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 void *s_mem; /* including colour offset */
223 unsigned int inuse; /* num of objs active in slab */
225 unsigned short nodeid;
227 struct slab_rcu __slab_cover_slab_rcu;
235 * - LIFO ordering, to hand out cache-warm objects from _alloc
236 * - reduce the number of linked list operations
237 * - reduce spinlock operations
239 * The limit is stored in the per-cpu structure to reduce the data cache
246 unsigned int batchcount;
247 unsigned int touched;
250 * Must have this definition in here for the proper
251 * alignment of array_cache. Also simplifies accessing
254 * Entries should not be directly dereferenced as
255 * entries belonging to slabs marked pfmemalloc will
256 * have the lower bits set SLAB_OBJ_PFMEMALLOC
260 #define SLAB_OBJ_PFMEMALLOC 1
261 static inline bool is_obj_pfmemalloc(void *objp)
263 return (unsigned long)objp & SLAB_OBJ_PFMEMALLOC;
266 static inline void set_obj_pfmemalloc(void **objp)
268 *objp = (void *)((unsigned long)*objp | SLAB_OBJ_PFMEMALLOC);
272 static inline void clear_obj_pfmemalloc(void **objp)
274 *objp = (void *)((unsigned long)*objp & ~SLAB_OBJ_PFMEMALLOC);
278 * bootstrap: The caches do not work without cpuarrays anymore, but the
279 * cpuarrays are allocated from the generic caches...
281 #define BOOT_CPUCACHE_ENTRIES 1
282 struct arraycache_init {
283 struct array_cache cache;
284 void *entries[BOOT_CPUCACHE_ENTRIES];
288 * Need this for bootstrapping a per node allocator.
290 #define NUM_INIT_LISTS (3 * MAX_NUMNODES)
291 static struct kmem_cache_node __initdata init_kmem_cache_node[NUM_INIT_LISTS];
292 #define CACHE_CACHE 0
293 #define SIZE_AC MAX_NUMNODES
294 #define SIZE_NODE (2 * MAX_NUMNODES)
296 static int drain_freelist(struct kmem_cache *cache,
297 struct kmem_cache_node *n, int tofree);
298 static void free_block(struct kmem_cache *cachep, void **objpp, int len,
300 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
301 static void cache_reap(struct work_struct *unused);
303 static int slab_early_init = 1;
305 #define INDEX_AC kmalloc_index(sizeof(struct arraycache_init))
306 #define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
308 static void kmem_cache_node_init(struct kmem_cache_node *parent)
310 INIT_LIST_HEAD(&parent->slabs_full);
311 INIT_LIST_HEAD(&parent->slabs_partial);
312 INIT_LIST_HEAD(&parent->slabs_free);
313 parent->shared = NULL;
314 parent->alien = NULL;
315 parent->colour_next = 0;
316 spin_lock_init(&parent->list_lock);
317 parent->free_objects = 0;
318 parent->free_touched = 0;
321 #define MAKE_LIST(cachep, listp, slab, nodeid) \
323 INIT_LIST_HEAD(listp); \
324 list_splice(&(cachep->node[nodeid]->slab), listp); \
327 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
329 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
330 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
331 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
334 #define CFLGS_OFF_SLAB (0x80000000UL)
335 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
337 #define BATCHREFILL_LIMIT 16
339 * Optimization question: fewer reaps means less probability for unnessary
340 * cpucache drain/refill cycles.
342 * OTOH the cpuarrays can contain lots of objects,
343 * which could lock up otherwise freeable slabs.
345 #define REAPTIMEOUT_CPUC (2*HZ)
346 #define REAPTIMEOUT_LIST3 (4*HZ)
349 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
350 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
351 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
352 #define STATS_INC_GROWN(x) ((x)->grown++)
353 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
354 #define STATS_SET_HIGH(x) \
356 if ((x)->num_active > (x)->high_mark) \
357 (x)->high_mark = (x)->num_active; \
359 #define STATS_INC_ERR(x) ((x)->errors++)
360 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
361 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
362 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
363 #define STATS_SET_FREEABLE(x, i) \
365 if ((x)->max_freeable < i) \
366 (x)->max_freeable = i; \
368 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
369 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
370 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
371 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
373 #define STATS_INC_ACTIVE(x) do { } while (0)
374 #define STATS_DEC_ACTIVE(x) do { } while (0)
375 #define STATS_INC_ALLOCED(x) do { } while (0)
376 #define STATS_INC_GROWN(x) do { } while (0)
377 #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
378 #define STATS_SET_HIGH(x) do { } while (0)
379 #define STATS_INC_ERR(x) do { } while (0)
380 #define STATS_INC_NODEALLOCS(x) do { } while (0)
381 #define STATS_INC_NODEFREES(x) do { } while (0)
382 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
383 #define STATS_SET_FREEABLE(x, i) do { } while (0)
384 #define STATS_INC_ALLOCHIT(x) do { } while (0)
385 #define STATS_INC_ALLOCMISS(x) do { } while (0)
386 #define STATS_INC_FREEHIT(x) do { } while (0)
387 #define STATS_INC_FREEMISS(x) do { } while (0)
393 * memory layout of objects:
395 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
396 * the end of an object is aligned with the end of the real
397 * allocation. Catches writes behind the end of the allocation.
398 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
400 * cachep->obj_offset: The real object.
401 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
402 * cachep->size - 1* BYTES_PER_WORD: last caller address
403 * [BYTES_PER_WORD long]
405 static int obj_offset(struct kmem_cache *cachep)
407 return cachep->obj_offset;
410 static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
412 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
413 return (unsigned long long*) (objp + obj_offset(cachep) -
414 sizeof(unsigned long long));
417 static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
419 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
420 if (cachep->flags & SLAB_STORE_USER)
421 return (unsigned long long *)(objp + cachep->size -
422 sizeof(unsigned long long) -
424 return (unsigned long long *) (objp + cachep->size -
425 sizeof(unsigned long long));
428 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
430 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
431 return (void **)(objp + cachep->size - BYTES_PER_WORD);
436 #define obj_offset(x) 0
437 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
438 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
439 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
444 * Do not go above this order unless 0 objects fit into the slab or
445 * overridden on the command line.
447 #define SLAB_MAX_ORDER_HI 1
448 #define SLAB_MAX_ORDER_LO 0
449 static int slab_max_order = SLAB_MAX_ORDER_LO;
450 static bool slab_max_order_set __initdata;
452 static inline struct kmem_cache *virt_to_cache(const void *obj)
454 struct page *page = virt_to_head_page(obj);
455 return page->slab_cache;
458 static inline struct slab *virt_to_slab(const void *obj)
460 struct page *page = virt_to_head_page(obj);
462 VM_BUG_ON(!PageSlab(page));
463 return page->slab_page;
466 static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab,
469 return slab->s_mem + cache->size * idx;
473 * We want to avoid an expensive divide : (offset / cache->size)
474 * Using the fact that size is a constant for a particular cache,
475 * we can replace (offset / cache->size) by
476 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
478 static inline unsigned int obj_to_index(const struct kmem_cache *cache,
479 const struct slab *slab, void *obj)
481 u32 offset = (obj - slab->s_mem);
482 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
485 static struct arraycache_init initarray_generic =
486 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
488 /* internal cache of cache description objs */
489 static struct kmem_cache kmem_cache_boot = {
491 .limit = BOOT_CPUCACHE_ENTRIES,
493 .size = sizeof(struct kmem_cache),
494 .name = "kmem_cache",
497 #define BAD_ALIEN_MAGIC 0x01020304ul
499 #ifdef CONFIG_LOCKDEP
502 * Slab sometimes uses the kmalloc slabs to store the slab headers
503 * for other slabs "off slab".
504 * The locking for this is tricky in that it nests within the locks
505 * of all other slabs in a few places; to deal with this special
506 * locking we put on-slab caches into a separate lock-class.
508 * We set lock class for alien array caches which are up during init.
509 * The lock annotation will be lost if all cpus of a node goes down and
510 * then comes back up during hotplug
512 static struct lock_class_key on_slab_l3_key;
513 static struct lock_class_key on_slab_alc_key;
515 static struct lock_class_key debugobj_l3_key;
516 static struct lock_class_key debugobj_alc_key;
518 static void slab_set_lock_classes(struct kmem_cache *cachep,
519 struct lock_class_key *l3_key, struct lock_class_key *alc_key,
522 struct array_cache **alc;
523 struct kmem_cache_node *n;
530 lockdep_set_class(&n->list_lock, l3_key);
533 * FIXME: This check for BAD_ALIEN_MAGIC
534 * should go away when common slab code is taught to
535 * work even without alien caches.
536 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
537 * for alloc_alien_cache,
539 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
543 lockdep_set_class(&alc[r]->lock, alc_key);
547 static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
549 slab_set_lock_classes(cachep, &debugobj_l3_key, &debugobj_alc_key, node);
552 static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
556 for_each_online_node(node)
557 slab_set_debugobj_lock_classes_node(cachep, node);
560 static void init_node_lock_keys(int q)
567 for (i = 1; i <= KMALLOC_SHIFT_HIGH; i++) {
568 struct kmem_cache_node *n;
569 struct kmem_cache *cache = kmalloc_caches[i];
575 if (!n || OFF_SLAB(cache))
578 slab_set_lock_classes(cache, &on_slab_l3_key,
579 &on_slab_alc_key, q);
583 static void on_slab_lock_classes_node(struct kmem_cache *cachep, int q)
585 if (!cachep->node[q])
588 slab_set_lock_classes(cachep, &on_slab_l3_key,
589 &on_slab_alc_key, q);
592 static inline void on_slab_lock_classes(struct kmem_cache *cachep)
596 VM_BUG_ON(OFF_SLAB(cachep));
598 on_slab_lock_classes_node(cachep, node);
601 static inline void init_lock_keys(void)
606 init_node_lock_keys(node);
609 static void init_node_lock_keys(int q)
613 static inline void init_lock_keys(void)
617 static inline void on_slab_lock_classes(struct kmem_cache *cachep)
621 static inline void on_slab_lock_classes_node(struct kmem_cache *cachep, int node)
625 static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
629 static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
634 static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
636 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
638 return cachep->array[smp_processor_id()];
641 static size_t slab_mgmt_size(size_t nr_objs, size_t align)
643 return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align);
647 * Calculate the number of objects and left-over bytes for a given buffer size.
649 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
650 size_t align, int flags, size_t *left_over,
655 size_t slab_size = PAGE_SIZE << gfporder;
658 * The slab management structure can be either off the slab or
659 * on it. For the latter case, the memory allocated for a
663 * - One kmem_bufctl_t for each object
664 * - Padding to respect alignment of @align
665 * - @buffer_size bytes for each object
667 * If the slab management structure is off the slab, then the
668 * alignment will already be calculated into the size. Because
669 * the slabs are all pages aligned, the objects will be at the
670 * correct alignment when allocated.
672 if (flags & CFLGS_OFF_SLAB) {
674 nr_objs = slab_size / buffer_size;
676 if (nr_objs > SLAB_LIMIT)
677 nr_objs = SLAB_LIMIT;
680 * Ignore padding for the initial guess. The padding
681 * is at most @align-1 bytes, and @buffer_size is at
682 * least @align. In the worst case, this result will
683 * be one greater than the number of objects that fit
684 * into the memory allocation when taking the padding
687 nr_objs = (slab_size - sizeof(struct slab)) /
688 (buffer_size + sizeof(kmem_bufctl_t));
691 * This calculated number will be either the right
692 * amount, or one greater than what we want.
694 if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size
698 if (nr_objs > SLAB_LIMIT)
699 nr_objs = SLAB_LIMIT;
701 mgmt_size = slab_mgmt_size(nr_objs, align);
704 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
708 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
710 static void __slab_error(const char *function, struct kmem_cache *cachep,
713 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
714 function, cachep->name, msg);
716 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
721 * By default on NUMA we use alien caches to stage the freeing of
722 * objects allocated from other nodes. This causes massive memory
723 * inefficiencies when using fake NUMA setup to split memory into a
724 * large number of small nodes, so it can be disabled on the command
728 static int use_alien_caches __read_mostly = 1;
729 static int __init noaliencache_setup(char *s)
731 use_alien_caches = 0;
734 __setup("noaliencache", noaliencache_setup);
736 static int __init slab_max_order_setup(char *str)
738 get_option(&str, &slab_max_order);
739 slab_max_order = slab_max_order < 0 ? 0 :
740 min(slab_max_order, MAX_ORDER - 1);
741 slab_max_order_set = true;
745 __setup("slab_max_order=", slab_max_order_setup);
749 * Special reaping functions for NUMA systems called from cache_reap().
