAIT)
3249 kmem_flagcheck(cache, flags);
3250 obj = kmem_getpages(cache, local_flags, numa_mem_id());
3251 if (local_flags & __GFP_WAIT)
3252 local_irq_disable();
3255 * Insert into the appropriate per node queues
3257 nid = page_to_nid(virt_to_page(obj));
3258 if (cache_grow(cache, flags, nid, obj)) {
3259 obj = ____cache_alloc_node(cache,
3260 flags | GFP_THISNODE, nid);
3263 * Another processor may allocate the
3264 * objects in the slab since we are
3265 * not holding any locks.
3269 /* cache_grow already freed obj */
3275 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !obj))
3281 * A interface to enable slab creation on nodeid
3283 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3286 struct list_head *entry;
3288 struct kmem_cache_node *n;
3292 VM_BUG_ON(nodeid > num_online_nodes());
3293 n = cachep->node[nodeid];
3298 spin_lock(&n->list_lock);
3299 entry = n->slabs_partial.next;
3300 if (entry == &n->slabs_partial) {
3301 n->free_touched = 1;
3302 entry = n->slabs_free.next;
3303 if (entry == &n->slabs_free)
3307 slabp = list_entry(entry, struct slab, list);
3308 check_spinlock_acquired_node(cachep, nodeid);
3309 check_slabp(cachep, slabp);
3311 STATS_INC_NODEALLOCS(cachep);
3312 STATS_INC_ACTIVE(cachep);
3313 STATS_SET_HIGH(cachep);
3315 BUG_ON(slabp->inuse == cachep->num);
3317 obj = slab_get_obj(cachep, slabp, nodeid);
3318 check_slabp(cachep, slabp);
3320 /* move slabp to correct slabp list: */
3321 list_del(&slabp->list);
3323 if (slabp->free == BUFCTL_END)
3324 list_add(&slabp->list, &n->slabs_full);
3326 list_add(&slabp->list, &n->slabs_partial);
3328 spin_unlock(&n->list_lock);
3332 spin_unlock(&n->list_lock);
3333 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
3337 return fallback_alloc(cachep, flags);
3344 * kmem_cache_alloc_node - Allocate an object on the specified node
3345 * @cachep: The cache to allocate from.
3346 * @flags: See kmalloc().
3347 * @nodeid: node number of the target node.
3348 * @caller: return address of caller, used for debug information
3350 * Identical to kmem_cache_alloc but it will allocate memory on the given
3351 * node, which can improve the performance for cpu bound structures.
3353 * Fallback to other node is possible if __GFP_THISNODE is not set.
3355 static __always_inline void *
3356 slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3357 unsigned long caller)
3359 unsigned long save_flags;
3361 int slab_node = numa_mem_id();
3363 flags &= gfp_allowed_mask;
3365 lockdep_trace_alloc(flags);
3367 if (slab_should_failslab(cachep, flags))
3370 cachep = memcg_kmem_get_cache(cachep, flags);
3372 cache_alloc_debugcheck_before(cachep, flags);
3373 local_irq_save(save_flags);
3375 if (nodeid == NUMA_NO_NODE)
3378 if (unlikely(!cachep->node[nodeid])) {
3379 /* Node not bootstrapped yet */
3380 ptr = fallback_alloc(cachep, flags);
3384 if (nodeid == slab_node) {
3386 * Use the locally cached objects if possible.
3387 * However ____cache_alloc does not allow fallback
3388 * to other nodes. It may fail while we still have
3389 * objects on other nodes available.
3391 ptr = ____cache_alloc(cachep, flags);
3395 /* ___cache_alloc_node can fall back to other nodes */
3396 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3398 local_irq_restore(save_flags);
3399 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3400 kmemleak_alloc_recursive(ptr, cachep->object_size, 1, cachep->flags,
3404 kmemcheck_slab_alloc(cachep, flags, ptr, cachep->object_size);
3406 if (unlikely((flags & __GFP_ZERO) && ptr))
3407 memset(ptr, 0, cachep->object_size);
3412 static __always_inline void *
3413 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3417 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3418 objp = alternate_node_alloc(cache, flags);
3422 objp = ____cache_alloc(cache, flags);
3425 * We may just have run out of memory on the local node.