750 * These take care of doing round robin flushing of alien caches (containing
751 * objects freed on different nodes from which they were allocated) and the
752 * flushing of remote pcps by calling drain_node_pages.
754 static DEFINE_PER_CPU(unsigned long, slab_reap_node);
756 static void init_reap_node(int cpu)
760 node = next_node(cpu_to_mem(cpu), node_online_map);
761 if (node == MAX_NUMNODES)
762 node = first_node(node_online_map);
764 per_cpu(slab_reap_node, cpu) = node;
767 static void next_reap_node(void)
769 int node = __this_cpu_read(slab_reap_node);
771 node = next_node(node, node_online_map);
772 if (unlikely(node >= MAX_NUMNODES))
773 node = first_node(node_online_map);
774 __this_cpu_write(slab_reap_node, node);
778 #define init_reap_node(cpu) do { } while (0)
779 #define next_reap_node(void) do { } while (0)
783 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
784 * via the workqueue/eventd.
785 * Add the CPU number into the expiration time to minimize the possibility of
786 * the CPUs getting into lockstep and contending for the global cache chain
789 static void start_cpu_timer(int cpu)
791 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
794 * When this gets called from do_initcalls via cpucache_init(),
795 * init_workqueues() has already run, so keventd will be setup
798 if (keventd_up() && reap_work->work.func == NULL) {
800 INIT_DEFERRABLE_WORK(reap_work, cache_reap);
801 schedule_delayed_work_on(cpu, reap_work,
802 __round_jiffies_relative(HZ, cpu));
806 static struct array_cache *alloc_arraycache(int node, int entries,
807 int batchcount, gfp_t gfp)
809 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
810 struct array_cache *nc = NULL;
812 nc = kmalloc_node(memsize, gfp, node);
814 * The array_cache structures contain pointers to free object.
815 * However, when such objects are allocated or transferred to another
816 * cache the pointers are not cleared and they could be counted as
817 * valid references during a kmemleak scan. Therefore, kmemleak must
818 * not scan such objects.
820 kmemleak_no_scan(nc);
824 nc->batchcount = batchcount;
826 spin_lock_init(&nc->lock);
831 static inline bool is_slab_pfmemalloc(struct slab *slabp)
833 struct page *page = virt_to_page(slabp->s_mem);
835 return PageSlabPfmemalloc(page);
838 /* Clears pfmemalloc_active if no slabs have pfmalloc set */
839 static void recheck_pfmemalloc_active(struct kmem_cache *cachep,
840 struct array_cache *ac)
842 struct kmem_cache_node *n = cachep->node[numa_mem_id()];
846 if (!pfmemalloc_active)
849 spin_lock_irqsave(&n->list_lock, flags);
850 list_for_each_entry(slabp, &n->slabs_full, list)
851 if (is_slab_pfmemalloc(slabp))
854 list_for_each_entry(slabp, &n->slabs_partial, list)
855 if (is_slab_pfmemalloc(slabp))
858 list_for_each_entry(slabp, &n->slabs_free, list)
859 if (is_slab_pfmemalloc(slabp))
862 pfmemalloc_active = false;
864 spin_unlock_irqrestore(&n->list_lock, flags);
867 static void *__ac_get_obj(struct kmem_cache *cachep, struct array_cache *ac,
868 gfp_t flags, bool force_refill)
871 void *objp = ac->entry[--ac->avail];
873 /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
874 if (unlikely(is_obj_pfmemalloc(objp))) {
875 struct kmem_cache_node *n;
877 if (gfp_pfmemalloc_allowed(flags)) {
878 clear_obj_pfmemalloc(&objp);
882 /* The caller cannot use PFMEMALLOC objects, find another one */
883 for (i = 0; i < ac->avail; i++) {
884 /* If a !PFMEMALLOC object is found, swap them */
885 if (!is_obj_pfmemalloc(ac->entry[i])) {
887 ac->entry[i] = ac->entry[ac->avail];
888 ac->entry[ac->avail] = objp;
894 * If there are empty slabs on the slabs_free list and we are
895 * being forced to refill the cache, mark this one !pfmemalloc.
897 n = cachep->node[numa_mem_id()];
898 if (!list_empty(&n->slabs_free) && force_refill) {
899 struct slab *slabp = virt_to_slab(objp);
900 ClearPageSlabPfmemalloc(virt_to_head_page(slabp->s_mem));
901 clear_obj_pfmemalloc(&objp);
902 recheck_pfmemalloc_active(cachep, ac);
906 /* No !PFMEMALLOC objects available */
914 static inline void *ac_get_obj(struct kmem_cache *cachep,
915 struct array_cache *ac, gfp_t flags, bool force_refill)
919 if (unlikely(sk_memalloc_socks()))
920 objp = __ac_get_obj(cachep, ac, flags, force_refill);
922 objp = ac->entry[--ac->avail];
927 static void *__ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
930 if (unlikely(pfmemalloc_active)) {
931 /* Some pfmemalloc slabs exist, check if this is one */
932 struct slab *slabp = virt_to_slab(objp);
933 struct page *page = virt_to_head_page(slabp->s_mem);
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 inline int slabs_tofree(struct kmem_cache *cachep,
1184 struct kmem_cache_node *n)
1186 return (n->free_objects + cachep->num - 1) / cachep->num;
1189 static void cpuup_canceled(long cpu)
1191 struct kmem_cache *cachep;
1192 struct kmem_cache_node *n = NULL;
1193 int node = cpu_to_mem(cpu);
1194 const struct cpumask *mask = cpumask_of_node(node);
1196 list_for_each_entry(cachep, &slab_caches, list) {
1197 struct array_cache *nc;
1198 struct array_cache *shared;
1199 struct array_cache **alien;
1201 /* cpu is dead; no one can alloc from it. */
1202 nc = cachep->array[cpu];
1203 cachep->array[cpu] = NULL;
1204 n = cachep->node[node];
1207 goto free_array_cache;
1209 spin_lock_irq(&n->list_lock);
1211 /* Free limit for this kmem_cache_node */
1212 n->free_limit -= cachep->batchcount;
1214 free_block(cachep, nc->entry, nc->avail, node);
1216 if (!cpumask_empty(mask)) {
1217 spin_unlock_irq(&n->list_lock);
1218 goto free_array_cache;
1223 free_block(cachep, shared->entry,
1224 shared->avail, node);
1231 spin_unlock_irq(&n->list_lock);
1235 drain_alien_cache(cachep, alien);
1236 free_alien_cache(alien);
1242 * In the previous loop, all the objects were freed to
1243 * the respective cache's slabs, now we can go ahead and
1244 * shrink each nodelist to its limit.
1246 list_for_each_entry(cachep, &slab_caches, list) {
1247 n = cachep->node[node];
1250 drain_freelist(cachep, n, slabs_tofree(cachep, n));
1254 static int cpuup_prepare(long cpu)
1256 struct kmem_cache *cachep;
1257 struct kmem_cache_node *n = NULL;
1258 int node = cpu_to_mem(cpu);
1262 * We need to do this right in the beginning since
1263 * alloc_arraycache's are going to use this list.
1264 * kmalloc_node allows us to add the slab to the right
1265 * kmem_cache_node and not this cpu's kmem_cache_node
1267 err = init_cache_node_node(node);
1272 * Now we can go ahead with allocating the shared arrays and
1275 list_for_each_entry(cachep, &slab_caches, list) {
1276 struct array_cache *nc;
1277 struct array_cache *shared = NULL;
1278 struct array_cache **alien = NULL;
1280 nc = alloc_arraycache(node, cachep->limit,
1281 cachep->batchcount, GFP_KERNEL);
1284 if (cachep->shared) {
1285 shared = alloc_arraycache(node,
1286 cachep->shared * cachep->batchcount,
1287 0xbaadf00d, GFP_KERNEL);
1293 if (use_alien_caches) {
1294 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
1301 cachep->array[cpu] = nc;
1302 n = cachep->node[node];
1305 spin_lock_irq(&n->list_lock);
1308 * We are serialised from CPU_DEAD or
1309 * CPU_UP_CANCELLED by the cpucontrol lock
1320 spin_unlock_irq(&n->list_lock);
1322 free_alien_cache(alien);
1323 if (cachep->flags & SLAB_DEBUG_OBJECTS)
1324 slab_set_debugobj_lock_classes_node(cachep, node);
1325 else if (!OFF_SLAB(cachep) &&
1326 !(cachep->flags & SLAB_DESTROY_BY_RCU))
1327 on_slab_lock_classes_node(cachep, node);
1329 init_node_lock_keys(node);
1333 cpuup_canceled(cpu);
1337 static int cpuup_callback(struct notifier_block *nfb,
1338 unsigned long action, void *hcpu)
1340 long cpu = (long)hcpu;
1344 case CPU_UP_PREPARE:
1345 case CPU_UP_PREPARE_FROZEN:
1346 mutex_lock(&slab_mutex);
1347 err = cpuup_prepare(cpu);
1348 mutex_unlock(&slab_mutex);
1351 case CPU_ONLINE_FROZEN:
1352 start_cpu_timer(cpu);
1354 #ifdef CONFIG_HOTPLUG_CPU
1355 case CPU_DOWN_PREPARE:
1356 case CPU_DOWN_PREPARE_FROZEN:
1358 * Shutdown cache reaper. Note that the slab_mutex is
1359 * held so that if cache_reap() is invoked it cannot do
1360 * anything expensive but will only modify reap_work
1361 * and reschedule the timer.
1363 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1364 /* Now the cache_reaper is guaranteed to be not running. */
1365 per_cpu(slab_reap_work, cpu).work.func = NULL;
1367 case CPU_DOWN_FAILED:
1368 case CPU_DOWN_FAILED_FROZEN:
1369 start_cpu_timer(cpu);
1372 case CPU_DEAD_FROZEN:
1374 * Even if all the cpus of a node are down, we don't free the
1375 * kmem_cache_node of any cache. This to avoid a race between
1376 * cpu_down, and a kmalloc allocation from another cpu for
1377 * memory from the node of the cpu going down. The node
1378 * structure is usually allocated from kmem_cache_create() and
1379 * gets destroyed at kmem_cache_destroy().
1383 case CPU_UP_CANCELED:
1384 case CPU_UP_CANCELED_FROZEN:
1385 mutex_lock(&slab_mutex);
1386 cpuup_canceled(cpu);
1387 mutex_unlock(&slab_mutex);
1390 return notifier_from_errno(err);
1393 static struct notifier_block cpucache_notifier = {
1394 &cpuup_callback, NULL, 0
1397 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1399 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1400 * Returns -EBUSY if all objects cannot be drained so that the node is not
1403 * Must hold slab_mutex.
1405 static int __meminit drain_cache_node_node(int node)
1407 struct kmem_cache *cachep;
1410 list_for_each_entry(cachep, &slab_caches, list) {
1411 struct kmem_cache_node *n;
1413 n = cachep->node[node];
1417 drain_freelist(cachep, n, slabs_tofree(cachep, n));
1419 if (!list_empty(&n->slabs_full) ||
1420 !list_empty(&n->slabs_partial)) {
1428 static int __meminit slab_memory_callback(struct notifier_block *self,
1429 unsigned long action, void *arg)
1431 struct memory_notify *mnb = arg;
1435 nid = mnb->status_change_nid;
1440 case MEM_GOING_ONLINE:
1441 mutex_lock(&slab_mutex);
1442 ret = init_cache_node_node(nid);
1443 mutex_unlock(&slab_mutex);
1445 case MEM_GOING_OFFLINE:
1446 mutex_lock(&slab_mutex);
1447 ret = drain_cache_node_node(nid);
1448 mutex_unlock(&slab_mutex);
1452 case MEM_CANCEL_ONLINE:
1453 case MEM_CANCEL_OFFLINE:
1457 return notifier_from_errno(ret);
1459 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1462 * swap the static kmem_cache_node with kmalloced memory
1464 static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list,
1467 struct kmem_cache_node *ptr;
1469 ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid);
1472 memcpy(ptr, list, sizeof(struct kmem_cache_node));
1474 * Do not assume that spinlocks can be initialized via memcpy:
1476 spin_lock_init(&ptr->list_lock);
1478 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1479 cachep->node[nodeid] = ptr;
1483 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1484 * size of kmem_cache_node.
1486 static void __init set_up_node(struct kmem_cache *cachep, int index)
1490 for_each_online_node(node) {
1491 cachep->node[node] = &init_kmem_cache_node[index + node];
1492 cachep->node[node]->next_reap = jiffies +
1494 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
1499 * The memory after the last cpu cache pointer is used for the
1502 static void setup_node_pointer(struct kmem_cache *cachep)
1504 cachep->node = (struct kmem_cache_node **)&cachep->array[nr_cpu_ids];
1508 * Initialisation. Called after the page allocator have been initialised and
1509 * before smp_init().
1511 void __init kmem_cache_init(void)
1515 kmem_cache = &kmem_cache_boot;
1516 setup_node_pointer(kmem_cache);
1518 if (num_possible_nodes() == 1)
1519 use_alien_caches = 0;
1521 for (i = 0; i < NUM_INIT_LISTS; i++)
1522 kmem_cache_node_init(&init_kmem_cache_node[i]);
1524 set_up_node(kmem_cache, CACHE_CACHE);
1527 * Fragmentation resistance on low memory - only use bigger
1528 * page orders on machines with more than 32MB of memory if
1529 * not overridden on the command line.
1531 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
1532 slab_max_order = SLAB_MAX_ORDER_HI;
1534 /* Bootstrap is tricky, because several objects are allocated
1535 * from caches that do not exist yet:
1536 * 1) initialize the kmem_cache cache: it contains the struct
1537 * kmem_cache structures of all caches, except kmem_cache itself:
1538 * kmem_cache is statically allocated.
1539 * Initially an __init data area is used for the head array and the
1540 * kmem_cache_node structures, it's replaced with a kmalloc allocated
1541 * array at the end of the bootstrap.
1542 * 2) Create the first kmalloc cache.
1543 * The struct kmem_cache for the new cache is allocated normally.
1544 * An __init data area is used for the head array.
1545 * 3) Create the remaining kmalloc caches, with minimally sized
1547 * 4) Replace the __init data head arrays for kmem_cache and the first
1548 * kmalloc cache with kmalloc allocated arrays.
1549 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1550 * the other cache's with kmalloc allocated memory.
1551 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1554 /* 1) create the kmem_cache */
1557 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1559 create_boot_cache(kmem_cache, "kmem_cache",
1560 offsetof(struct kmem_cache, array[nr_cpu_ids]) +
1561 nr_node_ids * sizeof(struct kmem_cache_node *),
1562 SLAB_HWCACHE_ALIGN);
1563 list_add(&kmem_cache->list, &slab_caches);
1565 /* 2+3) create the kmalloc caches */
1568 * Initialize the caches that provide memory for the array cache and the
1569 * kmem_cache_node structures first. Without this, further allocations will
1573 kmalloc_caches[INDEX_AC] = create_kmalloc_cache("kmalloc-ac",
1574 kmalloc_size(INDEX_AC), ARCH_KMALLOC_FLAGS);
1576 if (INDEX_AC != INDEX_NODE)
1577 kmalloc_caches[INDEX_NODE] =
1578 create_kmalloc_cache("kmalloc-node",
1579 kmalloc_size(INDEX_NODE), ARCH_KMALLOC_FLAGS);
1581 slab_early_init = 0;
1583 /* 4) Replace the bootstrap head arrays */
1585 struct array_cache *ptr;
1587 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1589 memcpy(ptr, cpu_cache_get(kmem_cache),
1590 sizeof(struct arraycache_init));
1592 * Do not assume that spinlocks can be initialized via memcpy:
1594 spin_lock_init(&ptr->lock);
1596 kmem_cache->array[smp_processor_id()] = ptr;
1598 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1600 BUG_ON(cpu_cache_get(kmalloc_caches[INDEX_AC])
1601 != &initarray_generic.cache);
1602 memcpy(ptr, cpu_cache_get(kmalloc_caches[INDEX_AC]),
1603 sizeof(struct arraycache_init));
1605 * Do not assume that spinlocks can be initialized via memcpy:
1607 spin_lock_init(&ptr->lock);
1609 kmalloc_caches[INDEX_AC]->array[smp_processor_id()] = ptr;
1611 /* 5) Replace the bootstrap kmem_cache_node */
1615 for_each_online_node(nid) {
1616 init_list(kmem_cache, &init_kmem_cache_node[CACHE_CACHE + nid], nid);
1618 init_list(kmalloc_caches[INDEX_AC],
1619 &init_kmem_cache_node[SIZE_AC + nid], nid);
1621 if (INDEX_AC != INDEX_NODE) {
1622 init_list(kmalloc_caches[INDEX_NODE],
1623 &init_kmem_cache_node[SIZE_NODE + nid], nid);
1628 create_kmalloc_caches(ARCH_KMALLOC_FLAGS);
1631 void __init kmem_cache_init_late(void)
1633 struct kmem_cache *cachep;
1637 /* 6) resize the head arrays to their final sizes */
1638 mutex_lock(&slab_mutex);
1639 list_for_each_entry(cachep, &slab_caches, list)
1640 if (enable_cpucache(cachep, GFP_NOWAIT))
1642 mutex_unlock(&slab_mutex);
1644 /* Annotate slab for lockdep -- annotate the malloc caches */
1651 * Register a cpu startup notifier callback that initializes
1652 * cpu_cache_get for all new cpus
1654 register_cpu_notifier(&cpucache_notifier);
1658 * Register a memory hotplug callback that initializes and frees
1661 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1665 * The reap timers are started later, with a module init call: That part
1666 * of the kernel is not yet operational.
1670 static int __init cpucache_init(void)
1675 * Register the timers that return unneeded pages to the page allocator
1677 for_each_online_cpu(cpu)
1678 start_cpu_timer(cpu);
1684 __initcall(cpucache_init);
1686 static noinline void
1687 slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1689 struct kmem_cache_node *n;
1691 unsigned long flags;
1695 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1697 printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n",
1698 cachep->name, cachep->size, cachep->gfporder);
1700 for_each_online_node(node) {
1701 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1702 unsigned long active_slabs = 0, num_slabs = 0;
1704 n = cachep->node[node];
1708 spin_lock_irqsave(&n->list_lock, flags);
1709 list_for_each_entry(slabp, &n->slabs_full, list) {
1710 active_objs += cachep->num;
1713 list_for_each_entry(slabp, &n->slabs_partial, list) {
1714 active_objs += slabp->inuse;
1717 list_for_each_entry(slabp, &n->slabs_free, list)
1720 free_objects += n->free_objects;
1721 spin_unlock_irqrestore(&n->list_lock, flags);
1723 num_slabs += active_slabs;
1724 num_objs = num_slabs * cachep->num;
1726 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1727 node, active_slabs, num_slabs, active_objs, num_objs,
1733 * Interface to system's page allocator. No need to hold the cache-lock.
1735 * If we requested dmaable memory, we will get it. Even if we
1736 * did not request dmaable memory, we might get it, but that
1737 * would be relatively rare and ignorable.
1739 static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags,
1748 * Nommu uses slab's for process anonymous memory allocations, and thus
1749 * requires __GFP_COMP to properly refcount higher order allocations
1751 flags |= __GFP_COMP;
1754 flags |= cachep->allocflags;
1755 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1756 flags |= __GFP_RECLAIMABLE;
1758 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1760 if (!(flags & __GFP_NOWARN) && printk_ratelimit())
1761 slab_out_of_memory(cachep, flags, nodeid);
1765 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1766 if (unlikely(page->pfmemalloc))
1767 pfmemalloc_active = true;
1769 nr_pages = (1 << cachep->gfporder);
1770 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1771 add_zone_page_state(page_zone(page),
1772 NR_SLAB_RECLAIMABLE, nr_pages);
1774 add_zone_page_state(page_zone(page),
1775 NR_SLAB_UNRECLAIMABLE, nr_pages);
1776 for (i = 0; i < nr_pages; i++) {
1777 __SetPageSlab(page + i);
1779 if (page->pfmemalloc)
1780 SetPageSlabPfmemalloc(page);
1782 memcg_bind_pages(cachep, cachep->gfporder);
1784 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1785 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1788 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1790 kmemcheck_mark_unallocated_pages(page, nr_pages);
1797 * Interface to system's page release.
1799 static void kmem_freepages(struct kmem_cache *cachep, struct page *page)
1801 unsigned long i = (1 << cachep->gfporder);
1802 const unsigned long nr_freed = i;
1804 kmemcheck_free_shadow(page, cachep->gfporder);
1806 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1807 sub_zone_page_state(page_zone(page),
1808 NR_SLAB_RECLAIMABLE, nr_freed);
1810 sub_zone_page_state(page_zone(page),
1811 NR_SLAB_UNRECLAIMABLE, nr_freed);
1813 __ClearPageSlabPfmemalloc(page);
1815 BUG_ON(!PageSlab(page));
1816 __ClearPageSlab(page);
1820 memcg_release_pages(cachep, cachep->gfporder);
1821 if (current->reclaim_state)
1822 current->reclaim_state->reclaimed_slab += nr_freed;
1823 __free_memcg_kmem_pages(page, cachep->gfporder);
1826 static void kmem_rcu_free(struct rcu_head *head)
1828 struct slab_rcu *slab_rcu = (struct slab_rcu *)head;
1829 struct kmem_cache *cachep = slab_rcu->cachep;
1831 kmem_freepages(cachep, slab_rcu->page);
1832 if (OFF_SLAB(cachep))
1833 kmem_cache_free(cachep->slabp_cache, slab_rcu);
1838 #ifdef CONFIG_DEBUG_PAGEALLOC
1839 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1840 unsigned long caller)
1842 int size = cachep->object_size;
1844 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1846 if (size < 5 * sizeof(unsigned long))
1849 *addr++ = 0x12345678;
1851 *addr++ = smp_processor_id();
1852 size -= 3 * sizeof(unsigned long);
1854 unsigned long *sptr = &caller;
1855 unsigned long svalue;
1857 while (!kstack_end(sptr)) {
1859 if (kernel_text_address(svalue)) {
1861 size -= sizeof(unsigned long);
1862 if (size <= sizeof(unsigned long))
1868 *addr++ = 0x87654321;
1872 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1874 int size = cachep->object_size;
1875 addr = &((char *)addr)[obj_offset(cachep)];
1877 memset(addr, val, size);
1878 *(unsigned char *)(addr + size - 1) = POISON_END;
1881 static void dump_line(char *data, int offset, int limit)
1884 unsigned char error = 0;
1887 printk(KERN_ERR "%03x: ", offset);
1888 for (i = 0; i < limit; i++) {
1889 if (data[offset + i] != POISON_FREE) {
1890 error = data[offset + i];
1894 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1895 &data[offset], limit, 1);
1897 if (bad_count == 1) {
1898 error ^= POISON_FREE;
1899 if (!(error & (error - 1))) {
1900 printk(KERN_ERR "Single bit error detected. Probably "
1903 printk(KERN_ERR "Run memtest86+ or a similar memory "
1906 printk(KERN_ERR "Run a memory test tool.\n");
1915 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1920 if (cachep->flags & SLAB_RED_ZONE) {
1921 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
1922 *dbg_redzone1(cachep, objp),
1923 *dbg_redzone2(cachep, objp));
1926 if (cachep->flags & SLAB_STORE_USER) {
1927 printk(KERN_ERR "Last user: [<%p>](%pSR)\n",
1928 *dbg_userword(cachep, objp),
1929 *dbg_userword(cachep, objp));
1931 realobj = (char *)objp + obj_offset(cachep);
1932 size = cachep->object_size;
1933 for (i = 0; i < size && lines; i += 16, lines--) {
1936 if (i + limit > size)
1938 dump_line(realobj, i, limit);
1942 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1948 realobj = (char *)objp + obj_offset(cachep);
1949 size = cachep->object_size;
1951 for (i = 0; i < size; i++) {
1952 char exp = POISON_FREE;
1955 if (realobj[i] != exp) {
1961 "Slab corruption (%s): %s start=%p, len=%d\n",
1962 print_tainted(), cachep->name, realobj, size);
1963 print_objinfo(cachep, objp, 0);
1965 /* Hexdump the affected line */
1968 if (i + limit > size)
1970 dump_line(realobj, i, limit);
1973 /* Limit to 5 lines */
1979 /* Print some data about the neighboring objects, if they
1982 struct slab *slabp = virt_to_slab(objp);
1985 objnr = obj_to_index(cachep, slabp, objp);
1987 objp = index_to_obj(cachep, slabp, objnr - 1);
1988 realobj = (char *)objp + obj_offset(cachep);
1989 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1991 print_objinfo(cachep, objp, 2);
1993 if (objnr + 1 < cachep->num) {
1994 objp = index_to_obj(cachep, slabp, objnr + 1);
1995 realobj = (char *)objp + obj_offset(cachep);
1996 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1998 print_objinfo(cachep, objp, 2);
2005 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
2008 for (i = 0; i < cachep->num; i++) {
2009 void *objp = index_to_obj(cachep, slabp, i);
2011 if (cachep->flags & SLAB_POISON) {
2012 #ifdef CONFIG_DEBUG_PAGEALLOC
2013 if (cachep->size % PAGE_SIZE == 0 &&
2015 kernel_map_pages(virt_to_page(objp),
2016 cachep->size / PAGE_SIZE, 1);
2018 check_poison_obj(cachep, objp);
2020 check_poison_obj(cachep, objp);
2023 if (cachep->flags & SLAB_RED_ZONE) {
2024 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2025 slab_error(cachep, "start of a freed object "
2027 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2028 slab_error(cachep, "end of a freed object "
2034 static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp)
2040 * slab_destroy - destroy and release all objects in a slab
2041 * @cachep: cache pointer being destroyed
2042 * @slabp: slab pointer being destroyed
2044 * Destroy all the objs in a slab, and release the mem back to the system.