3426 * ____cache_alloc_node() knows how to locate memory on other nodes
3429 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3436 static __always_inline void *
3437 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3439 return ____cache_alloc(cachep, flags);
3442 #endif /* CONFIG_NUMA */
3444 static __always_inline void *
3445 slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
3447 unsigned long save_flags;
3450 flags &= gfp_allowed_mask;
3452 lockdep_trace_alloc(flags);
3454 if (slab_should_failslab(cachep, flags))
3457 cachep = memcg_kmem_get_cache(cachep, flags);
3459 cache_alloc_debugcheck_before(cachep, flags);
3460 local_irq_save(save_flags);
3461 objp = __do_cache_alloc(cachep, flags);
3462 local_irq_restore(save_flags);
3463 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3464 kmemleak_alloc_recursive(objp, cachep->object_size, 1, cachep->flags,
3469 kmemcheck_slab_alloc(cachep, flags, objp, cachep->object_size);
3471 if (unlikely((flags & __GFP_ZERO) && objp))
3472 memset(objp, 0, cachep->object_size);
3478 * Caller needs to acquire correct kmem_list's list_lock
3480 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3484 struct kmem_cache_node *n;
3486 for (i = 0; i < nr_objects; i++) {
3490 clear_obj_pfmemalloc(&objpp[i]);
3493 slabp = virt_to_slab(objp);
3494 n = cachep->node[node];
3495 list_del(&slabp->list);
3496 check_spinlock_acquired_node(cachep, node);
3497 check_slabp(cachep, slabp);
3498 slab_put_obj(cachep, slabp, objp, node);
3499 STATS_DEC_ACTIVE(cachep);
3501 check_slabp(cachep, slabp);
3503 /* fixup slab chains */
3504 if (slabp->inuse == 0) {
3505 if (n->free_objects > n->free_limit) {
3506 n->free_objects -= cachep->num;
3507 /* No need to drop any previously held
3508 * lock here, even if we have a off-slab slab
3509 * descriptor it is guaranteed to come from
3510 * a different cache, refer to comments before
3513 slab_destroy(cachep, slabp);
3515 list_add(&slabp->list, &n->slabs_free);
3518 /* Unconditionally move a slab to the end of the
3519 * partial list on free - maximum time for the
3520 * other objects to be freed, too.
3522 list_add_tail(&slabp->list, &n->slabs_partial);
3527 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3530 struct kmem_cache_node *n;
3531 int node = numa_mem_id();
3533 batchcount = ac->batchcount;
3535 BUG_ON(!batchcount || batchcount > ac->avail);
3538 n = cachep->node[node];
3539 spin_lock(&n->list_lock);
3541 struct array_cache *shared_array = n->shared;
3542 int max = shared_array->limit - shared_array->avail;
3544 if (batchcount > max)
3546 memcpy(&(shared_array->entry[shared_array->avail]),
3547 ac->entry, sizeof(void *) * batchcount);
3548 shared_array->avail += batchcount;
3553 free_block(cachep, ac->entry, batchcount, node);
3558 struct list_head *p;
3560 p = n->slabs_free.next;
3561 while (p != &(n->slabs_free)) {
3564 slabp = list_entry(p, struct slab, list);
3565 BUG_ON(slabp->inuse);
3570 STATS_SET_FREEABLE(cachep, i);
3573 spin_unlock(&n->list_lock);
3574 ac->avail -= batchcount;
3575 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3579 * Release an obj back to its cache. If the obj has a constructed state, it must
3580 * be in this state _before_ it is released. Called with disabled ints.
3582 static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3583 unsigned long caller)
3585 struct array_cache *ac = cpu_cache_get(cachep);
3588 kmemleak_free_recursive(objp, cachep->flags);
3589 objp = cache_free_debugcheck(cachep, objp, caller);
3591 kmemcheck_slab_free(cachep, objp, cachep->object_size);
3594 * Skip calling cache_free_alien() when the platform is not numa.