2045 * Before calling the slab must have been unlinked from the cache. The
2046 * cache-lock is not held/needed.
2048 static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp)
2050 struct page *page = virt_to_head_page(slabp->s_mem);
2052 slab_destroy_debugcheck(cachep, slabp);
2053 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
2054 struct slab_rcu *slab_rcu;
2056 slab_rcu = (struct slab_rcu *)slabp;
2057 slab_rcu->cachep = cachep;
2058 slab_rcu->page = page;
2059 call_rcu(&slab_rcu->head, kmem_rcu_free);
2061 kmem_freepages(cachep, page);
2062 if (OFF_SLAB(cachep))
2063 kmem_cache_free(cachep->slabp_cache, slabp);
2068 * calculate_slab_order - calculate size (page order) of slabs
2069 * @cachep: pointer to the cache that is being created
2070 * @size: size of objects to be created in this cache.
2071 * @align: required alignment for the objects.
2072 * @flags: slab allocation flags
2074 * Also calculates the number of objects per slab.
2076 * This could be made much more intelligent. For now, try to avoid using
2077 * high order pages for slabs. When the gfp() functions are more friendly
2078 * towards high-order requests, this should be changed.
2080 static size_t calculate_slab_order(struct kmem_cache *cachep,
2081 size_t size, size_t align, unsigned long flags)
2083 unsigned long offslab_limit;
2084 size_t left_over = 0;
2087 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
2091 cache_estimate(gfporder, size, align, flags, &remainder, &num);
2095 if (flags & CFLGS_OFF_SLAB) {
2097 * Max number of objs-per-slab for caches which
2098 * use off-slab slabs. Needed to avoid a possible
2099 * looping condition in cache_grow().
2101 offslab_limit = size - sizeof(struct slab);
2102 offslab_limit /= sizeof(kmem_bufctl_t);
2104 if (num > offslab_limit)
2108 /* Found something acceptable - save it away */
2110 cachep->gfporder = gfporder;
2111 left_over = remainder;
2114 * A VFS-reclaimable slab tends to have most allocations
2115 * as GFP_NOFS and we really don't want to have to be allocating
2116 * higher-order pages when we are unable to shrink dcache.
2118 if (flags & SLAB_RECLAIM_ACCOUNT)
2122 * Large number of objects is good, but very large slabs are
2123 * currently bad for the gfp()s.
2125 if (gfporder >= slab_max_order)
2129 * Acceptable internal fragmentation?
2131 if (left_over * 8 <= (PAGE_SIZE << gfporder))
2137 static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
2139 if (slab_state >= FULL)
2140 return enable_cpucache(cachep, gfp);
2142 if (slab_state == DOWN) {
2144 * Note: Creation of first cache (kmem_cache).
2145 * The setup_node is taken care
2146 * of by the caller of __kmem_cache_create
2148 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2149 slab_state = PARTIAL;
2150 } else if (slab_state == PARTIAL) {
2152 * Note: the second kmem_cache_create must create the cache
2153 * that's used by kmalloc(24), otherwise the creation of
2154 * further caches will BUG().
2156 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2159 * If the cache that's used by kmalloc(sizeof(kmem_cache_node)) is
2160 * the second cache, then we need to set up all its node/,
2161 * otherwise the creation of further caches will BUG().
2163 set_up_node(cachep, SIZE_AC);
2164 if (INDEX_AC == INDEX_NODE)
2165 slab_state = PARTIAL_NODE;
2167 slab_state = PARTIAL_ARRAYCACHE;
2169 /* Remaining boot caches */
2170 cachep->array[smp_processor_id()] =
2171 kmalloc(sizeof(struct arraycache_init), gfp);
2173 if (slab_state == PARTIAL_ARRAYCACHE) {
2174 set_up_node(cachep, SIZE_NODE);
2175 slab_state = PARTIAL_NODE;
2178 for_each_online_node(node) {
2179 cachep->node[node] =
2180 kmalloc_node(sizeof(struct kmem_cache_node),
2182 BUG_ON(!cachep->node[node]);
2183 kmem_cache_node_init(cachep->node[node]);
2187 cachep->node[numa_mem_id()]->next_reap =
2188 jiffies + REAPTIMEOUT_LIST3 +
2189 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
2191 cpu_cache_get(cachep)->avail = 0;
2192 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2193 cpu_cache_get(cachep)->batchcount = 1;
2194 cpu_cache_get(cachep)->touched = 0;
2195 cachep->batchcount = 1;
2196 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2201 * __kmem_cache_create - Create a cache.
2202 * @cachep: cache management descriptor
2203 * @flags: SLAB flags
2205 * Returns a ptr to the cache on success, NULL on failure.
2206 * Cannot be called within a int, but can be interrupted.
2207 * The @ctor is run when new pages are allocated by the cache.
2211 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2212 * to catch references to uninitialised memory.
2214 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2215 * for buffer overruns.
2217 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2218 * cacheline. This can be beneficial if you're counting cycles as closely
2222 __kmem_cache_create (struct kmem_cache *cachep, unsigned long flags)
2224 size_t left_over, slab_size, ralign;
2227 size_t size = cachep->size;
2232 * Enable redzoning and last user accounting, except for caches with
2233 * large objects, if the increased size would increase the object size
2234 * above the next power of two: caches with object sizes just above a
2235 * power of two have a significant amount of internal fragmentation.
2237 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2238 2 * sizeof(unsigned long long)))
2239 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2240 if (!(flags & SLAB_DESTROY_BY_RCU))
2241 flags |= SLAB_POISON;
2243 if (flags & SLAB_DESTROY_BY_RCU)
2244 BUG_ON(flags & SLAB_POISON);
2248 * Check that size is in terms of words. This is needed to avoid
2249 * unaligned accesses for some archs when redzoning is used, and makes
2250 * sure any on-slab bufctl's are also correctly aligned.
2252 if (size & (BYTES_PER_WORD - 1)) {
2253 size += (BYTES_PER_WORD - 1);
2254 size &= ~(BYTES_PER_WORD - 1);
2258 * Redzoning and user store require word alignment or possibly larger.
2259 * Note this will be overridden by architecture or caller mandated
2260 * alignment if either is greater than BYTES_PER_WORD.
2262 if (flags & SLAB_STORE_USER)
2263 ralign = BYTES_PER_WORD;
2265 if (flags & SLAB_RED_ZONE) {
2266 ralign = REDZONE_ALIGN;
2267 /* If redzoning, ensure that the second redzone is suitably
2268 * aligned, by adjusting the object size accordingly. */
2269 size += REDZONE_ALIGN - 1;
2270 size &= ~(REDZONE_ALIGN - 1);
2273 /* 3) caller mandated alignment */
2274 if (ralign < cachep->align) {
2275 ralign = cachep->align;
2277 /* disable debug if necessary */
2278 if (ralign > __alignof__(unsigned long long))
2279 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2283 cachep->align = ralign;
2285 if (slab_is_available())
2290 setup_node_pointer(cachep);
2294 * Both debugging options require word-alignment which is calculated
2297 if (flags & SLAB_RED_ZONE) {
2298 /* add space for red zone words */
2299 cachep->obj_offset += sizeof(unsigned long long);
2300 size += 2 * sizeof(unsigned long long);
2302 if (flags & SLAB_STORE_USER) {
2303 /* user store requires one word storage behind the end of
2304 * the real object. But if the second red zone needs to be
2305 * aligned to 64 bits, we must allow that much space.
2307 if (flags & SLAB_RED_ZONE)
2308 size += REDZONE_ALIGN;
2310 size += BYTES_PER_WORD;
2312 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2313 if (size >= kmalloc_size(INDEX_NODE + 1)
2314 && cachep->object_size > cache_line_size()
2315 && ALIGN(size, cachep->align) < PAGE_SIZE) {
2316 cachep->obj_offset += PAGE_SIZE - ALIGN(size, cachep->align);
2323 * Determine if the slab management is 'on' or 'off' slab.
2324 * (bootstrapping cannot cope with offslab caches so don't do
2325 * it too early on. Always use on-slab management when
2326 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2328 if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init &&
2329 !(flags & SLAB_NOLEAKTRACE))
2331 * Size is large, assume best to place the slab management obj
2332 * off-slab (should allow better packing of objs).
2334 flags |= CFLGS_OFF_SLAB;
2336 size = ALIGN(size, cachep->align);
2338 left_over = calculate_slab_order(cachep, size, cachep->align, flags);
2343 slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t)
2344 + sizeof(struct slab), cachep->align);
2347 * If the slab has been placed off-slab, and we have enough space then
2348 * move it on-slab. This is at the expense of any extra colouring.
2350 if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) {
2351 flags &= ~CFLGS_OFF_SLAB;
2352 left_over -= slab_size;
2355 if (flags & CFLGS_OFF_SLAB) {
2356 /* really off slab. No need for manual alignment */
2358 cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab);
2360 #ifdef CONFIG_PAGE_POISONING
2361 /* If we're going to use the generic kernel_map_pages()
2362 * poisoning, then it's going to smash the contents of
2363 * the redzone and userword anyhow, so switch them off.
2365 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2366 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2370 cachep->colour_off = cache_line_size();
2371 /* Offset must be a multiple of the alignment. */
2372 if (cachep->colour_off < cachep->align)
2373 cachep->colour_off = cachep->align;
2374 cachep->colour = left_over / cachep->colour_off;
2375 cachep->slab_size = slab_size;
2376 cachep->flags = flags;
2377 cachep->allocflags = 0;
2378 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2379 cachep->allocflags |= GFP_DMA;
2380 cachep->size = size;
2381 cachep->reciprocal_buffer_size = reciprocal_value(size);
2383 if (flags & CFLGS_OFF_SLAB) {
2384 cachep->slabp_cache = kmalloc_slab(slab_size, 0u);
2386 * This is a possibility for one of the malloc_sizes caches.