3595 * This will avoid cache misses that happen while accessing slabp (which
3596 * is per page memory reference) to get nodeid. Instead use a global
3597 * variable to skip the call, which is mostly likely to be present in
3600 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3603 if (likely(ac->avail < ac->limit)) {
3604 STATS_INC_FREEHIT(cachep);
3606 STATS_INC_FREEMISS(cachep);
3607 cache_flusharray(cachep, ac);
3610 ac_put_obj(cachep, ac, objp);
3614 * kmem_cache_alloc - Allocate an object
3615 * @cachep: The cache to allocate from.
3616 * @flags: See kmalloc().
3618 * Allocate an object from this cache. The flags are only relevant
3619 * if the cache has no available objects.
3621 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3623 void *ret = slab_alloc(cachep, flags, _RET_IP_);
3625 trace_kmem_cache_alloc(_RET_IP_, ret,
3626 cachep->object_size, cachep->size, flags);
3630 EXPORT_SYMBOL(kmem_cache_alloc);
3632 #ifdef CONFIG_TRACING
3634 kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
3638 ret = slab_alloc(cachep, flags, _RET_IP_);
3640 trace_kmalloc(_RET_IP_, ret,
3641 size, cachep->size, flags);
3644 EXPORT_SYMBOL(kmem_cache_alloc_trace);
3648 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3650 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3652 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3653 cachep->object_size, cachep->size,
3658 EXPORT_SYMBOL(kmem_cache_alloc_node);
3660 #ifdef CONFIG_TRACING
3661 void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
3668 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3670 trace_kmalloc_node(_RET_IP_, ret,
3675 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3678 static __always_inline void *
3679 __do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
3681 struct kmem_cache *cachep;
3683 cachep = kmalloc_slab(size, flags);
3684 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3686 return kmem_cache_alloc_node_trace(cachep, flags, node, size);
3689 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3690 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3692 return __do_kmalloc_node(size, flags, node, _RET_IP_);
3694 EXPORT_SYMBOL(__kmalloc_node);
3696 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3697 int node, unsigned long caller)
3699 return __do_kmalloc_node(size, flags, node, caller);
3701 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3703 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3705 return __do_kmalloc_node(size, flags, node, 0);
3707 EXPORT_SYMBOL(__kmalloc_node);
3708 #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3709 #endif /* CONFIG_NUMA */
3712 * __do_kmalloc - allocate memory
3713 * @size: how many bytes of memory are required.
3714 * @flags: the type of memory to allocate (see kmalloc).
3715 * @caller: function caller for debug tracking of the caller
3717 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3718 unsigned long caller)
3720 struct kmem_cache *cachep;
3723 /* If you want to save a few bytes .text space: replace
3725 * Then kmalloc uses the uninlined functions instead of the inline
3728 cachep = kmalloc_slab(size, flags);
3729 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3731 ret = slab_alloc(cachep, flags, caller);
3733 trace_kmalloc(caller, ret,
3734 size, cachep->size, flags);
3740 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3741 void *__kmalloc(size_t size, gfp_t flags)
3743 return __do_kmalloc(size, flags, _RET_IP_);
3745 EXPORT_SYMBOL(__kmalloc);
3747 void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3749 return __do_kmalloc(size, flags, caller);
3751 EXPORT_SYMBOL(__kmalloc_track_caller);
3754 void *__kmalloc(size_t size, gfp_t flags)
3756 return __do_kmalloc(size, flags, 0);
3758 EXPORT_SYMBOL(__kmalloc);
3762 * kmem_cache_free - Deallocate an object
3763 * @cachep: The cache the allocation was from.
3764 * @objp: The previously allocated object.
3766 * Free an object which was previously allocated from this
3769 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3771 unsigned long flags;
3772 cachep = cache_from_obj(cachep, objp);
3776 local_irq_save(flags);
3777 debug_check_no_locks_freed(objp, cachep->object_size);
3778 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3779 debug_check_no_obj_freed(objp, cachep->object_size);
3780 __cache_free(cachep, objp, _RET_IP_);
3781 local_irq_restore(flags);
3783 trace_kmem_cache_free(_RET_IP_, objp);
3785 EXPORT_SYMBOL(kmem_cache_free);
3788 * kfree - free previously allocated memory
3789 * @objp: pointer returned by kmalloc.
3791 * If @objp is NULL, no operation is performed.
3793 * Don't free memory not originally allocated by kmalloc()
3794 * or you will run into trouble.
3796 void kfree(const void *objp)
3798 struct kmem_cache *c;
3799 unsigned long flags;
3801 trace_kfree(_RET_IP_, objp);
3803 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3805 local_irq_save(flags);
3806 kfree_debugcheck(objp);
3807 c = virt_to_cache(objp);
3808 debug_check_no_locks_freed(objp, c->object_size);
3810 debug_check_no_obj_freed(objp, c->object_size);
3811 __cache_free(c, (void *)objp, _RET_IP_);
3812 local_irq_restore(flags);
3814 EXPORT_SYMBOL(kfree);
3817 * This initializes kmem_cache_node or resizes various caches for all nodes.
3819 static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
3822 struct kmem_cache_node *n;
3823 struct array_cache *new_shared;
3824 struct array_cache **new_alien = NULL;
3826 for_each_online_node(node) {
3828 if (use_alien_caches) {
3829 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
3835 if (cachep->shared) {
3836 new_shared = alloc_arraycache(node,
3837 cachep->shared*cachep->batchcount,
3840 free_alien_cache(new_alien);
3845 n = cachep->node[node];
3847 struct array_cache *shared = n->shared;
3849 spin_lock_irq(&n->list_lock);
3852 free_block(cachep, shared->entry,
3853 shared->avail, node);
3855 n->shared = new_shared;
3857 n->alien = new_alien;
3860 n->free_limit = (1 + nr_cpus_node(node)) *
3861 cachep->batchcount + cachep->num;
3862 spin_unlock_irq(&n->list_lock);
3864 free_alien_cache(new_alien);
3867 n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
3869 free_alien_cache(new_alien);
3874 kmem_cache_node_init(n);
3875 n->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3876 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3877 n->shared = new_shared;
3878 n->alien = new_alien;
3879 n->free_limit = (1 + nr_cpus_node(node)) *
3880 cachep->batchcount + cachep->num;
3881 cachep->node[node] = n;
3886 if (!cachep->list.next) {
3887 /* Cache is not active yet. Roll back what we did */
3890 if (cachep->node[node]) {
3891 n = cachep->node[node];
3894 free_alien_cache(n->alien);
3896 cachep->node[node] = NULL;
3904 struct ccupdate_struct {
3905 struct kmem_cache *cachep;
3906 struct array_cache *new[0];
3909 static void do_ccupdate_local(void *info)
3911 struct ccupdate_struct *new = info;
3912 struct array_cache *old;
3915 old = cpu_cache_get(new->cachep);
3917 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3918 new->new[smp_processor_id()] = old;
3921 /* Always called with the slab_mutex held */
3922 static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
3923 int batchcount, int shared, gfp_t gfp)
3925 struct ccupdate_struct *new;
3928 new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *),
3933 for_each_online_cpu(i) {
3934 new->new[i] = alloc_arraycache(cpu_to_mem(i), limit,
3937 for (i--; i >= 0; i--)
3943 new->cachep = cachep;
3945 on_each_cpu(do_ccupdate_local, (void *)new, 1);
3948 cachep->batchcount = batchcount;
3949 cachep->limit = limit;
3950 cachep->shared = shared;
3952 for_each_online_cpu(i) {
3953 struct array_cache *ccold = new->new[i];
3956 spin_lock_irq(&cachep->node[cpu_to_mem(i)]->list_lock);
3957 free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i));
3958 spin_unlock_irq(&cachep->node[cpu_to_mem(i)]->list_lock);
3962 return alloc_kmemlist(cachep, gfp);
3965 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3966 int batchcount, int shared, gfp_t gfp)
3969 struct kmem_cache *c = NULL;
3972 ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3974 if (slab_state < FULL)
3977 if ((ret < 0) || !is_root_cache(cachep))
3980 VM_BUG_ON(!mutex_is_locked(&slab_mutex));
3981 for_each_memcg_cache_index(i) {
3982 c = cache_from_memcg(cachep, i);
3984 /* return value determined by the parent cache only */
3985 __do_tune_cpucache(c, limit, batchcount, shared, gfp);
3991 /* Called with slab_mutex held always */
3992 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
3999 if (!is_root_cache(cachep)) {
4000 struct kmem_cache *root = memcg_root_cache(cachep);
4001 limit = root->limit;
4002 shared = root->shared;
4003 batchcount = root->batchcount;
4006 if (limit && shared && batchcount)
4009 * The head array serves three purposes:
4010 * - create a LIFO ordering, i.e. return objects that are cache-warm
4011 * - reduce the number of spinlock operations.