2387 * But since we go off slab only for object size greater than
2388 * PAGE_SIZE/8, and malloc_sizes gets created in ascending order,
2389 * this should not happen at all.
2390 * But leave a BUG_ON for some lucky dude.
2392 BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache));
2395 err = setup_cpu_cache(cachep, gfp);
2397 __kmem_cache_shutdown(cachep);
2401 if (flags & SLAB_DEBUG_OBJECTS) {
2403 * Would deadlock through slab_destroy()->call_rcu()->
2404 * debug_object_activate()->kmem_cache_alloc().
2406 WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU);
2408 slab_set_debugobj_lock_classes(cachep);
2409 } else if (!OFF_SLAB(cachep) && !(flags & SLAB_DESTROY_BY_RCU))
2410 on_slab_lock_classes(cachep);
2416 static void check_irq_off(void)
2418 BUG_ON(!irqs_disabled());
2421 static void check_irq_on(void)
2423 BUG_ON(irqs_disabled());
2426 static void check_spinlock_acquired(struct kmem_cache *cachep)
2430 assert_spin_locked(&cachep->node[numa_mem_id()]->list_lock);
2434 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2438 assert_spin_locked(&cachep->node[node]->list_lock);
2443 #define check_irq_off() do { } while(0)
2444 #define check_irq_on() do { } while(0)
2445 #define check_spinlock_acquired(x) do { } while(0)
2446 #define check_spinlock_acquired_node(x, y) do { } while(0)
2449 static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
2450 struct array_cache *ac,
2451 int force, int node);
2453 static void do_drain(void *arg)
2455 struct kmem_cache *cachep = arg;
2456 struct array_cache *ac;
2457 int node = numa_mem_id();
2460 ac = cpu_cache_get(cachep);
2461 spin_lock(&cachep->node[node]->list_lock);
2462 free_block(cachep, ac->entry, ac->avail, node);
2463 spin_unlock(&cachep->node[node]->list_lock);
2467 static void drain_cpu_caches(struct kmem_cache *cachep)
2469 struct kmem_cache_node *n;
2472 on_each_cpu(do_drain, cachep, 1);
2474 for_each_online_node(node) {
2475 n = cachep->node[node];
2477 drain_alien_cache(cachep, n->alien);
2480 for_each_online_node(node) {
2481 n = cachep->node[node];
2483 drain_array(cachep, n, n->shared, 1, node);
2488 * Remove slabs from the list of free slabs.
2489 * Specify the number of slabs to drain in tofree.
2491 * Returns the actual number of slabs released.
2493 static int drain_freelist(struct kmem_cache *cache,
2494 struct kmem_cache_node *n, int tofree)
2496 struct list_head *p;
2501 while (nr_freed < tofree && !list_empty(&n->slabs_free)) {
2503 spin_lock_irq(&n->list_lock);
2504 p = n->slabs_free.prev;
2505 if (p == &n->slabs_free) {
2506 spin_unlock_irq(&n->list_lock);
2510 slabp = list_entry(p, struct slab, list);
2512 BUG_ON(slabp->inuse);
2514 list_del(&slabp->list);
2516 * Safe to drop the lock. The slab is no longer linked
2519 n->free_objects -= cache->num;
2520 spin_unlock_irq(&n->list_lock);
2521 slab_destroy(cache, slabp);
2528 /* Called with slab_mutex held to protect against cpu hotplug */
2529 static int __cache_shrink(struct kmem_cache *cachep)
2532 struct kmem_cache_node *n;
2534 drain_cpu_caches(cachep);
2537 for_each_online_node(i) {
2538 n = cachep->node[i];
2542 drain_freelist(cachep, n, slabs_tofree(cachep, n));
2544 ret += !list_empty(&n->slabs_full) ||
2545 !list_empty(&n->slabs_partial);
2547 return (ret ? 1 : 0);
2551 * kmem_cache_shrink - Shrink a cache.
2552 * @cachep: The cache to shrink.
2554 * Releases as many slabs as possible for a cache.
2555 * To help debugging, a zero exit status indicates all slabs were released.
2557 int kmem_cache_shrink(struct kmem_cache *cachep)
2560 BUG_ON(!cachep || in_interrupt());
2563 mutex_lock(&slab_mutex);
2564 ret = __cache_shrink(cachep);
2565 mutex_unlock(&slab_mutex);
2569 EXPORT_SYMBOL(kmem_cache_shrink);
2571 int __kmem_cache_shutdown(struct kmem_cache *cachep)
2574 struct kmem_cache_node *n;
2575 int rc = __cache_shrink(cachep);
2580 for_each_online_cpu(i)
2581 kfree(cachep->array[i]);
2583 /* NUMA: free the node structures */
2584 for_each_online_node(i) {
2585 n = cachep->node[i];
2588 free_alien_cache(n->alien);
2596 * Get the memory for a slab management obj.
2597 * For a slab cache when the slab descriptor is off-slab, slab descriptors
2598 * always come from malloc_sizes caches. The slab descriptor cannot
2599 * come from the same cache which is getting created because,
2600 * when we are searching for an appropriate cache for these
2601 * descriptors in kmem_cache_create, we search through the malloc_sizes array.
2602 * If we are creating a malloc_sizes cache here it would not be visible to
2603 * kmem_find_general_cachep till the initialization is complete.
2604 * Hence we cannot have slabp_cache same as the original cache.
2606 static struct slab *alloc_slabmgmt(struct kmem_cache *cachep,
2607 struct page *page, int colour_off,
2608 gfp_t local_flags, int nodeid)
2611 void *addr = page_address(page);
2613 if (OFF_SLAB(cachep)) {
2614 /* Slab management obj is off-slab. */
2615 slabp = kmem_cache_alloc_node(cachep->slabp_cache,
2616 local_flags, nodeid);
2618 * If the first object in the slab is leaked (it's allocated
2619 * but no one has a reference to it), we want to make sure
2620 * kmemleak does not treat the ->s_mem pointer as a reference
2621 * to the object. Otherwise we will not report the leak.
2623 kmemleak_scan_area(&slabp->list, sizeof(struct list_head),
2628 slabp = addr + colour_off;
2629 colour_off += cachep->slab_size;
2632 slabp->s_mem = addr + colour_off;
2633 slabp->nodeid = nodeid;
2638 static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp)
2640 return (kmem_bufctl_t *) (slabp + 1);
2643 static void cache_init_objs(struct kmem_cache *cachep,
2648 for (i = 0; i < cachep->num; i++) {
2649 void *objp = index_to_obj(cachep, slabp, i);
2651 /* need to poison the objs? */
2652 if (cachep->flags & SLAB_POISON)
2653 poison_obj(cachep, objp, POISON_FREE);
2654 if (cachep->flags & SLAB_STORE_USER)
2655 *dbg_userword(cachep, objp) = NULL;
2657 if (cachep->flags & SLAB_RED_ZONE) {
2658 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2659 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2662 * Constructors are not allowed to allocate memory from the same
2663 * cache which they are a constructor for. Otherwise, deadlock.
2664 * They must also be threaded.
2666 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2667 cachep->ctor(objp + obj_offset(cachep));
2669 if (cachep->flags & SLAB_RED_ZONE) {
2670 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2671 slab_error(cachep, "constructor overwrote the"
2672 " end of an object");
2673 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2674 slab_error(cachep, "constructor overwrote the"
2675 " start of an object");
2677 if ((cachep->size % PAGE_SIZE) == 0 &&
2678 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2679 kernel_map_pages(virt_to_page(objp),
2680 cachep->size / PAGE_SIZE, 0);
2685 slab_bufctl(slabp)[i] = i + 1;
2687 slab_bufctl(slabp)[i - 1] = BUFCTL_END;
2690 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2692 if (CONFIG_ZONE_DMA_FLAG) {
2693 if (flags & GFP_DMA)
2694 BUG_ON(!(cachep->allocflags & GFP_DMA));
2696 BUG_ON(cachep->allocflags & GFP_DMA);
2700 static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp,
2703 void *objp = index_to_obj(cachep, slabp, slabp->free);
2707 next = slab_bufctl(slabp)[slabp->free];
2709 slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE;
2710 WARN_ON(slabp->nodeid != nodeid);
2717 static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp,
2718 void *objp, int nodeid)
2720 unsigned int objnr = obj_to_index(cachep, slabp, objp);
2723 /* Verify that the slab belongs to the intended node */
2724 WARN_ON(slabp->nodeid != nodeid);
2726 if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) {
2727 printk(KERN_ERR "slab: double free detected in cache "
2728 "'%s', objp %p\n", cachep->name, objp);
2732 slab_bufctl(slabp)[objnr] = slabp->free;
2733 slabp->free = objnr;
2738 * Map pages beginning at addr to the given cache and slab. This is required
2739 * for the slab allocator to be able to lookup the cache and slab of a
2740 * virtual address for kfree, ksize, and slab debugging.
2742 static void slab_map_pages(struct kmem_cache *cache, struct slab *slab,
2748 if (likely(!PageCompound(page)))
2749 nr_pages <<= cache->gfporder;
2752 page->slab_cache = cache;
2753 page->slab_page = slab;
2755 } while (--nr_pages);
2759 * Grow (by 1) the number of slabs within a cache. This is called by
2760 * kmem_cache_alloc() when there are no active objs left in a cache.
2762 static int cache_grow(struct kmem_cache *cachep,
2763 gfp_t flags, int nodeid, struct page *page)
2768 struct kmem_cache_node *n;
2771 * Be lazy and only check for valid flags here, keeping it out of the
2772 * critical path in kmem_cache_alloc().
2774 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2775 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2777 /* Take the node list lock to change the colour_next on this node */
2779 n = cachep->node[nodeid];
2780 spin_lock(&n->list_lock);
2782 /* Get colour for the slab, and cal the next value. */
2783 offset = n->colour_next;
2785 if (n->colour_next >= cachep->colour)
2787 spin_unlock(&n->list_lock);
2789 offset *= cachep->colour_off;
2791 if (local_flags & __GFP_WAIT)
2795 * The test for missing atomic flag is performed here, rather than
2796 * the more obvious place, simply to reduce the critical path length
2797 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2798 * will eventually be caught here (where it matters).
2800 kmem_flagcheck(cachep, flags);
2803 * Get mem for the objs. Attempt to allocate a physical page from
2807 page = kmem_getpages(cachep, local_flags, nodeid);
2811 /* Get slab management. */
2812 slabp = alloc_slabmgmt(cachep, page, offset,
2813 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2817 slab_map_pages(cachep, slabp, page);
2819 cache_init_objs(cachep, slabp);
2821 if (local_flags & __GFP_WAIT)
2822 local_irq_disable();
2824 spin_lock(&n->list_lock);
2826 /* Make slab active. */
2827 list_add_tail(&slabp->list, &(n->slabs_free));
2828 STATS_INC_GROWN(cachep);
2829 n->free_objects += cachep->num;
2830 spin_unlock(&n->list_lock);
2833 kmem_freepages(cachep, page);
2835 if (local_flags & __GFP_WAIT)
2836 local_irq_disable();
2843 * Perform extra freeing checks:
2844 * - detect bad pointers.