4012 * - reduce the number of linked list operations on the slab and
4013 * bufctl chains: array operations are cheaper.
4014 * The numbers are guessed, we should auto-tune as described by
4017 if (cachep->size > 131072)
4019 else if (cachep->size > PAGE_SIZE)
4021 else if (cachep->size > 1024)
4023 else if (cachep->size > 256)
4029 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
4030 * allocation behaviour: Most allocs on one cpu, most free operations
4031 * on another cpu. For these cases, an efficient object passing between
4032 * cpus is necessary. This is provided by a shared array. The array
4033 * replaces Bonwick's magazine layer.
4034 * On uniprocessor, it's functionally equivalent (but less efficient)
4035 * to a larger limit. Thus disabled by default.
4038 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
4043 * With debugging enabled, large batchcount lead to excessively long
4044 * periods with disabled local interrupts. Limit the batchcount
4049 batchcount = (limit + 1) / 2;
4051 err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
4053 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
4054 cachep->name, -err);
4059 * Drain an array if it contains any elements taking the node lock only if
4060 * necessary. Note that the node listlock also protects the array_cache
4061 * if drain_array() is used on the shared array.
4063 static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
4064 struct array_cache *ac, int force, int node)
4068 if (!ac || !ac->avail)
4070 if (ac->touched && !force) {
4073 spin_lock_irq(&n->list_lock);
4075 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4076 if (tofree > ac->avail)
4077 tofree = (ac->avail + 1) / 2;
4078 free_block(cachep, ac->entry, tofree, node);
4079 ac->avail -= tofree;
4080 memmove(ac->entry, &(ac->entry[tofree]),
4081 sizeof(void *) * ac->avail);
4083 spin_unlock_irq(&n->list_lock);
4088 * cache_reap - Reclaim memory from caches.
4089 * @w: work descriptor
4091 * Called from workqueue/eventd every few seconds.
4093 * - clear the per-cpu caches for this CPU.
4094 * - return freeable pages to the main free memory pool.
4096 * If we cannot acquire the cache chain mutex then just give up - we'll try
4097 * again on the next iteration.
4099 static void cache_reap(struct work_struct *w)
4101 struct kmem_cache *searchp;
4102 struct kmem_cache_node *n;
4103 int node = numa_mem_id();
4104 struct delayed_work *work = to_delayed_work(w);
4106 if (!mutex_trylock(&slab_mutex))
4107 /* Give up. Setup the next iteration. */
4110 list_for_each_entry(searchp, &slab_caches, list) {
4114 * We only take the node lock if absolutely necessary and we
4115 * have established with reasonable certainty that
4116 * we can do some work if the lock was obtained.
4118 n = searchp->node[node];
4120 reap_alien(searchp, n);
4122 drain_array(searchp, n, cpu_cache_get(searchp), 0, node);
4125 * These are racy checks but it does not matter
4126 * if we skip one check or scan twice.
4128 if (time_after(n->next_reap, jiffies))
4131 n->next_reap = jiffies + REAPTIMEOUT_LIST3;
4133 drain_array(searchp, n, n->shared, 0, node);
4135 if (n->free_touched)
4136 n->free_touched = 0;
4140 freed = drain_freelist(searchp, n, (n->free_limit +
4141 5 * searchp->num - 1) / (5 * searchp->num));
4142 STATS_ADD_REAPED(searchp, freed);
4148 mutex_unlock(&slab_mutex);
4151 /* Set up the next iteration */
4152 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
4155 #ifdef CONFIG_SLABINFO
4156 void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
4159 unsigned long active_objs;
4160 unsigned long num_objs;
4161 unsigned long active_slabs = 0;
4162 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4166 struct kmem_cache_node *n;
4170 for_each_online_node(node) {
4171 n = cachep->node[node];
4176 spin_lock_irq(&n->list_lock);
4178 list_for_each_entry(slabp, &n->slabs_full, list) {
4179 if (slabp->inuse != cachep->num && !error)
4180 error = "slabs_full accounting error";
4181 active_objs += cachep->num;
4184 list_for_each_entry(slabp, &n->slabs_partial, list) {
4185 if (slabp->inuse == cachep->num && !error)
4186 error = "slabs_partial inuse accounting error";
4187 if (!slabp->inuse && !error)
4188 error = "slabs_partial/inuse accounting error";
4189 active_objs += slabp->inuse;
4192 list_for_each_entry(slabp, &n->slabs_free, list) {
4193 if (slabp->inuse && !error)
4194 error = "slabs_free/inuse accounting error";
4197 free_objects += n->free_objects;
4199 shared_avail += n->shared->avail;
4201 spin_unlock_irq(&n->list_lock);
4203 num_slabs += active_slabs;
4204 num_objs = num_slabs * cachep->num;
4205 if (num_objs - active_objs != free_objects && !error)
4206 error = "free_objects accounting error";
4208 name = cachep->name;
4210 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4212 sinfo->active_objs = active_objs;
4213 sinfo->num_objs = num_objs;
4214 sinfo->active_slabs = active_slabs;
4215 sinfo->num_slabs = num_slabs;
4216 sinfo->shared_avail = shared_avail;
4217 sinfo->limit = cachep->limit;
4218 sinfo->batchcount = cachep->batchcount;
4219 sinfo->shared = cachep->shared;
4220 sinfo->objects_per_slab = cachep->num;
4221 sinfo->cache_order = cachep->gfporder;
4224 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
4228 unsigned long high = cachep->high_mark;
4229 unsigned long allocs = cachep->num_allocations;
4230 unsigned long grown = cachep->grown;
4231 unsigned long reaped = cachep->reaped;
4232 unsigned long errors = cachep->errors;
4233 unsigned long max_freeable = cachep->max_freeable;
4234 unsigned long node_allocs = cachep->node_allocs;
4235 unsigned long node_frees = cachep->node_frees;
4236 unsigned long overflows = cachep->node_overflow;
4238 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4239 "%4lu %4lu %4lu %4lu %4lu",
4240 allocs, high, grown,
4241 reaped, errors, max_freeable, node_allocs,
4242 node_frees, overflows);
4246 unsigned long allochit = atomic_read(&cachep->allochit);
4247 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4248 unsigned long freehit = atomic_read(&cachep->freehit);
4249 unsigned long freemiss = atomic_read(&cachep->freemiss);
4251 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4252 allochit, allocmiss, freehit, freemiss);
4257 #define MAX_SLABINFO_WRITE 128
4259 * slabinfo_write - Tuning for the slab allocator
4261 * @buffer: user buffer
4262 * @count: data length
4265 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4266 size_t count, loff_t *ppos)
4268 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4269 int limit, batchcount, shared, res;
4270 struct kmem_cache *cachep;
4272 if (count > MAX_SLABINFO_WRITE)
4274 if (copy_from_user(&kbuf, buffer, count))
4276 kbuf[MAX_SLABINFO_WRITE] = '\0';
4278 tmp = strchr(kbuf, ' ');
4283 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4286 /* Find the cache in the chain of caches. */
4287 mutex_lock(&slab_mutex);
4289 list_for_each_entry(cachep, &slab_caches, list) {