2845 * - POISON/RED_ZONE checking
2847 static void kfree_debugcheck(const void *objp)
2849 if (!virt_addr_valid(objp)) {
2850 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2851 (unsigned long)objp);
2856 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2858 unsigned long long redzone1, redzone2;
2860 redzone1 = *dbg_redzone1(cache, obj);
2861 redzone2 = *dbg_redzone2(cache, obj);
2866 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2869 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2870 slab_error(cache, "double free detected");
2872 slab_error(cache, "memory outside object was overwritten");
2874 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2875 obj, redzone1, redzone2);
2878 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2879 unsigned long caller)
2885 BUG_ON(virt_to_cache(objp) != cachep);
2887 objp -= obj_offset(cachep);
2888 kfree_debugcheck(objp);
2889 page = virt_to_head_page(objp);
2891 slabp = page->slab_page;
2893 if (cachep->flags & SLAB_RED_ZONE) {
2894 verify_redzone_free(cachep, objp);
2895 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2896 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2898 if (cachep->flags & SLAB_STORE_USER)
2899 *dbg_userword(cachep, objp) = (void *)caller;
2901 objnr = obj_to_index(cachep, slabp, objp);
2903 BUG_ON(objnr >= cachep->num);
2904 BUG_ON(objp != index_to_obj(cachep, slabp, objnr));
2906 #ifdef CONFIG_DEBUG_SLAB_LEAK
2907 slab_bufctl(slabp)[objnr] = BUFCTL_FREE;
2909 if (cachep->flags & SLAB_POISON) {
2910 #ifdef CONFIG_DEBUG_PAGEALLOC
2911 if ((cachep->size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
2912 store_stackinfo(cachep, objp, caller);
2913 kernel_map_pages(virt_to_page(objp),
2914 cachep->size / PAGE_SIZE, 0);
2916 poison_obj(cachep, objp, POISON_FREE);
2919 poison_obj(cachep, objp, POISON_FREE);
2925 static void check_slabp(struct kmem_cache *cachep, struct slab *slabp)
2930 /* Check slab's freelist to see if this obj is there. */
2931 for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) {
2933 if (entries > cachep->num || i >= cachep->num)
2936 if (entries != cachep->num - slabp->inuse) {
2938 printk(KERN_ERR "slab: Internal list corruption detected in "
2939 "cache '%s'(%d), slabp %p(%d). Tainted(%s). Hexdump:\n",
2940 cachep->name, cachep->num, slabp, slabp->inuse,
2942 print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 16, 1, slabp,
2943 sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t),
2949 #define kfree_debugcheck(x) do { } while(0)
2950 #define cache_free_debugcheck(x,objp,z) (objp)
2951 #define check_slabp(x,y) do { } while(0)
2954 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags,
2958 struct kmem_cache_node *n;
2959 struct array_cache *ac;
2963 node = numa_mem_id();
2964 if (unlikely(force_refill))
2967 ac = cpu_cache_get(cachep);
2968 batchcount = ac->batchcount;
2969 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2971 * If there was little recent activity on this cache, then
2972 * perform only a partial refill. Otherwise we could generate
2975 batchcount = BATCHREFILL_LIMIT;
2977 n = cachep->node[node];
2979 BUG_ON(ac->avail > 0 || !n);
2980 spin_lock(&n->list_lock);
2982 /* See if we can refill from the shared array */
2983 if (n->shared && transfer_objects(ac, n->shared, batchcount)) {
2984 n->shared->touched = 1;
2988 while (batchcount > 0) {
2989 struct list_head *entry;
2991 /* Get slab alloc is to come from. */
2992 entry = n->slabs_partial.next;
2993 if (entry == &n->slabs_partial) {
2994 n->free_touched = 1;
2995 entry = n->slabs_free.next;
2996 if (entry == &n->slabs_free)
3000 slabp = list_entry(entry, struct slab, list);
3001 check_slabp(cachep, slabp);
3002 check_spinlock_acquired(cachep);
3005 * The slab was either on partial or free list so
3006 * there must be at least one object available for
3009 BUG_ON(slabp->inuse >= cachep->num);
3011 while (slabp->inuse < cachep->num && batchcount--) {
3012 STATS_INC_ALLOCED(cachep);
3013 STATS_INC_ACTIVE(cachep);
3014 STATS_SET_HIGH(cachep);
3016 ac_put_obj(cachep, ac, slab_get_obj(cachep, slabp,
3019 check_slabp(cachep, slabp);
3021 /* move slabp to correct slabp list: */
3022 list_del(&slabp->list);
3023 if (slabp->free == BUFCTL_END)
3024 list_add(&slabp->list, &n->slabs_full);
3026 list_add(&slabp->list, &n->slabs_partial);
3030 n->free_objects -= ac->avail;
3032 spin_unlock(&n->list_lock);
3034 if (unlikely(!ac->avail)) {
3037 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
3039 /* cache_grow can reenable interrupts, then ac could change. */
3040 ac = cpu_cache_get(cachep);
3041 node = numa_mem_id();
3043 /* no objects in sight? abort */
3044 if (!x && (ac->avail == 0 || force_refill))
3047 if (!ac->avail) /* objects refilled by interrupt? */
3052 return ac_get_obj(cachep, ac, flags, force_refill);
3055 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3058 might_sleep_if(flags & __GFP_WAIT);
3060 kmem_flagcheck(cachep, flags);
3065 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3066 gfp_t flags, void *objp, unsigned long caller)
3070 if (cachep->flags & SLAB_POISON) {
3071 #ifdef CONFIG_DEBUG_PAGEALLOC
3072 if ((cachep->size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
3073 kernel_map_pages(virt_to_page(objp),
3074 cachep->size / PAGE_SIZE, 1);
3076 check_poison_obj(cachep, objp);
3078 check_poison_obj(cachep, objp);
3080 poison_obj(cachep, objp, POISON_INUSE);
3082 if (cachep->flags & SLAB_STORE_USER)
3083 *dbg_userword(cachep, objp) = (void *)caller;
3085 if (cachep->flags & SLAB_RED_ZONE) {
3086 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3087 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3088 slab_error(cachep, "double free, or memory outside"
3089 " object was overwritten");
3091 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3092 objp, *dbg_redzone1(cachep, objp),
3093 *dbg_redzone2(cachep, objp));
3095 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3096 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3098 #ifdef CONFIG_DEBUG_SLAB_LEAK
3103 slabp = virt_to_head_page(objp)->slab_page;
3104 objnr = (unsigned)(objp - slabp->s_mem) / cachep->size;
3105 slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE;
3108 objp += obj_offset(cachep);
3109 if (cachep->ctor && cachep->flags & SLAB_POISON)
3111 if (ARCH_SLAB_MINALIGN &&
3112 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
3113 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3114 objp, (int)ARCH_SLAB_MINALIGN);
3119 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3122 static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
3124 if (cachep == kmem_cache)
3127 return should_failslab(cachep->object_size, flags, cachep->flags);
3130 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3133 struct array_cache *ac;
3134 bool force_refill = false;
3138 ac = cpu_cache_get(cachep);
3139 if (likely(ac->avail)) {
3141 objp = ac_get_obj(cachep, ac, flags, false);
3144 * Allow for the possibility all avail objects are not allowed
3145 * by the current flags
3148 STATS_INC_ALLOCHIT(cachep);
3151 force_refill = true;
3154 STATS_INC_ALLOCMISS(cachep);
3155 objp = cache_alloc_refill(cachep, flags, force_refill);
3157 * the 'ac' may be updated by cache_alloc_refill(),
3158 * and kmemleak_erase() requires its correct value.
3160 ac = cpu_cache_get(cachep);
3164 * To avoid a false negative, if an object that is in one of the
3165 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3166 * treat the array pointers as a reference to the object.
3169 kmemleak_erase(&ac->entry[ac->avail]);
3175 * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY.
3177 * If we are in_interrupt, then process context, including cpusets and
3178 * mempolicy, may not apply and should not be used for allocation policy.
3180 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3182 int nid_alloc, nid_here;
3184 if (in_interrupt() || (flags & __GFP_THISNODE))
3186 nid_alloc = nid_here = numa_mem_id();
3187 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3188 nid_alloc = cpuset_slab_spread_node();
3189 else if (current->mempolicy)
3190 nid_alloc = slab_node();
3191 if (nid_alloc != nid_here)
3192 return ____cache_alloc_node(cachep, flags, nid_alloc);
3197 * Fallback function if there was no memory available and no objects on a
3198 * certain node and fall back is permitted. First we scan all the
3199 * available node for available objects. If that fails then we
3200 * perform an allocation without specifying a node. This allows the page
3201 * allocator to do its reclaim / fallback magic. We then insert the
3202 * slab into the proper nodelist and then allocate from it.
3204 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3206 struct zonelist *zonelist;
3210 enum zone_type high_zoneidx = gfp_zone(flags);
3213 unsigned int cpuset_mems_cookie;
3215 if (flags & __GFP_THISNODE)
3218 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3221 cpuset_mems_cookie = get_mems_allowed();
3222 zonelist = node_zonelist(slab_node(), flags);
3226 * Look through allowed nodes for objects available
3227 * from existing per node queues.
3229 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3230 nid = zone_to_nid(zone);
3232 if (cpuset_zone_allowed_hardwall(zone, flags) &&
3234 cache->node[nid]->free_objects) {
3235 obj = ____cache_alloc_node(cache,
3236 flags | GFP_THISNODE, nid);
3244 * This allocation will be performed within the constraints
3245 * of the current cpuset / memory policy requirements.
3246 * We may trigger various forms of reclaim on the allowed
3247 * set and go into memory reserves if necessary.
3251 if (local_flags & __GFP_WAIT)
3253 kmem_flagcheck(cache, flags);
3254 page = kmem_getpages(cache, local_flags, numa_mem_id());
3255 if (local_flags & __GFP_WAIT)
3256 local_irq_disable();
3259 * Insert into the appropriate per node queues
3261 nid = page_to_nid(page);
3262 if (cache_grow(cache, flags, nid, page)) {
3263 obj = ____cache_alloc_node(cache,
3264 flags | GFP_THISNODE, nid);
3267 * Another processor may allocate the
3268 * objects in the slab since we are
3269 * not holding any locks.
3273 /* cache_grow already freed obj */
3279 if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !obj))
3285 * A interface to enable slab creation on nodeid
3287 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3290 struct list_head *entry;
3292 struct kmem_cache_node *n;
3296 VM_BUG_ON(nodeid > num_online_nodes());
3297 n = cachep->node[nodeid];
3302 spin_lock(&n->list_lock);
3303 entry = n->slabs_partial.next;
3304 if (entry == &n->slabs_partial) {
3305 n->free_touched = 1;
3306 entry = n->slabs_free.next;
3307 if (entry == &n->slabs_free)
3311 slabp = list_entry(entry, struct slab, list);
3312 check_spinlock_acquired_node(cachep, nodeid);
3313 check_slabp(cachep, slabp);
3315 STATS_INC_NODEALLOCS(cachep);
3316 STATS_INC_ACTIVE(cachep);
3317 STATS_SET_HIGH(cachep);
3319 BUG_ON(slabp->inuse == cachep->num);
3321 obj = slab_get_obj(cachep, slabp, nodeid);
3322 check_slabp(cachep, slabp);
3324 /* move slabp to correct slabp list: */
3325 list_del(&slabp->list);
3327 if (slabp->free == BUFCTL_END)
3328 list_add(&slabp->list, &n->slabs_full);
3330 list_add(&slabp->list, &n->slabs_partial);
3332 spin_unlock(&n->list_lock);
3336 spin_unlock(&n->list_lock);
3337 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
3341 return fallback_alloc(cachep, flags);
3347 static __always_inline void *
3348 slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3349 unsigned long caller)
3351 unsigned long save_flags;
3353 int slab_node = numa_mem_id();
3355 flags &= gfp_allowed_mask;
3357 lockdep_trace_alloc(flags);
3359 if (slab_should_failslab(cachep, flags))
3362 cachep = memcg_kmem_get_cache(cachep, flags);
3364 cache_alloc_debugcheck_before(cachep, flags);
3365 local_irq_save(save_flags);
3367 if (nodeid == NUMA_NO_NODE)
3370 if (unlikely(!cachep->node[nodeid])) {
3371 /* Node not bootstrapped yet */
3372 ptr = fallback_alloc(cachep, flags);
3376 if (nodeid == slab_node) {
3378 * Use the locally cached objects if possible.
3379 * However ____cache_alloc does not allow fallback
3380 * to other nodes. It may fail while we still have
3381 * objects on other nodes available.
3383 ptr = ____cache_alloc(cachep, flags);
3387 /* ___cache_alloc_node can fall back to other nodes */
3388 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3390 local_irq_restore(save_flags);
3391 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3392 kmemleak_alloc_recursive(ptr, cachep->object_size, 1, cachep->flags,
3396 kmemcheck_slab_alloc(cachep, flags, ptr, cachep->object_size);
3398 if (unlikely((flags & __GFP_ZERO) && ptr))
3399 memset(ptr, 0, cachep->object_size);
3404 static __always_inline void *
3405 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3409 if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) {
3410 objp = alternate_node_alloc(cache, flags);
3414 objp = ____cache_alloc(cache, flags);
3417 * We may just have run out of memory on the local node.
3418 * ____cache_alloc_node() knows how to locate memory on other nodes
3421 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3428 static __always_inline void *
3429 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3431 return ____cache_alloc(cachep, flags);
3434 #endif /* CONFIG_NUMA */
3436 static __always_inline void *
3437 slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
3439 unsigned long save_flags;
3442 flags &= gfp_allowed_mask;
3444 lockdep_trace_alloc(flags);
3446 if (slab_should_failslab(cachep, flags))
3449 cachep = memcg_kmem_get_cache(cachep, flags);
3451 cache_alloc_debugcheck_before(cachep, flags);
3452 local_irq_save(save_flags);
3453 objp = __do_cache_alloc(cachep, flags);
3454 local_irq_restore(save_flags);
3455 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3456 kmemleak_alloc_recursive(objp, cachep->object_size, 1, cachep->flags,
3461 kmemcheck_slab_alloc(cachep, flags, objp, cachep->object_size);
3463 if (unlikely((flags & __GFP_ZERO) && objp))
3464 memset(objp, 0, cachep->object_size);
3470 * Caller needs to acquire correct kmem_list's list_lock
3472 static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3476 struct kmem_cache_node *n;
3478 for (i = 0; i < nr_objects; i++) {
3482 clear_obj_pfmemalloc(&objpp[i]);
3485 slabp = virt_to_slab(objp);
3486 n = cachep->node[node];
3487 list_del(&slabp->list);
3488 check_spinlock_acquired_node(cachep, node);
3489 check_slabp(cachep, slabp);
3490 slab_put_obj(cachep, slabp, objp, node);
3491 STATS_DEC_ACTIVE(cachep);
3493 check_slabp(cachep, slabp);
3495 /* fixup slab chains */
3496 if (slabp->inuse == 0) {
3497 if (n->free_objects > n->free_limit) {
3498 n->free_objects -= cachep->num;
3499 /* No need to drop any previously held
3500 * lock here, even if we have a off-slab slab
3501 * descriptor it is guaranteed to come from
3502 * a different cache, refer to comments before
3505 slab_destroy(cachep, slabp);
3507 list_add(&slabp->list, &n->slabs_free);
3510 /* Unconditionally move a slab to the end of the
3511 * partial list on free - maximum time for the
3512 * other objects to be freed, too.
3514 list_add_tail(&slabp->list, &n->slabs_partial);
3519 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3522 struct kmem_cache_node *n;
3523 int node = numa_mem_id();
3525 batchcount = ac->batchcount;
3527 BUG_ON(!batchcount || batchcount > ac->avail);
3530 n = cachep->node[node];
3531 spin_lock(&n->list_lock);
3533 struct array_cache *shared_array = n->shared;
3534 int max = shared_array->limit - shared_array->avail;
3536 if (batchcount > max)
3538 memcpy(&(shared_array->entry[shared_array->avail]),
3539 ac->entry, sizeof(void *) * batchcount);
3540 shared_array->avail += batchcount;
3545 free_block(cachep, ac->entry, batchcount, node);
3550 struct list_head *p;
3552 p = n->slabs_free.next;
3553 while (p != &(n->slabs_free)) {
3556 slabp = list_entry(p, struct slab, list);
3557 BUG_ON(slabp->inuse);
3562 STATS_SET_FREEABLE(cachep, i);
3565 spin_unlock(&n->list_lock);
3566 ac->avail -= batchcount;
3567 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3571 * Release an obj back to its cache. If the obj has a constructed state, it must
3572 * be in this state _before_ it is released. Called with disabled ints.
3574 static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3575 unsigned long caller)
3577 struct array_cache *ac = cpu_cache_get(cachep);
3580 kmemleak_free_recursive(objp, cachep->flags);
3581 objp = cache_free_debugcheck(cachep, objp, caller);
3583 kmemcheck_slab_free(cachep, objp, cachep->object_size);
3586 * Skip calling cache_free_alien() when the platform is not numa.
3587 * This will avoid cache misses that happen while accessing slabp (which
3588 * is per page memory reference) to get nodeid. Instead use a global
3589 * variable to skip the call, which is mostly likely to be present in
3592 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3595 if (likely(ac->avail < ac->limit)) {
3596 STATS_INC_FREEHIT(cachep);
3598 STATS_INC_FREEMISS(cachep);
3599 cache_flusharray(cachep, ac);
3602 ac_put_obj(cachep, ac, objp);
3606 * kmem_cache_alloc - Allocate an object
3607 * @cachep: The cache to allocate from.
3608 * @flags: See kmalloc().
3610 * Allocate an object from this cache. The flags are only relevant
3611 * if the cache has no available objects.
3613 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3615 void *ret = slab_alloc(cachep, flags, _RET_IP_);
3617 trace_kmem_cache_alloc(_RET_IP_, ret,
3618 cachep->object_size, cachep->size, flags);
3622 EXPORT_SYMBOL(kmem_cache_alloc);
3624 #ifdef CONFIG_TRACING
3626 kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
3630 ret = slab_alloc(cachep, flags, _RET_IP_);
3632 trace_kmalloc(_RET_IP_, ret,
3633 size, cachep->size, flags);
3636 EXPORT_SYMBOL(kmem_cache_alloc_trace);
3641 * kmem_cache_alloc_node - Allocate an object on the specified node
3642 * @cachep: The cache to allocate from.
3643 * @flags: See kmalloc().
3644 * @nodeid: node number of the target node.
3646 * Identical to kmem_cache_alloc but it will allocate memory on the given
3647 * node, which can improve the performance for cpu bound structures.
3649 * Fallback to other node is possible if __GFP_THISNODE is not set.
3651 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3653 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3655 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3656 cachep->object_size, cachep->size,
3661 EXPORT_SYMBOL(kmem_cache_alloc_node);
3663 #ifdef CONFIG_TRACING
3664 void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
3671 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3673 trace_kmalloc_node(_RET_IP_, ret,
3678 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3681 static __always_inline void *
3682 __do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
3684 struct kmem_cache *cachep;
3686 cachep = kmalloc_slab(size, flags);
3687 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3689 return kmem_cache_alloc_node_trace(cachep, flags, node, size);
3692 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3693 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3695 return __do_kmalloc_node(size, flags, node, _RET_IP_);
3697 EXPORT_SYMBOL(__kmalloc_node);
3699 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3700 int node, unsigned long caller)
3702 return __do_kmalloc_node(size, flags, node, caller);
3704 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3706 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3708 return __do_kmalloc_node(size, flags, node, 0);
3710 EXPORT_SYMBOL(__kmalloc_node);
3711 #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3712 #endif /* CONFIG_NUMA */
3715 * __do_kmalloc - allocate memory
3716 * @size: how many bytes of memory are required.
3717 * @flags: the type of memory to allocate (see kmalloc).
3718 * @caller: function caller for debug tracking of the caller
3720 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3721 unsigned long caller)
3723 struct kmem_cache *cachep;
3726 /* If you want to save a few bytes .text space: replace
3728 * Then kmalloc uses the uninlined functions instead of the inline
3731 cachep = kmalloc_slab(size, flags);
3732 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3734 ret = slab_alloc(cachep, flags, caller);
3736 trace_kmalloc(caller, ret,
3737 size, cachep->size, flags);
3743 #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3744 void *__kmalloc(size_t size, gfp_t flags)
3746 return __do_kmalloc(size, flags, _RET_IP_);
3748 EXPORT_SYMBOL(__kmalloc);
3750 void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3752 return __do_kmalloc(size, flags, caller);
3754 EXPORT_SYMBOL(__kmalloc_track_caller);
3757 void *__kmalloc(size_t size, gfp_t flags)
3759 return __do_kmalloc(size, flags, 0);
3761 EXPORT_SYMBOL(__kmalloc);
3765 * kmem_cache_free - Deallocate an object
3766 * @cachep: The cache the allocation was from.
3767 * @objp: The previously allocated object.
3769 * Free an object which was previously allocated from this
3772 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3774 unsigned long flags;
3775 cachep = cache_from_obj(cachep, objp);
3779 local_irq_save(flags);
3780 debug_check_no_locks_freed(objp, cachep->object_size);
3781 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3782 debug_check_no_obj_freed(objp, cachep->object_size);
3783 __cache_free(cachep, objp, _RET_IP_);
3784 local_irq_restore(flags);
3786 trace_kmem_cache_free(_RET_IP_, objp);
3788 EXPORT_SYMBOL(kmem_cache_free);
3791 * kfree - free previously allocated memory
3792 * @objp: pointer returned by kmalloc.
3794 * If @objp is NULL, no operation is performed.
3796 * Don't free memory not originally allocated by kmalloc()
3797 * or you will run into trouble.
3799 void kfree(const void *objp)
3801 struct kmem_cache *c;
3802 unsigned long flags;
3804 trace_kfree(_RET_IP_, objp);
3806 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3808 local_irq_save(flags);
3809 kfree_debugcheck(objp);
3810 c = virt_to_cache(objp);
3811 debug_check_no_locks_freed(objp, c->object_size);
3813 debug_check_no_obj_freed(objp, c->object_size);
3814 __cache_free(c, (void *)objp, _RET_IP_);
3815 local_irq_restore(flags);
3817 EXPORT_SYMBOL(kfree);
3820 * This initializes kmem_cache_node or resizes various caches for all nodes.
3822 static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp)
3825 struct kmem_cache_node *n;
3826 struct array_cache *new_shared;
3827 struct array_cache **new_alien = NULL;
3829 for_each_online_node(node) {
3831 if (use_alien_caches) {
3832 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
3838 if (cachep->shared) {
3839 new_shared = alloc_arraycache(node,
3840 cachep->shared*cachep->batchcount,
3843 free_alien_cache(new_alien);
3848 n = cachep->node[node];
3850 struct array_cache *shared = n->shared;
3852 spin_lock_irq(&n->list_lock);
3855 free_block(cachep, shared->entry,
3856 shared->avail, node);
3858 n->shared = new_shared;
3860 n->alien = new_alien;
3863 n->free_limit = (1 + nr_cpus_node(node)) *
3864 cachep->batchcount + cachep->num;
3865 spin_unlock_irq(&n->list_lock);
3867 free_alien_cache(new_alien);
3870 n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
3872 free_alien_cache(new_alien);
3877 kmem_cache_node_init(n);
3878 n->next_reap = jiffies + REAPTIMEOUT_LIST3 +
3879 ((unsigned long)cachep) % REAPTIMEOUT_LIST3;
3880 n->shared = new_shared;
3881 n->alien = new_alien;
3882 n->free_limit = (1 + nr_cpus_node(node)) *
3883 cachep->batchcount + cachep->num;
3884 cachep->node[node] = n;
3889 if (!cachep->list.next) {
3890 /* Cache is not active yet. Roll back what we did */
3893 if (cachep->node[node]) {
3894 n = cachep->node[node];
3897 free_alien_cache(n->alien);
3899 cachep->node[node] = NULL;
3907 struct ccupdate_struct {
3908 struct kmem_cache *cachep;
3909 struct array_cache *new[0];
3912 static void do_ccupdate_local(void *info)
3914 struct ccupdate_struct *new = info;
3915 struct array_cache *old;
3918 old = cpu_cache_get(new->cachep);
3920 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3921 new->new[smp_processor_id()] = old;
3924 /* Always called with the slab_mutex held */
3925 static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
3926 int batchcount, int shared, gfp_t gfp)
3928 struct ccupdate_struct *new;
3931 new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *),
3936 for_each_online_cpu(i) {
3937 new->new[i] = alloc_arraycache(cpu_to_mem(i), limit,
3940 for (i--; i >= 0; i--)
3946 new->cachep = cachep;
3948 on_each_cpu(do_ccupdate_local, (void *)new, 1);
3951 cachep->batchcount = batchcount;
3952 cachep->limit = limit;
3953 cachep->shared = shared;
3955 for_each_online_cpu(i) {
3956 struct array_cache *ccold = new->new[i];
3959 spin_lock_irq(&cachep->node[cpu_to_mem(i)]->list_lock);
3960 free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i));
3961 spin_unlock_irq(&cachep->node[cpu_to_mem(i)]->list_lock);
3965 return alloc_kmemlist(cachep, gfp);
3968 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3969 int batchcount, int shared, gfp_t gfp)
3972 struct kmem_cache *c = NULL;
3975 ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3977 if (slab_state < FULL)
3980 if ((ret < 0) || !is_root_cache(cachep))
3983 VM_BUG_ON(!mutex_is_locked(&slab_mutex));
3984 for_each_memcg_cache_index(i) {
3985 c = cache_from_memcg(cachep, i);
3987 /* return value determined by the parent cache only */
3988 __do_tune_cpucache(c, limit, batchcount, shared, gfp);
3994 /* Called with slab_mutex held always */
3995 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
4002 if (!is_root_cache(cachep)) {
4003 struct kmem_cache *root = memcg_root_cache(cachep);
4004 limit = root->limit;
4005 shared = root->shared;
4006 batchcount = root->batchcount;
4009 if (limit && shared && batchcount)
4012 * The head array serves three purposes:
4013 * - create a LIFO ordering, i.e. return objects that are cache-warm
4014 * - reduce the number of spinlock operations.