4290 if (!strcmp(cachep->name, kbuf)) {
4291 if (limit < 1 || batchcount < 1 ||
4292 batchcount > limit || shared < 0) {
4295 res = do_tune_cpucache(cachep, limit,
4302 mutex_unlock(&slab_mutex);
4308 #ifdef CONFIG_DEBUG_SLAB_LEAK
4310 static void *leaks_start(struct seq_file *m, loff_t *pos)
4312 mutex_lock(&slab_mutex);
4313 return seq_list_start(&slab_caches, *pos);
4316 static inline int add_caller(unsigned long *n, unsigned long v)
4326 unsigned long *q = p + 2 * i;
4340 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4346 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4352 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->size) {
4353 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4355 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4360 static void show_symbol(struct seq_file *m, unsigned long address)
4362 #ifdef CONFIG_KALLSYMS
4363 unsigned long offset, size;
4364 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4366 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4367 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4369 seq_printf(m, " [%s]", modname);
4373 seq_printf(m, "%p", (void *)address);
4376 static int leaks_show(struct seq_file *m, void *p)
4378 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
4380 struct kmem_cache_node *n;
4382 unsigned long *x = m->private;
4386 if (!(cachep->flags & SLAB_STORE_USER))
4388 if (!(cachep->flags & SLAB_RED_ZONE))
4391 /* OK, we can do it */
4395 for_each_online_node(node) {
4396 n = cachep->node[node];
4401 spin_lock_irq(&n->list_lock);
4403 list_for_each_entry(slabp, &n->slabs_full, list)
4404 handle_slab(x, cachep, slabp);
4405 list_for_each_entry(slabp, &n->slabs_partial, list)
4406 handle_slab(x, cachep, slabp);
4407 spin_unlock_irq(&n->list_lock);
4409 name = cachep->name;
4411 /* Increase the buffer size */
4412 mutex_unlock(&slab_mutex);
4413 m->private = kzalloc(x[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4415 /* Too bad, we are really out */
4417 mutex_lock(&slab_mutex);
4420 *(unsigned long *)m->private = x[0] * 2;
4422 mutex_lock(&slab_mutex);
4423 /* Now make sure this entry will be retried */
4427 for (i = 0; i < x[1]; i++) {
4428 seq_printf(m, "%s: %lu ", name, x[2*i+3]);
4429 show_symbol(m, x[2*i+2]);
4436 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4438 return seq_list_next(p, &slab_caches, pos);
4441 static void s_stop(struct seq_file *m, void *p)
4443 mutex_unlock(&slab_mutex);
4446 static const struct seq_operations slabstats_op = {
4447 .start = leaks_start,
4453 static int slabstats_open(struct inode *inode, struct file *file)
4455 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4458 ret = seq_open(file, &slabstats_op);
4460 struct seq_file *m = file->private_data;
4461 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4470 static const struct file_operations proc_slabstats_operations = {
4471 .open = slabstats_open,
4473 .llseek = seq_lseek,
4474 .release = seq_release_private,
4478 static int __init slab_proc_init(void)
4480 #ifdef CONFIG_DEBUG_SLAB_LEAK
4481 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4485 module_init(slab_proc_init);
4489 * ksize - get the actual amount of memory allocated for a given object
4490 * @objp: Pointer to the object
4492 * kmalloc may internally round up allocations and return more memory
4493 * than requested. ksize() can be used to determine the actual amount of
4494 * memory allocated. The caller may use this additional memory, even though
4495 * a smaller amount of memory was initially specified with the kmalloc call.
4496 * The caller must guarantee that objp points to a valid object previously
4497 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4498 * must not be freed during the duration of the call.
4500 size_t ksize(const void *objp)
4503 if (unlikely(objp == ZERO_SIZE_PTR))
4506 return virt_to_cache(objp)->object_size;
4508 EXPORT_SYMBOL(ksize);