4015 * - reduce the number of linked list operations on the slab and
4016 * bufctl chains: array operations are cheaper.
4017 * The numbers are guessed, we should auto-tune as described by
4020 if (cachep->size > 131072)
4022 else if (cachep->size > PAGE_SIZE)
4024 else if (cachep->size > 1024)
4026 else if (cachep->size > 256)
4032 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
4033 * allocation behaviour: Most allocs on one cpu, most free operations
4034 * on another cpu. For these cases, an efficient object passing between
4035 * cpus is necessary. This is provided by a shared array. The array
4036 * replaces Bonwick's magazine layer.
4037 * On uniprocessor, it's functionally equivalent (but less efficient)
4038 * to a larger limit. Thus disabled by default.
4041 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
4046 * With debugging enabled, large batchcount lead to excessively long
4047 * periods with disabled local interrupts. Limit the batchcount
4052 batchcount = (limit + 1) / 2;
4054 err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
4056 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
4057 cachep->name, -err);
4062 * Drain an array if it contains any elements taking the node lock only if
4063 * necessary. Note that the node listlock also protects the array_cache
4064 * if drain_array() is used on the shared array.
4066 static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
4067 struct array_cache *ac, int force, int node)
4071 if (!ac || !ac->avail)
4073 if (ac->touched && !force) {
4076 spin_lock_irq(&n->list_lock);
4078 tofree = force ? ac->avail : (ac->limit + 4) / 5;
4079 if (tofree > ac->avail)
4080 tofree = (ac->avail + 1) / 2;
4081 free_block(cachep, ac->entry, tofree, node);
4082 ac->avail -= tofree;
4083 memmove(ac->entry, &(ac->entry[tofree]),
4084 sizeof(void *) * ac->avail);
4086 spin_unlock_irq(&n->list_lock);
4091 * cache_reap - Reclaim memory from caches.
4092 * @w: work descriptor
4094 * Called from workqueue/eventd every few seconds.
4096 * - clear the per-cpu caches for this CPU.
4097 * - return freeable pages to the main free memory pool.
4099 * If we cannot acquire the cache chain mutex then just give up - we'll try
4100 * again on the next iteration.
4102 static void cache_reap(struct work_struct *w)
4104 struct kmem_cache *searchp;
4105 struct kmem_cache_node *n;
4106 int node = numa_mem_id();
4107 struct delayed_work *work = to_delayed_work(w);
4109 if (!mutex_trylock(&slab_mutex))
4110 /* Give up. Setup the next iteration. */
4113 list_for_each_entry(searchp, &slab_caches, list) {
4117 * We only take the node lock if absolutely necessary and we
4118 * have established with reasonable certainty that
4119 * we can do some work if the lock was obtained.
4121 n = searchp->node[node];
4123 reap_alien(searchp, n);
4125 drain_array(searchp, n, cpu_cache_get(searchp), 0, node);
4128 * These are racy checks but it does not matter
4129 * if we skip one check or scan twice.
4131 if (time_after(n->next_reap, jiffies))
4134 n->next_reap = jiffies + REAPTIMEOUT_LIST3;
4136 drain_array(searchp, n, n->shared, 0, node);
4138 if (n->free_touched)
4139 n->free_touched = 0;
4143 freed = drain_freelist(searchp, n, (n->free_limit +
4144 5 * searchp->num - 1) / (5 * searchp->num));
4145 STATS_ADD_REAPED(searchp, freed);
4151 mutex_unlock(&slab_mutex);
4154 /* Set up the next iteration */
4155 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC));
4158 #ifdef CONFIG_SLABINFO
4159 void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
4162 unsigned long active_objs;
4163 unsigned long num_objs;
4164 unsigned long active_slabs = 0;
4165 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4169 struct kmem_cache_node *n;
4173 for_each_online_node(node) {
4174 n = cachep->node[node];
4179 spin_lock_irq(&n->list_lock);
4181 list_for_each_entry(slabp, &n->slabs_full, list) {
4182 if (slabp->inuse != cachep->num && !error)
4183 error = "slabs_full accounting error";
4184 active_objs += cachep->num;
4187 list_for_each_entry(slabp, &n->slabs_partial, list) {
4188 if (slabp->inuse == cachep->num && !error)
4189 error = "slabs_partial inuse accounting error";
4190 if (!slabp->inuse && !error)
4191 error = "slabs_partial/inuse accounting error";
4192 active_objs += slabp->inuse;
4195 list_for_each_entry(slabp, &n->slabs_free, list) {
4196 if (slabp->inuse && !error)
4197 error = "slabs_free/inuse accounting error";
4200 free_objects += n->free_objects;
4202 shared_avail += n->shared->avail;
4204 spin_unlock_irq(&n->list_lock);
4206 num_slabs += active_slabs;
4207 num_objs = num_slabs * cachep->num;
4208 if (num_objs - active_objs != free_objects && !error)
4209 error = "free_objects accounting error";
4211 name = cachep->name;
4213 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4215 sinfo->active_objs = active_objs;
4216 sinfo->num_objs = num_objs;
4217 sinfo->active_slabs = active_slabs;
4218 sinfo->num_slabs = num_slabs;
4219 sinfo->shared_avail = shared_avail;
4220 sinfo->limit = cachep->limit;
4221 sinfo->batchcount = cachep->batchcount;
4222 sinfo->shared = cachep->shared;
4223 sinfo->objects_per_slab = cachep->num;
4224 sinfo->cache_order = cachep->gfporder;
4227 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
4231 unsigned long high = cachep->high_mark;
4232 unsigned long allocs = cachep->num_allocations;
4233 unsigned long grown = cachep->grown;
4234 unsigned long reaped = cachep->reaped;
4235 unsigned long errors = cachep->errors;
4236 unsigned long max_freeable = cachep->max_freeable;
4237 unsigned long node_allocs = cachep->node_allocs;
4238 unsigned long node_frees = cachep->node_frees;
4239 unsigned long overflows = cachep->node_overflow;
4241 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4242 "%4lu %4lu %4lu %4lu %4lu",
4243 allocs, high, grown,
4244 reaped, errors, max_freeable, node_allocs,
4245 node_frees, overflows);
4249 unsigned long allochit = atomic_read(&cachep->allochit);
4250 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4251 unsigned long freehit = atomic_read(&cachep->freehit);
4252 unsigned long freemiss = atomic_read(&cachep->freemiss);
4254 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4255 allochit, allocmiss, freehit, freemiss);
4260 #define MAX_SLABINFO_WRITE 128
4262 * slabinfo_write - Tuning for the slab allocator
4264 * @buffer: user buffer
4265 * @count: data length
4268 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4269 size_t count, loff_t *ppos)
4271 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4272 int limit, batchcount, shared, res;
4273 struct kmem_cache *cachep;
4275 if (count > MAX_SLABINFO_WRITE)
4277 if (copy_from_user(&kbuf, buffer, count))
4279 kbuf[MAX_SLABINFO_WRITE] = '\0';
4281 tmp = strchr(kbuf, ' ');
4286 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4289 /* Find the cache in the chain of caches. */
4290 mutex_lock(&slab_mutex);
4292 list_for_each_entry(cachep, &slab_caches, list) {
4293 if (!strcmp(cachep->name, kbuf)) {
4294 if (limit < 1 || batchcount < 1 ||
4295 batchcount > limit || shared < 0) {
4298 res = do_tune_cpucache(cachep, limit,
4305 mutex_unlock(&slab_mutex);
4311 #ifdef CONFIG_DEBUG_SLAB_LEAK
4313 static void *leaks_start(struct seq_file *m, loff_t *pos)
4315 mutex_lock(&slab_mutex);
4316 return seq_list_start(&slab_caches, *pos);
4319 static inline int add_caller(unsigned long *n, unsigned long v)
4329 unsigned long *q = p + 2 * i;
4343 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4349 static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s)
4355 for (i = 0, p = s->s_mem; i < c->num; i++, p += c->size) {
4356 if (slab_bufctl(s)[i] != BUFCTL_ACTIVE)
4358 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4363 static void show_symbol(struct seq_file *m, unsigned long address)
4365 #ifdef CONFIG_KALLSYMS
4366 unsigned long offset, size;
4367 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4369 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4370 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4372 seq_printf(m, " [%s]", modname);
4376 seq_printf(m, "%p", (void *)address);
4379 static int leaks_show(struct seq_file *m, void *p)
4381 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
4383 struct kmem_cache_node *n;
4385 unsigned long *x = m->private;
4389 if (!(cachep->flags & SLAB_STORE_USER))
4391 if (!(cachep->flags & SLAB_RED_ZONE))
4394 /* OK, we can do it */
4398 for_each_online_node(node) {
4399 n = cachep->node[node];
4404 spin_lock_irq(&n->list_lock);
4406 list_for_each_entry(slabp, &n->slabs_full, list)
4407 handle_slab(x, cachep, slabp);
4408 list_for_each_entry(slabp, &n->slabs_partial, list)
4409 handle_slab(x, cachep, slabp);
4410 spin_unlock_irq(&n->list_lock);
4412 name = cachep->name;
4414 /* Increase the buffer size */
4415 mutex_unlock(&slab_mutex);
4416 m->private = kzalloc(x[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4418 /* Too bad, we are really out */
4420 mutex_lock(&slab_mutex);
4423 *(unsigned long *)m->private = x[0] * 2;
4425 mutex_lock(&slab_mutex);
4426 /* Now make sure this entry will be retried */
4430 for (i = 0; i < x[1]; i++) {
4431 seq_printf(m, "%s: %lu ", name, x[2*i+3]);
4432 show_symbol(m, x[2*i+2]);
4439 static const struct seq_operations slabstats_op = {
4440 .start = leaks_start,
4446 static int slabstats_open(struct inode *inode, struct file *file)
4448 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4451 ret = seq_open(file, &slabstats_op);
4453 struct seq_file *m = file->private_data;
4454 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4463 static const struct file_operations proc_slabstats_operations = {
4464 .open = slabstats_open,
4466 .llseek = seq_lseek,
4467 .release = seq_release_private,
4471 static int __init slab_proc_init(void)
4473 #ifdef CONFIG_DEBUG_SLAB_LEAK
4474 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4478 module_init(slab_proc_init);
4482 * ksize - get the actual amount of memory allocated for a given object
4483 * @objp: Pointer to the object
4485 * kmalloc may internally round up allocations and return more memory
4486 * than requested. ksize() can be used to determine the actual amount of
4487 * memory allocated. The caller may use this additional memory, even though
4488 * a smaller amount of memory was initially specified with the kmalloc call.
4489 * The caller must guarantee that objp points to a valid object previously
4490 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4491 * must not be freed during the duration of the call.
4493 size_t ksize(const void *objp)
4496 if (unlikely(objp == ZERO_SIZE_PTR))
4499 return virt_to_cache(objp)->object_size;
4501 EXPORT_SYMBOL(ksize);